Mathematical Methods for Robust and Nonlinear Control: EPSRC Summer School (Lecture Notes in Control and Information Sciences)

Lecture Notes in Control and Information Sciences Editors: M. Thoma, M. Morari 367 Matthew C. Turner, Declan G. Bate...

Author: Matthew C. Turner | Declan G. Bates

120 downloads 1187 Views 8MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

Lecture Notes in Control and Information Sciences Editors: M. Thoma, M. Morari

367

Matthew C. Turner, Declan G. Bates (Eds.)

Mathematical Methods for Robust and Nonlinear Control EPSRC Summer School

ABC

Series Advisory Board F. Allgöwer, P. Fleming, P. Kokotovic, A.B. Kurzhanski, H. Kwakernaak, A. Rantzer, J.N. Tsitsiklis

Editors Matthew C. Turner Department of Engineering University of Leicester University Road LE1 7RH Leicester UK Email: [email protected]

Declan G. Bates Department of Engineering University of Leicester University Road LE1 7RH Leicester UK Email: [email protected]

Library of Congress Control Number: 2007932607 ISSN print edition: 0170-8643 ISSN electronic edition: 1610-7411 ISBN-10 1-84800-024-3 Springer Berlin Heidelberg New York ISBN-13 978-1-84800-024-7 Springer Berlin Heidelberg New York This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable for prosecution under the German Copyright Law. Springer is a part of Springer Science+Business Media springer.com c Springer-Verlag Berlin Heidelberg 2007 MATLAB and Simulink are registered trademarks of The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA 01760-2098, U.S.A. http://www.mathworks.com and Scilab is a trademark of INRIA, Domaine de c 1989Voluceau, Rocquencourt - B.P. 105, 78153 Le Chesnay cedex, France. www.scilab.org Copyright 2007. INRIA. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Typesetting: by the authors and SPS using a Springer LATEX macro package Printed on acid-free paper

SPIN: 12057016

89/SPS

543210

Preface

The underlying theory on which much modern robust and nonlinear control is based can often be difficult for the student to grasp. In particular, the mathematical aspects can be problematic for students from a standard engineering background. The EPSRC sponsored Summer School which was held in Leicester in September 2006 attempted to “fill the gap” in students’ appreciation the theory relevant to several important areas of control. This book is a collection of lecture notes which were presented at that workshop and consists of, broadly, two parts. The first nine chapters are devoted to the theory behind several areas of robust and nonlinear control and are aimed at introducing fundamental concepts to the reader. The last six chapters contain detailed case studies which aim to demonstrate the use and effectiveness of these modern techniques in real engineering applications. It is hoped that this book will provide a useful introduction to many of the more common robust and nonlinear control techniques and serve as a valuable reference for the more adept practitioner.

Leicester, May 2007

Matthew C. Turner Declan G. Bates

Contents

List of Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . XV

Part I Theory of Robust and Nonlinear Control 1 H∞ Control Design Declan G. Bates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Design Specifications and Fundamental Trade-offs . . . . . . . . . . . . . . . . . . 1.2.1 Linear Design Specifications for Robust Control Systems . . . . . 1.2.2 Frequency Domain Design Specifications and Fundamental Trade-offs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Mixed-sensitivity H∞ Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.1 Formulating the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.2 Weighting Function Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.3 Solution of the H∞ Control Problem . . . . . . . . . . . . . . . . . . . . . . 1.3.4 Design Example: Control Law Design for the Bell 205 Helicopter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 H∞ Loop-shaping Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.1 Fundamental Trade-offs in Terms of L . . . . . . . . . . . . . . . . . . . . . 1.4.2 The H∞ Loop-shaping Design Procedure . . . . . . . . . . . . . . . . . . 1.4.3 Advantages of H∞ Loop-shaping . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.4 Design Example: Control Law Design for the Harrier V/STOL Aircraft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Structural Methods for Linear Systems: An Introduction Nicos Karcanias and Efstathios Milonidis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Classification of System Representations . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.1 State Space Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.2 Polynomial Models: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.3 Transfer Function Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Background on Polynomial Matrices and Matrix Pencils . . . . . . . . . . . . .

3 3 5 6 6 8 8 11 11 13 24 24 26 29 33 43 47 48 50 50 51 51 52

VIII

Contents

2.3.1 Matrix Divisors and Minimal Bases . . . . . . . . . . . . . . . . . . . . . . . 2.3.2 Strict Equivalence Invariants of Matrix Pencils [13] . . . . . . . . . . 2.4 Dynamics, Stability, Controllability and Observability . . . . . . . . . . . . . . . 2.4.1 Solution of State Space Equations . . . . . . . . . . . . . . . . . . . . . . . . 2.4.2 Internal-External and Total Stability . . . . . . . . . . . . . . . . . . . . . . . 2.4.3 Controllability and Observability . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.4 System Minimality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5 Poles and Zeros of State Space Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.1 Eigenvalues, Eigenvectors and Free Rectilinear Motions . . . . . . 2.5.2 Forced Rectilinear Motions and Frequency Transmission . . . . . 2.5.3 Frequency Transmission Blocking and State Space Zeros . . . . . 2.5.4 Right Regular Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.5 Properties of Zero Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.6 Right Singular Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.7 Frequency Transmission Blocking for Infinite Frequencies . . . . 2.5.8 Zero Structure and System Transformations . . . . . . . . . . . . . . . . 2.5.9 The Zero Pencil of Strictly Proper System . . . . . . . . . . . . . . . . . . 2.5.10 Decoupling Zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6 Poles and Zeros of Transfer Function Models . . . . . . . . . . . . . . . . . . . . . . 2.6.1 Dynamic Characterisation of Transfer Function Poles and Zeros 2.6.2 Smith–McMillan Form Characterisation of Poles and Zeros . . . 2.6.3 Matrix Fraction Descriptions, and Poles and Zeros . . . . . . . . . . . 2.6.4 Infinite Poles and Zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.5 Smith–McMillan Form at Infinity: Infinite Poles and Zeros . . . 2.6.6 Impulsive Dynamics and Properties of Infinite Poles and Zeros [57] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.7 System Algebraic Functions and Generalised Nyquist and Root Locus . 2.7.1 Characteristic Gain, Frequency Functions . . . . . . . . . . . . . . . . . . 2.7.2 Poles and Zeros of the System Algebraic Functions . . . . . . . . . . 2.7.3 Root Locus and the Output Zeroing Problem . . . . . . . . . . . . . . . 2.8 The Feedback Configuration and Structural Properties . . . . . . . . . . . . . . . 2.8.1 Structural Properties of the Feedback Configuration . . . . . . . . . 2.8.2 Closed-loop Performance and the Return Ratio Difference and Sensitivity Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.9 Determinantal Assignment Problems: Exterior Algebra-Algebraic Geometry Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.9.1 Determinantal Assignment Problems . . . . . . . . . . . . . . . . . . . . . . 2.9.2 The General Determinantal Assignment Problem . . . . . . . . . . . . 2.9.3 Grassmann- Plucker Invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A Invariants and Canonical Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B List of Symbols, Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

53 55 55 55 57 58 60 61 61 62 63 65 65 66 67 69 70 71 72 72 73 74 74 75 76 77 77 78 79 80 80 83 84 85 88 89 92 92 94 94

Contents

IX

3 Modelling and Model Reduction—State-Space Truncation† David J. N. Limebeer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 3.2 State-space Truncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 3.2.1 The Truncation Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 3.2.2 Singular Perturbation Approximation . . . . . . . . . . . . . . . . . . . . . . 104 Main Points of the Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 3.3 Balanced Realization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 3.3.1 Model Reduction Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 3.3.2 Balanced Realization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 Main Points of the Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 3.4 Balanced Truncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 3.4.1 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 3.4.2 Error Bound for “one-step” Truncation . . . . . . . . . . . . . . . . . . . . 112 3.4.3 The Error Bound for Balanced Truncation . . . . . . . . . . . . . . . . . . 113 Tightness of the Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 Frequency Dependence of the Error . . . . . . . . . . . . . . . . . . . . . . . 115 Main Points of the Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 3.5 Balanced Singular Perturbation Approximation . . . . . . . . . . . . . . . . . . . . . 116 Main Point of the Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 3.6 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 3.7 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 3.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 4 Linear Matrix Inequalities in Control Guido Herrmann, Matthew C. Turner and Ian Postlethwaite . . . . . . . . . . . . . . . 123 4.1 Introduction to LMI Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 4.1.1 Fundamental LMI Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 4.1.2 Systems of LMIs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 4.1.3 Types of LMI Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 LMI Feasibility Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 Linear Objective Minimization Problems . . . . . . . . . . . . . . . . . . 127 Generalized Eigenvalue Problems . . . . . . . . . . . . . . . . . . . . . . . . . 127 4.2 Tricks in LMI Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 4.2.1 Change of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 4.2.2 Congruence Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 4.2.3 Schur Complement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 4.2.4 The S-procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 4.2.5 The Projection Lemma and Finsler’s Lemma . . . . . . . . . . . . . . . 132 4.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 4.3.1 Lyapunov Stability for Continuous-time Systems . . . . . . . . . . . . 134 4.3.2 L2 Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 4.3.3 Lyapunov Stability for Discrete-time Systems . . . . . . . . . . . . . . 135 4.3.4 l2 Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

X

Contents

4.3.5 Sector Boundedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.6 A Slightly More Detailed Example . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

137 138 141 141

5 Anti-windup Compensation and the Control of Input-constrained Systems Matthew C. Turner, Guido Herrmann and Ian Postlethwaite . . . . . . . . . . . . . . . . 143 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 5.1.1 Input Constraints in Control Systems . . . . . . . . . . . . . . . . . . . . . . 143 5.1.2 Constrained System Description . . . . . . . . . . . . . . . . . . . . . . . . . . 144 5.1.3 Constrained Control and Anti-windup . . . . . . . . . . . . . . . . . . . . . 146 5.2 Problems Due to Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 5.2.1 Clues From Classical Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 5.3 Stability of Systems with Input Saturation . . . . . . . . . . . . . . . . . . . . . . . . . 152 5.3.1 Definitions of Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152 5.3.2 Saturation Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 An Equivalent Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 Sector Bounding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 5.3.3 The Multivariable Circle Criterion . . . . . . . . . . . . . . . . . . . . . . . . 157 5.4 Anti-windup Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 5.5 An Anti-windup Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 5.5.1 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 5.5.2 Full Order Compensators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 5.6 Simple Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 5.6.1 Simple 2nd-order Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 The Nominal System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 The Constrained System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 The Constrained System and Anti-windup . . . . . . . . . . . . . . . . . . 166 5.6.2 Lockheed Martin F104 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 168 The Nominal System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 5.6.3 The Constrained System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 5.6.4 The Constrained System and Anti-windup . . . . . . . . . . . . . . . . . . 169 5.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 5.7.1 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 5.8 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 6 Output Feedback H∞ Loop-shaping Controller Synthesis Emmanuel Prempain and Ian Postlethwaite . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 6.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 6.2.1 LMI Formulation of Performance Specifications . . . . . . . . . . . . 176 H∞ Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 H2 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

Contents

XI

6.2.2 Normalized Left Coprime Factorization for LTI Systems . . . . . 177 H∞ Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 H∞ Loop-shaping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181 6.4.1 LMI Formulation of the H∞ Loop-shaping Controller Synthesis 181 6.4.2 Controller Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 6.4.3 Design Procedure for a Static H∞ Loop-shaping Controller . . . 184 6.4.4 Static H∞ Flight Control System Design for the Bell 205 Helicopter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 Plant Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 Static H∞ Helicopter Controller Design . . . . . . . . . . . . . . . . . . . 184 6.5 H∞ Loop-shaping for Polytopic Systems . . . . . . . . . . . . . . . . . . . . . . . . . 186 6.5.1 Left Coprime Factors for Polytopic Systems . . . . . . . . . . . . . . . . 187 6.6 LMI Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 6.6.1 Illustrative Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 6.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192 6.3 6.4

7 Stability and Asymptotic Behaviour of Nonlinear Systems: An Introduction Hartmut Logemann and Eugene P. Ryan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 7.2 Terminology and Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 7.3 Background Concepts in Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 7.4 Initial-value Problems: Existence of Solutions . . . . . . . . . . . . . . . . . . . . . . 199 7.4.1 Ordinary Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 199 7.4.2 Autonomous Differential Inclusions . . . . . . . . . . . . . . . . . . . . . . . 201 7.4.3 ω -limit Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 7.5 Barb˘alat’s Lemma, LaSalle’s Invariance Principle, and Lyapunov Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 7.6 Generalizations of Barb˘alat’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 7.7 Nonautonomous Ordinary Differential equations . . . . . . . . . . . . . . . . . . . . 211 7.8 Autonomous Differential Inclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 8 Sliding-mode Observers† Christopher Edwards, Sarah K. Spurgeon, Chee P. Tan and Nitin Patel . . . . . . . 221 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 8.2 A Discontinuous Observer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222 8.3 Observers with Linear and Discontinuous Injection . . . . . . . . . . . . . . . . . 225 ˙ Observer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 8.4 The Walcott and Zak 8.4.1 Synthesizing the Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228 8.5 A Convex Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228 8.6 A Case Study: Road Tyre Friction Estimation . . . . . . . . . . . . . . . . . . . . . . 232 8.6.1 Tyre/Road Friction and Vehicle Modelling . . . . . . . . . . . . . . . . . 232 8.6.2 Observer Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234

XI I

Contents

8.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.8 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.9 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

238 240 240 240

9 Sliding-mode Control in Systems with Output Time Delay Alan S.I. Zinober, G. Liu and Yuri B. Shtessel . . . . . . . . . . . . . . . . . . . . . . . . . . . 243 9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243 9.2 Sliding-mode Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 9.2.1 Regulator System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 9.2.2 Model-Following Control System . . . . . . . . . . . . . . . . . . . . . . . . . 244 9.2.3 Sliding-mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245 9.2.4 Feedback Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 9.2.5 Second-order Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 9.3 Application Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250 9.3.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 9.3.2 Pad´e Approximations and Time Delay Systems . . . . . . . . . . . . . 252 9.3.3 System Centre Method and Sliding-mode Control . . . . . . . . . . . 254 9.3.4 Numerical Example and Simulations . . . . . . . . . . . . . . . . . . . . . . 255 9.3.5 Feedback by yˆ and Describing Function . . . . . . . . . . . . . . . . . . . . 257 9.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 Part II Applications of Robust and Nonlinear Control 10 Control Engineering and Systems Biology Burton W. Andrews and Pablo A. Iglesias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267 10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267 10.2 Negative Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268 10.2.1 Negative Feedback: Regulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 268 10.2.2 Negative Feedback: Sensitivity and Robustness . . . . . . . . . . . . . 273 10.3 Positive Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277 10.3.1 Positive Feedback: Amplification . . . . . . . . . . . . . . . . . . . . . . . . . 277 10.3.2 Positive Feedback: Switching and Memory . . . . . . . . . . . . . . . . . 278 10.3.3 Positive Feedback: Oscillations . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 10.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284 11 Robust Control of a Distillation Column Da-Wei Gu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 11.2 Dynamic Model of the Distillation Column . . . . . . . . . . . . . . . . . . . . . . . . 290 11.3 Uncertainty Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293 11.4 Closed-loop System Performance Specifications . . . . . . . . . . . . . . . . . . . . 296 11.5 Open-loop and Closed-loop System Interconnections . . . . . . . . . . . . . . . . 300

Contents

11.6

Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.6.1 Loop Shaping Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.6.2 µ -synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.7 Nonlinear System Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

X III

301 301 307 314 318 319

12 Robust Control of a Hard-disk Drive Da-Wei Gu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329 12.1 Hard Disk Drive Servo System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329 12.2 Derivation of Uncertainty Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335 12.3 Closed-loop System Design Specifications . . . . . . . . . . . . . . . . . . . . . . . . . 340 12.4 System Interconnections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342 12.5 Controller Design in Continuous Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343 12.5.1 µ -design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 12.5.2 H∞ Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351 12.5.3 H∞ Loop-shaping Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351 12.6 Comparison of Designed Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352 12.7 Controller Order Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359 12.8 Design of Discrete-time Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361 12.9 Nonlinear System Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365 12.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369 13 Modelling and Control of Railway Vehicle Suspensions Argyrios C. Zolotas and Roger M. Goodall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373 13.1 Overview of Railway Vehicle Dynamics and Control . . . . . . . . . . . . . . . . 373 13.1.1 Railway Vehicles: Conventional Configuration . . . . . . . . . . . . . . 373 13.1.2 Suspension Design Requirements . . . . . . . . . . . . . . . . . . . . . . . . . 374 13.1.3 Modelling of Suspensions (for Applying Control) . . . . . . . . . . . 375 13.1.4 Control Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378 Tilting Trains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378 Active Secondary Suspensions . . . . . . . . . . . . . . . . . . . . . . . . . . . 379 Active Primary Suspensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380 13.2 Case Study: Control of Secondary Suspensions - Tilting Trains . . . . . . . 381 13.2.1 Historical Facts on Tilt Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 381 13.2.2 Tilting Vehicle Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382 13.2.3 Tilt Control Requirements and Assessment Approach . . . . . . . . 385 Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385 Tilt Control Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386 Track Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387 13.2.4 Conventional Tilt Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387 Classical Nulling Control Strategy . . . . . . . . . . . . . . . . . . . . . . . . 387 Command-driven with Precedence Control . . . . . . . . . . . . . . . . . 390 13.2.5 Nulling-type Tilt via Robust Control Techniques . . . . . . . . . . . . 394

XIV

Contents

LQG/LTR Nulling-type Tilt Control . . . . . . . . . . . . . . . . . . . . . . . 13.2.6 Multi-objective H∞ /H2 Nulling-type Control via LMIs . . . . . . . 13.2.7 Case Study Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13.3 Appendix A- Tilting Train Parameter Values and Notation . . . . . . . . . . . . 13.4 Appendix B- H∞ Based Controllers: Preliminaries . . . . . . . . . . . . . . . . . . 13.4.1 Basic Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . • Frequency Domain Spaces and Norms . . . . . . . . . . . . . . . . . . . • Linear Fractional Transformations . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

394 399 406 407 407 407 408 409 411

14 Case Study on Anti-windup Compensation - Micro-actuator Control in a Hard-disk Drive Guido Herrmann, Matthew C. Turner and Ian Postlethwaite . . . . . . . . . . . . . . . 413 14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413 14.2 The Micro-actuator Control loop and Windup Problems . . . . . . . . . . . . . 415 14.3 Anti-windup Compensation for Discrete Linear Control Systems . . . . . . 419 14.4 Anti-windup Compensation for the Micro-actuator . . . . . . . . . . . . . . . . . . 424 14.5 The Micro-actuator Control Loop as Part of a Hard-disk-drive Servo-system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425 14.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429 15 Enhancing Immune System Response Through Optimal Control Robert F. Harrison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431 15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431 15.2 Lotka-Volterra Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431 15.2.1 Analysis of Equilibria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433 15.3 Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436 15.3.1 Linear, Time-varying Quadratic Optimal Control . . . . . . . . . . . . 436 15.4 Immune System Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438 15.4.1 Optimal Enhancement of the Immune Response . . . . . . . . . . . . . 440 15.4.2 Some Practical Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . 442 15.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443

List of Contributors

Declan G. Bates University of Leicester University Road Leicester, LE1 7RH [email protected]

Matthew C. Turner University of Leicester University Road Leicester, LE1 7RH [email protected]

Nicos Karcanias City University Northampton Square London, EC1V 0HB [email protected]

Ian Postlethwaite University of Leicester University Road Leicester, LE1 7RH [email protected]

Efstathios Milonidis City University Northampton Square London, EC1V 0HB [email protected]

Emmanuel Prempain University of Leicester University Road Leicester, LE1 7RH [email protected]

David J.N. Limebeer Imperial College London South Kensington Campus London, SW7 2AZ [email protected]

Eugene P. Ryan Mathematical Sciences University of Bath Bath, BA2 7AY [email protected]

Guido Herrmann University of Leicester University Road Leicester, LE1 7RH [email protected]

Hartmut Logemann Mathematical Sciences University of Bath Bath, BA2 7AY [email protected]

XVI

List of Contributors

Christopher Edwards University of Leicester University Road Leicester, LE1 7RH [email protected] Sarah K. Spurgeon University of Leicester University Road Leicester, LE1 7RH [email protected] Nitin Patel University of Leicester University Road Leicester, LE1 7RH [email protected] Chee P. Tan Monash University Malaysia 46150 Petaling Jaya Malaysia [email protected] Alan S.I. Zinober University of Sheffield Mappin Street Sheffield, S10 2TN [email protected] G. Liu University of Sheffield Mappin Street Sheffield, S10 2TN [email protected] Yuri B. Shtessel The University of Alabama in Huntsville Huntsville, AL 35899 USA [email protected]

Pablo A. Iglesias The Johns Hopkins University Baltimore, MD 21218 USA [email protected] Burton W. Andrews The Johns Hopkins University Baltimore, MD 21218 USA [email protected] Da-Wei Gu University of Leicester University Road Leicester, LE1 7RH [email protected] Argyrios C. Zolotas Loughborough University Leicestershire Loughborough, LE11 2TJ [email protected] Roger M. Goodall Loughborough University Leicestershire Loughborough, LE11 2TJ [email protected] Robert F. Harrison The University of Sheffield Mappin Street Sheffield, S1 3JD [email protected]

1 H∞ Control Design Declan G. Bates

Summary. This chapter summarises the basic ideas behind H∞ control theory.

1.1 Introduction Classical design and analysis techniques, many of which date back to the 1950s, are still widely used in industry for the design and analysis of automatic control systems. The con tinued success and popularity of these techniques is particularly impressive considering the radical advances in systems and electronics made over this period. Clearly, an understand ing of both the advantages and limitations of these methods is necessary in order to properly evaluate the more modern techniques for the design and analysis of robust control systems de scribed in this chapter. We therefore begin by brieﬂy summarising the philosophy, advantages and limitations of the classical approach to control system design and analysis. Philosophy •

•

•

Classical control system design is based on the philosophy of successive loop closure. This means that even when more than one plant output is required to be controlled simul taneously, single-input-single-output (SISO) design techniques such as root-locus, Bode and Nyquist plots are used in a sequential design process, with individual controllers for each feedback loop being designed one at a time. As will be clear from the above, classical control systems are highly structured. Indeed, the actual structure of the control system may be completely determined in advance, with the control law designer only left to determine the values of particular gains and ﬁlters. Although the reasons for adopting a particular controller structure are generally quite com plex, at least three important factors can be identiﬁed. The ﬁrst is the limitations imposed by the various hardware components required to implement the controller on the actual system. The second is the natural tendency to exploit previous design experience and ‘lessons learned’ by basing each successive generation of control systems on the previous one, with only those modiﬁcations which are shown to be absolutely necessary being im plemented. The third reason is to assist in the process of certiﬁcation of the overall system it is easier to demonstrate the correct operation (in normal and failure cases ) of the con trol system in the case of highly structured systems, where every element of the control system has a high level of visibility and functionality. Classical design and analysis techniques are predominantly linear, and are based around linearised models of the plant generated at particular operating conditions. Modiﬁcations

M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 3-46, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

4

•

D.G. Bates to the control law to cope with non-linear effects such as actuator rate and magnitude saturation are usually added after the initial linear design is complete, based on non linear closed loop simulations. Extension of the control law to cover the full non linear operating envelope is usually achieved by scheduling the gains of the controller as functions of one or more parameters such as speed, dynamic pressure or angle of attack. Finally, analysis of the control system is usually carried out using linear stability metrics such as Gain and Phase margins, before final checks are conducted using non linear simulations. Classical control techniques make use of predominately frequency domain tools (Bode and Nyquist plots, Root locus design, Gain and Phase margins, etc.) to achieve designs which satisfy closed loop specifications which are themselves often given in the time as well as the frequency domain. Despite this apparent gap between specifications and design tools, there is clearly a consensus that frequency domain tools offer a level of design transparency and intuition that cannot be matched by purely time domain techniques. This fact, together with the questionable robustness properties of some time-domain ‘optimal’ control techniques, has probably been responsible for the relative lack of success of these optimal techniques in real applications to date.

A key theme of this chapter is that the robust controller design and analysis methods it describes are direct and natural developments of the classical frequency domain control techniques still in common use in industry today. This fact is often obscured, for two main reasons. The first is largely a historical one. Before the advent of robust control theory in the early 1980s, much attention and publicity was given to so-called ‘optimal’ or modern control theory and its associated methods - state-space equations and linear quadratic optimisation for controller synthesis, [1]. These concepts were thought of, and taught as, time-domain methods. This fact, together with the introduction of ‘optimisation’ as a tool for controller design, represented a complete break from the traditional classical methods. Unfortunately, however, the fundamental problem of system uncertainty was largely ignored by modern control theorists. The not altogether surprising result was that supposedly ‘optimal’ controllers were often found to give poor performance, or even instability, in real-world applications because of high sensitivity to modelling errors. The well known and well-publicised 60-degree guaranteed phase margins of LQR full state feedback, for example, were found to disappear when Kalman filters were used to estimate unmeasurable states in the LQG control method [2]. Although certain industries have successfully applied modern control methods to control problems through careful attention to robustness margins [3], modern control techniques are still not widely used. The resulting lack of successful applications of modern control techniques has seriously hampered their acceptance by industry, with the result that many industrial control system designers today still rely on classical methods for both analysis and design. Robust control theory was born out of a belated realisation of the flaws inherent in the time domain modern control methods. In particular, the lack of a sensible way of handling system uncertainty, and the loss of design transparency associated with time-domain methods were the starting point for the development of H∞ control, [4]. Thus, robust control should properly be seen as a return to the frequency domain philosophy of classical control, with the primary development being to allow for truly multivariable design in the presence of system and signal uncertainty. The second issue obscuring the link between robust and classical control techniques is the relative complexity of the mathematical machinery required to solve H∞ optimisation problems. The generation of efficient and elegant solutions to these challenging problems has been the subject of intensive research over the past twenty years, and therefore it is not surprising that much of the robust control literature seems from the outside to have been written by, and for,

H

∞

Control Design

5

system theorists and mathematicians, rather than practising control engineers. With many of these computational problems now settled, however, more attention is being devoted to the development and application of robust control methods as practical engineering tools, rather than as mathematical concepts, [38]. The recent significant increase in successful implementations of robust controllers in a wide variety of applications strongly suggests that these techniques are reaching maturity and have finally begun to achieve widespread acceptance in industry.

1.2 Design Speciﬁcations and Fundamental Trade-Offs In this section we will describe the basic set of design specifications that must be satisfied by robust control systems. Since many of these specifications are to a greater or lesser extent contradictory, we will also analyse the various fundamental trade-offs which exist in all multivariable control systems. Note that the following development is based on the assumption that many of the most important specifications for control systems can be addressed in a linear controller synthesis framework. In this approach, any additional non-linear specifications (e.g., limiting of safety critical signals, mode-switching and scheduling logic, anti-windup for actuator saturation, etc.) are addressed via conditioning and scheduling schemes, after the initial linear controller (or set of controllers) has been designed. This approach follows exactly the approach used in classical control system design, i.e., linear controller synthesis followed by the addition of scheduling, anti-windup and safety-limiting schemes. The fundamental difference between the two approaches lies in the additional complexity of the nonlinear schemes needed in the case of multivariable linear controllers. In particular, scheduling and anti-windup schemes for classical SISO controllers are relatively easy to design and implement - in the case of multivariable (especially dynamic multivariable) linear controllers these two tasks become more difficult. However, significant progress has been made in these areas in recent years, with the result that systematic and powerful techniques are now available to solve these problems. An alternative approach to that outlined above is to employ non-linear controller design methods which would allow non-linear controllers to be synthesised directly, and indeed much attention has been devoted to developing such techniques in recent years. Two of the most important of these techniques are the so-called Sliding Mode, [5], and Non-Linear Dynamic Inversion (NDI), [6,7], approaches. In addition, non-linear versions of the H∞ controller design techniques discussed in this text have also been developed, [8, 9] and are starting to be applied to real applications. Although all these methods show great promise, they also have some inherent disadvantages. NDI techniques, for example, provide no robustness guarantees, and it is not even clear how to ‘tune’ an NDI design to improve its robustness (although see [10] for some significant progress in this direction). Indeed, a fundamental problem with all non-linear controller synthesis methods is that corresponding non-linear robustness analysis methods are in general much less well developed. Sliding Mode and non-linear H∞ techniques do offer sensible robustness guarantees, but employ mathematical techniques of a complexity far beyond what is generally used in industry. Finally, it can be argued that attempts to address simultaneously linear and non-linear aspects of the controller synthesis process actually lead to additional complexity and reduced visibility in the final design - one step at a time can be faster and simpler. The above discussion motivates our approach to the problem of synthesising controllers as set out in the following sections.

6

D.G. Bates

1.2.1 Linear Design Speciﬁcations for Robust Control Systems We summarise the most important linear design speciﬁcations for a robust control system below. Note that in the following we consider the design of a single linear controller at a particular point in the plant’s operating envelope. 1. Robust stability to unstructured system uncertainty: the linear controller must provide adequate levels of stability against unstructured uncertainty (as defined by some robustness measure) for the nominal linearised model of the plant. 2. Robust stability to unstructured and structured system uncertainty: the linear controller must provide adequate levels of stability against unstructured uncertainty (as defined by some robustness measure) for all linear models of the plant in some set defined by a structured uncertainty model. 3. Nominal performance: the linear controller must provide adequate attenuation of disturbance signals (at the plant sensors and actuators), adequate attenaution of measurement noise, and adequate tracking and decoupling of reference commands, for the nominal linearised model of the plant. 4. Robust performance: the linear controller must provide adequate attenuation of disturbance signals (at the plant sensors and actuators), adequate attenaution of measurement noise, and adequate tracking and decoupling of reference commands, for all linear models of the plant in some set defined by a structured uncertainty model. Items 3 and 4 above are, in effect, specifications on signal uncertainty (recall that good tracking of reference demands amounts to attenuating the effect of the uncertain command signal on the tracking error signal). The specifications must be satisfied nominally (in the absence of system uncertainty) and robustly (in the presence of system uncertainty). For our current purposes we will adopt the classical approach and consider adequate performance with respect to reference demands to imply well damped responses with appropriate rise/settling times, zero steady-state error and minimal coupling into other controller variables. In the following subsection , we will show how each of the design speciﬁcations given above can be formulated in terms of bounds on the minimum or maximum singular values of various open and closed-loop transfer function matrices.

1.2.2 Frequency Domain Design Speciﬁcations and Fundamental Trade-Offs Consider the multivariable feedback control system shown in Figure 1.1. Assuming that the system is internally stable, the following fundamental equations hold: y = TO r + SO GdI + SO dO − TO m,

u = KSO r − KSO dO + SI dI − KSO m

(1.1) (1.2)

Thus, nominal speciﬁcations on signal uncertainty (i.e., speciﬁcations in the absence of system uncertainty) can be deﬁned in terms of the maximum singular values of the input and output sensitivity and complementary sensitivity functions SI , SO , TI and TO . In particular, we can formulate the following design objectives: For attenuation of output disturbance signals at the plant output, σ (SO ) should be small. For attenuation of output disturbance signals at the plant input, σ (KSO ) should be small.

H

∞

Control Design

dI

r

��

� K

7

dO

�� G u

� ��

�y ��

m

Fig. 1.1. A typical feedback control system For attenuation of input disturbance signals at the plant input, σ (SI ) should be small. For attenuation of input disturbance signals at the plant output, σ (SO G) should be small. For attenuation of measurement noise signals at the plant output, σ (TO ) should be small. For attenuation of measurement noise signals at the plant input, σ (KSO ) should be small. For good reference tracking σ (TO ) and σ (TO ) should be ≈ 1. Thus, good reference track ing is equivalent to attenuation of output disturbances at the plant output, since forcing SO to 0 forces TO to the identity. For avoidance of large control signals due to reference demands (and hence actuator sat uration), σ (KSO ) should be small. Requirements for robust stability in the presence of unstructured system uncertainty can also be stated in terms of the maximum singular values of these sensitivity functions: For robust stability to input multiplicative uncertainty, σ (TI ) should be small. For robust stability to output multiplicative uncertainty, σ (TO ) should be small. For robust stability to additive uncertainty, σ (KSO ) should be small. Now, since SI + TI = I

(1.3)

SO + TO = I

(1.4)

it is obvious that all of the above requirements cannot be satisfied simultaneously. Feedback controller design is therefore a problem of managing trade-offs between conflicting design objectives. In practice, this problem is made easier by the fact that different design objectives are often important over different frequency ranges. For example, disturbance attenuation is generally required at low frequencies, while the attenuation of measurement noise becomes important at high frequencies. Finally we note that the constraints outlined above do not constitute a complete list of the limitations on the achievable performance of a feedback control system. In particular, right-half-plane zeros in the plant, and constraints on control signal rates and amplitudes can seriously limit closed-loop performance - see [38] for a full discussion. As we shall see in the next section, the so-called mixed sensitivity H∞ controller design method provides a powerful technique for synthesising controllers which optimally satisfy the various design objectives described above. Before examining this method in detail, we note at this stage that the singular value design objectives given above correspond to the case of unstructured plant uncertainty. For control design problems where the plant uncertainty is

8

D.G. Bates

structured (e.g., diagonal), the structured singular value µ , and its associated synthesis techniques, [38], may be used to obtain less conservative designs - see [19] for further details.

1.3 Mixed-Sensitivity H∞ Controller Design In this section we describe the mixed sensitivity H∞ controller design method, and illustrate its application via a control law design for a model of the Canadian NRC’s Bell 205 variable stability helicopter.

1.3.1 Formulating the Problem Consider the general feedback conﬁguration shown in Figure 1.2. The various signals in the ﬁgure are: u the control variables, v the signals which the controller has access to, w the exogenous signals such as disturbances, reference demands, measurement noise, etc., and z the

� �

w

z� P

u K

v

�

Fig. 1.2. A general conﬁguration for H∞ controller design

controlled variables, typically error signals and control signals which are to be minimised in some sense to meet the control objectives. Now, if we partition the plant P compatibly with K, the closed-loop transfer matrix from w to z is given by a lower linear fractional transformation (LFT): z = {P11 + P12 K(I − P22 K)−1 P21 } w

(1.5)

= Fl (P, K) w

(1.6)

The standard H∞ optimal control problem is then the minimisation of the H∞ -norm of Fl (P, K) over all stabilising controllers, i.e. inf k Fl (P, K)( jω ) k∞ = inf max σ (Fl (P, K)( jω )) = inf max K

K

ω

K

w= 6 0

k z( jω ) k2 k w( jω ) k2

(1.7)

In practice, it is usually not necessary to compute the optimal H∞ controller, and it is com putationally (and theoretically) simpler to design a sub-optimal controller via the following iterative procedure. Let γmin be the minimum value of k Fl (P, K)( jω ) k∞ over all stabilising

H

∞

Control Design

9

controllers K. Then the H∞ sub-optimal control problem is: given a γ > γmin , find all stabilising controllers K such that k Fl (P, K)( jω ) k∞ < γ

R This problem can be solved efficiently using standard MATLAB software which uses the algorithm of Doyle et al., [20] - see below. By iteratively reducing the value of γ an optimal solution is approached. So far, we have described the H∞ control problem in its most general setting. We now need to show how the various singular value control objectives discussed in the previous subsection can be represented under the H∞ control framework. This is the task of formulating the H∞ control problem: for the conﬁguration of Figure 1.2, how do we construct the generalised plant P and which signals do we include in z and w, in order to reﬂect our design objectives on SI , SO , TI and TO ? This task clearly depends on which particular design objectives we are trying to satisfy, and is best illustrated by example.

Example 1.1. S/KS Mixed-sensitivity design Consider the problem of synthesising a con troller K to satisfy the following design objectives: • • •

Attenuation of low frequency disturbances at the plant output Minimisation of high frequency actuator usage Robust stability to additive uncertainty at high frequencies

These control objectives correspond to a regulation problem, where we simply want to maintain the output of the plant at some set-point in the presence of signal and system uncertainty. Now, the first two design objectives given above amount to requiring σ (SO ) to be small at low frequencies and σ (KSO ) to be small at high frequencies, since as we have seen, SO is the transfer matrix between dO and the output, and KSO is the transfer matrix between dO and the control signals. Note that minimisation of KSO at high frequencies will also satisfy the third objective of providing good robustness to additive uncertainty. We are thus interested in shaping the closed-loop transfer matrices SO and KSO as a function of frequency. We can do this by selecting frequency dependent weighting functions W1 (s) and W2 (s) and computing a stabilising controller K which minimises the cost function

W1 SO

W2 KSO ∞

where W1 (s) is a low pass ﬁlter with a bandwidth equal to that of the disturbance and W2 (s) is a high pass ﬁlter with a crossover frequency approximately equal to that of the desired closed-loop bandwidth. Note that in general the weighting functions will be (usually diagonal) transfer matrices. The S/KS mixed sensitivity minimisation problem given above can be put into the standard H∞ control conﬁguration as shown in Figure 1.3. In this set-up, the output disturbance signal dO is chosen as the single exogenous input w, while the vector of controlled variables is given by z = [zT1 zT2 ]T , where z1 = W1 y and z2 = −W2 u. It is then easy to show that z1 = W1 SO w and z2 = W2 KSO w. Thus, the generalised plant P is given by the relation   z1 w  z2  = P11 P12 (1.8) P21 P22 u v

10

D.G. Bates

P

w=d

�

G

u

�� +

� W1

�

z1

� −W2

�

z2

-

+

y

�� + �

v

r=0

K

�

Fig. 1.3. S/KS mixed sensitivity optimisation in standard form (regulation) where P11 = and

W1 W1 G , P12 = , P21 = −I, P22 = −G 0 −W2 Fl (P, K) =

W1 S W2 KS

(1.9)

(1.10)

The S/KS mixed sensitivity optimisation can also be formulated in terms of a tracking prob lem, as shown in Figure 1.4. In this formulation, the exogenous input signal w is a reference command r, and the vector of controlled variables z is made up of z1 = −W1 e = W1 (r − y) and z2 = W2 u. It is left to the reader to verify that for this set-up we again have that z1 = W1 SO w and z2 = W2 KSO w. Example 1.2. Consider the following set of controller design objectives: Attenuation of low frequency disturbances at the plant output Attenuation of high frequency measurement noise Tracking of low frequency reference commands Robust stability to output multiplicative uncertainty at high frequencies Formulate a mixed sensitivity H∞ control problem to satisfy these objectives. Construct the generalised plant P required to cast this problem in the standard H∞ control conﬁguration. Note that when performing numerical calculationsthe generalised plant P can be easily calcu R lated, either directly by using the appropriate functions in MATLAB or by constructing the R R plant as a Simulink block diagram and then using the MATLAB function linmod.

1.3.2 Weighting Function Selection As will by now be apparent, the process of designing controllers using the H∞ mixed sensitiv ity method essentially revolves around choosing (and usually iteratively modifying) weighting

H

P

w=r

� u

G

��

∞

Control Design

11

� W1

�

z1

� W2

�

z2

+

y

-

K

v

�

Fig. 1.4. S/KS mixed sensitivity optimisation in standard form (tracking) functions in order to ‘shape’ certain closed-loop transfer matrices over frequency. In general, weighting function selection becomes difﬁcult when the ‘stacked’ cost functions contain more than two terms. For two terms, e.g., S/KS or S/T, the process is relatively simple, however, since the bandwidth requirements on each term are usually complementary, and simple, stable low-pass and high-pass ﬁlters are sufﬁcient to carry out the required shaping over frequency. Note that the weights Wi in mixed sensitivity H∞ control must all be stable and proper (this is due to some assumptions made about the generalised plant by the algorithm to compute the H∞ controller - see the next section). Thus, for example, if we wish to provide zero steadystate tracking error by weighting S with a term including integral action, we have to approx imate 1s by s+1 ε , where ε << 1. Similarly, in order to ensure that KS rolls-off at frequencies above the desired system bandwidth we cannot use a non-proper weight of the form (1 + τ1 s). Instead we use a weight of the form (1 + τ1 s)/(1 + τ2 s) where τ2 << τ1 . A useful discussion of the ‘tricks’ involved in using unstable and non-proper weights in H∞ design is contained in [21]. As will be clear from the design example below, however, the restrictions on the weighting functions described above do not generally pose a problem in practice - selection and iteration of H∞ weighting functions is straightforward because the relationship between these functions and the controller design objectives is relatively transparent, [22]. This con trasts markedly with the situation in, for example, H2 and L1 design methods.

1.3.3 Solution of the H∞ Control Problem Before considering a design example using mixed sensitivity H∞ optimisation, we give a brief outline of the algorithms used to compute the corresponding H∞ controllers. The most gen eral algorithms for H∞ (and H2 ) control problems are the solutions proposed by Glover and Doyle, [18], and Doyle et al., [20]. These algorithms are widely used and are available in, for R example, the relevant MATLAB toolboxes. The simplest form of the solution for the sub optimal H∞ control problem is given as follows:

12

D.G. Bates

Consider the general control configuration of Figure 1.5, with the generalised plant P given in state-space form as



 A B1 B2 P =  C1 D11 D12  C2 D21 D22

� �

w

(1.11)

z� P

u K

�

y

Fig. 1.5. General H∞ control conﬁguration Then there exists a stabilising controller K(s) such that k Fl (P, K) k∞ < γ if and only if (i) X∞ ≥ 0 is a solution to the algebraic Riccati equation AT X∞ + X∞ A +C1T C1 + X∞ (γ −2 B1 BT1 − B2 BT2 )X∞ = 0

(1.12)

such that Re λi [A + (γ −2 B1 BT1 − B2 BT2 )X∞ ] < 0, ∀i; and (ii) Y∞ ≥ 0 is a solution to the algebraic Riccati equation AY∞ +Y∞ AT + B1 BT1 +Y∞ (γ −2C1T C1 −C2T C2 )Y∞ = 0

(1.13)

such that Re λi [A +Y∞ (γ −2C1T C1 −C2T C2 )] < 0, ∀i; and (iii) ρ (X∞Y∞ ) < γ 2 All such controllers are then given by K = Fl (Kc , Q), where 

 A∞ −Z∞ L∞ Z∞ B2 Kc =  F∞ 0 I  −C2 I 0

(1.14)

F∞ = −BT2 X∞ , L∞ = −Y∞C2T , Z∞ = (I − γ −2Y∞ X∞ )−1

(1.15)

with

H

∞

Control Design

A∞ = A + γ −2 B1 BT1 X∞ + B2 F∞ + Z∞ L∞C2

13 (1.16)

and Q(s) is any stable proper transfer function such that k Q k∞ < γ . For Q(s) = 0, we get K(s) = Kc11 (s) = −Z∞ L∞ (sI − A∞ )−1 F∞

(1.17)

This is the so-called ‘central’ controller, and it has the same number of states as the generalised plant P. To obtain a controller which achieves γmin , to within a speciﬁed tolerance, we can perform a bisection on γ until its value approaches γmin sufﬁciently accurately - note that this R bisection is automated in the MATLAB function hinfsyn. The simplified solution to the sub-optimal H∞ control problem outlined above requires several assumptions about the generalised plant P - see pp. 363 of [38] for the full list. More general solutions are also available which allow some of the assumptions to be relaxed. In practice, however, these assumptions are non-restrictive to the point that most sensibly posed control problems will meet them. Thus if the software being used to compute K complains, then it probobly means that the control problem is not well formulated (e.g., the weighting functions are not ‘sensible’) and you should think again.

1.3.4 Design Example: Control Law Design for the Bell 205 Helicopter Small general purpose helicopters such as the Bell 205 have a broad spectrum of tasks to perform, often in bad weather and low level environments. In order to achieve the performance required during these tasks without excessive pilot workload, helicopters need to be agile yet easy to control. Typical rotorcraft are, however, naturally unstable, and also exhibit signiﬁcant inter-axis coupling. This coupling is inherent to the vehicle and is due to the asymmetric forces and moments produced by the main and tail rotor systems. Thus, even relatively simple manoeuvres require complicated multivariable control inputs, resulting in high pilot workload. Consider, for example, a side-step task, where the pilot attempts to laterally translate the rotocraft using mainly rolling commands. Roll inputs will dominate the response, but at the same time the pilot needs to coordinate heading angle (with the pedals), height (with the collective lever) and keep track over the ground (using longitudinal inputs). Heading excursions are caused by the vertical offset of the tail rotor thrust and the vertical fin sideforce from the centre of mass. Height does not remain constant as the main rotor thrust is directed sideways and thus lift no longer counter-balances the vertically directed rotocraft’s weight. Also, any lateral demands cause longitudinal motion as the thrust is not always normal to the main rotor disk, and the inclination of the lift vectors on individual blade sections contributes a significant amount of off-axis inputs. From the above discussion it will be clear that the design of full-authority control laws for helicopters is a particularly challenging task, requiring a multivariable strategy. Additional difficulties are introduced by the absence, in general, of accurate models of rotocraft dynamics which can be used in the controller design process, since the equations governing the motion of helicopters are complex, and difficult to formulate with high levels of precision. For example, the rotor dynamics are particularly difficult to model. The use of robust multivariable controller design methods is therefore essential in order to generate control laws which will provide adequate stability and handling qualities on the real vehicle.

14

D.G. Bates

Research on the application of H∞ robust control methods to rotocraft control problems has been on-going in the Control and Instrumentation research group at the University of Leicester for over 10 years. Among the specific vehicles used in these studies have been the GKN Westland Lynx and the NRC Bell 205. Designs using a variety of H∞ control methods have been successfully implemented and evaluated in piloted simulations trials (at DERA Bedford’s fixed-based simulation facility) and in test-flights at the Canadian National Research Council Flight Research Labs., Ottawa. More details of this research is available in the following references, [16, 23, 24]. In this (much simplified) design example, we consider the design of a full-authority controller for the Bell 205 at a low speed flight condition using a simple S/KS mixed sensitivity approach. The vehicle model used in this example is a linearised representation of the Bell 205 dynamics at a 10 knot, sea level flight condition with a central centre of gravity configuration, [25]. The corresponding state-space model is given by

x˙ = Ax + Bu

(1.18)

y = Cx

(1.19)

with u 

−0.0036  −0.1841   0.0062   0.0149 A=  0.0195   −0.0184   0 0 DC 

0.0741  −1.1351   0.0027   −0.0270 B=  −0.0309   0.1570   0 0

w

q

0.0300 −0.4456 −0.0091 −0.0016 −0.0116 −0.0204 0 0

0.2490 0.3393 −0.2695 −0.4157 −0.8566 0.0274 1.0000 0

DB

DA

0.1236 0.0594 −0.0673 0.0027 0.0054 0.0011 0 0

−0.0007 0.0010 0.0003 0.1062 0.2216 0.0318 0 0

v

p

−0.0056 −0.0512 0.0066 −0.0544 −0.0396 0.0692 0 0

−0.4154 −0.1812 0.2333 −0.3341 −0.6855 −0.3037 0 1.0000

r θ φ  −0.0795 6.7753 0 0.6229 3.8927 5.6643    0.0250 0 0  0.2726 −5.9312 −3.7175    0.1429 0 0   −0.7329 0 0   0 0 0 0 0 0

DP  −0.0025 −0.0034    0.0001  0  0 0.1927  C =  0 0.1625    −0.4636  0  0 0

1 0 0 0

0 0 0 0

0 0 0 0

0 0 0 0

0 0 0 1

0 1 0 0

 0 w(m/sec) 0  θ (rads) 1  φ (rads) 0 r(rads/sec)

Comparisons of this model with ﬂight test data suggests that it captures the salient rigid body modes relatively well, but that the absense of the rotor dynamics introduces signiﬁcant (un structured) uncertainty over a wide frequency range, [24, 26]. The controller (or pilot in man ual control) generates four blade angle commands which are effectively the helicopter inputs, since the actuators (which are typically modelled as ﬁrst order lags) are modelled as unity gains in this example. The blade angles are thus DC, main rotor collective, DB, longitudinal

H

∞

Control Design

15

cyclic, DA, lateral cyclic and DP, tail rotor collective. The action of each of these blade angles can briefly be described as follows. The main rotor collective changes all the blades of the main rotor by an equal amount and so roughly speaking controls lift. The longitudinal and lateral cyclic inputs change the main rotor blade angles differently thereby tilting the lift vector to give longitudinal and lateral motion respectively. The tail rotor is used to balance the torque generated by the main rotor and so stops the helicopter spinning round; it is also used to give lateral motion. This description outlines the primary effect of each input - as discussed above, however, in reality all inputs affect all outputs since the vehicles dynamics are highly coupled. At this flight condition the helicopter is also open-loop unstable. We consider the problem of synthesising a controller K to satisfy the following design objectives:

• • •

Close tracking and decoupling of pilot commands on all four outputs Minimisation of control signals to avoid actuator saturation Robust stability to additive uncertainty at high frequencies - maximum closed-loop band width of 4 rads/sec

These speciﬁcations can be addressed under the framework of mixed sensitivity S/KS H∞ op timisation as follows. Consider the problem formulation illustrated in Figure 1.6. The process

w=r (pilot demands)

u (blade angles)

P

� G

� W1

� z1

� W2

� z2

�+ �� y (measured variables) K

v tracking errors

�

Fig. 1.6. S/KS mixed sensitivity formulation for Bell 205 design of designing a control law to meet the speciﬁcations given above can now be broken down into three stages. First we need to choose diagonal weighting matrices W1 and W2 which capture the dynamic requirements on the closed-loop system. Then we compute a stabilising controller K which minimises the cost function

W1 SO

W2 KSO ∞ Finally we analyse the performance and robustness properties of the resulting closed-loop sys tem to check if our design speciﬁcations have been satisﬁed. If they have not been, we need to adjust the weighting functions and repeat the process.

16

D.G. Bates

Initial Weighting function selection We ﬁrst of all consider the choice of W1 , recalling that the inverse of this weighting func tion deﬁnes a desired upper bound on the magnitude of S. For tracking problems of the type considered here, a convenient choice for this weighting function is W1 = diag{w1i }, where w1i =

s/M + ω ∗

B

s + ωB∗ A

(1.20)

This weighting function is parameterised by the terms M,A and ωB∗ which have the following clear meanings: 1/|w1i | equals A ≤ 1 at low frequencies, equals M ≥ 1 at high frequencies, and crosses the 0 dB line at approximately ωB∗ . For the problem considered here we require approximate integral action for tracking of pilot commands, and thus we choose A = 0.001 to force S to be ≈ 0 at low frequencies. A typical initial value for M is 2, while ωc∗ is set equal to 4 to reflect our desired bandwidth for the closed-loop system. For simplicity we set equal weights on all four channels. A Bode magnitude plot of W1 is shown in Figure 1.7. The weighting function W2 now needs to be chosen to reflect our specifications on limiting of control signals as well as the robust stability requirements. Since the maximum allowable values for all four blade angles are similar, a reasonable initial choice for W2 is W2 = I4×4 . H∞ Controller synthesis In order to compute a stabilising controller to minimise our S/KS cost function we need to conR struct the generalised plant P. This is easily achieved using the MATLAB function sysic. The function hinfsyn then computes a stabilising controller using a bisection algorithm to minimise γ between initial upper and lower bounds supplied by the user (typical values are 10 and 0.1). For our initial weighting functions given above a stabilising controller returning a value of γ less than 10 could not be found and so it was necessary to reduce the weight on KS. For W2 = diag{0.01} a stabilising controller was computed which returned a value of γmin = 2.8. Although this value of γ is still greater than 1 (indicating that our weighting function specifications for upper bounds on S and KS have not been completely satisfied), at this point we proceed to the closed-loop analysis to get an idea of how the controller is performing. Initial closed-loop analysis Initial time domain simulations for the closed-loop system indicate good nominal performance properties - the responses for a 25◦ demand in pitch attitude shown in Figure 1.8 show a fast response, with small overshoot and little coupling into other controlled variables. This good performance is explained by inspection of the singular value plot of S, which shows that the weighting function upper bound has almost been met at most frequencies - Figure 1.9. Analysis of the KS and T closed-loop transfer matrices (Figures 1.10 and 1.11) shows, however, that the sizes of the control signals produced by the controller are quite large, and that the closed-loop bandwidth is around 10 rads/sec - much higher than the 4 rads/sec required to satisfy our robust stability specification. This situation is verified by Figure 1.12 which shows the large, high-frequency movement of the blade angles required to produce the responses shown in Figure 1.8. Since the maximum allowable deflection of the longitudinal cyclic is only 24◦ , we can see that for aggresive pilot demands actuator saturation could easily occur with this controller.

H

∞

Control Design

17

Adjustment of weighting functions From the analysis above, we can see that our design is ‘un-balanced’ - too much emphasis is being given to the nominal performance specifications and not enough to the control signal limiting / robust stability specifications. We could try to address this problem by simply increasing the size of the static weight on KS, (or by reducing the magnitude of W1 ). A more refined strategy, however, is to make W2 dynamic, in order to penalise more high frequency signals and thus also force down the closed-loop bandwidth. We thus set W2 = diag{w2i }, where w2i is equal to the high pass filter

w2i =

2s + 0.002 s+5

(1.21)

The singular value plot of W2 is shown in Figure 1.13. H∞ controller synthesis For our new weighting functions, hinfsyn returned a stabilising controller with γmin = 7.72. This higher value of γ implies that we have ‘tightened’ our speciﬁcations, i.e., increased (over certain frequencies) the weighting on KS, without changing the weight on S. This should result in a more balanced design, with some degradation in nominal performance but better robust stability and smaller control signals. Closed-loop analysis The response of the system to the same 25◦ demand in pitch attitude is shown in Figure 1.14. The response is more sluggish, with greater coupling into the other controlled variables, but is still acceptable. As expected, the singular value plot of S (Figure 1.15) shows that the upper bound speciﬁed by the inverse of W1 has been violated more, with the resulting decrease in nominal performance. The singular value plots for KS and T shown in Figures 1.16 and 1.17, hovever, show that the dynamic weight W2 has had the intended effect of ‘pushing down’ the control signals at high frequencies. Note in particular from the T plot that the closed-loop bandwidth is now almost exactly 4 rads/sec. The effect of W2 is also clearly seen in the time domain plot of the control signals produced by the demand in pitch attitude, Figure 1.18. Note in particular that the high frequency longitudinal cyclic signal has been dramatically reduced in magnitude, whereas the lower frequency tail rotor collective signal has been left largely unchanged. The controller design described above is obviously just a ‘ﬁrst attempt’ and is intended mainly to illustrate the way in which weighting functions are used to satisfy design speciﬁcations in the H∞ mixed sensitivity method. Many reﬁnements could be considered in order to produce a ﬁnal design. Different weights could be used in each channel of W1 and W2 to set performance versus robustness trade-offs for each controlled variable individually. Extra measurement sig nals (such as pitch and roll rates) could be fed back to the controller. Finally, a two-degree-of freedom control structure could be adopted in order to improve the time-domain responses for particular pilot demands.

18

D.G. Bates

W1 weight − multivariable 60

50

Singular Values (dB)

40

30

20

10

0

−10 −4 10

−3

−2

10

−1

10

0

10

1

10

2

10

10

Frequency (rad/sec)

Fig. 1.7. Singular value plot of W1 weighting matrix 30

theta, phi (degrees), r (deg/sec), w (feet/sec)

25

20

15

10

5

0

−5

0

2

4

6

8

10 time (secs)

12

14

16

18

20

Fig. 1.8. Closed-loop response of controlled variables to a pilot demand on θ

H

∞

Control Design

19

Singular Values 10

1/(W1) 0

−10

Singular Values (dB)

−20

−30

−40

−50

−60

−70

−80 −4 10

−3

−2

10

−1

10

0

10

1

10

2

10

10

3

10

Frequency (rad/sec)

Fig. 1.9. Singular value plot of S with inverse of W1 weighting matrix Singular Values 50

1/(W2) 40

Singular Values (dB)

30

20

10

0

−10

−20 −3 10

−2

10

−1

10

0

10

1

10

2

10

3

10

Frequency (rad/sec)

Fig. 1.10. Singular value plot of KS with inverse of W2 weighting matrix

20

D.G. Bates

Complemenatry−Sensitivity T 20

0

−20

−40

Gain

−60

−80

−100

−120

−140

−160 −3 10

−2

−1

10

0

10

1

10

2

10

3

10

10

Frequency − rad/s

Fig. 1.11. Singular value plot of T 20

15

DC, DB, DA, DP (degrees)

10

5

0

−5

−10

DB DP −15

−20

0

2

4

6

8

10 time (secs)

12

14

Fig. 1.12. Control signals for θ demand

16

18

20

H

∞

Control Design

21

W2 weight − multivariable 10

0

Singular Values (dB)

−10

−20

−30

−40

−50

−60

−70 −4 10

−3

−2

10

−1

10

0

10

1

10

2

10

10

Frequency (rad/sec)

Fig. 1.13. Singular value plot of dynamic W2 weighting matrix 30

theta, phi (degrees), r (deg/sec), w (feet/sec)

25

20

15

10

5

0

−5

0

2

4

6

8

10 time (secs)

12

14

16

18

20

Fig. 1.14. Closed-loop response of controlled variables to a pilot demand on θ

D.G. Bates

Sensitivity S 10

1/(W1) 0

−10

Gain

−20

−30

−40

−50

−60 −3 10

−2

−1

10

0

10

10

1

10

2

10

3

10

Frequency − rad/s

Fig. 1.15. Singular value plot of S with inverse of W1 weighting matrix KS 80

60

1/(W2) 40

20

Gain

22

0

−20

−40

−60 −3 10

−2

10

−1

10

0

10

1

10

2

10

3

10

KS − rad/s

Fig. 1.16. Singular value plot of KS with inverse of W2 weighting matrix

H

Control Design

∞

23

Complemenatry−Sensitivity T 20

0

−20

Gain

−40

−60

−80

−100

−120

−140 −2 10

−1

0

10

1

10

2

10

10

Frequency − rad/s

Fig. 1.17. Singular value plot of T 4

2

0

DC, DB, DA, DP (degrees)

−2

DB −4

−6

−8

DP −10

−12

−14

−16

0

2

4

6

8

10 time (secs)

12

14

Fig. 1.18. Control signals for θ demand

16

18

20

24

D.G. Bates

1.4 H∞ Loop-Shaping Controller Design In this section, we describe an alternative method for designing robust multivariable control systems using H∞ optimisation. The method was ﬁrst proposed by MacFarlane and Glover in [39], and has subsequently been developed and applied to a wide number of applications, see for example [14, 15, 27–31] and the references therein. The method uses weighting func tions to shape the system’s open-loop transfer matrix L, rather than the closed-loop functions S, KS and T , and focusses on maximising robustness to coprime factor uncertainty, rather than multiplicative or additive uncertainty. This leads to a very attractive controller design procedure which yields controllers with strong robustness properties.

1.4.1 Fundamental Trade-Offs in Terms of L In this section we revisit the fundamental design tradeoffs derived in Section 2.1, this time writing them in terms of the input and output open-loop transfer matrices LI and LO . We will make use of the following properties of the singular values of a matrix (see pp. 505 of [38] for the proofs):

σ (A−1 ) = 1/σ (A) 1 σ (A) − 1 ≤ ≤ σ (A) + 1 σ (I + A)−1

(1.22) (1.23)

Thus, in the case that σ (KG) > 1 and σ (GK) > 1, then 1 1 ≤ σ (SI ) σ (KG) + 1 σ (KG) − 1 1 1 ≤ σ (SO ) ≤ σ (GK) + 1 σ (GK) − 1

(1.24) (1.25)

This gives the following set of equivalencies:

σ (SO ) << 1 ⇔ σ (GK) >> 1 σ (SI ) << 1 ⇔ σ (KG) >> 1

(1.26) (1.27)

Also, from the fact that TI = KG(I + KG)−1 and TO = GK(I + GK)−1 , it follows that

σ (TI ) ≈ σ (KG) at frequencies where σ (KG) is << 1

(1.28)

σ (TO ) ≈ σ (GK) at frequencies where σ (GK) is << 1

(1.29)

σ (KSO ) ≈ 1/σ (G) at frequencies where σ (GK) is >> 1

(1.31)

σ (SO G) ≈ 1/σ (K) at frequencies where σ (GK) is >> 1

(1.30)

From the above relationships, we can formulate the following equivalent set of design objec tives to those listed in Section 2.1 in terms of the input and output open-loop transfer matrices KG and GK:

H •

•

• • • •

• •

∞

Control Design

25

For attenuation of output disturbance signals at the plant output, σ (GK) should be large; valid for frequencies at which σ (GK) >> 1 For attenuation of output disturbance signals at the plant input, σ (K) should be small; valid for frequencies at which σ (GK) << 1 . At frequencies where σ (GK) >> 1, atten uation of output disturbance signals at the plant input requires 1/σ (G) to be small, i.e., it is a limitation imposed by the plant and cannot be set by the designer. For attenuation of input disturbance signals at the plant input, σ (KG) should be large; valid for frequencies at which σ (KG) >> 1 For attenuation of input disturbance signals at the plant output, σ (K) should be large; valid for frequencies at which σ (GK) >> 1 For attenuation of measurement noise signals at the plant output, σ (GK) should be small; valid for frequencies at which σ (GK) << 1 For attenuation of measurement noise signals at the plant input, σ (K) should be small; valid for frequencies at which σ (GK) << 1 . At frequencies where σ (GK) >> 1, attenu ation of measurement noise signals at the plant input requires 1/σ (G) to be small, i.e., it is a limitation imposed by the plant and cannot be set by the designer. For good reference tracking σ (GK) should be large; valid for frequencies at which σ (GK) >> 1, and equivalent to attenuation of output disturbances at the plant output. For avoidance of large control signals due to reference demands (and hence actuator satu ration), σ (K) should be small; valid for frequencies at which σ (GK) << 1 . At frequen cies where σ (GK) >> 1, avoidance of large control signals due to reference demands requires 1/σ (G) to be small, i.e., it is a limitation imposed by the plant and cannot be set by the designer.

Recalling the requirements for robust stability in the presence of unstructured system uncer tainty given in Section 2.1, we have the following additional objectives: • • •

For robust stability to input multiplicative uncertainty, σ (KG) should be small; valid for frequencies at which σ (KG) << 1. For robust stability to output multiplicative uncertainty, σ (GK) should be small; valid for frequencies at which σ (GK) << 1. For robust stability to additive uncertainty, σ (K) should be small; valid for frequencies at which σ (GK) << 1 . At frequencies where σ (GK) >> 1, robust stability to additive uncertainty requires 1/σ (G) to be small, i.e., it is a limitation imposed by the plant and cannot be set by the designer.

Now, since we usually have that disturbance attenuation and reference tracking properties are important at low frequencies, and measurement noise attenuation, control signal reduction and robust stability properties are important at high frequencies, the above specifications can be met by shaping singular values of the open-loop transfer matrices GK and KG as shown in Figure 1.19 for GK. Choosing K to shape the singular values of GK and KG in this way is a relatively easy task. However, to do so in a way that also guarantees closed-loop stability is in general difficult, since closed-loop stability cannot be determined from open-loop singular values. Furthermore, for SISO systems Bode [33] showed that closed-loop stability is closely related to the open-loop gain and phase near the crossover frequency ωc , where |GK( jωc )| = 1. In particular, the roll-off rate from high to low gain at crossover is limited by phase requirements for stability, and in practice this corresponds to a roll-off rate of less than 40 dB/decade. An immediate consequence of this is that there is a lower limit on the difference between ωl and ωh in Figure 1.19. For MIMO systems a similar gain-phase relationship

26

D.G. Bates

Fig. 1.19. Design speciﬁcations for singular values of L holds around the crossover frequency region, but this is given in terms of the eigenvalues of GK, not the singular values, [34]. Thus, stability constraints are even more difﬁcult to handle in multivariable loop shaping than in classical SISO loop shaping. The approach adopted in the H∞ loop shaping design method is to ﬁrst shape the singular values of the plant G using weighting matrices to satisfy performance objectives, without explicitly considering the issue of closed-loop stability. Once this step is complete, an H∞ robust stabilisation problem is solved to maximise the robust stability properties of the shaped plant. This ‘‘two-step’’ design procedure is described in detail in the following subsection.

1.4.2 The H∞ Loop-Shaping Design Procedure The H∞ loop shaping design method is essentially a two stage process. First, the open-loop plant is augmented by (generally diagonal) weighting matrices to give a desired shape to the singular values of the open-loop frequency response. Then, the resulting shaped plant is ro bustly stabilised with respect to coprime factor uncertainty using H∞ optimisation. An imple mentation structure for H∞ loop shaping controllers is shown in Figure 1.20. With reference to this ﬁgure, the weighting matrix W1 (s) is chosen to add integral action and ensure reasonable roll-off rates for the open-loop singular values around the desired crossover frequencies. The constant weighting matrix k is then used to adjust control actuation requirements to respect the various actuator rate and magnitude limits. Note that in this conﬁguration the plant G is assumed to be scaled so as to be approximately normalised with respect to maximum allow able input signals. Scaling may also be required on the plant outputs to ensure that units of appropriate relative size are used for all output signals. The scalar diagonal matrix W2 is used to prioritise certain controlled variables over others.

H

Zc

�

� �� -

K∞ (0)W2

�

k

W1 (s)

�

K∞ (s)

∞

W2

Control Design

� G(s)

27

Z

�

�

Fig. 1.20. Implementation structure of the H∞ Loop-Shaping controller K(s) The second stage of the H∞loop shaping design method involves the use of H∞ optimisation to compute a controller block K∞ which robustly stabilises the shaped plant against a partic ular type of uncertainty description, based on stable perturbations to each of the factors in a coprime factorisation of the shaped plant. For a shaped plant G with a normalised left coprime factorisation, [38], G = M −1 N

(1.32)

an uncertain plant model G p can then be written as G p = (M + ΔM )−1 (N + ΔN )

(1.33)

where ΔM and ΔN are stable unknown transfer function matrices which represent the uncer tainty in the nominal plant model G - see Figure 1.21. The objective of robust stabilisation is

�

ΔN

+

� ��

ΔM

�

φ

� u

Nl

� ��

+

�

+

K

Ml−1

�

Fig. 1.21. Normalised coprime factor uncertainty description

to stabilise the class of perturbed plants deﬁned by

y

28

D.G. Bates G p = {(M + ΔM )−1 (N + ΔN ) : k [ΔN ΔM ] k∞ < ε

(1.34)

Now, as shown in [39], the largest possible class of such systems, i.e., the maximum of ε , εmax , is given by: 1 K γmin = inf (I − GK)−1 M −1 k∞ (1.35) I K εmax where γmin is then the stability margin. Note that γmin is the H∞ norm of the transfer function u matrix from φ to in Figure 1.21. Given a minimal realisation [A, B,C, D] of a controllable y and observable plant, the solutions X and Z of the two Riccati equations Az Z + ZA∗z − ZC∗ R−1CZ + BS−1 B = 0

A∗z X + XAz − XBS−1 B∗ X +C∗ R−1C

(1.36)

= 0

(1.37)

where Az = A − BS−1 D∗C

R = I + DD∗ , S = I + D∗ D give the optimal γ 1

1

γmin = {1− k [N M] k2H }− 2 = (1 + ρ (XZ)) 2

(1.38)

where k . kH denotes Hankel norm, [38], and ρ denotes the spectral radius. The central con troller K∞ which guarantees that K (I − GK)−1 M −1 k∞ ≤ γ (1.39) I for a speciﬁed γ > γmin , is then given by K∞ =

Ak Bk Ck Dk

Ak = A + BF + γ 2 (L∗ )−1 ZC∗ (C + DF) 2

∗ −1

Bk = γ (L )

(1.40)

(1.41) (1.42)

ZC

Ck = B∗ X

(1.43)

Dk = −D

(1.44)

where F = −S−1 (D∗C + B∗ X) 2

L = (1 − γ )I + XZ

(1.45) (1.46)

For γ < 4 (i.e., allowable coprime factor uncertainty of at least 25%) it can be shown theoretically, [39], that the controller K∞ (s) does not significantly change the shapes of the open-loop singular values. Thus robust stability is achieved without significant degradation in the origi-

H

∞

Control Design

29

nal performance characteristics specified by the plant weighting functions. If γ is greater than 4, this indicates that the chosen loop shapes are incompatible with robust stability (e.g., the roll-off rate is too big around crossover), and further adjustment of the weighting functions is then required.

The ﬁnal step of the design procedure is to add the constant preﬁlter K∞(0)W2 in order to ensure zero steady state tracking error, assuming integral action in W1 . Note that the K∞ con troller block is a dynamic compensator of order equal to that of the shaped plant - thus the order of the complete H∞ loop shaping controller is equal to the order of the original plant plus twice the order of the weighting functions.

1.4.3 Advantages of H∞ Loop-Shaping The H∞ loop shaping design method has a number of important advantages, which are sum marised below. •

•

•

No gamma iteration - the solution of the H∞ loop shaping controller synthesis equations is particularly attractive in that the optimal γ can be found without recourse to the γ -iteration which is normally required to solve H∞ control problems. Provides robustness to unstable perturbations, and uncertainty in the location of lightly damped resonant poles, by maximising robustness to normalised coprime-factor uncertainty. No pole-zero cancellations - cancellation of the stable plant poles by the controller is a common phenomonon in mixed sensitivity H∞ designs, [35, 40]. Consider, for example, the controller designed for the Bell 205 helicopter in the previous section. The poles of the helicopter model G are given by:

Poles = {−0.67 ± 0.60 j, +0.40, −0.13 ± 0.44 j, +0.19, −0.79, −0.40 The zeros of the controller K are given by Zeros = {−0.67 ± 0.60 j, −0.13 ± 0.44 j, −0.79, −0.14 ± 0.12 j, −0.40, −0.5, −0.5, −0.5, −0.5

which includes all the stable poles of G. A physical interpretation of this phenomenon is that in the S/KS mixed sensitivity problem formulation, only disturbances at one point in the loop (i.e., the plant output) are considered by the controller. The optimisation exploits the fact that there is no disturbance at the plant input which can excite plant poles, and tries to cancel the feedback path from the output disturbance to the plant input with ze ros (see [40] for a more complete discussion of this phenomenon). Cancellation of stable plant poles by the controller presents a particular problem for systems with slow stable poles compared with the bandwidth of the system, e.g., lightly damped resonances. This is because the controller will only have cancelled these poles in the closed-loop trans fer function between certain exogenous inputs and the controlled outputs; these lightlydamped poles will then still appear in other important closed-loop transfer functions. A second problem with the cancellation of such poles is that the closed-loop system will ex hibit poor robust performance properties. This is because, as shown in Section 3.4.5, even small amounts of system uncertainty can allow signiﬁcant movement in the locations of lightly damped poles. Thus, the controller may fail to cancel the poles exactly, leading to poor robust performance. This problem can be avoided in mixed sensitivity designs by introducing extra weighting functions to allow for disturbances at more than one point in

30

D.G. Bates the loop - the choice of the resulting weights is then more complicated, however, since the various trade-offs become much less clear. In contrast, complete pole/zero cancellations do not occur in controllers p roduced using the H∞ loop shaping method. The reason for this becomes clear on noting that the coprime factor singular value robustness measure γ defined as

γ =

K (I + GK)−1 M −1 k∞ I

can also be written as, (see [4], pp. 320),

w1 K −1 γ = (I + GK) [I G] k∞ =

w2 I

(1.47)

z1

z2 ∞

(1.48)

where w1 and w2 are disturbances at the output and input of the shaped plant as shown in Figure 1.22. The H∞ loop shaping controller takes account of disturbances at more than one point in the loop, and thus does not attempt to cancel all the stable plant poles. z1

�

w2

� ��

w1

�

��

Gs

−K

z2

�

Fig. 1.22. Alternative H∞ loop shaping problem formulation •

•

Clear management of conﬂicting speciﬁcations around crossover - in mixed sensitivity optimisation separate weights are chosen for S and T , or S and KS. At low and high frequencies, it is relatively easy to make sure that the weights do not conﬂict with each other, and indeed it can be argued that one of the great advantages of the mixed sensi tivity approach is that the designer can directly shape many different closed-loop transfer functions. Around the desired loop bandwidth frequency, however, it is not always easy to manage the trade-offs between conﬂicting objectives. This problem does not arise when selecting weighting functions in the H∞ loop shaping procedure, because shaping the open-loop transfer matrix L simultaneously shapes both S and T , thus removing the pos sibility of choosing conﬂicting weights. Provides balanced robustness and performance properties at the plant input and output the coprime factor singular value robustness measure γ provides bounds on all input and output closed-loop transfer functions, since, from Theorem 3.1 in [36] we have that K G γmin = inf k (I + GK)−1 [I G] k∞ = inf k (I + KG)−1 [K I] k∞ (1.49) I I K K

H

∞

Control Design

31

In particular, we have that, ( [36] Theorem 5.4), −1 σ ((I + GK)−1 ) = σ (SO ) ≤ εmax σ (Ms )c(W2 ) −1

σ ((I + KG)

) = σ (SI ) ≤

σ (GK(I + GK)−1 ) = σ (TO ) ≤ σ (KG(I + KG)−1 ) = σ (TI ) ≤

σ (K(I + GK)−1 ) = σ (KSO ) ≤

−1 εmax σ (Ms )c(W1 ) −1 εmax σ (Ns )c(W2 ) −1 εmax σ (Ns )c(W1 ) −1 εmax σ (Ms )σ (W1 )σ (W2 )

(1.50) (1.51) (1.52) (1.53) (1.54)

where

σ 2 (W2 GW1 ) 1 + σ 2 (W2 GW1 )

σ (Ns ) =

1 1 + σ 2 (W2 GW1 ) σ (W1 ) c(W1 ) = σ (W1 ) σ (W2 ) c(W2 ) = σ (W2 )

σ (Ms ) =

•

!1 2

1

≤1

(1.55)

≤1

(1.56)

2

(1.58)

From the above we can see that γmin provides balanced performance and robustness guar antees at the plant input and output. This is in sharp contrast to mixed sensitivity H∞ optimisation which generally only provides robustness guarantees at the plant input or output. Guaranteed simultaneous gain/phase margins - From [37], we have that the coprime factor singular value robustness measure γ provides the following guaranteed symmetric multi variable gain and phase margins: r r 1+ε 1+ε −20log10 dB ≤ GMcp ≤ +20log10 dB (1.59) 1−ε 1−ε − sin−1 ε degrees ≤ PMcp ≤ + sin−1 ε degrees

•

(1.57)

(1.60)

These gain and phase variations are allowed simultaneously at each input and output of the plant. In the SISO case the allowable gain and phase variation are double those given above. Exact observer implementation - In [40], it was shown that the controller resulting from the H∞ loop shaping procedure can be written as an exact plant observer plus state feed back. Assuming, purely for notational convenience, a strictly proper shaped plant, with a stabilisable and detectable state-space realisation As Bs Gs = (1.61) Cs 0 the relevant equations are x˙ˆs = As xˆs + Hs (Cs xˆs − ys ) + Bs us

us = Ks xˆs

(1.62) (1.63)

32

D.G. Bates

where xˆs is the observer state, us and ys are respectively the input and output of the shaped plant, and

Hs = −ZsCs∗

(1.64)

Ks = −B∗s [I − γ −2 I − γ −2 Xs Zs ]−1 Xs

(1.65)

where Zs and Xs are the appropriate solutions to the generalised algebraic Riccati equations for Gs given previously. In general, H∞ controllers cannot be written as exact plant state observers, as there will be a worst-case disturbance term entering the observer state equation, [20]. However, for the controllers produced by the H∞ loop shaping method it is possible, and this clear structure lends itself to gain-scheduling in that the controller matrices Ks and Hs can be simply scheduled as a function of one or more aircraft parameters. Figure 1.23 shows the implementation structure of the H∞ loop shaping controller in observer form.

r

� K0

�+ � + �

us

�

u

�y

� G

kW1

� W2

�

Hs

Bs + � � �+ � R + �

�

ys

� �+ �

� Cs

As � Ks �

Fig. 1.23. Observer-form implementation of H∞ loop shaping controller

Note that the observer is for the shaped plant Gs and that the constant prefilter must be recalculated to take account of the different structure of the observer-form implementa tion. As well as greatly simplifying the task of controller scheduling, the observer/state feedback structure gives a clearer (more ‘classical’) functionality to the different ‘blocks’ of the H∞ loop shaping controller - this extra transparency can be of great help in the controller certiﬁcation process.

H

∞

Control Design

33

1.4.4 Design Example: Control Law Design for the Harrier V/STOL Aircraft A significant advance in aerospace engineering in recent years has seen the development of actuator technologies which allow much greater ﬂexibility in the use of propulsion generated thrusts as primary ﬂight control effectors. Pitch, roll and yaw moments necessary for attitude control, as well as various thrust and lift forces necessary for trim and manoeuvring control, can now all be partially or exclusively generated by vectoring nozzles and various propulsive lift systems such as ejectors, remote augmented lift systems (RALS), and various techniques of boundary-layer control using engine air, [41]. The development of these new technologies means that there are now many combinations of aerodynamic and propulsion effectors which can be employed to enhance control authority during STOVL and conventional operations, to minimise fuel consumption and pilot workload throughout different ﬂight modes, and to permit reconﬁguration in the event of malfunction of one or more components. The use of propulsion effectors for aerodynamic control, however, creates signiﬁcant interaction between the two (heretofore independent) airframe and engine subsytems, necessitating the use of robust multivariable control methods to design integrated ﬂight and propulsion control (IFPC) systems. In [41] a methodology for IFPC design was developed, based on a ‘centralised’ approach to the problem. This approach consists of ﬁrst designing a centralised controller considering the airframe and propulsion subsystems as one integrated system, and then partitioning the centralised controller into decentralised subcontrollers with a speciﬁed coupling structure. The subcontrollers will in general be of lower order than the centralised controller, and will thus be easier to implement. They can also be independently tested and validated at subsystem level. Independent integrity of the subcontrollers is necessary not just for control theoretic reasons (different types of control to be applied to systems with different dynamics, uncertainties, nonlinearities etc. ) but for practical and ‘political’ reasons as well. Since different subsystems may be designed and built by different and independent suppliers, design accountability and commercial issues will often dictate that each manufacturer retains a high degree of control over the particular subsystem (and its controller) for which it is responsible. The example described in this section investigates the application of the H ∞ loop shaping design method to the problem of designing a centralised integrated ﬂight and propulsion control system for the VAAC Harrier aircraft. As well as integrating the airframe and engine control problems, the longitudinal and lateral/directional control problems are also addressed simul taneously. This is to ensure minimal coupling between longitudinal and lateral dynamics for manoeuvres at high angles of attack. Results of piloted simulations with an IFPC system de signed using this approach, as well as more details on stategies for partitioning centralised IFPC systems can be found in [17, 42–44]. The aircraft model used in this example is based on the Harrier Wide Enve lope Model (WEM): a full envelope non-linear lateral plus longitudinal representation of the aerodynamics, engine and actuator characteristics of the VAAC Harrier aircraft. The WEM software incorporates a full thermodynamic powerplant simulation of a Pegasus engine, which is integrated with the airframe system dynamics: thus in principle all interactions between these two systems are captured in this model. Representative actuation systems, including rate and saturation limits, have been placed on all control motivators. Saturation limits for control motivators are as follows: elevator (etad): -10.25 to 11.25 deg, aileron (xid): +/- 14 deg, rudder(zetad): +/- 15 deg, nozzle angle (thejd): 0 to 98.5 deg, throttle position (pthtp): 0.26 to 1 dimensionless.

34

D.G. Bates

R The WEM software is coded in fortran and run under the Simulink simulation environment. The overall software package represents a highly detailed model of the aircraft, producing 22 aerodynamic outputs and 12 engine outputs. The package was conﬁgured to allow the extrac tion of linear state space models at various points over the ﬂight envelope for the purposes of control system design. These (lower order) linear models are of the form

x˙ = Ax + Bu; y = Cx + Du, where the state vector are given by x = [θ , φ , ψ , q, p, r, v f , ve, vd, f np, hnp, qe f ] where v f , ve and vd are forward, east and down velocities respectively, f np is engine fan speed, hnp is compressor speed, qe f is fuel ﬂow and the other symbols have their usual mean ings. The control inputs are given by u = [etad, xid, zetad,the jd, ptht p] where etad is elevator angle, xid is aileron angle, zetad is rudder angle, the jd is thrust vector angle and ptht p is thrust magnitude. All of the states defined above are available as outputs. Both the linear and non-linear WEM models cover the full flight envelope for the VAAC, from fully airborne flight at 200 knots down to hover. In this example, the linear model corresponding to the 80 knot transition flight phase is used for controller design. At this point in the flight envelope, the aircraft is longitudinally unstable and propulsion system generated forces and moments are taking over control of the aircraft from the aerodynamic effectors. The controlled variables z are given by z = [Qq, Pp, r, v f , vd] where Qq = q + 0.3θ , Pp = p + 0.3φ , r = yaw rate, v f = velocity forward and vd = velocity down. The above choice of controlled variables was based on control configuration mode 4 of the NASA V/STOL systems research aircraft [45], and provides response types that are generally desirable for good handling qualities in transition flight, [41]. Control of Qq and Pp corresponds to tracking of transient rate commands and steady state attitude commands - a so-called rate command-attitude hold system. The corresponding demands on q and θ for a demand on Qq are shown in Figures 1.24 and 1.25 respectively (this can be verified by solving the first order differential equation θ˙ + 0.3θ = Qq). Performance specifications, based on [46], for the closed-loop system are given as:

1. 90 per cent of demanded rate/attitude to be reached within 2 seconds for the pitch, roll and yaw outputs. 2. Bandwidth of the pitch, roll and yaw channels to be approximately 7 rads/sec. 3. 90 per cent of demanded velocity to be reached within 4 seconds. 4. Maximise decoupling of command tracking over all channels, ensuring: (a) 10 degrees attitude demand causes less than 1 degree change in other attitudes, and less than 1 knot change in velocities,

H

∞

Control Design

35

(b) 10 knots velocity demand causes less than 1 knot change in other velocities, and less than 1 degree change in attitudes. 5. Control signals to stay within rate/saturation limits at all times. The structure of the closed-loop system with a one degree-of-freedom H ∞ Loop-Shaping controller K(s) is shown in Figure 1.20. As discussed previously, the design parameters in the H ∞ Loop-Shaping procedure are the transfer function matrix W1 , and the two constant matrices W2 and k. For this example all three matrices are chosen to be diagonal. An essential initial step in the design procedure is the proper scaling of the plant inputs and outputs, [14]. In this example, input scaling was applied to the plant to approximately normalise the actuator signals by their maximum allowable values. Cross coupling between all outputs was considered of equal importance, so output scaling was used only to convert the units of the first three outputs from radians to degrees. Note that the scaled plant G(s) in this example also includes linearised models of the dynamics of the plant actuators - the singular values of the scaled plant are shown in Figure 1.26. The design procedure itself then consists of two basic steps. First the weighting matrices W1 , W2 , and k are chosen to shape the singular values of the open-loop plant so that the given performance specifications are satisfied. Then the feedback controller K∞ (s) is calculated so that the shaped plant is robustly stabilised against normalised coprime factor uncertainty. For this example, initial values for the weighting matrices were chosen as:

W1 (s) =

s+5 I , W2 = I5×5 , k = diag [0.1 0.1 0.1 0.1 0.1] s 5×5

W1 was chosen to ensure good tracking properties with zero steady state error, good distur bance rejection at low frequencies, and a moderate roll-off rate around crossover. The matrix W2 is generally used to reﬂect the relative importance of the outputs to be controlled and was therefore chosen as the identity matrix for this example. Finally, the matrix k was used to set the bandwidth of the open-loop singular values, and to adjust the relative magnitudes of the various actuator signals. The singular values of the shaped plant are shown in Figure 1.27. The implementation structure shown in Figure 1.20 has the advantage that reference signals do not directly excite the dynamics of K∞ (s) - since this block has been designed for stabilisation and not performance purposes, its presence in the forward loop can result in large amounts of overshoot (classical derivative kick). The constant preﬁlter K∞ (0)W2 is included to ensure zero steady state tracking error, assuming integral action in W1 . The second step in the design procedure is the calculation of the H ∞ robust stabilisation controller K∞ (s). For the choice of weighting matrices given above, a robust stabilisation controller K∞ (s) was calculated which returned an εmax of only 0.233, indicating poor robust stability properties for the system. Thus, although the performance of the controller in time domain simulation was quite acceptable (see Figures 1.28 and 1.29), adjustment of the weighting functions is required in order to improve the robustness of the IFPC system. In order to improve the robust stability of the closed-loop, we need to decrease the rate of roll-off of the open-loop singular values of the shaped plant around cross-over. This can easily be achieved in the H∞ loop shaping provedure by adjusting the weighting matrix W1 so that

36

D.G. Bates

the zero at 5 rads/sec takes effect at a slightly lower frequency, say 1 rads/sec, thus reducing the roll-off of the singular values around crossover. For the new weighting matrix W1 (s) =

s+1 I , s 5×5

with W2 and k unchanged, a new robust stabilisation controller K∞ (s) was calculated which returned an εmax of 0.279, indicating much improved robust stability properties for the system, i.e., allowable coprime factor uncertainty of > 27%. Analysis of the closed-loop system in time domain simulation, however, revealed that the tracking response for pilot demands was now a little sluggish, with increased coupling between the different controlled variables. In particular the response to demands on Qq (Figure 1.30) no longer meets the specification of achieving 90% of the reference demand within 2 seconds. In order to improve the closed-loop performance, the ﬁrst element of the W1 weighting matrix was changed to s+s 5 (to improve the Qq tracking response at low frequencies) while the gain matrix k was set to

k = diag [0.4 0.2 0.2 0.4 0.2] in order to emphasis the use of the elevator and nozzle angle (for improved control of Qq), and to increase the closed-loop bandwidth slightly in all loops. Robust stabilisation of this shaped plant gave a (still quite acceptable) εmax of 0.269, with improved tracking response for all controlled variables, and Qq in particular (Figure 1.31). The responses for all other variables now also meet command tracking, decoupling and control usage specifications - see for example Figures 1.32 to 1.35. Finally, the achieved open-loop singular values after robust stabilisation (i.e., including K∞ ) are compared with those specified for the shaped plant in Figure 1.36. As shown by the figure, and as expected from the ‘good’ value of γmin /εmax , the controller K∞ has not significantly altered the desired loop shapes. As with the mixed sensitivity H∞ synthesis example, the design described above is not complete, and is intended primarily to illustrate the process by which performance and robustness trade-offs can be managed via adjustment of the weighting functions in the design process. It does, however, demonstrate some of the more appealing features of the H∞ loop shaping method - in particular, the classical nature of the approach, the transparency of the relationship between the design specifications and the plant weighting functions, and the robustness of the controllers produced by the H∞ robust stabilisation optimisation.

H

∞

Control Design

Qqdemand 12

10

magnitude

8

6

4

2

0

0

1

2

3

4

5 time(secs)

6

7

8

9

10

Fig. 1.24. Pilot demand on Qq

20

theta_c and theta

15

magnitude

10

5 q_c and q

0

−5 0

1

2

3

4

5 time(secs)

6

7

8

9

10

Fig. 1.25. Corresponding demands on q and θ - rate-command attitude-hold response

37

38

D.G. Bates

Nominal plant 80

60

40

mag (dB)

20

0

−20

−40

−60

−80

−100 −3

−2.5

−2

−1.5

−1

−0.5 log10(w)

0

0.5

1

1.5

2

1.5

2

Fig. 1.26. Open-loop singular values of the scaled plant

Shaped plant 150

100

mag (dB)

50

0

−50

−100

−150 −3

−2.5

−2

−1.5

−1

−0.5 log10(w)

0

0.5

1

Fig. 1.27. Open-loop singular values of the shaped plant

H

−3

10

6

8

6

4

−3

Pp

0

5

Design

r

10

4

2

−2

10

x 10

15

5

0

0

2

0

x 10

degrees/sec

8

magnitude

magnitude

Qq 12

∞ Control

0

5

vf

10

−5

0

5 time(secs)

10

vd

80.5

0.35

0.3

80.4

0.25

0.2

ft/s

ft/s

80.3

0.15

80.2

0.1

80.1

80

0.05

0

5 time(secs)

0

10

0

5 time(secs)

10

Fig. 1.28. Responses for a demand on Qq - initial weighing functions

−3

5

degrees (away from trim)

degrees (away from trim)

1

0

−1

−2

−3

−4

0

5

10

x 10

XID

0

−5

−10

−15

ZETAD 0.015

degrees (away from trim)

ETAD 2

0

THEJD

5

10

0.01

0.005

0

−0.005

−0.01

0

5 time(secs)

10

PTHTP

40

magnitude (away from trim)

degrees (away from trim)

0

20

0

−20

−40

−60

0

5 time(secs)

10

−0.05

−0.1

−0.15

−0.2

−0.25

−0.3

−0.35

0

5 time(secs)

10

Fig. 1.29. Actuator responses for a demand on Qq - initial weighing functions

39

40

D.G. Bates

Pp

8

0.008

6

4

2

0

r 0.015

degrees/sec

0.01

magnitude

magnitude

Qq 10

0.006

0.004

0.01

0.005

0.002

0

5

0

10

0

5

vf

0

5 time(secs)

10

vd 0.4

80.8

0.3

80.6

0.2

ft/s

81

ft/s

0

10

80.4

0.1

80.2

0

80

0

5 time(secs)

−0.1

10

0

5 time(secs)

10

Fig. 1.30. Degraded responses for a demand on Qq

−3

8

3

6

4

x 10

−3

Pp 6

r

x 10

5

degrees/sec

4

magnitude

magnitude

Qq 10

2

1

4

3

2

1

2

0

0

−1

0

5

10

0

0

vf

5

10

−1

0

5 time(secs)

vd 0.3

80.35

80.3

0.2

0.1

80.2

ft/s

ft/s

80.25

80.15

0

80.1

−0.1

80.05

80

0

5 time(secs)

10

−0.2

0

5 time(secs)

10

Fig. 1.31. Improved responses for a demand on Qq

10

H

−3

8

−0.2

6

4

2

−4

Pp 2

0

5

−0.4

−0.6

−1

10

x 10

r

0

−0.8

0

−2

x 10

degrees/sec

0

magnitude

magnitude

Qq 10

Control Design

∞

−2

−4

−6

0

5

vf

10

−8

0

5 time(secs)

10

vd

80.4

12

10

80.3

8

ft/s

ft/s

80.2

6

4

80.1

2

80

79.9

0

0

5 time(secs)

−2

10

0

5 time(secs)

10

Fig. 1.32. Responses for a demand on vd −3

ETAD 4

0

−5

−10

−15

0

5

10

x 10

XID

2

0

−2

−4

−6

−8

−10

ZETAD 0.015

degrees (away from trim)

degrees (away from trim)

degrees (away from trim)

5

0

THEJD

5

10

0.01

0.005

0

−0.005

−0.01

0

PTHTP magnitude (away from trim)

degrees (away from trim)

0

40

20

0

−20

−40

−60

0

5 time(secs)

10

−0.05

−0.1

−0.15

−0.2

−0.25

−0.3

−0.35

0

5 time(secs)

10

Fig. 1.33. Actuator responses for demand on vd

5 time(secs)

10

41

D.G. Bates

−3

x 10

Qq

Pp

r

12

0.3

4

10

0.2

3

8

magnitude

magnitude

5

2

1

6

4

0

2

−1

0

0

5

10

degrees/sec

42

1

0

−0.1

−0.2

−0.3

0

5

−4

vf 80.0002

0.1

x 10

10

−0.4

0

5 time(secs)

10

vd

0

80.0001

−1

ft/s

ft/s

80.0001

−2

−3

80

−4

80

79.9999

−5

0

5 time(secs)

−6

10

0

5 time(secs)

10

Fig. 1.34. Responses for a demand on Pp

ETAD

XID

0.03

0.02

0.01

0

5

10

10

5

0

−5

magnitude (away from trim)

degrees (away from trim)

0

20

0

−20

−40

−60

0

5 time(secs)

0

5

−4

THEJD 40

1

degrees (away from trim)

degrees (away from trim)

degrees (away from trim)

0.04

0

ZETAD

15

0.05

10

x 10

10

0.5

0

−0.5

−1

−1.5

0

PTHTP

−0.2

−0.4

−0.6

−0.8

−1

−1.2

0

5 time(secs)

10

Fig. 1.35. Actuator responses demand on Pp

5 time(secs)

10

H

∞

Control Design

43

Loop gain frequency response 100

50

dB

0

Specified loop gain − − −50

Achieved loop gain −

−100

−150 −2 10

−1

10

0

10 rad/s

1

10

2

10

Fig. 1.36. Comparison of achieved versus speciﬁed open-loop singular values

References 1. F.L. Lewis, Optimal Control, Wiley, 1986. 2. J.C. Doyle, “Guaranteed margins for LQG regulators”, IEEE Transactions on Automatic Control, 23(4), pp. 756-757, 1978. 3. J.D. Blight, R.L. Dailey and D. Gangsaas, “Practical control law design for aircraft using multivariable techniques”, International Journal of Control, 59(1), pp. 93-137, 1994. 4. K. Zhou and J.C. Doyle Essentials of robust control, Prentice Hall, 1998. 5. C. Edwards and S.K. Spurgeon Sliding mode control: theory and applications, Taylor and francis, London, 1996. 6. D. Enns, D. Bugajski, R. Hendrick and G. Stein “Dynamic inversion: an evolving method ology for ﬂight control design”, International Journal of Control 59(1), pp. 71-79, 1994. 7. J. Reiner, G. Balas and W.L. Garrard “Robust dynamic inversion for control of highly maneuverable aircraft”, AIAA Journal of Guidance Control and Dynamics, 18(1), pp. 18 24, 1995. 8. C.D. Yang and C.C. Kung, “Nonlinear H∞ ﬂight control of general six-degree-of-freedom motions”, AIAA Journal of Guidance, Control and Dynamics, 23(2), pp. 278-288, 2000. 9. C.S. Wu, B.S. Chen and Y.W. Jan “Uniﬁed design for H2 , H∞ , and mixed control of spacecraft”, AIAA Journal of Guidance, Control and Dynamics, 22(6), pp. 884-896, 1999. 10. G. Papageorgiou and R. Hyde “Analysing the stability of NDI-based ﬂight controllers with LPV methods”, Proc. of the AIAA Conference on Guidance, Navigation and Control, Montreal, 2001.

44

D.G. Bates

11. G.E. Cooper and R.P. Harper “The use of pilot rating in the evaluation of aircraft handling qualities”, NASA TN D-5153, April 1969. 12. S. Bennani and G. Looye “Design of ﬂight control laws for a civil aircraft using µ synthesis”, Proc. of the AIAA Conference on Guidance, Navigation and Control, pp. 314 324, 1998. 13. I. Fialho, G. Balas, A. Packard, J. Renfrow and C. Mullaney “Linear Fractional Transfor mation Control of the F-14 Aircraft Lateral-Directional Axis during Powered Approach Landing”, Proc. of the Amrican Control Conference pp. 128-132, 1997. 14. R.A. Hyde H∞ Aerospace Control Design - A VSTOL Flight Application, Springer Verlag, 1995. 15. A. Smerlas, I. Postlethwaite, D.J. Walker, M.E. Strange, J. Howitt, R.I. Horton, A.W. Gubbels, and S.W. Baillie “Design and Flight Testing of an H∞ Controller for the NRC Bell 205 experiemntal Fly-By-Wire Helicopter”, Proceedings of the AIAA Conference on Guidance, Navigation and Control, Boston, 1998. 16. I. Postlethwaite, A. Smerlas, D.J. Walker, A.W. Gubbels, S.W. Baillie, M.E. Strange and J. Howitt, “H∞ control of the NRC Bell 205 ﬂy-by-wire helicopter”, Journal of the Amer ican Helicopter Society, 44(4), pp. 276-284, 1999. 17. D.G. Bates, S.L. Gatley, I. Postlethwaite and A.J. Berry “Design and Piloted Simulation of a Robust Integrated Flight and Propulsion Controller”, AIAA Journal of Guidance, Navigation and Control, 23(2), pp. 269-277, 2000. 18. K. Glover and J.C. Doyle “State-space formulae for all stabilizing controllers that satisfy an H∞ norm bound and relations to risk sensitivity”, Systems and Control Letters, 11, pp. 167-172, 1988. 19. G. J. Balas, J. C. Doyle, K. Glover, A. Packard and R. Smith, µ -Analysis and Synthesis Toolbox User’s Guide, The Mathworks, 1995. 20. J.C. Doyle, K. Glover, P.P. Khargonekar and B.A. Francis “State-space solutions to stan dard H2 and H∞ control problems”, IEEE Transactions on Automatic Control, AC-34(8), pp. 831-847, 1989. 21. G. Meinsma “Unstable and non-proper weights in H∞ control”, Automatica, 31(1), 1655 1658, 1995. 22. J.S. Freudenberg and D.P. Looze “An analysis of H∞ optimization design methods”, IEEE Transactions on Automatic Control, AC-31(3), pp. 194-200, 1986. 23. A. Smerlas, I. Postlethwaite and D.J. Walker “H∞ loop shaping: the Bell 205 helicopter case study”, Proceedings of the AIAA Conference on Guidance, Naviagation and Control, pp. 1605-1612, 1999. 24. A. Smerlas Robust multivariable control of helicopters: from mathematical models to ﬂight tests, Ph.D. Thesis, Department of Engineering, University of Leicester, March 1999. 25. R.K. Hefﬂey, W.F. Jewell, J.M. Lehman and R.A. Van Winkle A compilation and analysis of helicopter handling qualities data, NASA Contractor Report 3144, 1979. 26. M. Strange, J. Howitt Conﬁguration of the DERA HELISIM model to represent the ﬂight dynamics of the NRC Bell 205 ﬂy-by-wire helicopter, DERA Technical Report TR97459 1, 1997. 27. D.G. Bates, S.L. Gatley, I. Postlethwaite and A.J. Berry “Integrated Flight and Propulsion Control Design using H∞ loop-shaping Techniques”, Proc. IEEE Conference on Decision and Control, pp. 1523-1528, 1999. 28. S. Le Ballois and G. Duc “ H∞ Control of an Earth Observation Satelite”, AIAA Journal of Guidance Navigation and Control, 19(3), pp. 628-635, 1996. 29. G. Ferreres, and M. M’Saad “Parametric Robustness Analysis of a Missile Autopilot”, AIAA journal of Guidance Control and Dynamics, 19(3), pp. 621-627, 1996.

H

∞

Control Design

45

30. R. Samar Robust Multi-Mode Control of High Performance Aero-Engines, Ph.D. Thesis, Department of Engineering, University of Leicester, UK, 1995. 31. G. Papageorgiou and K. Glover “Design of a robust gain scheduled controller for the high incidence research model”, Proceedings of the AIAA Conference on Guidance, Naviagation and Control, pp. 1613-1623, 1999. 32. G. Papageorgiou, K. Glover, G. D’Mello, G. and Y. Patel “Taking robust LPV control into flight on the VAAC Harrier”, Proc. of the IEEE Conference on Decision and Control, Sydney, 2000. 33. H.W. Bode Network analysis and feedback amplifier design, D. Van Nostrand Co. New York, 1945. 34. Doyle, J., C. and Stein, G., “Multivariable feedback design: concepts for a classical/modern synthesis”, IEEE Transactions on Automatic Control, AC-26(1), pp. 4-16, 1981. 35. I. Postlethwaite, M.-C. Tsai and D.-W. Gu “Weighting function selection in H∞ design”, in proceedings of the IFAC World Congress on Automatic Control, 1990. 36. K. Glover, J. Sefton, J. and D.C. Mcfarlane “A tutorial on loop shaping and H-infinity robust stabilization”, Proc. of the IFAC World Congress on Automatic Control, pp. 94103, 1990. 37. K. Glover, G. Vinnicombe, G. and G. Papageorgiou “Guaranteed multi-loop stability margins and the gap metric”, Proc. of the IEEE Conference on Decision and Control, Sydney, 2000. 38. S. Skogestad and I. Postlethwaite Multivariable feedback control, Wiley, 1996. 39. D. McFarlane and K. Glover “A Loop Shaping Design Procedure using H ∞ Synthesis”, IEEE Trans. on Automatic Control, AC-36, pp. 759-769, 1992. 40. J. Seftonand K. Glover “Pole-zero cancellations in the general H∞ problem with reference to a two block design”, Systems and Control Letters, 14, pp. 295-306, 1990. 41. S. Garg “Robust Integrated Flight/Propulsion Control Design for a STOVL Aircraft using H∞ Control Design Techniques”, Automatica 29(1), pp. 129-145, 1993. 42. S.L. Gatley, D.G. Bates, and I. Postlethwaite “A Partitioned Integrated Flight and Propulsion Control System with Engine Safety Limiting”, IFAC Journal of Control Engineering Practice, 8, pp. 845-859, 2000. 43. S.L. Gatley, and D.G. Bates and I. Postlethwaite “Partitioning and re-design of H∞ loop shaping integrated flight and propulsion control systems”, Proc. of the AIAA Conference on Guidance, Naviagation and Control, Montreal, 2001. 44. S.L. Gatley, and D.G. Bates and I. Postlethwaite “An engine-limiting and anti-windup scheme for a partitioned integrated flight and propulsion control system”, Proc. of the AIAA Conference on Guidance, Naviagation and Control, Montreal, 2001. 45. J.A. Franklin, M.W. Stortz, P.F. Borchers and E. Moralez III, Flight evaluation of advanced controls and displays for transition and landing on the NASA V/STOL systems research aircraft, NASA TP 3607, 1996. 46. J.A. Franklin, J., A. “Design criteria for integrated flight/propulsion control systems for STOVL fighter aircraft”, proceedings of Piloting Vertical Flight Aircraft: A conference on flying qualities and human factors, San Francisco, CA, pp. 37-57, 1993.

2 Structural Methods for Linear Systems: An Introduction Nicos Karcanias and Efstathios Milonidis

Summary. This paper assumes familiarity with the basic Control and Dynamics, as covered in undergraduate courses. It introduces the different alternative system representations for lin ear systems and provides a quick review of the fundamental mathematical tools, which are essential for the treatment of the more advanced notions in Linear Systems. The paper focuses on some fundamental concepts underpinning the study of linear systems and dynamics and which play a crucial role in the analysis and design of control systems; thus, the paper deals with notions such as those of Controllability, Observability, Stability, poles and zeros and re lated dynamics and their properties under different compensation schemes. There is a concrete flavour in the current approach which runs through this presentation and this is that of the un derlying algebraic structure. The term “structure” refers to aspects of the state space/transfer function description, which remain invariant under a variety of transformations. The set of transformations considered here are of the compensation type and include state feedback and output injection, dynamic compensation, as well as of the representation type transformations that include state, input, and output co-ordinate transformations. This structure stems from the system description and defines the nature of the dynamics and the related geometric proper ties and these in turn define what it is possible to achieve under feedback; such an approach is known as a structural approach. Central to our analysis are the notions of poles and zeros. The poles of a system are crucial characteristics of the internal system dynamics, characterise system free response, stability and general aspects of the performance of a system. The poles of a system are affected by the different compensation schemes and their assignment is the subject of many design methodologies aiming at shaping the internal system dynamics under different compensation schemes. The notion of zeros is more complex, since they express the interaction between internal dynamics and the effort to control and observe the system and they are thus products of overall system design, that apart from process synthesis involves selection of actuation and measurement schemes for the system. The significance of zeros is mainly due to that they remain invariant under a large set of compensation schemes, as well as that they define limits of what can be achieved under compensation. This makes zeros crucial for design, since they are part of those factors characterising the potential of a given system to achieve certain design objectives under compensation. The invariance of zeros implies that their design is an issue that has to be addressed outside the traditional control design; this requires understanding of the zero formation process and involves early design stages mech anisms such as process instrumentation. Poles and zeros are conceptually inverse concepts (resonances, antiresonances) and such mechanisms are highlighted throughout the paper. The M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 47-98, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

48

N. Karcanias and E. Milonidis

role of system structure in characterising different system properties is central to this paper and it is defined by a set of discrete and continuous invariants; these invariants characterise a vari ety of key system properties and their type/values define the structure of canonical forms and determine somehow the potential of a given system for compensation. Invariants and canoni cal forms under the general transformation group are linked to compensation theory, whereas those associated with representation transformations play a key role in system identification. The emphasis in this article is to provide an overview of the fundamentals concepts, the back ground mathematical tools, explain their dynamic significance and link them to problems of control and systems design.

2.1

Introduction

The study of problems of Control Analysis, Control Synthesis – Design and Model Identi fication, heavily relies on the Theory of Linear Systems. The notion of “system structure” is dominant in describing system properties and the term implies those aspects of the sys tem model, which remain invariant under a set of transformations that may be applied on the system. The system structure is a generic term that refers to system aspects such as intercon nection graph for system components and functions that may be defined on the system model and are referred to as invariants, which are further classified into discrete (integer, real num bers) and continuous (polynomials, rational functions etc.). The notion of structure is referred to both input-output and state space models. The sets of transformations are clearly different for the two cases, but relationships between the structure of state space and transfer function models exist. The notion of system structure is intimately related to that of canonical form, which corresponds to a system description, derived under the given group of transformations, expresses a specialisation of the system invariants in the set of considered models and involves the minimal number of parameters. System structure is essential for the Model Identification problem and canonical forms, corresponding to representation transformations, provide a ve hicle for Model Identification. For the Control Analysis and Synthesis, aspects of the structure (as it is expressed by the system invariants) characterise the presence, or absence of certain system properties and the type and values of invariants provide criteria for solvability of a number of control synthesis problems. In the area of control design, the types and values of invariants frequently impose limitations in what it is possible to achieve. Although the link be tween system structure and achievable performance, under certain forms of compensation is not explicitly known, system structure expresses in a way the potential of a system to provide certain solutions to posed control problems. This paper considers the link between fundamental system notions, which are defined in a dynamic sense and the system structure, as this is expressed by invariants of the state space, or transfer function description. We consider different state space representations and trans fer function representations which lead to unifying descriptions, such as the matrix pencils, polynomial and rational matrices which provide the set up for studying system properties and their link to system structure. The unifying operator for all these state space descriptions is that of a matrix pencil. The theory of invariants and canonical forms of matrix pencils under the general group of the strict equivalence transformations is known as the Kronecker theory and underpins the theory of invariants and canonical forms under any combination of state space transformations. For the case of polynomial and rational models, the theory of Smith and Smith–McMillan forms provide the natural vehicle for defining structure and then link ing it the fundamental system properties. The paper focuses on some fundamental concepts

Structural Methods for Linear Systems: An Introduction

49

underpinning the study of linear systems and their dynamics and which play a crucial role in the analysis and design of control systems; thus, we focus on fundamentals such as the notions of Controllability, Observability, Stability, poles and zeros and related dynamics and try to establish their links to the underlying system structure. It is these links which allow the establishment of connections between structural tools and methodologies and control design. It is assumed that the fundamentals of Control and State Space Analysis are known from the undergraduate courses. The structural characterisation of system properties allows the developmentof an understand ing on the effect of feedback on them and allows the means for the development of com pensation schemes. Simple structural tests for Controllability and Observability properties are given, which allow their characterisation with algebraic means and enable the establishment of simple tests for their invariance under different forms of feedback. The concepts of pole, zero take a considerable attention. These notions have emerged as the key tools of the classi cal methods of Nyquist-Bode and root locus for the analysis and design of linear, single-input, single –output (SISO) feedback systems. The development of the state space S(A, B,C, D) de scription and transfer function G(s) description for linear multivariable systems has led to a variety of definitions for the zeros and poles for the multivariable case. Loosely speaking, multivariable poles and zeros are resonant and antiresonant frequencies respectively, that is to say they are frequencies whose transmission explodes with time, or whose transmission is completely blocked. The inversion of roles of poles and zeros suggested by their classical complex analysis definition (resonance, antiresonance) motivates the dynamic (in terms of tra jectories) properties of zeros. The physical problem used to define multivariable zeros is the “output zeroing problem”, which deals with defining appropriate non zero input exponential signal vectors and initial conditions which result in identically zero output. Such a problem is the dual of the “zero input” problem defining poles, which deals with defining appropriate initial conditions, such that with zero input the output is a nonzero exponential vector signal. Those two physical problems emphasize the duality of the roles of poles and zeros. Apart from its natural dynamic appeal such definitions for poles and zeros have the addi tional advantages that they reveal the geometric dimension of such concepts, as well as their link with fundamental structural invariants of the system. The poles-eigenvalues have a well defined geometry introduced by the eigenvectors and the corresponding spaces (the A-invariant spaces of the state space). Similarly, the geometry of zeros is linked to generalised eigenvalue eigenvector problems and corresponding spaces (types of (A, B)-invariant spaces). The Jordan form of the state matrix reveals the invariant structure of poles in the state space set-up. The Smith form and in some more detail the Kronecker form of the state space system matrix in troduce the zero structure of the state space models; for transfer function models the pole zero structure is introduced by the Smith–McMillan form. Such links reveal the poles as invariants of the alternative system representations under a variety of representation and feedback trans formations. The strong invariance of zeros (large set of transformations) makes them critical structural characteristics, which strongly influence the potential of systems to achieve perfor mance improvements under compensation. The dynamic characterisation of zeros leads to algebraic characterisations, which reveal them as byproducts of the interaction of the internal dynamics and the model, input, output struc ture. This is contrary to the pole characterisation, which shows that they express the internal dynamics. Such observations lead to that zero design is a task associated with overall selection of inputs, outputs and thus belongs to the earlier system design stage of process instrumen tation. In the paper we consider both finite and infinite zeros and examine them in both state space and transfer function context. The relationships between the corresponding notions for

50

N. Karcanias and E. Milonidis

state space and transfer function models reveal interesting links to the fundamental system properties of controllability and observability. Every square system (same number of input and outputs) has zeros (finite and/or infinite); however, nonsquare systems generically do not have zeros and this is an important difference with the poles which exist independent from input, output dimensionalities. Closing feedback loops creates square systems and this involves creation of zeros; such phenomena are within the area of designing zeros and the mechanisms for zero formation are examined. The effect of feedback and other forms of transformations on the pole and zero structure of the system is finally considered. We examine the general feedback configuration under dynamic pre-compensation (or feedback compensation) and this leads to a pole assignment (stabilization) problem linked to the solution of a matrix Diofantine equation. The study of such equations over “relevant” for control rings is briefly examined, their solvability is linked to Minimality of representations and the solution of such equations is given in parametric forms. Finally, an alternative structural framework for generalised design is introduced through the formulation of the Determinantal Assignment Problem (DAP). DAP provides a unifying formulation to all frequency assignment problems (pole, zero) and introduces new types of algebraic structure. This alternative approach provides a powerful machinery for studying problems of “structure assignment” based on tools from exterior algebra and classical algebraic geometry. The zero assignment by a constant squaring down may be studied and solved within this framework. Throughout the paper we will use the following notation: R, C are the fields of real, complex numbers respectively, R[s] is the ring of polynomials in s with coefficients in R, R(s) is the field of rational functions in s with coefficients in R, Rmxn is the set of m × n matrices with real coefficients Rmxn (s) is the set of m × n rational matrices, V denotes a vector space over R, or C, or R(s), by V we denote a basis matrix and by v a vector of V . We denote by deg the degree of a polynomial, det{A} is the determinant of a square matrix A, which is also denoted by |A|. Finally, if i takes values from the set {1, 2, . . . , n}, we shall denote this by i ∈∼n.

2.2 2.2.1

Classification of System Representations State Space Descriptions

Linear time invariant multivariable systems are represented in the time domain by state variable model S(A, B,C, D) : x˙ = Ax + Bu, y = Cx + Du (2.1) where x is an n-vector of the state variables, u is a p-vector of inputs and y is an m-vector outputs. A, B,C, D are respectively n × n, n × p, m × n, m × p matrices. The above description may be represented in an autonomous or implicit form as:     x x˙ A B O   I OO   u (2.2) u˙ = S(Φ , Ω ) : C D −I OOO | {z } y | {z } y˙ | {z } | {z } ,Φ

,ξ˙

,Ω

= [xt , ut , yt ]t

,ξ

is the composite vector, or implicit where Φ , Ω are the coefficient matrices and ξ vector of the state space description. The vector ξ contains the state, input and output vectors

Structural Methods for Linear Systems: An Introduction

51

and makes no distinction between them. The S(Φ , Ω ) description belongs to the general class of generalised autonomous differential descriptions [29], [31] S(F, G) : F z˙ = Gz

(2.3)

where F, G are r × k matrices and z is a k-vector. The above system is characterised by the matrix pencil pF − G, where p = d/dt denotes the derivative operator; pF − G completely characterises the state space description and it is referred to as the implicit system pencil . An alternative matrix pencil form for the state space description is obtained by taking Laplace transforms of (2.1), which lead to the s-domain description sx(s) ˜ − x(0) = Ax(s) ˜ + Bu(s) ˜ y(s) ˜ = Cx(s) ˜ + Du(s) ˜ where x(s), ˜ u(s), ˜ y(s) ˜ denote the Laplace transforms of x(t), u(t), y(t) vectors respectively and x(0) the initial value of x(t). We may express the above in a matrix form as sI − A −B x(s) ˜ x(0) sI − A −B = , P(s) = (2.4) −C −D u(s) ˜ −y(s) ˜ −C −D and the matrix coefficient P(s) is a matrix pencil entirely characterising the state space model and it is known as the Rosenbrock System Matrix Pencil [58].

2.2.2

Polynomial Models

The state space description of a linear system assumes that the system is described in terms of first order differential equations; however, this is not the most general internal description for linear systems. For a number of processes the most natural description is that defined by the general differential system [58], [6]:

Σp

¯ ¯ d A(p)v(t) = B(p)u(t) p= ¯ ¯ y(t) = C(p)v(t) + D(p)u(t) dt

(2.5)

¯ ¯ ¯ ¯ where A(p), B(p), C(p), D(p) are polynomial matrices in p of dimensions n¯ × n, ¯ n¯ × p, ¯ m × n¯ respectively, and v(t) is a vector valued function with values in Rn¯ known as pseudo-state vector [6]. The above description is known as Polynomial Model Description (PMD) and may be also represented as ¯ ¯ A(p) −B(p) 0 v(t) 0 ξ (t) = (2.6) = ⇔ T (p) ¯ ¯ −y(t) −C(p) −D(p) u(t) −y(t) where T (p) is known as the Rosenbrock’s System matrix. The relationship between PMDs and state space models in extensively treated in [58]. Such models are very important for electromechanical systems, or models derived by the Lagrange methodology.

2.2.3

Transfer Function Descriptions

The input-output, or transfer function model is described by y(s) ˜ = G(s)u(s), ˜ G(s) = C(sI − A)−1 B + D

(2.7)

52

N. Karcanias and E. Milonidis

where G(s) is an m× p rational matrix. The transfer function may also be described in a matrix fraction description form as G(s) = Nr (s)Dr (s)−1 = Dℓ (s)−1 Nℓ (s)

(2.8)

where Nr (s), Nℓ (s) are the m × p right, left polynomial matrix numerators respectively and Dr (s), Dℓ (s) are the p × p, m × m polynomial matrix denominators correspondingly. It will be assumed that Dr (s), Nr (s) are right coprime and Dℓ (s), Nℓ (s) are left coprime (see background on polynomial matrices below) [1], [6], [18]. Using (2.7) and the factorisation (2.8) we can readily obtain the following description ˜ = Nℓ (s)u(s) ˜ Dℓ (s)y(s)

(2.9)

˜ = Dℓ (s)−1 u(s) ˜ the description and by introducing the vector h(s) ˜ ˜ y(s) ˜ = Nr (s)h(s), u(s) ˜ = Dr (s)h(s) The above two lead to the following input output type representations for the system y(s) ˜ = 0, Tℓ (s) = [Dℓ (s)Nℓ (s)] [Dℓ (s), Nℓ (s)] −u(s) ˜ y(s) ˜ Nr (s) ˜ Nr (s) = h(s), Tr (s) = u(s) ˜ Dr (s) Dr (s)

(2.10)

(2.11)

(2.12)

The first, based on Tℓ (s), is referred to as Kernel input-output description, whereas the second based on Tr (s) as a parametric input-output description. The matrices Tℓ (s), Tr (s) based on coprime MFDs will be called right-, left- composite matrices.

2.3

Background on Polynomial Matrices and Matrix Pencils

The study of algebraic structure of linear systems represented by state space, or transfer function models heavily relies on the theory of polynomial matrices [13], [18] and matrix pencils [13]; here we review some of the fundamentals of their structure and introduce some useful notation. We consider matrices T (s) of dimension q × r with elements from the field of rational functions R(s), or the ring of polynomials R[s]; such matrices are called respectively rational, polynomial. The rank of T (s) over R(s) is denoted by ρ = rank{T (s)} and will be called the normal rank of T (s). T (s) may be viewed as a function of the complex variable s and thus for some s = z, rank{T (s)} = ρz < ρ ; such values s = z are called zeros of T (s) and ρz is called the local rank of T (s). The structure of zeros of T (s) is linked to study of certain form of equivalence defined on such matrices, which reveals the zeros as roots of invariant polynomials. Let T1 (s), T2 (s) be q × r polynomial matrices. These matrices are called R[s] -unimodular equivalent [13], or simply R[s]-equivalent, if there exist polynomial matrices Uℓ (s),Ur (s) of dimension q × q, r × r respectively with the property |Ur (s)| = c1 6= 0, |Uℓ (s)| = c2 6= 0 ( | · | denotes determinant) and called R[s]-unimodular such that: T1 (s) = Uℓ (s)T2 (s)Ur (s)

(2.13)

Structural Methods for Linear Systems: An Introduction

53

The above relation introduces an equivalence and for any matrix T (s) we have an equivalence class and associated invariants which are described by the following result. Smith Form Theorem [13]: If T (s) is a q × r polynomial matrix with normal rank ρ ≤ min(q, r) there exist unimodular matrices Uℓ (s),Ur (s) such that   0 0 f1 (s)   ..   .     (2.14) .. Uℓ (s)T (s)Ur (s) =   = S(s)   .    0 fρ (s)  0 0

where S(s) is q × r polynomial matrix f1 (s), . . . , fρ (s) are uniquely defined and f1 (s)/ f2 (s)/ . . . / fρ (s) denotes the successive divisibility (i.e., f1 (s) divides f2 (s) etc.).

The polynomials fi (s) are called invariant polynomials of T (s) and the set { fi (s), i = 1, . . . , ρ } is a complete invariant under R[s]-equivalence. The roots of fi (s) (including multiplicities) define finite zeros of T (s). The structure of these zeros (multiplicities and groupings) is defined by factorising the fi (s) into irreducible factors over the real, or complex numbers; for every zero z we define the set of z-elementary divisors by grouping all factors with root at z. The set of elementary divisors (for all zeros) is also a complete invariant under R[s]-equivalence. Note that although the Smith form under R[s]-equivalence defines the finite zero structure (finite frequencies), it does not convey any information on the structure at infinity; an alternative form is required and will be considered in a later section.

2.3.1

Matrix Divisors and Minimal Bases

Definition 2.1. For a matrix P(s) ∈ R p×m [s], rank{P(s)} = m we define [39], [18], [1]: (i) A matrix R(s) ∈ Rm×m [s] such that P(s) = P′ (s)R(s) is called a right matrix divisor (RMD) of P(s). If R(s) is any other RMD and R(s) = W (s)R(s) then R(s) is called a right greatest matrix divisor (RGMD) of P(s). If rank{P(s)} = p, the notions of left matrix divisors (LMD) and greatest left matrix divisors (LGMD) are defined similarly. (ii) Let P(s) = [p1 (s), . . . , pm (s)] ∈ R p×m [s] and rank{P(s)} = m. The set IP = {δi : δi = ∂ (pi (s)), i ∈ m} ˜ of the degrees of the columns pi (s) is defined as the set of column degrees and cP = ∑m i=1 δi as the column complexity of P(s). Row degrees and row complexity are defined in a similar manner. If pi (s) = pi,h sδi + . . . + pi,0 , then we may express it as

ˆ P(s) = [p1,h , . . . , pm,h ]diag{sδ1 , . . . , sδm } + P(s)

(2.15)

ˆ have degrees less than δi and Ph = [p1,h , . . . , pm,h ] = [P(s)]h ∈ R p×m The columns of P(s) is referred to as the high column coefficient matrix of P(s). If rank{Ph } = m, then P(s) is called column reduced. The high row coefficient matrix and row reducedness notions are defined similarly. (iii) A matrix P(s) ∈ R p×m [s] with rank{P(s)} = m is called right irreducible, or least degree, if all RMDs are R[s]-unimodular. A left irreducible matrix is defined in a similar manner. P(s) ∈ R p×m [s] with rank{P(s)} = m (or p) is called a minimal basis [12] if it is: (a) Right (left) irreducible (b) Column reduced.

54

N. Karcanias and E. Milonidis

If Pr , {Pi (s) ∈ Rki ×m [s], i ∈ ν˜ } is a set of matrices, then the matrix Pr   P1 (s) v   Tpr (s) ,  ...  ∈ Rk×m [s], k = ∑ ki Pv (s)

i=1

is called a matrix representative of Pr . Pr is right regular, if ρ (Tpr (s)) = m. If Pr is right regular, then a right common matrix divisors (RCMD) and a right greatest common matrix divisor (RGCMD) of Pr is defined as a RMD and a RGMD of Tρr (s) respectively. The set Pr is called right coprime (RC), if it is right regular and Tρr (s) is right irreducible. For a set of matrices with the same number of rows, the notions of a matrix representative, left regularity, left common matrix divisors (LCMD), left greatest common matrix divisor (LGCMD) and left coprimeness are defined in a similar manner. The above definitions for matrices over R[s] have their counterparts for matrices over R pr (s), R p (s) the ring proper and Ω −stable rational functions respectively with the appropriate changes [76], [77], [80]. Proposition 2.1. For an m × p rational matrix G(s) consider the matrix fraction description G(s) = Nr (s)Dr (s)−1 = Dℓ (s)−1 Nℓ (s) where Nr (s), Nℓ (s) are the m × p right, left polynomial matrix numerators respectively and Dr (s), Dℓ (s) are the p × p, m × m polynomial matrix denominators correspondingly. Then, (i) The pair Dr (s), Nr (s) is right coprime, iff the matrix Tr (s) = [Nr (s)t , Dr (s)t ]t has full rank and no zeros. (ii) The pair Dℓ (s), Nℓ (s) is left coprime, iff the matrix Tℓ (s) = [Dℓ (s), Nℓ (s)] has full rank and no zeros. With a polynomial, or rational matrix T (s) we may associate two important rational vector spaces (vector spaces of rational vectors and with scalars the rational functions R(s)) Nr (T ) = {x(s) : T (s)x(s) = 0, x(s) ∈ Rr [s]}

(2.16)

Nℓ (T ) = {yt (s) : yt (s)T (s) = 0, y(s)t ∈ R1×q [s]} Nr (T ), Nℓ (T ) are called respectively right-, left-rational vector spaces, where dimNr (T ) = r − ρ , dimNℓ (T ) = q − ρ ; with such spaces we can always define polynomial bases. If X(s) is an (r − ρ ) × r polynomial basis for Nr (T ), or of any rational vector space X with dimX = r − ρ , then it is called least degree if it has no zeros [60]. A polynomial basis X(s) = [x1 (s), . . . , xr−ρ (s)] with column degrees {d1 , . . . , dr−ρ } is said to be of least complexity, if Σ di = δ (X). Note that δ (X) denotes the degree of X(s), which is defined as the maximal of the degrees of all maximal order minors of X(s). A least degree and least complexity polynomial basis of Nr (T ) is a minimal basis and the ordered set of degrees {d1 , . . . , dr−ρ } are called right minimal indices and δr (T ) = Σ di as the right-order of T (s) [12], [60]. The notion of left minimal indices and left order are defined similarly on Nℓ (T ). Note that the sets of minimal indices are invariants of the corresponding rational vector spaces. A special case of a polynomial matrix is that of a matrix pencil [13], sF − G, where F, G are q × r real (or complex) matrices and s is an independent complex variable taking values on the compactified complex plane (that includes the point at infinity). For such matrices we define the notion of strict equivalence in the following way: Two pencils sF − G, sF ′ − G′ of dimension q × r are strict equivalent, if there exist real matrices Q, R of dimension q × q, r × r respectively such that

Structural Methods for Linear Systems: An Introduction sF ′ − G′ = Q(sF − G)R, |Q|, |R| 6= 0

55 (2.17)

The above introduces the notion of strict equivalence of matrix pencils and the equivalence classes are characterised by a set of invariants which will be defined subsequently. Pencils may ˆ ′ where s, sˆ are independent complex varibe represented in a homogenous form as sF ′ − sG ables. Frequencies on the compactified complex plane are represented as ordered pairs (α , β ), where at least one of the α , β is 6= 0. Pairs (α , β ) : β 6= 0 correspond to finite frequencies. With the homogeneous pencil sF − sG ˆ we associate the single variable pencils sF − G, sF − sG. ˆ The sets of invariants that may be defined under the strict equivalence are known as Kronecker Invariants and are summarised below:

2.3.2

Strict Equivalence Invariants of Matrix Pencils [13]

Elementary Divisors: The Smith form of the homogeneous pencil sF − sG ˆ defines a set of ˆ τ , sˆq . The set of elementary divisors elementary divisors of the following type: s p , (s − α s) ˆ τ , are called finite elementary divisors (fed) of sF − G, whereas those of the sˆq type s p , (α − s) are called infinite elementary divisors (ied) of sF − G.

Minimal Indices: A matrix pencil sF − G, where at least one of Nr {sF − G}, or Nℓ {sF − G} are non trivial {6= 0} are called singular, otherwise they are called regular. If Nr {sF − G} = 6 0, then the minimal indices of this rational space are denoted Ic (F, G) = {εi , i = 1, . . . , µ } and are referred to as column minimal indices (cmi) of the pencil. Similarly, if Nℓ {sF − G} then the minimal indices of this rational vector space are denoted by Ir (F, G) = {η j , j = 1, . . . , ν } and referred to as row minimal indices. This set of invariants are complete [36] for the strict equivalence of matrix pencils, that is they uniquely characterise the strict equivalence class of a matrix pencil. There is a uniquely defined element by the invariants which is referred to as Kronecker canonical form [13]. Kronecker Canonical Form of a matrix Pencil: Consider a matrix pencil sF − G and assume that its Kronecker invariants are of the types: elementary divisors of the type: {(s − α )τ , . . . ; sˆq } column minimal indices: {ε1 = . . . = εt = 0, ε j > 0, j = t + 1, . . . µ }, row minimal indices: {η1 = . . . = ησ = 0, ηi > 0, i = σ + 1, . . . , ν }. There always exists a pair of strict equivalence transformation Q, R such that Q(sF − R)R = blockdiag{Oσ ,t ; . . . Lε (s), . . . , Lη (s), . . . ; sFw − Gw } I 0 : (η +1)× η block (2.18) Lε (s) = s[Iε 0]−[0, Iε ] : ε ×(ε +1) block, Lη (s) = s η − 0 Iη sFw − Gw = block − −diag{sI − J(a); . . . ; sHq − Iq ; . . .}

where J(a) is the τ × τ Jordan block associated with (s − α )τ and Hq is a q × q nilpotent block (1s on the first super diagonal and the rest zero).

2.4 2.4.1

Dynamics, Stability, Controllability and Observability Solution of State Space Equations

The eigenvalues and eigenvectors of the matrix A define the internal dynamics of the system. The set of eigenvalues of the matrix A will be referred to as the system poles and with every

56

N. Karcanias and E. Milonidis

eigenvalue λ of A we have two eigenvalue-eigenvector problems Au = λ u, vt A = vt λ , vt u = 1

(2.19)

where u is a right - and vt a left-eigenvector and the triple (λ , u, vt ) is called a system mode. If φ (A) is the set of distinct eigenvalues then the structure of λ ∈ φ (A) is defined by the λ -Segre characteristic, S(λ ) = {νi , i ∈ q} ˜ that is the dimensions of λ -Jordan blocks in the Jordan form of A. Alternatively, S(λ ) is defined by the set of degrees of the (s − λ )ν type of the Smith form q of sIn −A, ∑i=1 νi = ρ is called the algebraic multiplicity and q the geometric multiplicity of λ . The maximal of all geometric multiplicities of the eigenvalues of A is referred to as the Segre index of A. The structure of eigenvalues defines the type of free motions and has implications on internal stability. For the state space model S(A, B,C, D), which is excited by an initial condition x(0) = x0 and a control input u(t) the corresponding solutions for the state and output trajectories are given by [1] Z x(t) = eAt x0 +

y(t) = CeAt x0 +

Z t 0

t

0

eA(t−τ ) Bu(τ )d τ ,

CeA(t−τ ) Bu(τ )d τ + Du(t)

(2.20)

The solutions above consist of two parts: The integral parts of (2.20) defining the forced response, which are contributed from inputs (and disturbances) and the remaining part of the solutions, which are the free response contributions where the matrix eA(t−t0 ) is called the state transition matrix. Taking Laplace transforms of (2.1), the solutions are expressed as: x(s) = (sI − A)−1 x0 + (sI − A)−1 Bu(s)

y(s) = C(sI − A)−1 x0 + {C(sI − A)−1 B + D}u(s)

(2.21)

where G(s) = C(sI − A)−1 B + D is the transfer function matrix of the system. Assuming for the shake of simplicity that A = U Λ V, where Λ is diagonal, and using the dyadic expansion of eAt , the spectral form of the state trajectory x(t) for a zero input response can be written as x(t) = eAt x(0) = eλ1 t u1 v1 x0 + . . . + eλn t un vtn x0 ⇒ x(t) =

n

∑ eλ t hvi , x0 iui 1

(2.22)

i=1

where h·, ·i denotes the inner product. The spectral form of the state trajectory when x(0) = 0 (initial condition) and for a nonzero input is x(t) =

Z t 0

eA(t−τ ) Bu(τ )d τ =

Z t n

∑ ui eλ (t−τ ) vti Bu(τ )d τ i

(2.23)

0 i=1

and if ∗ denotes convolution, then the total output response is then n

n

y(t) =

∑ γi eλ t hvi , x0 i + ∑ γi eλ t ∗ hβi , u(t)i

i=1

where

i

i

(2.24)

i=1

 t  vt1 β1  ..   ..  CU = C[u1 . . . u1 ] = [γ1 . . . γn ], V B =  .  B =  .  βnt vtn 

(2.25)

Structural Methods for Linear Systems: An Introduction

2.4.2

57

Internal-External and Total Stability

For linear, time invariant systems the notions of stability, which are more frequently used are defined next. We consider stability of equilibrium points, whereas stability of motion is always reduced to the previous case. Note that the origin (x = 0) is always an equilibrium point for S(A, B,C, D) models. Definition 2.2 ([1). [7]] The state space model S(A, B,C, D) will be called: (i) Internally stable in the sense of Lyapunov (LIS), if for any initial x(0) the zero input response (free motion, u(t) = 0) remain bounded for all t ≥ 0. (ii) Asymptotically internally stable, if for any initial state x(0) the zero input response remains bounded for all t ≥ 0 and tends to zero as t → ∞. This property will be referred to in short as internal stability (IS). (iii) Bounded Input Bounded Output Stable (BIBO), if for any bounded input the zero state output response x(0) = 0 is bounded. (iv) Totally stable (TS) if for any initial state x(0) and any bounded input u(t), the output, as well as all state variables are bounded. The notion of BIBO stability refers to the transfer function description and may also be referred to as external stability. A number of criteria for these properties, based on eigenvaluespoles, are summarised below [7]. Theorem 2.1. Consider the system S(A, B,C, D) with G(s) transfer function and let {λi = σi + jωi , i ∈∼n}, {p j = σ j + jω j , j ∈∼}v be the sets of eigenvalues, poles respectively. The system has the following properties:

(i) Lyapunov internally stable, iff σi ≤ 0, all i ∈∼n, and those with σi = 0 have a simple structure (algebraic multiplicity is equal to the geometric multiplicity). (ii) Asymptotically internally stable, iff σi < 0, all i ∈∼n. (iii) BIBO stable, iff σ¯ j < 0 all i ∈∼υ. (iv) Totally stable, if it is Lyapunov internally stable and BIBO stable. Note that IS implies BIBO-S and thus TS. BIBO stability does not always implies IS, since transfer function and state space are not always equivalent. If the two representations are equivalent (when system is both controllable and observable), then BIBO-stability is equivalent to IS and thus TS. Remark 2.1. Eigenvalues and poles are indicators of stability. Equivalent tests for stability, without computing the eigenvalues, poles are defined on the characteristic, pole polynomial by the Routh-Hurwitz conditions and equivalent tests. Tests for BIBO stability and corresponding indicators may be formulated on the impulse response of the system. For discrete time systems, the stability definitions and criteria are similar to those discussed here with the obvious changes (open closed right half place, becomes open, closed unit circle etc.). A stronger notion of stability is the notion of Finite Settling Time Stability (reaching the steady state in finite number of discrete time steps), which does not occur in continuous time system and it is related to Dead-Beat control [39].

58

N. Karcanias and E. Milonidis

2.4.3

Controllability and Observability

Some of the most fundamental concepts in control theory characterising the coupling of internal mechanism to its environment are those of controllability and observability [19], [20], [1], [7], [65]. These concepts are defined below. Definition 2.3. Consider the state space model S(A, B,C, D) and let X be its state space (Rn ). Then the system is called: (i) State controllable, or simply controllable, if there exists of finite time T > 0, such that for initial state x(0) ∈ X and any x1 ∈ X , there exists an input u(t), defined on [0, T ] that will transfer x(0) to x1 at time T (x(T ) = x1 ). Otherwise, it is called uncontrollable. (ii) State observable, or simply observable, if there exists of finite time T > 0, such that for initial state x(0), the knowledge of the input u(t) and output over the time interval [0, T ] suffices to determine the state x(0); otherwise, the system is called unobservable. (iii) Output controllable, if for any output vector y1 there exists T > 0 and an input u(t) defined over [0, T ] that transfers the output y(0) = 0 to y(T ) = y1 . Furthermore, if it is output controllable and the output can be steered over any interval of time on a pre-assigned curve, it will be called output function controllable. Note that strictly speaking, transferring to the zero state any initial state x(0) is referred to as controllability, whereas transferring the origin x(0) = 0 to any final state in the state space is referred to as reachability. These differences become essential in more general families of models than those considered here. The problem of observability has to do with reconstruction of initial conditions (measurement) and thus reconstruction of state trajectories; thus, we may refer to observability as state reconstructibility. A summary of tests for the above properties is given below. (see for instance [1], [7], [58], [22], [34] etc.) Theorem 2.2. (Controllability/Observability criteria): For the state space model S(A, B,C, D) with n, m, p number of states, outputs, inputs and with (λ , u, vt ) as the triple of eigenvalues, right and left-eigenvector respectively we have the following results: (a) Controllability : The system is state controllable, iff either of the equivalent conditions hold true: (i) All rows of e−At B are linearly independent on [0, +∞) over C. (ii) All row of (sI − A)−1 B are linearly independent over C. R tτ (iv) The controllability Grammian, Wc = 0T eAτ BBt eA d τ is nonsingular for all T (iii) If A has distinct eigenvalues then for every λ , vt B 6= 0t , any T > 0. (v) The n × (np) controllability matrix Qc = [B, AB, . . . , An−1 B] has rank n. (vi) The controllability pencil , Pc (s) = [sI − A, −B] has rank n for all s ∈ C, or equivalently it has no fed. (vii)The restricted controllability pencil, Rc (s) = sN − NA ( N is a left annihilator of B) has rank n for all s ∈ C, or equivalently, it has no fed. (b) Observability : The system is state observable, iff either of the equivalent conditions hold true. (i) All columns of CeAt are linearly independent on [0, +∞) over C. (ii) All columns of C(sI − A)−1 are linearly independent over C. (iii) If A has distinct eigenvalues then forR every λ , Cu 6= 0. t (iv) The observability Grammian, Wo = 0T eA τ Ct CeAτ d τ is nonsingular for any T > 0.

Structural Methods for Linear Systems: An Introduction

59

(v) The nm × n observability matrix, Qo = [Ct , At Ct , . . . , (At )n+1Ct ]t has rank n. (vi) The observability pencil, Pot (s) = [sI − At , −Ct ]t has rank n for all s ∈ C, or equivalently, it has no fed. (vii)The restricted observability pencil, Ro (s) = sM − AM (M is a right annihilator of C), has rank n for all s ∈ C, or equivalently, it has no fed. (c) Output Controllability: Output controllable, iff either of the equivalent conditions hold true: (i) All rows of G(s) are linearly independent over C. (ii) The matrix Qoc = [D,CB,CAB, . . . ,CAn−1 B] has rank m. (iii) The system is output function controllable, iff rank of G(s) is equal to m, over R(s). The above tests define equivalent indicators for the controllability, observability properties and demonstrate that both controllability, observability properties express the interaction of internal mechanism with the environment represented by the inputs, outputs and are shaped at the instrumentation stage of the process design. Remark 2.2. A generic system is always state controllable and observable. Thus, if all parameters in S(A, B,C, D) with fixed dimensions, are free the cases of uncontrollability, unobservability are nongeneric. Note however, that because of the process interconnections, we frequently deal with S(A, B,C, D) families with a fixed underlined graph and it is then the graph that may determine the controllability, observability properties. These graph based notions of controllability, observability are referred to as “structural”. Remark 2.3. Controllability, observability tests are based on the notions of rank of matrices, which generically is full. The degree of nonsingularity, singularity, measured by the smallest singular value, or the condition number is important indicators of “how well” the system is controllable, observable. Such issues are essential in design problems when either there are restriction on the size of gains in feedback matrices, or when there are limitation in the energy of control signals etc. The questions of controllability, observability may be equivalently interpreted as questions of controlling, or observing the system eigenvalues. This interpretation highlights the internal structure, illustrates certain aspects of “mode connectibility” to inputs, outputs and provides some structural tests for controllability, observability. Using the Jordan description of state equations (A is in Jordan canonical form) alternative tests for controllability, observability may be stated and are found in [7]. Some important implications of these tests are. Corollary 2.1. Let q be the Segre index of A (the maximal of all geometric multiplicities of the eigenvalues) and let Σ be the family of n, m, p fixed dimension systems, having the same q, but with otherwise arbitrary parameters. (i) Necessary condition for every S(A, B,C, D) ∈ Σ to the controllable and observable is that p ≥ q and m ≥ q respectively. (ii) Every system in Σ for which p < q, m < q is uncontrollable, unobservable respectively. Note that the value of q is frequently a property that may be inferred from the structure (diagram, graph) of the process and the nature of subprocesses and thus q may serve as a prime indicators on the necessary minimum number of inputs and outputs.

60

N. Karcanias and E. Milonidis

2.4.4

System Minimality

The concepts of uncontrollability, observability are essential in the establishment of the relationships between internal and external descriptions. This is illustrated by the following important result [20]. Theorem 2.3. For the system S(A, B,C, D), there is a coordinate transformation x′ = Ux such that the corresponding state space description S′ (A′ , B′ ,C′ , D′ ) has the form  ′   ′  ′   ′  x˙co Aco A′12 A′13 A′14 xco Bco ′  ′ ′ ′   B′   x˙co   xco 0  co   ′  =  0 Aco A23     (2.26)  x˙   0 0 A′ 0   x′  +  0  u co co co ′ ′ ′ x˙co 0 0 0 Aco xco 0

′ ′ y = [0,Cco ,Cco , 0]x′ + Du ′ , x′ , x′ , x′ are subvectors of x′ which are: controllable-unobservable, controllablewhere xco co co co observable, uncontrollable-observable and uncontrollable-unobservable states. Furthermore, the systems S(A, B,C, D) and S′ (A′ , B′C′ , D′ ) have the same transfer function which is expressed as: ′ G(s) = Cco (sI − A′co )−1 B′co + D (2.27)

This result expresses the Kalman decomposition of the state space, which shows that the transfer function represents only the controllable and observable subsystem, but not in general the whole S(A, B,C, D) system. Remark 2.4. The transfer function and the state space descriptions are completely equivalent, iff the system is both controllable and observable. The dimension of the controllable and observable subsystem is defined as the McMillan degree of G(s) and it is denoted by δM (G). We will see later on that δM (G) is the degree of the pole polynomial of G(s). Furthermore, any realization S(A, B,C, D) is observable and controllable iff its dimension is equal to δM (G). Two important concepts related to controllability, observability are those of stabilizability, detectability [74]: Definition 2.4. The system S(A, B,C, D) will be called: (i) Stabilizable, if the unstable eigenspace of A is contained in the controllable subspace of the system. (ii) Detectable , if the unobservable subspace of the system is contained in the stable eigen space A. Corollary 2.2. The system S(A, B,C, D) is: (i) Stabilisable, iff its uncontrollable eigenvalues (in the modal sense) are stable. (ii) Detectable, iff its unobservable eigenvalues (in the modal sense) are stable. Those two more relaxed conditions (than controllability, observability) have implications as far as using transfer functions as design models. In fact, if the system is both stabilisable and detectable, then the transfer function may be used for feedback design, but not otherwise. The uncontrollable, unobservable, uncontrollable and unobservable eigenvalues are also referred to as input-output, input-output decoupling zeros (idz, odz, i-odz) [58] and the corresponding sets, including multiplicities, will be denoted by ZID , ZOD , ZIOD respectively. These sets are computed by the following property [58], [36].

Structural Methods for Linear Systems: An Introduction

61

Corollary 2.3. For the sets of decoupling zeros we have the properties: (i) ZID is defined by the roots of the f-ed of Pc (s) = [sI − A, −B], or Rc (s) = sN − NA. (ii) ZOD is defined by the roots of f-ed of Po (s) = [sI − At , −Ct ]t , or Ro (s) = sM − AM. (iii) ZIIOD = ZID ∩ ZOD . Remark 2.5. The system is stabilisable and detectable iff the f-ed of Pc (s) (or Rc (s)) and Po (s) (or Ro (s)) are stable. Output controllability is only concerned with getting to y(T ) = y1 final value, but nothing in the definition states that the output will stay at y(t) = y1 value, or track a specified function c(t). Output function controllability addresses this property, but the required u(t) might contain impulses. Output function Controllability with a smooth input u(t) will be referred to as smooth output controllability [60].

2.5 2.5.1

Poles and Zeros of State Space Model Eigenvalues, Eigenvectors and Free Rectilinear Motions

For a single input, single output (SISO) system represented by a rational transfer function g(s) where g(s) = n(s)/d(s), where n(s), d(s) are coprime polynomials with deg{n(s)} = r and deg{d(s)} = n we define as finite poles the roots of d(s) and as finite zeros the zeros of n(s). If r < n we say that g(s) has an infinite zero of order n − r, and if r > n then g(s) has an infinite pole with order n − r. Finite and infinite poles have the property that the gain of the transfer function becomes unbounded (tends to infinity) and finite and infinite zeros are those frequencies for which the gain vanishes. In this sense, the notion of poles and zeros are dual since the first characterises resonance and the second antiresonances. It is this basic property that motivates a number of the definitions and problems that relate to multivariable poles and zeros. For a state space model the internal natural dynamics are defined by the zero input response (u(t) = 0) which in turn is characterised by the eigenvalues and eigenvectors of the state matrix A (see System Description in Time Domain), which determine the solution space of the autonomous system. Using S(Φ , Ω ) description for u(s) ≡ 0 we get the zero input differential description. A 0 x(t) I O x(t) ˙ S(H, Θ ) : (2.28) = C −I y(t) OO y(t) ˙ which of course is equivalent to

x(t) ˙ = Ax(t), y = Cx(t)

(2.29)

Definition 2.5. We define as poles of the state space model the eigenvalues of A and as pole directions the corresponding eigenvectors. If (λ , u) is a pair of an eigenvalue and eigenvector of A, then for every initial condition x(0) = cu (c constant) the corresponding solution of S(H, Θ ) is

62

N. Karcanias and E. Milonidis x(t) = exp(λ t)cu, y(t) = exp(λ t)Cu

(2.30)

which are λ exponential trajectories in the state and output spaces along the constant direction xλ = u and yλ = Cu. Such straight line exponential motions are frequently called rectilinear motions and xλ , yλ are called state pole and output pole directions correspondingly. The fundamental operator describing the pole structure is the pencil W (s) = sI − A

(2.31)

which is called the state pole pencil. The eigenstructure of A is defined by its Jordan form and this is defined algebraically by the elementary divisors of the Smith form of W (s) [13]. The presence of elementary divisors with degree higher than one implies the existence of Jordan blocks for the corresponding eigenvalues. The dynamic characterisation of such multiple eigenvalues is in terms of generalised rectilinear motions (involving terms of the type exp(λ t)t i , i = 1, 2, . . .). The pole-zero duality motivates the definition of the output zeroing problem, that is investigation of the type of solutions of the system for which the output is identically zero (the output now is zero instead of the input for the case of pole). When y(t) ≡ 0 the S(Φ , Ω ) description is reduced to the system I 0 x(t) ˙ A B x(t) = (2.32) S(Γ , ∆ ) : 0 0 u(t) ˙ C D u(t) the solutions of which describe the zero dynamics of the original system. The fundamental pencil of S(Γ , ∆ ) is the Rosenbrock system matrix pencil [58] and its structure describe the zero structure of the system [46]. The dynamic characterisation of the zero structure benefits by extending first the notion of free rectilinear motion to that of the forced rectilinear motion.

2.5.2

Forced Rectilinear Motions and Frequency Transmission

Free rectilinear motions are characterised by frequencies, which are fixed and are the eigenvalues of the matrix A. We consider now the generation of rectilinear motions for arbitrary complex frequencies λ , but in a forced manner. This is described next [46]: Proposition 2.2. For the state space model S(A, B) a rectilinear motion in the state x(t) = exp(λ (t))xλ , t ≥ 0, where x(0) = xλ and λ a given complex number can be generated, if and only if the input is also rectilinear of the same type, i.e. u(t) = exp(λ t)uλ , t ≥ 0 and the triple ( λ , xλ , uλ ) satisfy the condition (λ I − A)xλ = Buλ

(2.33)

This result establishes the property that every complex frequency can be transmitted in a simple rectilinear way through the input and state space of the system. The totality of xλ vector solutions of (2.33) is denoted by T (λ ) and it is called the transmission space of λ . Note that (2.33) may also be written as: x [λ I − A, −B] λ = 0 (2.34) uλ

and this demonstrates that the triple (λ , xλ , uλ ) satisfies a generalised eigenvalue eigenvector problem on the pencil Pc (s) = [sI − A, −B] which is known as input-state space pencil. A number of interesting properties are listed below [22], [26].

Structural Methods for Linear Systems: An Introduction

63

Proposition 2.3. For every λ ∈ C we have the properties: (i) The set of xλ solutions by T (λ ) define a p-dimensional linear vector space. (ii) If λ is not an eigenvalue of A then for any arbitrary uλ the corresponding xλ is defined by xλ = (λ I − A)−1 uλ . (iii) If N is an (n − p) × n left annihilator of B (NB = 0 and N is a basis matrix for left null space of B) and B† is a p × n left inverse of B (B† B = I p ), then the computation of ( λ , xλ , uλ ) may be reduced to the decoupled set of conditions (λ N − NA)xλ = 0, uλ = B† (λ I − A)xλ

(2.35)

(iv) For any triple (λ , xλ , uλ ) there exists a p × n matrix L, which may be interpreted as a state feedback such that uλ = Lxλ . Under this condition equation (2.34) becomes (λ I − A − BL)xλ = 0

(2.36)

and thus (λ , xλ ) becomes a pair of closed loop eigenvalue and eigenvector under an appropriate state feedback. The triples (λ , xλ , uλ ) defines closed loop eigenvalues and eigenvector and the space Tx (λ ) defines the family of closed loop eigenvectors for a given λ which may be generated under suitable state feedback. The pencil Rc (s) = sN − NA that provides the decoupled characterisation in (2.35) is referred to as restricted input state pencil. The structure of Pc (s), Rc (s) describe the controllability properties of the system and are instrumental in generating the closed loop eigenstructure. The result for the frequency transmission may now be extended to a general frequency transmission from input to state and then output spaces [46], [32]: Theorem 2.4. For the state space model S(A, B,C, D) a pair of forced rectilinear motions in the state x(t) = exp(λ t)xλ ,t ≥ 0, where x(0) = xλ and output y(t) = exp(λ t)yλ ,t ≥ 0, where λ is a given complex frequency can be generated in a forced mode, if and only if the input is also rectilinear of the same type, i.e. u(t) = exp(λ t)uλ ,t ≥ 0 and the quadruple (λ , xλ , uλ , yλ ) satisfies the condition λ I − A −B xλ 0 = (2.37) −C −D uλ −yλ For every complex λ the above condition is always satisfied for a triple (xλ , uλ , yλ ). The family of solutions of all xλ is the space Tx (λ ) and the corresponding set of solutions yλ is denoted by Ty (λ ) and called the output transmission space of λ . Note that the transmission in the input space Tu (λ ) covers the entire input space. A diagram illustrating the general frequency transmission is shown below. It is readily seen that condition (2.37) represents a generalised eigenvalue eigenvector problem for λ and vector [xtλ , utλ , ytλ ]t which is defined on the implicit system pencil sΦ − Ω .

2.5.3

Frequency Transmission Blocking and State Space Zeros

The general frequency transmission problem enables the formulation of “frequency transmission blocking” [46], [32] as shown below:

64

N. Karcanias and E. Milonidis

INPUT SPACE U

ue u(

D y(

)

OUTPUT SPACE Y

t

)

B

ye

C

t

STATE SPACE X

( )

xe

t

A

Fig. 2.1. General Frequency Transmission Theorem 2.5. For the state space model S(A, B,C, D) there exists a forced rectilinear motion in the state x(t) = exp(z(t))xz ,t ≥ 0, where x(0) = xz such that the output is identically zero, i.e. y(t) ≡ 0, if and only if the input is also rectilinear of the same type, i.e. u(t) = exp(zt)uz ,t ≥ 0 and the triple (z, xz , uz ) satisfies the condition zI − A −B xz =0 (2.38) −C −D uz This result establishes a link of the output zeroing problem with the existence of complex numbers z for which the generalised eigenvalue eigenvector problem has a solution; such solutions exist as long as the right null space of P(z), Nr {P(z)} = 6 0 for some complex number z. The nature of solutions depends on the existence of the rational vector space defined by Nr = Nr {P(s)} = {ξ (s) polynomial vector : P(s)ξ (s) = 0}

(2.39)

and referred to as right system space. Systems may be classified as follows: Definition 2.6. A system S(A, B,C, D) is called right regular if Nr = 0 and it is called right singular if Nr 6= 0. A right regular system implies that the rational space Nr is the 0 space; however, Nr {P(z)} may be nonzero for some specific complex number z. For right singular systems, condition (2.38) is satisfied for any complex z. Remark 2.6. Note that if ρ = rankR(s) {P(s)} ≤ min(m + n, p + n) is the normal rank of P(s) we have the following characterisation: System is Right Regular ⇔ ρ = n + p,

(2.40)

System is Right Singular ⇔ ρ < n + p

(2.41)

Structural Methods for Linear Systems: An Introduction

2.5.4

65

Right Regular Systems

For this class of systems the Output Zeroing Problem establishes the link with the Smith structure of P(s) as indicated below: Corollary 2.4. If S(A, B,C, D) is right regular, then the Output Zeroing Problem has a solution for some complex number z, if and only if z is a zero of the Smith form of P(s). Definition 2.7. For any system S(A, B,C, D) we define as its set of state space zeros the roots of the invariant polynomials of the Smith form of P(s) (including multiplicities). For right regular systems to every zero z there corresponds a pair (xz , uz ) which are called respectively state and input zero directions. For the case of elementary divisors with degree higher than one, higher order rectilinear motions exist (based on P(s) pencil) which characterise such multiplicities. In the following we shall assume for the sake of simplicity that all elementary divisors of P(s) are linear (degree one). The properties of the zeros and zero directions are summarised below.

2.5.5

Properties of Zero Directions

Theorem 2.6. [32] For a right regular system having a set of distinct zeros Λz = {z1 , . . . , zk } the following properties hold true: (i) For every zero z the pair (xz , uz ) is uniquely defined modulo c, constant. The triple (z, xz , uz ) satisfies the generalised eigenvector condition and the output zeroing condition Axz = zxz + Buz , Cxz + Duz = 0

(2.42)

(ii) The set of 1-dimensional spaces {xzi }, i = 1, 2, . . . , k are linearly independent and their direct sum Vz = {xzi }, ⊕ . . . ⊕ {xzk }, is a maximal dimension space for which the equivalent conditions hold true: (a) AVz ⊂ Vz + B, Vz ∩ B = 0, B = Im {B} (2.43)

(b) For any n × k basis matrix V for V there exists a p × k matrix U and a k × k matrix A¯ z , where A¯ z has a Jordan form uniquely defined by the set of elementary divisors of the Smith form of P(s) : AV = V A¯ z +U (2.44)

The above results stated for distinct zeros also generalises to the general case with the difference that the 1-dimensional spaces {xz } now become Jordan structure type spaces [17]. The space Vz associated with zeros is referred to as the Fixed spectrum output Nulling space.

66

N. Karcanias and E. Milonidis

2.5.6

Right Singular Systems

The class of right singular systems includes those with more inputs than outputs (p > m), as well as those for which ρ < min(m + n, p + n). The latter condition always implies right singularity and such systems are called degenerate; if ρ = min(m + n, p + n) the system is called nondegenerate. Degeneracy always implies right singularity but not vice versa. For right singular systems the rational vector space Nr defined in (2.39), we may always define minimal polynomial bases, i.e., matrices with no zeros and full rank high coefficient matrix [31], [23]. Such matrices are referred to as Right Nulling Minimal Bases and if r = dimNr {P(s)}, any such a basis matrix may be represented as: X(s) Tζ (s) = , X(s) = [x1 (s), . . . , xr (s)], U(s) = [u1 (s), . . . , ur (s)] (2.45) U(s) where xi (s), ui (s) are n-, p-dimensional polynomial vectors respectively with corresponding degrees deg{xi (s)} = εi , deg{ui (s)} = εi + 1. Any pair defined by the partitioning of the columns of Tζ (s) (x(s), u(s)) is called a right pair and if x(s) = a0 + a1 + . . . + sε aε , u(s) = b0 + sb1 + . . . + sε bε + sε +1 bε +1 , then {ai }, {bi } are the supporting sets of x(s), u(s) respectively and the linear space Rε = span{a0 , a1 , . . . , aε } is the supporting linear space of xε (s). The significance of right pairs in characterising the zero dynamics of systems is discussed next: Theorem 2.7. [22], [23] Let R be a subspace of the state space of the right singular system S(A, B,C, D) and let δ i (t) denote the i-th order impulse. If x(0− ) ∈ R, necessary and sufficient condition for a control input u(t) = δ (0) (t)b0 + δ (1) (t)b1 + . . . + δ (k) (t)bk to result in y(t) ≡ 0 is that the state trajectory is of the type x(t) = δ (0) (t)a0 + . . . + δ 1 (t)a1 + . . . δ k−1 (t)ak−1 , the space span{a0 , a1 , . . . , ak−l } ⊆ R, x(0− ) ∈ span{a0 , . . . , ak−1 } and the Laplace transforms x(s), ˜ u(s), ˜ satisfy the conditions sI − A −B x(s) ˜ =0 (2.46) −C −D u(s) ˜ The above result establishes an alternative form of output zeroing based on impulsive trajectories. Such characterisation is intimately linked to the minimal basis theory and provides an alternative, impulse based dimension, to the output nulling properties. The geometry of right nulling minimal bases has a number of properties, which are summarised below.

Geometric Properties of Right Nulling Minimal Bases Theorem 2.8. [23], [32], [70] For the minimal basis in (2.45) with the state vectors {x1 (s),. . .,xr (s)}, deg{xi (s)} = εi , {a j } is the supporting set and Rε i supporting space, the following properties hold true: (i) The vectors in the εi+1 set {a j }i are linear independent and {a j }i is a basis for the εi+1 dimensional space Rε i , which is a minimal dimension controllability subspace with the output nulling property. (ii) The set of subspaces {Rε1 , . . . , Rεr } is linearly independent and the subspace R = Rε1 ⊕ . . .⊕Rεr is the maximal output nulling controllability subspace of the system and dimR = ∑ri=1 εi + r.

Structural Methods for Linear Systems: An Introduction

67

(iii) Each subspace Rε satisfies the conditions ARε ⊆ Bε + B and Rε ∩ B 6= 0; furthermore, for every ε + 1 symmetric complex set Λ there exists a basis matrix Rε , a n × (ε + 1) matrix U and a (ε + 1) × (ε + 1) matrix A with spectrum Λ such that ARε = Rε A +U

(2.47)

The subspaces V satisfying the condition AV ⊆ V + B are known as (A, B)-invariant subspaces [3], [74]; for such spaces, for any initial condition x(0) there exists an input u(t) such that the trajectory is restricted in V for all t ≥ 0. Controllability subspaces [74] are special classes of (A, B)-invariant subspaces which have the additional property that any two points in the subspace may be joined in finite time by a trajectory (generated by some appropriate input) that lies entirely in the subspace. An important property of controllability subspaces is the assignment of the spectrum of the restriction map A in (2.47). This latter property is behind the following alternative form of the output zeroing problem. Corollary 2.5. Consider the right singular system S(A, B,C, D) and let λ be an arbitrary complex number. There always exists a control input u(t) = exp(λ t)uλ such that the state trajectory is x(t) = exp(λ t)xλ and y(t) ≡ 0. Furthermore, the triple (λ , xλ , uλ ) satisfies the conditions λ I − A −B xλ =0 (2.48) −C −D uλ The vectors xλ are always vectors of the maximal output nulling controllability subspace R and this suggests the following characterisation of zeros as unique solutions of output zeroing. Theorem 2.9. [32] For the right singular system S(A, B,C, D) there exists an output nulling solution x(t) = exp(zt)xz , u(t) = exp(zt)uz , t ≥ 0 such that x(0) = xz ∈ / R and y(t) ≡ 0, iff z is a zero of an invariant polynomial of P(s) and the triple (z, xz , uz ) satisfies condition (2.48). The above result states that although for initial conditions in the space R we have output zeroing problems with arbitrary frequencies, additional initial conditions may be found outside the space R that are linked to specific frequencies. These are the zeros of the Smith form and the analysis given for right regular systems applies to them. With such solutions a space V f , the output nulling fixed spectrum space, is defined which is independent of R. The space W = V f ⊕ R is the maximal output nulling (A, B)-invariant subspace of the system characterised by a finite spectrum and W .

2.5.7

Frequency Transmission Blocking for Infinite Frequencies

Frequency transmission blocking has been considered so far for finite frequencies. Such problems however, may be posed also for frequencies which tend to infinity. There are two ways such problems may be posed. The first is to use the notion of blocking of impulsive input and state trajectories (similarly to the treatment given for right singular systems) and the second is to use duality between 0 and infinite frequencies (described by the w = 1/s bilinear transformation). The second approach is used here, whereas the impulsive approach will be demonstrated on the input-output model formulation of frequency blocking later on. The essence of this duality may be described on the autonomous differential system S(F, G) in (2.4) as follows:

68

N. Karcanias and E. Milonidis

Differentiator – Integrator duality [29] : If F, G are r × k real matrices, we may associate the following two autonomous differential systems S(F, G) : F pz = Gz, p = d/d ˆ G) : F zˆ = pGˆz S(F,

(2.49) (2.50)

Those two systems have associated pencils (sF − G) and (F − sG) ˆ where s, sˆ are independent complex variables and they are special cases of the homogeneous matrix pencil sF − sG. ˆ As such the pencils are called algebraically dual, the systems are characterised by differentiator integrator duality and the following relationships hold between the set of their elementary divisors (ed) [29], [31]. Lemma 2.1. [29]: For the pencils (sF − G) and (F − sG) ˆ the following properties hold true: (i) A zero ed of sF − G, s p , is an infinite ed of F − sG. ˆ (ii) An infinite ed of sF − G, sˆq is a zero ed of F − sG. ˆ (iii) A finite non zero ed of sF − G, (s − α )τ is a finite nonzero ed (sˆ − α −1 )τ of F − sG. ˆ The above properties suggest that infinite frequency behaviour of S(A, B,C, D) may be studied as an equivalent zero frequency problems on the algebraically dual system ˆ B,C, D) : {x(t) ˙ˆ + Bu(t), ˙ˆ y(t) ˙ˆ + Du(t) ˙ˆ S(A, ˆ = Ax(t) ˆ = Cx(t)

(2.51)

ˆ B,C, D) be its algebraically Theorem 2.10. [30] Let S(A, B,C, D) be right regular and S(A, dual. If {xˆ2 , . . . , xˆk } is a chain of vectors, then for all x(0) ˆ = xˆi+1 , i = 1, 2, . . . , k − 1, there ˙ˆ = uˆi + t uˆi−1 + ti−1 uˆ1 ,t ≥ 0 such that x(t) ˙ˆ = xˆi + t xˆi−1 + . . . + t i−2 x2 exists a control input u(t) and y(t) ˆ ≡ 0, if and only if the vectors xˆi , uˆi satisfy. I 0 xˆi A B xˆi−1 = , i = 1, 2, . . . , k, xˆ0 = 0, uˆ0 = 0 (2.52) 0 0 uˆi C D uˆi−1 The significant of the above, as far as the characterisation of the infinite structure of P(s) is indicated by the following result: Corollary 2.6. The set of conditions (2.52) have a nontrivial solution if P(s) has an infinite elementary divisor sˆq , q > 1, where k ≤ q − 1. Furthermore, q − 1 is the maximal length chain of independent vectors {xˆ2 , . . . , xˆk } for which conditions (2.52) are satisfied. The set of uˆi and xˆi vectors are then defined by:    D 0 ··· 0 uˆ1    CB D 0   uˆ2   =0    .. . . . .   ..    . uˆi−1 CAi−3 CAi−4 · · · D and xˆi = Ai−2 Buˆ1 + Ai−3 Buˆ2 + . . . + Buˆi−1

(2.53)

Structural Methods for Linear Systems: An Introduction

69

The above results suggest that every infinite elementary divisor sˆq of P(s) is characterised by a q − 1 chain of vectors {xˆ2 , . . . , xˆq } linked to an output zeroing problem. The number q¯ = q − 1 is called an order of an infinite zero of the system. The vectors uˆi , xˆi are called input-state-infinite zero directions. The subspace Vq = span{xˆ2 , . . . , xˆq } has dimension q¯ = q − 1 [78] and it is called an output-zeroing asymptotic subspace [32], [17]. Such subspaces behave as the limits of transmission subspaces where the associated frequency λ becomes unbounded ( |λ | → ∞); for this reason this class is referred to as asymptotic, or sliding subspaces [71]. The set of subspaces Vqi associated with nonlinear ied are linearly independent and V∞ = Vq1 ⊕ . . . ⊕ Vqr is the maximal asymptotic output nulling space and the space Z = V f ⊕ P ⊕ V∞ is the general output nulling space of the system [32], [17].

The results in this section reveal that the invariants of the system matrix pencil such as finite and infinite elementary divisors and row minimal indices characterise different aspects of the dynamics of a system. The set of row minimal indices may have an interpretation similar to those of row minimal indices, but on the transposed dual system S′ (At , Bt ,Ct , Dt ).

2.5.8

Zero Structure and System Transformations

The significance of zeros is due to that they do not change under a number of transformations that may be applied on a state space model. The general set of such transformations is described in Figure (2.2) where the following notation is used: T : m × m Output Co-ordinate transformation, |T | 6= 0; Q, Q−1 : n × n pair of State Co-ordinate transformations, |Q| 6= 0; F : m × p Constant Output Feedback matrix; L : n × p State Feedback matrix; K : n × m Output Injection matrix. The above set of transformations (R, T, Q, Q−1 , L, K) when applied on the original system described by P(s), produce a system S′ (A′ , B′ ,C′ , D′ ) described by P′ (s) system matrix where P(s), P′ (s) are strict equivalent since −1 sI − A −B QO Q K P′ (s) = (2.54) −C −D L R O T Theorem 2.11. [34], [68], [52] The set of finite and infinite elementary divisor, column and row minimal indices of the system matrix pencil P(s) is invariant under the group of transformations that involves: state, input, output co-ordinate transformations, state feedback, output injection and output feedback. The zero structure as defined by the above set of invariants is therefore unaffected by this group of transformations which is usually referred to as Kronecker transformation group. This is why the state space zeros are also referred to as invariant zeros. The set of column, row minimal indices of P(s) are called respectively right -, left indices of the system. The significance of such invariants as far as the fundamental properties of controllability and observability under feedback transformations is described by the following result [34], [52], [74]. Theorem 2.12. Consider the system S(A, B,C). The following properties hold true: (i) (a) The controllability properties remain invariant under state feedback. (b) If the system has finite zeros and/or left indices then there exists output injection that can make the system maximally uncontrollable.

70

N. Karcanias and E. Milonidis

F

K

D

+

uc

~ u

+

u R

B +

+

+

+

+

~ x

Q

x

z

x Q-1

~ x

+

y

C

T

~ y

A L

Fig. 2.2. The general set of State Space Transformations (ii) (a) The observability properties remain invariant under output injection. (b) If the system has finite zeros and/or right indices, then there exists state feedback that can make the system maximally unobservable. The nature of the above results stems from that the presence of finite zeros and/or right indices implies the existence of a subspace W = V f ⊕ R which has to be in Nr (C). Given that W is (A, B)-invariant there exists state feedback L such that W becomes an eigenspace of (A + BL); moreover W is in Nr (C) which implies that the system becomes maximally unobservable. Of course, for a generic feedback L we have preservation of observability, but this property is not true for all L.

2.5.9

The Zero Pencil of Strictly Proper System

The computation of the invariants of the pencil P(s) is a problem reduced to finding the Kronecker form. Such computations may be considerably reduced for the case of strictly proper system by defining the zero pencil, which has considerably smaller dimensions. If D = 0 then the condition y(t) = Cx(t) = 0 implies that the differential system S(Γ , ∆ ) may be reduced to the following equivalent description S(NM, NAM) : NM v(t) ˙ = NAMv(t), x(t) = Mv(t), u(t) = B† {x(t) ˙ − Ax(t)}

(2.55)

where N, M are (n − p) × n, n × (n − m) respectively base matrices for Nℓ {B}, Nr {C} and B† is a p × n left inverse of B. The matrix pencil Z(s) = sNM − NAM associated with S(NM, NAM) has dimensions (n − p) × (n − m) (much smaller than (n + m) × (n − p)P(s)) and it is called the zero pencil of the system [32]. Z(s) characterises the zero dynamics of the system in an autonomous form (that does not involve the inputs) and its structure is related to that of P(s) as shown below [32], [44], [36].

Structural Methods for Linear Systems: An Introduction

71

′

Corollary 2.7. The pencils P(s) and Z(s) have the same sets of finite e.d. If sˆq is an infinite ′ e.d of Z(s), then q′ = q − 2, where sˆq is an infinite e.d. of P(s) for which q ≥ 3. If ε ′ , η are ′ row minimal indices of Z(s), then ε = ε − 1, η ′ = η − 1, where ε , η are the corresponding indices of P(s). The direct link between the invariants of the two pencils allows the computation of the zero structure from the Kronecker form of the smaller pencil Z(s). A direct link between properties of the state space parameters and the number of s, ˆ and sˆ2 infinite elementary divisors of P(s) [78], [30] allows the recovery of the full zero structure from Z(s). The smaller dimensions pencil Z(s) allows the study of properties related to the number, as well as computation of zeros and the results are summarised below [32]: Corollary 2.8. For the strictly proper system S(A, B,C) the number of finite zero σ f has the following properties: (i) If p 6= m and the system S(A, B,C) is generic, then σ f = 0. Nongeneric systems with p 6= m have a number of zeros σ f such that σ f ≤ min(n − p, n − m). (ii) If p = m, S(A, B,C) always has zeros (finite and infinite) which have the following properties (a) The zeros are defined as the roots of the zero polynomial z(s) = |Z(s)|. (b) The number of zeros σ f ≤ n − p and equality holds when CB has full rank. If CB is rank deficient, then some of the finite zeros migrate to infinity.

2.5.10

Decoupling Zeros

The term zero has been also used in the literature for values of complex numbers, which are not linked to output zeroing problems. The most common term used is that of decoupling zeros which are defined below: Definition 2.8. For the System S(A, B,C, D) we define: (i) The set of Input Decoupling Zeros as the set of eigenvalue (including multiplicities) of A, which are uncontrollable. (ii) The set of Output Decoupling zeros, as the set of eigenvalues (including multiplicities) of A, which are unobservable. (iii) The set of Input-Output Decoupling zeros, as the set of eigenvalues (including multiplicities) of A, which are both uncontrollable and unobservable. The above notions may be characterised algebraically in terms of the structure of sets of elementary divisors of the Smith form of matrix pencils. If M is an n × (n − m) right annihilator of C and N is an (n − p) × n left annihilator of B then [34]: Input Decoupling zeros: The structure of input decoupling zeros is defined by the set of elementary divisors of either of the pencils Pc (s) = [sI − A, −B], Rc (s) = sN − NA

(2.56)

Output Decoupling Zeros: The structure of output decoupling zeros is defined by the set of elementary divisors of either of the pencils

72

N. Karcanias and E. Milonidis P0 (s) = [sI − At , −Ct ]t , R0 (s) = sM − AM

(2.57)

The characterisation of input-output decoupling zeros structure may be achieved using the Kalman decomposition of the state space [20], [34]. There are certain links between decoupling zeros and invariant zeros and these are described below. Corollary 2.9. For the strictly proper system S(A, B,C) the following relationship hold between zeros and decoupling zeros: (a) If m > p, then every output decoupling zero is also an invariant zero; in general, input decoupling zeros are not invariant zeros. (b) If m < p, then every input decoupling zero is also an invariant zero; in general, output decoupling zeros are not invariant zeros. (c) If m = p, then every input and every output decoupling zero is also an invariant zero. (d) There always exists state feedback that makes all invariant zeros output decoupling zeros. Similarly, there always exists output injection that makes all invariant zeros input decoupling zeros.

2.6

Poles and Zeros of Transfer Function Models

The state space characterisation of poles links them with the eigenvalues, a subset of which appear as poles of the transfer function matrix G(s). Thus, in an input-output set-up we expect poles to be the frequencies for which the gain explodes in elements of G(s). On the other hand, the characterisation of zeros (right regular systems) implies that a zero s = z has an associated direction xz = (zI − A)−1 Buz and that G(z)uz = 0, G(s) = C(sI − A)−1 B + D

(2.58)

Thus, transfer function zeros correspond to frequencies where G(s) loses rank beyond its normal rank, θ = rankR(s) {G(s)}. The finite poles and zeros of transfer functions may also be characterised by corresponding dynamic problems.

2.6.1

Dynamic Characterisation of Transfer Function Poles and Zeros

Dynamic problems similar to those used for characterising state space poles and zeros may be used for transfer function matrix models as shown below [9]: Theorem 2.13. Given a transfer function matrix G(s), then a complex number p is a pole of G(s) if and only if there is an input u(t) = ∑ki ui δ (i) (t), δ (i) (t) is the ith order impulse, such that with zero initial condition the output is y(t) = y p exp(pt), t > 0

(2.59)

The above implies that p is a pole, if there exists an impulsive input that transfers the 0 state to an appropriate state at t = 0+ (point on an eigenvector) such that the resulting output is rectilinear. The relationship with the state space pole characterisation is clear.

Structural Methods for Linear Systems: An Introduction

73

Theorem 2.14. Given a right regular transfer function matrix G(s) (Nr {G(s)} = 0), then the complex number z is a zero of G(s), if and only if we can find an input u(t) = uz exp(zt) + ∑ki ui δ (i) (t), δ (i) (t) is the i-th order impulse, such that with zero initial condition the output y(t) ≡ 0 for t > 0. The above result expresses an output zeroing problem, where the impulsive part of the input transfer the zero state to an appropriate initial condition for t = 0+ (a state zero direction) which together with the exponential input make the output identically zero for t > 0. This clearly provides links with the state space output zeroing problem. The dynamic characterisation given here in terms of Theorems (13) and (14) can be satisfied as long as p is a pole of an entry of G(s) and z is a frequency where G(z) loses rank beyond the value of the normal rank. The analysis of the pole, zero structure is in terms of the Smith–McMillan form is considered next.

2.6.2

Smith–McMillan Form Characterisation of Poles and Zeros

The Smith–McMillan form of a rational matrix is a canonical form under R[s]-unimodular equivalence which has as invariants the polynomials characterising the finite pole and zero structure of the rational transfer function [58]. Theorem 2.15. Let G(s) be an m× p rational matrix and let θ = rankR(s) {G(s)} ≤ min(m, p). There exist R[s]-unimodular matrices Uℓ (s),Ur (s) of m × m, p × p dimensions respectively and an m × p matrix M(s) such that ∗ M (s) 0 G(s) = Uℓ (s)M(s)Ur (s), M(s) = 0 0 M ∗ (s) = diag{ε1 (s)/ψ1 (s), . . . , εθ (s)/ψθ (s)} = E(s)Ψ −1 (s) = Ψ −1 (s)E(s)

(2.60)

E(s) = diag{εi (s)}, Ψ (s) = diag{ψi (s)} where εi (s) divides εi+ j (s) and each ψi (s) divides ψi− j (s). The set of polynomials {εi (s), i = 1, . . . , θ }, {ψi (s) : i = 1, . . . , θ } completely characterise the R[s] equivalence class of G(s). The canonical matrix M(s) is called the Smith–McMillan form of G(s) and the polynomials εi (s), ψi (s) are the elementary zero, pole polynomial respectively and z(s) = ∏θi=l εi (s), p(s) = ∏θi=l ψi (s) are correspondingly the zero, pole polynomial of G(s). The roots of the εi (s), ψi (s) are respectively the zeros, poles of G(s) and they are the frequencies which satisfy the conditions of Theorems (13), (14) correspondingly. We define the number δM (G) = deg{p(s)} as the McMillan degree of G(s); this defines the number of states required for a minimal realisation of G(s) [20]. The zeros and poles defined through the Smith–McMillan form of G(s) are called transmission zeros, poles respectively. Their computation, apart from the derivation of the canonical form, may be achieved as described below [46]: Corollary 2.10. The polynomial p(s) is the least common denominator of all non-identically zero minors of all orders of G(s), z(s) is the greatest common divisors of the numerators of all non identically zero minors of maximal order of G(s), which have been adjusted to have p(s) as their common denominator.

74

N. Karcanias and E. Milonidis

2.6.3

Matrix Fraction Descriptions, and Poles and Zeros

An alternative input–output description for a system is that provided by coprime MFDs, or their corresponding composite representation, that is: G(s) = Nr (s)Dr (s)−1 = Dℓ (s)−1 Nℓ (s) (2.61) Nr (s) Tr (s) = , Tℓ (s) = [Dℓ (s), Nℓ (s)] (2.62) Dr (s) Such representations have a number of important properties and they provide alternative characterisations of poles, zeros which is equivalent to the Smith–McMillan characterisation [72], [18]. Theorem 2.16. Let G(s) be a rational transfer function matrix represented as in (2.61), (2.62). (i) Let (Nr (s), Dr (s)), (N˜ r (s), D˜ r (s)) (respectively (Dℓ (s), Nℓ (s)), (D˜ ℓ (s), N˜ ℓ (s))) be two right (left) coprime MFD pairs with the same dimensions. These two pairs correspond to the same transfer function matrix G(s), iff there exist Ur (s),Uℓ (s) R[s] unimodular matrices of appropriate dimensions such that T˜r (s) = Tr (s)Ur (s), T˜ℓ (s) = Uℓ (s)Tℓ (s)

(2.63)

(ii) If (Nr (s), Dr (s)) is a right coprime MFD pair and (Dℓ (s), Nℓ (s)) is a left coprime MFD pair of G(s) then, Nr (s), Nℓ (s) have essential parts of Smith form defined by E(s) = diag{ε1 (s), . . . , εθ (s)} and Dr (s), Dℓ (s) have essential parts of Smith form defined by Ψ (s) = diag{ψ1 (s), . . . , ψθ (s)}. (iii) Every transfer function matrix G(s) has a minimal factorisation. G(s) = N¯ r (s)Zr (s)Dr (s)−1 = Dℓ (s)−1 Zℓ (s)N¯ ℓ (s)

(2.64)

where (N¯ r (s)Zr (s), Dr (s)) are right coprime, (Dℓ (s), Zℓ (s)N¯ ℓ (s)) are left coprime Zr (s), Zℓ (s) are p × p, m × m matrices with Smith form defined by E(s), and Dr (s), Dℓ (s) are p × p, m × m matrices with Smith form defined by Ψ (s) and the matrices N¯ r (s), N¯ ℓ (s) have no zeros. Part (i) expresses the property that coprime MFDs of the same transfer function are right, or left R[s]-equivalent and thus define representations of the same system with respect to different co-ordinate frames. Part (ii) provides a characterisation of the pole, zero structure in terms of Smith forms of denominators, numerators respectively. Finally, part (iii) defines a factorisation of G(s) where the p × p, or m × m transfer functions Zr (s)Dr (s)−1 , Dℓ (s)−1 Zℓ (s) define completely the pole, zero structure and N¯ r (s), N¯ ℓ (s) are least degree bases of the column, or row space respectively of the system.

2.6.4

Infinite Poles and Zeros

The structure of Smith–McMillan form defined over R[s] [58] provides information on the finite poles and zeros. However, R[s]-equivalence does not preserve the structure of poles and zeros at infinity and another form of equivalence is required [79], which is based on the ring of proper rational functions R pr (s) [18], [76]. For state space models, properties at infinity were examined before based on the duality between zero and infinite elementary divisors. The current treatment provides also an alternative characterisation of the structure at infinity of state space models.

Structural Methods for Linear Systems: An Introduction

2.6.5

75

Smith–McMillan Form at Infinity: Infinite Poles and Zeros

The set of proper rational functions R pr (s) under the standard operations of addition and multiplication is a Euclidean ring with a degree function introduced by the notion of discrete valuation introduced below: Definition 2.9. [76]: (Discrete Valuation) Let t(s) = n(s)/d(s) ∈ R(s), where n(s), d(s) ∈ R[s], d(s) 6= 0 and define the map δ∞ (·) : R(s) → Z ∪ {∞} via:

δ∞ (t(s)) := degd(s) − degn(s),t(s) 6= 0, δ∞ (t(s)) := +∞,t(s) ≡ 0

(2.65)

This function satisfies the conditions: δ (xy) = δ (x) + δ (y) and δ (x + y) ≥ min{δ (x) + δ (y)} and it is referred to as discrete valuation of R(s). For proper rational functions, δ∞ (t(s)) ≥ 0, and thus the function introduces a Euclidean degree. The set of elements of R pr (s) with the property δ∞ (t(s)) = 0 are the biproper rational functions and they are the units (their inverse is proper) of R pr (s). Every t(s) ∈ R(s) can be written as: 1 n (s) t(s) = ( )q∞ 1 s d1 (s)

(2.66)

where q∞ := δ∞ (t(s)) and n1 (s), d1 (s) ∈ R[s] with degn1 (s) = degd1 (s). If q∞ > 0 we say that t(s) has a zero at s = ∞ of order q∞ and if q∞ < 0, then we say that t(s) has a pole at s = ∞ of order |q∞ |. Clearly, elements of R pr (s) have no poles at s = ∞. For a transfer function matrix G(s) ∈ Rmxp (s), the structure at infinity is studied under equivalence over the ring R pr (s) and it is defined by the Smith–McMillan form at infinity. We note that G(s) is called proper if lim G(s), as s → ∞, exists, strictly proper if the limit is zero and nonproper if the limit tends to infinity. The m × m rational matrix U(s) ∈ Rmxm pr (s) is said to be biproper, or R pr (s) unimodular, if |U(s)| = u(s) ∈ R pr (s) is a unit. Two m × p rational ˜ matrices G(s), G(s) are said to be R pr (s)-equivalent, if there exists Uℓ (s),Ur (s) such that ˜ G(s), G(s) are related as in (2.13). The matrices Uℓ (s),Ur (s) have no poles, or zeros at s = ∞, since their limit at s = ∞ are constant full rank matrices and thus they preserve the infinite pole zero structure of G(s). This structure is described by the following result [76], [79]. Theorem 2.17. Let G(s) ∈ Rmxp (s) with rankR(s) {G(s)} = θ . There exist biproper rational matrices Uℓ (s),Ur (s) such that ∗ M∞ 0 G(s) = Uℓ (s)M∞ (s)Ur (s), M∞ (s) = (2.67) 0 0 M∞∗ (s) = diag{sq1 , . . . , sqθ }, where q1 ≥ q2 ≥ . . . ≥ qk ≥ 0 ≥ qk+1 ≥ . . . ≥ qθ

The above result yields an immediate characterisation of poles and zeros at infinity of the rational matrix. If π∞ is the number of qi ’s in (2.67) with qi > 0 then G(s) has π∞ poles at infinity, each having order qi . Similarly, if τ∞ is the number of qi ’s in (2.67) with qi < 0 then G(s) has τ∞ zeros at infinity, each having order |qi |. Similar results on the characterisation of poles and zeros at infinity based on R pr (s) coprime MFDs may be given as for the case of finite poles and zeros. The above definition is equivalent to the definition of the structure at infinity based on bilinear transformation and use of the standard Smith–McMillan form. In fact, the infinite pole/zero structure of G(s) may be equivalently defined as the finite pole/zero structure of G(1/w) at w = 0. For state space models the structure at infinity is described by the infinite elementary divisors; the relationship between them and the infinite structure is summarised next [76], [78].

76

N. Karcanias and E. Milonidis

Corollary 2.11. If P(s) is the system matrix pencil of the systems S(A, B,C, D), then the following properties hold true: (i) P(s) has no infinite poles. (ii) P(s) has an infinite elementary divisor sˆq , q > 1, if and only if P(s) has an infinite zero with order q¯ = q − 1. (iii) The infinite elementary divisors s, ˆ i.e., with degree 1 have no dynamic significance and do not correspond to infinite poles, zeros. The notion of the Smith–McMillan form over R(s), or over R pr (s) may be generalised with respect to the region P = Ω ∪ {∞}, where Ω is any symmetric region with respect to the real axis that excludes at least one point of the real axis. The corresponding equivalence is over the ring of proper and Ω -stable rational functions, which is also a Euclidean ring with a degree function of the total number of zeros in the region P [77], [80].

2.6.6

Impulsive Dynamics and Properties of Infinite Poles and Zeros [57]

The infinite pole zero structure of a rational matrix may be also characterised dynamically. Two such results are given next. Theorem 2.18. Let G(s) ∈ Rmxp (s) and assume that it has a pole at infinity of order q. Then, there exists an input consisting solely of exponential modes u(t) = ∑ j ∑k u jk t k−1 exp(λ j t), where u jk are constant vectors such that the output is given by q−1

y(t) =

∑ yi δ (i) (t) + ∑ ∑ yg exp(λ j t), yq−1 6= 0 k

i=0

(2.68)

g k

The above suggests that the presence of infinite pole causes the transmittance to contain a linear combination of impulses, which were not present in the input. The property that yq−1 6= 0 implies that the presence of a q order infinite pole generates an impulse of order δ (q−1) (t). For infinite zeros we have the result: Theorem 2.19. Let G(s) ∈ Rmxp (s) and assume that it has a zero at infinity of order q. Then q−1 there exists an input of the form u(t) = ∑i=0 ui δ (i) (t) + ∑g ∑k u jk t k−1 exp(λ j t), where ui , u jk are constant vectors uq−1 6= 0, such that the corresponding output y(t) contain no impulses. The above suggest that the presence of an infinite zero in the transfer function causes the nontransmittance of the impulsive part of certain inputs. The condition that uq−1 6= 0 indicates that the presence of an infinite zero of order q does ensure the absorption of a particular impulse of order q. Remark 2.7. Although the standard Smith–McMillan form cannot be directly used for system compensation, due to that R[s]-equivalence transformation are not physically realisable, the Smith–McMillan form at infinity, derived under R pr (s)-equivalence has a compensation interpretation. For any G(s), condition (2.67) may also be expressed as Uˆ ℓ (s)G(s)Uˆ r (s) = M∞ (s), where Uˆ ℓ (s), Uˆ r (s) denote respectively the inverses of Uˆ ℓ (s), Uˆ r (s) which are also biproper matrices and thus they can be realised in terms of (A, B,C, D) parameters. In this case Uˆ ℓ (s)Uˆ r (s) express respectively proper post-, pre-compensation and M∞ (s) is the resulting transfer function.

Structural Methods for Linear Systems: An Introduction

77

M∞ (s) is the simplest transfer function that can be obtained under biproper post and pre compensation and the infinite poles and zeros represent the “strongest” set of invariants and is linked to the Kronecker canonical form of the state space system [36], [44], [34].

2.7

System Algebraic Functions and Generalised Nyquist and Root Locus

The pole, zero structure of a linear system plays a crucial role in Control analysis and design and it is particularly important in the study of Nyquist and Root Locus of Control Systems. The study of poles and zeros so far has been based on the algebraic properties of the structure of the system. In the case of Nyquist and Root Locus, however, the complex nature of the representation becomes dominant and this gives rise to alternative characterisations of poles and zeros which differs in a number of aspects from the algebraic nature. A brief introduction to the basics of multivariable Nyquist [47], [48], [56] and Root Locus [38], [47] is given next.

2.7.1

Characteristic Gain, Frequency Functions

We consider the general rectilinear motion problem for a frequency s and for the case of square systems m = p. If we define the scalar output feedback law y(t) = ku(t), k real, 6= 0, or equivalently u(t) = gy(t), g = k−1 and apply it on the system description S(Φ , Ω ) in (2.2) this leads to the closed-loop differential system I 0 x(t) ˙ A B x(t) S(Γ , ∆ (g)) : = (2.69) 0 0 u(t) ˙ C gI − D u(t) the solutions of which characterise the scalar output feedback solutions of the original system. The study of rectilinear motions according to the complex frequency s, i.e. [x(t)t , ut (t)]t = [xt , ut ]t exp(st) implies that condition (2.38) has to be satisfied, which leads to the eigenvalueeigenvector problem. sI − A −B x sI − A −B = 0, P(s, g) = (2.70) −C gI − D u −C gI − D where P(s, g) is known as the closed-loop system net [47]. The two variable nature of P(s, g) suggests that we may consider: (i) s as a function of g : For a fixed g, then u = (gI − D)−1Cx and {sI − S(g)}x = 0, S(g) = A + B(gI − D)−1C

(2.71)

(ii) g as a function of s : For a fixed s, then x = (sI − A)−1 BU and {gI − G(s)u = 0, G(s) = D +C(sI − A)−1 B

(2.72)

By (2.71) and (2.72) it is seen that the eigenvalue-eigenvector problems on S(g) and G(s) define the Root-locus and Nyquist diagrams respectively for the system S(A, B,C, D). S(g) is referred to as characteristic gain matrix and defines the closed loop eigenvalues for a scalar

78

N. Karcanias and E. Milonidis

output feedback k = g−1 . G(s) is the transfer function, which is now seen as the generator of Nyquist gains and directions and thus it is also called characteristic gain matrix. The characteristic equations for G(s) and S(g) are defined as:

∆ (g, s) = det{gI − G(s)] = 0, ∇(g, s) = det{sI − S(g)] = 0

(2.73)

and they are a pair of algebraic equations relating the complex variables s and g. ∆ (g, s) = 0 and ∇(s, g) = 0 define a pair of algebraic functions g(s) and s(g) known as characteristic gain and characteristic frequency functions [47], [48], [56] respectively g(s) and s(g) are the natural means of generalising Nyquist and Root-Locus theory to multivariable case correspondingly; the eigenvalue-eigenvector problems defined by (2.71), (2.72) demonstrate that both Root Locus and Nyquist in the multivariable case has an eigenframe dimension, which expresses the closed-loop eigenvectors and Nyquist directions respectively. Definition 2.10. For the system described by S(A, B,C, D), or G(s) we define: (i) If s takes values on the Nyquist contour (i.e., s = jω ), then the m branches gi ( jω ) of ∆ (g, s) are defined as the Nyquist diagrams, or characteristic gain loci of the system and the eigenvalue eigenvector decomposition (defined frequency by frequency). G( jω ) , W ( jω )Λ ( jω )W ( jW )−1 , Λ ( jω ) = diag{gi ( jω )}

(2.74)

expresses the characteristic gain decomposition of G( jω ). The columns of W ( jω ) are eigenvectors for the corresponding gi ( jω ) and they are called the Nyquist directions, or characteristic gain vectors. (ii) If g, (or k) takes values on the real axis including g = ∞(k = o), then the n branches si (k) of ∇(s, g) are defined as characteristic frequency loci or Root loci and the corresponding eigenvectors are the characteristics frequency vectors, or closed loop eigenvectors under the scalar negative feedback scheme.

2.7.2

Poles and Zeros of the System Algebraic Functions

The functions ∆ (g, s) and ∇(g, s) that define the two system algebraic equations are not independent. Expansion of the determinant of |P(s, g)| leads to the following identity det[sI − A]∆ (g, s) = deg[gI − D]∇(s, g)

(2.75)

/ σ (A) and g ∈ / σ (D) If σ (A), σ (D) denote the spectra of A, D respectively, then for all s ∈

∆ (g, s) = 0 ⇔ ∇(s, g)

(2.76)

Hence, a knowledge of characteristic gain as a function of frequency for G(s) is equivalent to a knowledge of characteristic frequency as a function of gain for S(g). If G(s) = N(s)D(s)−1 is a right coprime MFD, then ∆ (g, s) · |D(s)| = |gD(s) − N(s)| = ∆˜ (g, s) and

∆˜ (g, s) = |gD(s) − N(s)| = am (s)gm + am−1 (s)gm−1 + . . . + a0 (s) = 0

(2.77)

where am (s) = |D(s)| and a0 (s) = (−1)m |N(s)|. Clearly, the total number of poles and zeros of G(s) are given by the roots of am (s) and a0 (s) respectively ∆˜ (g, s) may be factorised as

∆˜ (g, s) = e(s)∆1 (g, s) . . . ∆υ (g, s)c(g)

(2.78)

Structural Methods for Linear Systems: An Introduction

79

where ∆i(g,s) are irreducible polynomials over C[g, s]. The terms e(s), c(g) are the products of all irreducible factors which are independent of g, s respectively. The roots of e(s), c(g) are called fixed poles, fixed gains correspondingly. In the following, ∆˜ (g, s) is assumed irreducible and that g(s) is the associated algebraic function. Equation (2.77) is called the characteristic gain-frequency equation. Remark 2.8. (McMillan Structure and Characteristic Equation) :If{εi (s), i =1, . . . , m},{ψi (s), i = 1, . . . , m} are the zero, pole elementary polynomials then the coefficients am (s),. . ., a0 (s) of the characteristic equation are expressed as: am (s) = ψ1 (s)ψ2 (s) . . . ψm (s) am−1 (s) = ε1 (s)ψ2 (s) . . . ψm (s)hm−1 (s), hm−1 (s) ∈ R[s] .. .

(2.79)

a1 (s) = ε1 (s)ε2 (s) . . . εm−1 (s)ψm (s)h1 (s), h1 (s) ∈ R[s] a0 (s) = ε1 (s)ε2 (s) . . . εm−1 (s)εm (s)h, h ∈ R 6= 0

Remark 2.9. The characteristic gain-frequency equation defines the characteristic gain algebraic function g(s) which defines the Nyquist diagrams; by re-arranging with respect to s (polynomials in g) the characteristic frequency algebraic function s(g) is obtained. The characterisation of poles and zeros of algebraic functions may be introduced by using their power series, and this is treated in [67].

2.7.3

Root Locus and the Output Zeroing Problem

The relationships between the current Root Locus formulation and zero structure are considered here. For square nondegenerate systems the system matrix pencil P(s) is characterised only be finite and infinite elementary divisors. For the asymptotic case when k → ∞, that is g = 0 the differential system S(Γ , ∆ (0)) defines the asymptotic solutions. Note S(Γ , ∆ (0)) = S(Γ , ∆ ) which is the output zeroing problem of Equation (2.22) and thus we are led to the following results on asymptotic properties [38], [17]. Theorem 2.20. (Asymptotic Properties of Root Locus)For square nondegenerate systems and for the unity feedback configuration with scalar gain k, the following hold true: (i) As k → ∞ the closed-loop poles of the system tend to the finite zeros, with the structure defined by the set of fed and the infinite zeros, with a structure defined by the orders of infinite zeros of P(s) (degrees of nonlinear ∞- e.d. minus one). (ii) The closed loop eigenstructure as k → ∞ defined by the structure of the decomposition of the maximal output nulling subspace Z , which is expressed as Z = Vz1 ⊕ . . . ⊕ Vzk ⊕ Vq1 ⊕ . . . ⊕ Vqµ

(2.80)

where Vz corresponds to (s − z)d , dimVz = d and Vq corresponds to sˆq , dimVq = q − 1. Conditions (2.80) define the structure of closed-loop eigenvectors as k → ∞. The rates at which closed-loop eigenvalues tend to infinity is defined by the order of the “Butterworth Patterns” of

80

N. Karcanias and E. Milonidis

the Asymptotic Root Locus; such rates are also referred to as Root-Locus Infinite Zeros and in general they differ from the algebraic definition previously given. A proper definition of such infinite zeros is given through the Laurent expansions of the Root Locus algebraic function and this lead to orders, which may be fractional (see [38], [56], [67]). The properties of root loci may be inferred from the Markov parameters (CB,CAB, ...,CAk B, . . .) [38] and the basic rules described there reveal the explicit relationships between the system structure (expressed in terms of finite, infinite zeros) and their asymptotic properties. The finite zeros define the terminal locations of root loci under high gain, whereas the orders of infinite zeros define the generic values of the asymptotic Butterworth pattern. For the Nyquist diagrams, the infinite zeros describe the generic values of the terminal phases of the Nyqusit diagrams. Apart from stability, the characteristic gain loci may reveal further closed-loop properties such as ability to track, reject signals under the assumption of relative normality of G(s) [56].

2.8

The Feedback Configuration and Structural Properties

In this section we consider the general Feedback Configuration for a multivariable system and establish its general structural properties and fundamental operators used in control design.

2.8.1

Structural Properties of the Feedback Configuration

The most commonly used feedback configuration in the context of algebraic synthesis method is that denoted in Figure (2.3), where P(s) ∈ Rm×p (s),C(s) ∈ R p×m (s) are the plant, controller transfer functions and w1 , w2 are external control, or disturbance signals, e1 , e2 are plant controller inputs and y1 , y2 are the plant, controller outputs. For the sake of simplicity we drop (s) in the descriptions of C(s), P(s) and their MFDs, as well as from the Laplace transforms of vector signals. Such a configuration is quite versatile and may accommodate several control problems. For instance, in a tracking problem, w1 would be a reference signal to be tracked by plant output y2 . In a disturbance rejection, or desensitisation to noise, w1 would be the disturbance/noise. Depending on whether w1 , or w2 is the externally applied control signal (as opposed to noise etc.) the configuration can represent either feedback, or cascade compensation. Such a configuration will be referred to as complete feedback configuration. The system equations are defined by [39], [80], [6]: e1 w1 0 P e1 y1 C 0 e1 = − , (2.81) −C 0 0 P e2 e2 w2 e2 y2 or, by

e1 w1 y , w= , y= 1 , e2 w2 y2 C 0 0 I G= ,F = 0 P −I 0

e = w − FGe, y = Ge, e =

(2.82)

The feedback configuration is well formed if |I + FG| = |I + PC| = |I + CP| = 6 0. Then we may define the transfer functions H(P,C),W (P,C) by

Structural Methods for Linear Systems: An Introduction

81

w2

+

w1 +

y1 +

C

e2

y2

P

-

controller

plant

Fig. 2.3. Complete Feedback Configuration e = (I + FG)−1 w = H(P,C)w, y = G(I + FG)−1 w = W (P,C)w,

where

W (P,C) = F{H(P,C) − I}

(2.83)

C(I + PC)−1 −CP(I +CP)−1 , W (P,C) = PC(I + PC)−1 P(I + PC)−1 (I + PC)−1 −P(I +CP)−1 H(P,C) = C(I + PC)−1 (I + PC)−1

Furthermore, if we also define the transfer functions y1 w y w = T (P,C) 1 , 2 = R(P,C) 2 e1 w2 e2 w1

(2.84)

(2.85)

then T (P,C), R(P,C) are expressed as : C(I + PC)−1 −C(I +CP)−1 T (P,C) = , −1 −1 (I + PC) (I + PC) P R(P,C) =

P(I + PC)−1 P(I +CP)−1 (I +CP)−1 (I +CP)−1C

(2.86)

In the above expressions a number of common terms appear which have particular significance for the study of feedback systems. Thus, we define Q = PC, Q′ = CP, F = I + PC, F ′ = I +CP, L = I + (PC)−1 , S = (I + PC)−1 , S′ = (I +CP)−1

(2.87)

where Q, Q′ are referred to as return ratio matrices, F, F ′ as return difference matrices, L as inverse-return difference matrix and S, S′ as sensitivity matrices. Remark 2.10. All transfer function matrices, W (P,C), H(P,C), T (P,C), R(P,C) associated with the complete feedback configuration have the same pole polynomial.

82

N. Karcanias and E. Milonidis

The above implies that any one of them may be used for the study of pole assignment and stabilisation of the feedback configuration. We shall use in the following, the simplest, i.e. ,H(P,C). −1 −1 Theorem 2.21. [39], [80] If P = A−1 B1 , = B2 A−1 2 ,C = D1 N1 = N2 D2 are left, right coprime MFDs (over R[s], or R p (s), then

(i) A left , right coprime MFD for H(P,C) (over R[s], or R p (s)) is defined by −1 −1 A1 B1 A1 0 D2 0 D2 B2 H(P,C) = = −N1 D1 0 D1 0 A2 −N2 A2

(2.88)

(ii) The pole polynomial (for R[s] MFDS), or pole function (for R p (s) MFDs) of H(P,C) is f ∼ |U1 | ∼ |U2 | , U1 = A1 D2 + B1 N2 , U2 = D1 A2 + N1 B2

(2.89)

where ∼ means equal modulo units (for R[s] or R p (s) cases). An important issue for the feedback system is that of well posedness i.e., properness of all transfer function associated with the feedback system. Proposition 2.4. [39], [80] If C, P are proper transfer functions, then the completed feedback configuration is well posed, iff |I +C(∞)P(∞)| = |I + P(∞)C(∞)| = 6 0

(2.90)

Remark 2.11. Note that C(∞), P(∞) are constant matrices, since C, P are proper and 0 matrices if they are strictly proper. The complete feedback configuration with P,C proper is generically well posed and if at least one of P,C is strictly proper, then this always holds true. Let S p , Sc be the plant, controller state space model and S f be the state space description of the complete feedback configuration. Theorem 2.22. [7], [80] : For the well posed complete feedback system S f . The following properties hold true: (i) S f is controllable, observable, iff both S p , Sc are controllable, observable. (ii) S f is stablizable, detectable, iff both S p , Sc are stablizable, detectable. (iii) If S p , Sc are both stabilizable and detectable, then S f is internally stable, iff H(P,C) is BIBO stable. The above result is fundamental, since it provides the means for studying internal stability in terms of BIBO stability of H(P,C) and it provides the basis for the study of stabilization and pole assignment of the feedback configuration. Corollary 2.12. Consider the well posed feedback system S f with S p , Sc stabilisable and detectable. −1 −1 (i) If P = A−1 B1 = B2 A−1 2 ,C = D1 N1 = N2 D2 are R p (s)-coprime MFD’s the S f is internally stable, iff either of the equivalent condition hold true

A1 D2 + B1 N2 = U1 , U1 : R p (s) − unimodular, D1 A2 + N1 B2 = U2 ,U2 : R p (s) − unimodular

(2.91)

Structural Methods for Linear Systems: An Introduction

83

−1 −1 (ii) If S p , Sc are controllable and observable and P = A−1 B1 = B2 A−1 2 ,C = D1 N1 = N2 D2 are R p (s)-coprime MFD’s, then the characteristic polynomial of S f is defined by sI − A f ∼ |A1 D2 + B1 N2 | ∼ |D1 A2 + N1 B2 | (2.92)

Furthermore, if F = I + PC, F ′ = I +CP are the return difference matrices and A p , Ac are the plant controller state matrices, then |I + PC| ∼ |I +CP| ∼ |sI − A f |/|sI − A p | · |sI − Ac |

(2.93)

In (2.91) we may assume that U1 ,U2 are identity matrices; such equation are known as Diofantine Equations (for solvability and parametrisation of family of solutions see [39]) and (A1 , B1 ), (D2 , N2 ) (also (A2 , B2 ), (D1 , N1 ) are referred to as mutually stabilising pairs. For a given P, these equations define the whole family of C compensators which stabilize the system; additional performance characteristics may then be assigned by selecting appropriate elements from the stabilising family. If C is given, the solution of these equations defines the family of plants which are stabilised by C; this has important implications in the parametrisation of those plants that may be stabilised by a given type controller, as well as in the study of robustness. Note that (2.91) may also be expressed as a Diofantine equations, where now U1 or U2 are polynomial matrices with given set of zeros; this provides the means for an algebraic formulation of the pole assignment problem. Finally, condition (2.93), establishes the important property of the determinant of the return difference as the ratio of closed-loop to open-loop pole polynomials of the feedback system; it is this property that underlies the theory of both Nyquist and root locus.

2.8.2

Closed-Loop Performance and the Return Ratio Difference and Sensitivity Matrices

An enlarged version of the complete feedback configuration frequently used in design problems is shown in Figure (2.4) and it is referred to as a Control Design Configuration, P,C, F are proper transfer function matrices of the plant, precompensator, feedback compensator respectively (with appropriate dimensions), and r, di , do , n are the reference, plant input disturbance, plant output disturbance, sensor noise vector signals respectively. We assume that P,C, F are stabilisable and detectable and that C, F stabilise the feedback configuration, and thus all transfer function matrices from any external signal to the output y are stable. Let us define the transfer functions from all external signals to the output by: Hr : r → y, Hod : do → y, Hid : di → y, Hn : n → y

Hr = PC(I + FPC)−1 = (I + PCF)−1 PC, Hod = (I + PCF)−1 Hid = (I + PCF)−1 P, Pn = (I + PCF)−1 PCF and because of linearity we have that y = Hr r + Hod d0 + Hid di + Hn n

(2.94) (2.95) (2.96)

For external signals with given bandwidth, the above equation describes their effect on the overall system response in terms of the frequency domain transfer functions Hr ( jω ), Hod ( jω ), Hid ( jω ), Hn ( jω ), which are expressed in terms of return ratio PCF, return difference (I + PCF) and sensitivity matrix (I + PCF)−1 . The frequency domain study of tracking, disturbance, noise rejection, sensitivity to plant parameter variations, robustness etc. involves the notion of gain of transfer function matrices which is defined in terms of the singular values [66], [11], [56].

84

N. Karcanias and E. Milonidis

d1

d0

+

r

+ -

e1

C

y1 +

e2

P

y2

+

y

+ +

n +

F Fig. 2.4. Control Design Configuration

2.9

Determinantal Assignment Problems: Exterior Algebra-Algebraic Geometry Methods

For the study of problems of linear feedback synthesis which are of the determinantal type (such as pole zero assignment, stabilisation) a specific framework has been developed which is specially suited to tackle such problems and it is referred to as algebrogeometric approach, since it relies on tools from algebra and algebraic geometry. The problems faced in this area are of a multilinear nature and the number of design parameters is not necessarily large. The main difficulty of the determinantal problems arises from that they are equivalent to finding real solutions to sets of nonlinear and linear equations; in the case of stabilisation, this is equivalent to determining solutions of nonlinear equations and nonlinear inequalities. The first of the two problems naturally belongs to the intersection theory of complex algebraic varieties, whereas, the latter belongs to the intersection theory of semialgebraic sets. Additional difficulties arise due to the need of studying design problems with specific dynamic structure (centralised or decentralised) which make the varieties involved not standard, and the compactification issues quite prominent. Determining real intersections is not an easy problem; furthermore, it is also important to be able to compute solutions whenever they exist. The DAP approach [27] has been developed for determinantal problems which are of multilinear nature and thus may be naturally split into a linear and multilinear problem (decomposability of multivectors). The final solution is thus reduced to the solvability of a set of linear equations (characterising the linear problem) together with quadratics (characterising the multilinear problem of decomposability). The approach heavily relies on exterior algebra and this has implications on the computability of solutions (reconstruction of solutions whenever they exist). New sets of invariants (of a projective character) are introduced [27] which, in turn, characterise the solvability of the problem. The distinct advantages of the DAP approach are that it provides the means for computing the solutions [42], [43] and can handle both generic and exact solvability investigations.

2.9.1

Determinantal Assignment Problems

Consider the linear system described by equations (2.1), or (2.8). We may define the following cluster of problems:

Structural Methods for Linear Systems: An Introduction

85

(I) Pole assignment by state feedback : Consider L ∈ Rn×p , where L is a state feedback applied on system (2.1). If we denote by Pc (s) = [sI − A, −B], L˜ = [In , Lt ]t , then the closed loop characteristic polynomial is given by ˜ pL (s) = det{sI − A − BL} = det{Pc (s) − L}

(2.97)

(II) Design of an n-state observer : Consider the problem of designing an n-state observer for the system of (2.1). If T ∈ Rn×m is a feedback, T˜ = [In , T ] and P0 (s) = [sI − At , −Ct ]t , then the characteristic polynomial of the observer is defined by pT (s) = det{sI − A − TC} = det{T˜ P0 (s)}

(2.98)

(III) Pole assignment by constant output feedback: Consider the system described by (2.2) under an output feedback K ∈ Rm×p . The closed loop characteristic polynomial pK (s) is given by [18]: pK (s) = det{Dℓ (s) + Nℓ (s)K} = det{Dr (s) + KNr (s)} (2.99) By defining the matrices Tℓ (s) = [Dℓ (s), Nℓ (s)] ∈ Rm×(m+p) [s], Tr (s) =

then,

Dr (s) ∈ R(m+p)×p [s] Nr (s)

K˜ ℓ = [Im , K t ]t ∈ R(m+p)×m , K˜ r = [I p , K] ∈ R p×(m+p)

(2.100)

pK (s) = det{Tℓ (s)K˜ ℓ } = det{K˜ r Tr (s)}

(2.101)

(IV) Zero assignment by squaring down [37], [32], [28], [61]: The characterisation of zeros in terms of the properties of P(s) system matrix pencil, or the transfer function matrix G(s), clearly demonstrate that zeros express the interaction between internal dynamics and the effort to control and observe the system. Earlier stages than Control Design affect directly the zero formation and typical problems that may be posed are: (i) Given a system described by either the (A, B), or (A,C) pairs, define C, or B matrices respectively such that the resulting triple (A, B,C) corresponds to a square system with zeros which can be arbitrarily assigned. (ii) Given a system described by the triple (A, B,C, D), or the m× p transfer function G(s), m > p, define a constant p × m matrix K, such that the resulting KG(s) p × p transfer function has a resulting set of zeros. The first class, and its variations (where C, or B are not entirely free but selected from a given set) are referred to as zero assignment by input, output structure design. Such problems may be studied either in a geometric, or algebraic setup and correspond to problems of transforming the Kronecker structure of the corresponding pencils; this is expressed for instance as: Problem 2.1. Given Rc (s) = sN − NA of (n − p) × n dimension, find a matrix M n × (n − p) constant, such that Z(s) = (sN − NA)M has a given zero structure (defined by the set of ed), or for a given φ (s) ∈ R[s] det{Z(s)} = |sNM − NAM| (2.102)

86

N. Karcanias and E. Milonidis

The second class represents problems of squaring down the system by selecting an appropriate linear combination of the outputs and it is an integral part of the process of closing feedback loops, or defining sets of outputs which may be controlled independently. In fact, for a system with m > p we can expect to have independent control over at most p linear combinations of m outputs. If c ∈ R p is the vector of the variables which are to be controlled, then c = Hy, where H ∈ R p×m is a squaring down post-compensator, and G′ (s) = HG(s) is the squared down transfer function matrix [37], [28]. This may be expressed formally as: Problem 2.2. Given the m × p transfer function G(s), m > p, expressed as the factorised MFD G(s) = N¯ r (s)Zr (s)Dr (s)−1 , define K constant, p × m such that the p × p square down transfer function Gσ (s) = KG(s) has a given zero structure, or for a given φ (s) ∈ R[s] det{Nσ (s)} = |K N¯ r (s)Zr (s)| = |K N¯ r (s)| · |Zr (s)| = φ (s)

(2.103)

The latter formulation clearly demonstrates that the existing zeros of the nonsquare system remain as zeros of the squared down system; the process of squaring down adds new zeros which are created by the matrix K N¯ r (s), and thus a simpler formulation is the study of K N¯ r (s) zero structure formation. A family of problems of determinantal type are associated with dynamic compensation [43] and are considered next. (V) Dynamic Compensation Problems: Consider the standard feedback configuration of Figure (2.3). If G(s) ∈ R pr (s)m×p ,C(s) ∈ R(s) p×m , and assume coprime MFD’s and C(s) = Aℓ (s)−1 Bℓ (s) = Br (s)Ar (s)−1

The closed loop characteristic polynomial may be expressed as [7], [39], [80] Ar (s) f (s) = det [Dℓ (s), Nℓ (s)] , Br (s) Dr (s) or f (s) = det [Aℓ (s), Bℓ (s)] Nr (s)

(2.104)

(2.105)

(i) if p ≤ m, then C(s) may be interpreted as feedback compensator and we will use the expression of the closed loop polynomial. (ii) if p ≥ m, the C(s) may be interpreted as precompensator and we will use the expression of the closed loop polynomial. The above general dynamic formulation covers a number of important families of C(s) compensators as: (a) Constant, (b) PI, (c) PD, (d) PID, (e) Bounded degree. In fact, (a) Constant Controllers: If p ≤ m, Aℓ = I p , Bℓ = K ∈ R p×m , then (100) expresses the constant output feedback case, whereas if p ≥ m, Ar = Im , Br = K ∈ R p×m expresses the constant precompensation formulation. (b) Proportional plus Integral Controllers: Such controllers are defined by 1 C(s) = K0 + K1 = [sI p ]−1 [sK0 + K1 ] s

(2.106)

where K0 , K1 ∈ R p×m and the left MFD for C(s) is coprime, iff rank(K1 ) = p. From the above the determinantal problem for the output feedback PI design is expressed as:    sDr (s)   Dr (s) f (s) = det [sI p , sK0 + K1 ] = det [I p , K0 , K1 ]  sNr (s)  (2.107) Nr (s)   Nr (s)

Structural Methods for Linear Systems: An Introduction

87

(c) Proportional plus Derivative Controllers: Such controllers are expressed as C(s) = sK0 + K1 = [I p ]−1 [sK0 + K1 ]

(2.108)

where K0 , K1 ∈ R p×m and the left MFD for C(s) is coprime for finite s and also for s = ∞ if rank(K0 ) = p. From the above the determinantal output PD feedback is expressed as:   Dr (s) Dr (s) f (s) = det{[I p , sK0 + K1 ] } = det{[I p , K1 , K0 ]  Nr (s) } (2.109) Nr (s) sNr (s)

(d) PID Controllers: These controllers are expressed as

1 C(s) = K0 + K1 + sK2 = [sI p ]−1 [s2 K2 + sK0 + K1 ] s

(2.110)

where K0 , K1 ∈ R p×m and the left MFD is coprime with the only exception possibly at s = 0, s = ∞ (coprimeness at s = 0 is guaranteed by rank(K1 ) = p and at s = ∞ by rank(K2 ) = p). From (2.109), the determinantal output PID feedback is expressed as : Dr (s) 2 } f (s) = det{[sI p , s K2 + sK0 + K1 ] Nr (s)   sDr (s)  sNr (s)   } = det{[I p , K0 , K1 , K2 ]  (2.111) Nr (s)  2 s Nr (s)

(e) Observability Index Bounded Dynamics (OBD) Controllers: These are defined by the property that their McMillan degree is equal to pk, where k is the observability index [74], [18] of the controller. Such controllers are expressed as [Aℓ (s), Bℓ (s)] = Tk sk + . . . + T0

(2.112)

Tk , Tk−1 , . . . , T0 ∈ R p×(p×m) and Tk = [I p , X]. Note that the above representation is not always coprime, and coprimeness has to be guaranteed first for McMillan degree to be pk; otherwise, the McMillan degree is less than pk. The dynamic determinantal OBD output feedback problem is expressed using as Tr (s) = R(s) in the form 

sk R(s)



 sk−1 R(s) 

 f (s) = det{(Tk sk + . . . + T0 )R(s)} = det{[Tk , Tk−1 , . . . , T0 ]  

.. . R(s)

 } 

(2.113)

Remark 2.12. The above formulation of the determinantal dynamic assignment problems is based on the assumption that p ≤ m and thus output feedback configuration is used. If p ≥ m, we can similarly formulate the corresponding problems as determinantal dynamic precompensation problems and use right coprime MFDs for C(s).

88

N. Karcanias and E. Milonidis

2.9.2

The General Determinantal Assignment Problem

All the problems introduced in the previous sections belong to the same problem family, i.e., the determinantal assignment problem (DAP) [28]. This problem is to solve the following equation with respect to polynomial matrix H(s) : ¯ det(H(s)M(s)) = f (s)

(2.114)

where f (s) is the polynomial of an appropriate degree d. The difficulty for the solution of DAP is mainly due to the multilinear nature of the problem as this is described by its determinantal character. We should note, however, that in all cases mentioned previously, all dynamics can ¯ be shifted from H(s) to M(s), which in turn transforms the problem to a constant DAP as it has been shown by the analysis in the previous section. This equivalent formulation of the problem may be described as follows: Let M(s) ∈ R[s] p×r [s], r ≤ p, such that rank(M(s)) = r, f (s) be a real polynomial of an appropriate degree d, and let H be a family of full rank r × p constant matrices having a certain structure. Solve with respect to H ∈ H the equation: fM (s, H) = det(HM(s)) = f (s)

(2.115)

Remark 2.13. The degree of the polynomial f (s) depends firstly upon the degree of M(s) and secondly, upon the structure of H. However, in most of our problems the degree of p(s) is equal to the degree of M(s). The determinantal assignment problem has two main aspects. The first has to do with the solvability conditions for the problem and the second, whenever this problem is solvable, to provide methods for constructing these solutions. We classify the solutions to two classes: exact and generic solvability conditions. Notation [50]: Let Qk,n denote the set of lexicographically ordered, strictly increasing sequences of k integers from 1, 2, . . . , n. If {xi1 , . . . , xik } is a set of vectors of V , ω = (i1 , . . . , ik ) ∈ Qk,n , then xi1 ∧ . . . ∧ xik = xω ∧ denotes the exterior product and by ∧r V we denote the r-th exterior power of V . If H ∈ F m×n and r ≤ min{m, n}, then by Cr (H) we denote the r-th compound matrix of H. If hti , mi (s), i = 1, . . . , r, we denote the rows of H, columns of M(s) respectively, then Cr (H) = ht1 ∧ . . . ∧ htr = ht ∧ ∈ R1×σ Cr (M(s)) = m1 (s) ∧ . . . ∧ mr (s) = m(s)∧ ∈ Rσ [s], σ = and by Binet-Cauchy theorem [50] we have that [28]: fM (s, H) = Cr (H)Cr (M(s)) = hh∧, m(s)∧i =

∑

p r

hω mω (s)

(2.116)

ω ∈Qr,p

where h, i denotes inner product, ω = (i1 , . . . , ir ) ∈ Qr,p , and hω , mω (s) are the coordinates of h∧, m(s)∧ respectively. Note that hω is the r × r minor of H which corresponds to the ω set of columns of H and thus hω is a multilinear alternating function of the entries hi j of H. The multilinear, skew symmetric nature of DAP suggests that the natural framework for its study is that of exterior algebra [49]. The study of the zero structure of the multilinear function fM (s, H) may thus be reduced to a linear subproblem and a standard multilinear algebra problem as it is shown below.

Structural Methods for Linear Systems: An Introduction

89

(i) Linear subproblem of DAP: Set m(s)∧ = p(s) ∈ Rσ [s]. Determine whether there exists a k ∈ Rσ , k 6= 0, such that fM (s, k) = kt p(s) = ∑ ki pi (s) = f (s), i ∈ ∼σ, f (s) ∈ R[s]

(2.117)

(ii) Multilinear subproblem of DAP: Assume that K is the family of solution vectors k of (2.118). Determine whether there exists H t = [h1 , . . . , hr ], where H t ∈ R p×r , such that h1 ∧ . . . ∧ hr = h∧ = k, k ∈ K

(2.118)

Polynomials defined by Equation (2.118) are called polynomial combinants [28] and the zero assignability of them provides necessary conditions for the solution of the DAP. The solution of the exterior Equation (2.118) is a standard problem of exterior algebra and it is known as decomposability of multivectors. Note that notions and tools from exterior algebra play also an important role in the linear subproblem, since fM (s, k) is generated by the decomposable multivector m(s) ∧ . The essence of the approach is projective, that is we use a natural embedding for determinantal problems to embed the space of the unknown, H, of DAP, into an appropriate projective space. In this way we can see our problem as search for common solutions of some set of linear equations and another set of second order polynomial equations. This also allows us to compactify H into H and then use algebraic geometric, or topological intersection theory methods to determine existence of solutions for the above sets of equations. The characteristic of the current framework is that it allows the use of algebraic geometry and topological methods [15] for the study of solvability conditions but also computations. Central to the latter is the solution of the linear system derived by (2.118) with the quadratics characterising the solvability of (2.118), which are known as Quadratic Pl¨ucker Relations (QPR) [49], [16].

2.9.3

Grassmann- Plucker Invariants

The importance of the DAP framework is that it uses the natural embedding of a Grassmannian into a projective space and this in turn defines new sets of invariants characterising the solvability of the different DAP problems [28]. We may summarise the results relating to the new invariants as follows: Let T (s) ∈ R p×r (s), T (s) = [t1 (s), . . . ,tr (s)], p ≥ r, rankR(s) {T (s)} = r and let Xt = spanR(s) (T (s)). If T (s) = M(s)D(s)−1 is a RCMFD of T (s), then M(s) is a polynomial −1 , where ˜ basis for Xt . If Q(s) is a greatest right divisor of M(s) then T (s) = M(s)Q(s)D(s) ˜ M(s) is a least degree polynomial basis of Xt [15]. A Grassmann Representative (GR) for Xt is defined by t(s)∧ = t1 (s) ∧ . . . ∧ tr (s) = m˜ 1 (s) ∧ . . . ∧ m˜ r (s) · zt (s)/pt (s)

(2.119)

where zt (s) = det{Q(s)}, pt (s) = det{D(s)} are the zero, pole polynomials of T (s) and p ˜ m˜ 1 (s) ∧ . . . ∧ m˜ r (s) = m(s)∧ ˜ ∈ Rσ [s], σ = , is also a GR of Xt . Since M(s) is a least r degree polynomial basis for Xt , the polynomials of m(s)∧ ˜ are coprime and m(s)∧ ˜ will be ˜ then δ is the referred to as a reduced polynomial GR (R - R[s]-GR) of Xt . If δ = deg(m(s)∧), Forney dynamical order [12] of Xt . The polynomial vector m(s)∧ ˜ may be expressed as

90

N. Karcanias and E. Milonidis

m(s)∧ ˜ = p(s) = p0 + sp1 + . . . + sδ pδ = Pδ eδ (s), Pδ ∈ Rσ ×(δ +1)

(2.120)

˜ and eδ (s) = [1, s, . . . , sδ ]t . It can be readily shown that all where Pδ is a basis matrix for m(s)∧ R[s]-GRs of Xt differ only by a nonzero scalar factor a ∈ R. By choosing an m(s)∧ ˜ for which ||pδ || = 1, a monic reduced Grassman representative (C-R[s]-GR) of Xt is defined, and shall be denoted by g(Xt ) and is referred to as canonical R[s]-GR; the basis matrix Pδ of g(Xt ) is defined as the Pl¨ucker matrix of Xt [28]. The following properties hold true [28]: Theorem 2.23. g(Xt ), or the associated Pl¨ucker matrix Pδ , is a complete (basis free) invariant of Xt . Corollary 2.13. Let T (s) ∈ R p×r (s), p ≥ r, rankR(s) {T (s)} = r, zt (s), pt (s) be the monic zero, pole polynomials of T (s) and let g(Xt ) = p(s) be the C-R[s]-GR of the column space Xt of T (s). The vector t(s)∧ may be uniquely decomposed as t(s)∧ = c · p(s) · zt (s)/pt (s), where c ∈ R − {0}

(2.121)

˜ ˜ If M(s) ∈ ≥ r, rankR(s) {M(s)} = r, then M(s) = M(s)Q(s), where M(s) is a least degree basis and Q(s) is a greatest right divisor of M(s) and thus R p×r [s], p

m(s)∧ = m(s) ˜ ∧ ·det{Q(s)} = p(s) · zm (s) = Pδ eδ (s) · zm (s)

(2.122)

The linear part of DAP is thus reduced to fM (s, k) = kt p(s)zm (s) = kt Pδ eδ (s) · zm (s)

(2.123)

Corollary 2.14. The zeros of M(s) are fixed zeros of all combinants of m(s) ∧ . ˜ k) = The zeros of fM (s, k) which may be freely assigned are those of the combinant f M(s, ˜ km(s)∧, ˜ where m(s)∧ ˜ is reduced. Given that the zeros of f M(s, k) are not affected by scaling with constants, we may always assume that m(s)∧ ˜ = Pδ eδ (s). In the following, the case of combinants generated by reduced m(s)∧ ˜ will be considered. If a(s) ∈ R[s] is the polynomial which has to be assigned, then max(deg(a(s)) = δ , where δ is the Forney dynamical order of Xt . If a(s) = atδ eδ (s) = a0 + a1 s + . . . + aδ sδ , where a ∈ Rδ +1 , then the problem of finding k, k ∈ Rσ , such that fM˜ (s, k) = a(s) is reduced to the solution of p Pδt k = a, Pδt ∈ R(δ +1)×σ , σ = (2.124) r The matrix M(s) ∈ R p×r [s] generating DAP will be called linearly assignable (LA), if (2.124) has a solution for all a; otherwise, it will be called linearly nonassignable (LNA). M(s) will be called completely assignable (CA), if it is LA and (2.118) has a solution for at least a solution of the linear problem defined by (2.124). An important family of nonassignable M(s) matrices are those for which there is no k such that fM (s, k) = c, c ∈ R; such M(s) are called strongly nonassignable (SNA) and they imply that they cannot assign all zeros at s = ∞. Some results characterising the above properties are stated below: Remark 2.14. Necessary condition for M(s) to be LA is that M(s) is a least degree matrix. Corollary 2.15. Let M(s) ∈ R p×r [s] be a least degree matrix, Pδ be the Pl¨ucker matrix of Xm and let π = rank(Pδ ). Then,

Structural Methods for Linear Systems: An Introduction

91

p (i) Necessary and sufficient condition for M(s) to be LA is that π = δ + 1 (i.e., ≥ δ +1 r and π = δ + 1). (ii) Let M(s) be LA and let P˜ † be the right inverse of Pδt and P˜δ⊥ a basis for Nr (Pδt ). For every a ∈ Rδ +1 , the solution of (eq:118) is given by k = P˜ † a + P˜δ⊥ c, where c ∈ Rσ −δ −1 arbitrary.

(2.125)

(iii) Let Pδ = [p0 , p1 , . . . , pδ ] = [p0 , . . . , pδ ] ∈ Rσ ×(δ +1) and let π = rank{Pδ }, π¯ = rank{P¯δ }. M(s) is strongly nonassignable, iff π¯ = σ . A number of Pl¨ucker type matrices associated with a linear system are: ¨ (a) Controllability Plucker Matrix [33] : For the pair (A, B), b(s)t ∧ denotes the exterior n+ p basis matrix product of the rows of Pc (s) = [sI − A, −B] and P(A, B) is the (n + 1) × n of b(s)t ∧ . P(A, B) will be called the Controllability Pl¨ucker matrix and its rank properties characterise the system controllability as shown below: Theorem 2.24. [33]: The system S(A, B) is controllable iff P(A, B) has full rank. ¨ (b) Observability Plucker Matrix [33] : For the pair (A,C), c(s)∧denotes the exterior prodn+m t t t uct of the columns of Po (s) = [sI − A , −C ] and P(A,C) is the × (n + 1) basis n matrix of c(s) ∧ . P(A,C) will be called observability Pl¨ucker matrix and its rank properties characterise system observability. Corollary 2.16. [33]: The system S(A,C) is observable, iff P(A,C) has full rank. ¨ (c) Transfer Function Matrix Plucker Matrices: For the transfer function matrix G(s) represented by the RCMFD, LCMFD we define by tr (s)∧, tℓ (s)t ∧ the exterior product of the m+ p columns of Tr (s), rows of Tℓ (s) respectively. By P(Tr ) we denote the × (n + 1) basis p m+ p matrix of tr (s)∧, and by P(Tℓ ) the (n + 1) × basis matrix for tℓ (s)t ∧ . P(Tr ), P(Tℓ ) p will be referred to as right, left fractional representation Pl¨ucker matrices respectively. Such matrices provide the prime indicators for the solution of the output feedback, or constant precompensation problem. Theorem 2.25. [41], [42]: For a generic system with mp > n, then the corresponding Pl¨ucker matrices P(Tr ), P(Tℓ ) have full rank. The full rank of these matrices is a necessary condition for the solvability of pole assignment problems and their singular values characterise the norm properties of the corresponding solutions. Given that Tr (s), Tℓ (s) uniquely characterise (modulo unimodular equivalence) the transfer function G(s), we may also refer to P(Tr ), P(Tℓ ) as the right-, left- transfer function Pl¨ucker matrices and we denote them simply by Pr (G), Pℓ (G). ¨ (d) Column, Row Plucker Matrices: For the transfer function G(s), m ≥ p, n(s)∧ denotes the exterior product of the columns of the numerator Nr (s), of a RCMFD and by P(N) the

92

N. Karcanias and E. Milonidis

m × (d + 1) basis matrix of n(s) ∧ . Note that d = δ , the Forney order of Xg , if G(s) has ℓ no finite zeros and d = δ + κ , where κ is the number of finite zeros of G(s), otherwise. If Nr (s) is least degree (has no finite zeros), then Pc (N) will be called the column space Pl¨ucker matrix of the system. For this case the row space Pl¨ucker matrix may be similarly defined and it is Pr (N) = 1. For systems with m ≤ p and full rank transfer functions Pc (N) = 1, whereas Pr (N) is a nontrivial matrix. Such matrices play a key role in problems such as the squaring down problem. Corollary 2.17. For a generic system with m > p, for which p(m − p) > δ + 1, where δ is the Forney order, Pc (N) has full rank. The above classes of problems are typical formulations amongst a very large class of problems of integrated system design [24], [25] which involve the design, or redesign of input, output structure of the system aiming at producing square systems with prescribed zero structure and thus good potential for control design. Such problems are referred to as zero structure assignment problems and may be studied with algebraic means (polynomial matrix theory, or matrix pencils) geometric theory (dynamic cover problems for invariant spaces), or exterior algebra and algebraic geometry tools (determinantal formulation and intersection theory of varieties) [4], [28]. The study of generic solvability of such problems and a method for computing solutions may be found in [41], [42], [43].

2.10

Conclusions

This overview paper provides an introduction to the algebraic structural aspects of Linear Systems with emphasis on the system structure and related system properties. Our viewpoint has been that the algebraic structure of the system model underpins the nature of dynamic and feedback properties. Poles and zeros and system invariants have thus been central to this approach. Poles express the internal dynamics of the system and they are directly related to stability and other aspects of performance and they are affected by the different types of feedback transformations. On the other hand, zeros are measures of the interaction between internal dynamics and their coupling to inputs and outputs. They are invariant under fundamental feedback transformations and their alternation may be achieved under dynamic compensation, or by design or re-design of the input, output system structure. The significance of poles and zeros for control design is discussed in the references cited in the bibliography. The theory of invariants and canonical forms of state space systems can be considered from the perspective defined by the theory of matrix pencils [13], [22], [44] and the theory of minimal bases of rational vector spaces [12], [75]. The role and significance of invariants for Control Synthesis and Control Design problems may be found in the literature [7], [12], [18], [43], [58], [72], [74] and references therein.

A Invariants and Canonical Forms The development of theory of matrix pencil invariants and canonical forms requires some basic results that may be found in algebra books, as well as some definitions on a number of important notions [36].

Structural Methods for Linear Systems: An Introduction

93

Definition 2.11. Let X be a set, E be an equivalence relation on X and let x ∈ X ; the equivalence class, or orbit of x under E is defined by E(x) = {y : y ∈ X : xEy}. The set of all equivalence classes is called the quotient set, or orbit set and it is denoted by X /E. Theorem 2.26. If E is an equivalence relation on a set X , then the family of all E-equivalence classes forms a partition of X , i.e.,there exist {xi ∈ X : i = 1, 2, . . .} such that X = E(x1 ) ∪ . . . ∪ E(xi ) ∪ . . . , E(xi ) ∩ E(x j ) 6= 0/

(2.126)

The set of elements {xi , i = 1, 2, . . .} for which conditions (2.126) hold true is called a system of distinct representatives for E and it is a subset D of X that contains precisely one element from each of the E- equivalence classes. Definition 2.12. Let X , T be sets, E an equivalence relation defined on X we define: (i) A function f : X → T is called an invariant of E, when ∀x, y ∈ X : xEy implies f (x) = f (y). (ii) f : X → T is called a complete invariant for E, when f (x) = f (y) implies xEy. (iii) A set of invariants { fi : fi : X → Ti , i = 1, 2, . . . , k} is a complete set for E, if the map f , f : X → T1 × . . . × Tk , where x → ( f1 (x), . . . , fk (x)) is a complete invariant for E on X. A complete invariant defines a one to one correspondence between the equivalence classes E(x) and the image of f . If f : X → T1 × . . . × Tk where x → ( f1 (x), . . . , fk (x)) is a complete invariant for E on X , then the set ( f1 (x), . . . , fk (x)) characterises uniquely E(x). The values fi (x) are often called invariants and this definition has been used in the paper. If we “specialise” the invariant f such that its image T ⊂ X expresses f in the “simplest” possible way, then we define the notion of a canonical element, or canonical form. Definition 2.13. A set of canonical forms for E equivalence on X is a subset C of X such that ∀x ∈ X there is a unique c ∈ X for which xEc. Equivalence relations may be introduced by the action of groups on sets and in particular by the action of transformation groups. If X is a set and G (X ) is the group of permutations of X , any subgroup H (X ) of G (X ) is called a transformation group of X . A transformation group is therefore a set H of mappings of X into X , which are bijective. If (G , ⋆) is a group operating on a set X , where ⋆ denotes the group operation and ◦ the operation of G on X , then the relation: there exists s ∈ G such that y = s ◦ x, is an equivalence relation and it is called the induced equivalence relation of G on X . G ◦ x is then the equivalence class of x under this relation.

B List of Symbols, Abbreviations R:

field of real numbers

C:

field of complex numbers

R[s] : ring of polynomials in s with coefficients in R

94

N. Karcanias and E. Milonidis

R(s) : field of rational functions in s with coefficients in R R p (s) : ring of proper and Ω -stable rational functions R pr (s) : ring of proper rational functions Rm×n : the set of m × n matrices with real coefficients Rm×n : the set of m × n rational matrices V :

vector space over R, or C, or R(s)

deg :

degree of a polynomial

det{A} : determinant of a square matrix A, also denoted by |A| f ed :

finite elementary divisors

ied :

infinite elementary divisors

cmi :

column minimal indices

rmi :

row minimal indices

∧:

exterior product of vectors

References 1. P. J. Antsaklis and A. N. Michael. Linear Systems. McGraw-Hill, New York. (1997) 2. J. D. Aplevich. Implicit Linear Systems. Lecture Notes in Control and Information Sciences. 152. Springer, Berlin. (1991) 3. G. Basile and G. Marro. Controlled and Conditioned Invariants in Linear Systems. Englewood Cliffs. New Jersey. Prentice Hall . (1992) 4. R.W. Brockett and C.I. Byrnes. Multivariable Nyquist Criterion, Root Loci and Pole Placement: A geometric viewpoint. IEEE Trans. Aut. Control. AC-26. 271-283. (1981) 5. P. Brunovsky. A classification of linear controllable systems. Kybernetica. 3. 173-187. (1970) 6. F.M. Callier and C.A. Desoer. Multivariable Feedback Systems. Springer-Verlag. New York. (1982) 7. C.T. Chen. Linear Systems Theory and Design. Holt-Rinehart and Winston. New York. (1984) 8. M.J. Denham. Canonical forms for the identification of multivariable linear systems. IEEE Trans. Aut. Control. 19. 645–656. (1974)

Structural Methods for Linear Systems: An Introduction

95

9. C.A. Desoer and J.D. Schulman. Zeros and Poles of Matrix Transfer Functions and their Dynamical Interpretation. IEEE Trans. Circ. and Systems. 21. 3–8. (1974) 10. B.W. Dickinson, T. Kailath, and M. Morf. Canonical matrix fraction and state-space descriptions for deterministic and stochastic linear systems. IEEE Trans. Automatic Control. AC-19. 656–667. (1974) 11. J C . .Doyle and G. Stein. Multivariable Feedback Design: Concepts for a Classical /Modern Synthesis.IEEE Trans. Aut. Control. AC-26. 4–16. (1981) 12. D.G. Forney. Minimal bases of rational vector spaces with applications to multivariable linear systems. SIAM J. Control. 13. 493–520. (1975) 13. G. Gantmacher. Theory of Matrices. Vol. 2. Chelsea. New York . (1975) 14. C. Giannacopoulos and N. Karcanias. Pole assignment of strictly proper systems by constant output feedback. Int. J. Control. 42. 543. (1985) 15. P. Griffiths. and J. Harris. Principles of Algebraic Geometry. Wiley Publications. (1978) 16. W V . D . Hodge and P.D Pedoe. Methods of Algebraic Geometry. Vol. 2. Cambridge Univ. Press. (1952) 17. S .Jaffe and N. Karcanias. Matrix Pencil Characterisation of Almost (A,B) – invariant subspaces: A classification of geometric concepts. Int. J. Control. 33. 51–53. (1981) 18. T. Kailath. Linear Systems. Englewood Cliffs. New Jersey. Prentice Hall. (1980) 19. R.E. Kalman. Canonical Structure of linear dynamical systems. Proc. Nat. Ac. Sci. 48. 596–600. (1962) 20. R.E. Kalman. Mathematical description of linear dynamical systems. SIAM J. Control, Vol. 1. 152–192. (1963) 21. R.E. Kalman. Kronecker invariants and Feedback. Ordinary Differential Equat. L. Weiss (ed). Academic Press. 459–471. New York. (1972) 22. N. Karcanias. Matrix pencil approach to geometric system theory. Proc. IEE, Vol. 126. 585–590. (1979) 23. N. Karcanias. Minimal Bases of Matrix Pencils: Algebraic, Toeplitz Structure and Geometric Properties. Linear Algebra and its Applications. Special Issue on Linear Systems. 205–206, 831–865. (1994) 24. N. Karcanias. Global Process Instrumentation: Issues and Problems of a System and Control Theory Framework. Measurement. 14, 103–113. (1994) 25. N. Karcanias. Control Problems in Global Process Instrumentation: A Structural Approach. Proc. of ESCAPE-6. Comp. Chem. Eng. 20. 1101–1106. (1996) 26. N. Karcanias. Multivariable Poles and Zeros. In Control Systems, Robotics and Automation, from Encyclopedia of Life Support Systems (EOLSS), Developed under the Auspices of the UNESCO, Eolss Publishers. Oxford. UK. [http://www.eolss.net] [Retrieved October 26, 2005]. (2002) 27. N. Karcanias and C. Giannakopoulos. On Grassman Invariants and Almost Zeros of Linear Systems and the Determinantal Zero, Pole Assignment Problem. Int. J. Control. 40(4). 673–698. (1984) 28. N. Karcanias and C. Giannakopoulos. Necessary and Sufficient Conditions for Zero Assignment by Constant Squaring Down. Linear Algebra and its Applications. Vol. 122/123. 415–446. (1989) 29. N. Karcanias and G. E. Hayton. Generalised Autonomous Dynamical Systems, Algebraic Duality and Geometric Theory. Proc of 8th IFAC World Congress. Kyoto. Japan. Pergamon Press. 289–294. (1981) 30. N. Karcanias and G.E. Hayton. State Space and Transfer Function Infinite Zeros; A Unified Approach. Proc. Joint Aut. Cont. Conf. Univ. of Virginia. (1981)

96

N. Karcanias and E. Milonidis

31. N .Karcanias and G .Kalogeropoulos .Geometric Theory and Feedback Invariants of Generalised Linear Systems: A Matrix Pencil Approach. Circuits, Systems and Signal Process. 8(3). 375–395. (1989) 32. N. Karcanias and B. Kouvaritakis. The Output Zeroing Problem and its Relationship to the Invariant Zero Structure. Int. J. Control. 30. 395–415. (1979) 33. N.Karcanias and J.Leventides: Grassman Invariants, Matrix pencils and Linear System Properties. Linear Algebra and its Applications. Vol. 241- 243. 705– 731. (1995) 34. N. Karcanias and P. MacBean. Structural invariants and canonical forms of linear multivariable systems. Proc. 3rd IMA Conf. on Control Theory. Academic Press. 257–282. London. (1980) 35. N. Karcanias and M. Mitrouli. Minimal Bases of Matrix Pencils and coprime Matrix Fraction Descriptions. IMA Journal of Math. Control Theory and Inform. Vol. 19. 245– 278. (2002) 36. N. Karcanias and D. Vafiadis. Canonical Forms for State Space Descriptions. In Control Systems, Robotics and Automation, from Encyclopedia of Life Support Systems (EOLSS), Developed under the Auspices of the UNESCO. Eolss Publishers. Oxford. UK. [http://www.eolss.net] [Retrieved October 26, 2005]. (2002) 37. B. Kouvaritakis and A.G.J. MacFarlane. Geometric approach to analysis and synthesis of system zeros: Part II: non-square systems. Int. J. Control. Vol. 23. 167–181. (1976) 38. B. Kouvaritakis and U. Shaked. Asymptotic Behavior of Root-Loci of Multivariable Systems. Int. J. Control. 23. 297– 340. (1976) 39. V. Kucera. Discrete Linear Control: The Polynomial Equation Approach. John Wiley & Sons. New York. (1979) 40. F.L. Lewis. A survey of Linear Singular Systems. Circuits, Systems and Signal Processes. 8. 375–397. (1989) 41. J. Leventides and N. Karcanias: The Pole Placement Map, its Properties and Relationships to System Invariants. IEEE Trans. on Aut. Control. AC-38. 1266–1270. (1993) 42. J. Leventides and N. Karcanias. Global asymptotic linearisation of the pole placement map: A closed form solution for the output feedback problem. Automatica. Vol. 31. 1303– 1309. (1993) 43. J. Leventides and N. Karcanias. Dynamic Pole Assignment using Global, Blow up Linearisation: Low Complexity Solutions. Journal of Optimisation Theory and Applications. 96. 57–86. (1998) 44. J.J. Loiseau, K. Ozcaldiran, M. Malabre and N. Karcanias. Feedback canonical Forms of Singular Systems. Kybernetica. 27. 289–305. (1991) 45. D.G. Luenberger. Canonical forms for linear multivariable systems. IEEE Trans. Aut. Control. 12. 290–293. (1967) 46. A.G.J. MacFarlane and N. Karcanias. Poles and Zeros of Linear Multivariable Systems: A survey of the Algebraic, Geometric and Complex Variable Theory. Int. J. Control. 24. 33–74. (1976) 47. A.G.J. MacFarlane and N. Karcanias. Relations Between State Space and Frequency Response Concepts. Proc. 7th IFAC Cong. Part 43B. Helsinki. Finland. (1978) 48. A.G.J. MacFarlane and I. Postlewaite. The generalised Nyquist stability Criterion and Multivariable Root Loci. Int. J. Control. Vol. 25. 581–622. (1977) 49. M. Marcus. Finite dimensional multilinear algebra (in two parts). Marcel Deker, New York. (1973)hfill 50. M. Marcus and H. Minc. A Survey of Matrix Theory and Matrix Inequalities. Allyn and Bacon. Bacon. (1964) 51. B.C. Moore. Principal Component Analysis in Linear Systems: Controllability, Observability and Model Reduction. IEEE Trans. Autom. Control. AC-26, 17–32. (1981)

Structural Methods for Linear Systems: An Introduction

97

52. A.S. Morse. Structural invariants of linear multivariable systems. SIAM J. Control. 11. 446–465. (1973) 53. D.H. Owens. Feedback and Multivariable Systems. Peter Peregrinus Ltd. IEE Publication. Stevenage. (1978) 54. V.M. Popov . Some properties of the control systems with irreducible matrix transfer functions. Lecture Notes in Mathematics. Springer. 144. 250–261. (1969) 55. V.M. Popov . Invariant descriptions of linear time invariant controllable systems. SIAM J. Control. 10. 252–264. (1972) 56. I. Postlewaite, J.M. Edmunds and A.G.J. MacFarlane. Principle Gains and Principle Phases in the analysis of Linear Multivariable Feedback Systems. IEEE Trans. Aut. Control. AC-26. 32–46. (1981) 57. A.C. Pugh and V. Krishnaswamy. Algebraic and dynamic characterisations of poles and zeros at infinity. Int. J. Control. 42. 1145–1153. (1985) 58. H H . .Rosenbrock . State Space and Multivariable Theory. Nelson. London. (1970) 59. H.H. Rosenbrock. Structural properties of linear dynamical systems. Int. J. Control. Vol. 20. 191–202. (1974) 60. H.H. Rosenbrock. Order, degree and complexity. Int. J.Control 19. 323–331. (1979) 61. A. Saberi and P. Sannuti. Squaring Down of non-strictly proper systems. Int. J. Control. Vol. 51. 621–629. (1990) 62. R. Saeks and R. A. DeCarlo. Interconnected Dynamical Systems. Marcel Dekker. New York. (1981) 63. M.G. Safonov, A.J. Laub and G.L. Hartmann. Feedback Properties of Multivariable Systems: The Role and Use of the Return Difference Matrix. IEEE Trans. Auto. Control. Vol. AC-26. 47–65. (1981) 64. U. Shaked and N. Karcanias. The use of zeros and zero directions in model reduction. Int. Journal Control. 23. 113–135. (1976) 65. R.E. Skelton. Dynamic Systems Control. John Wiley & Sons. (1988) 66. S. Skogestad and I. Postlethwaite. Multivariable Feedback Control. John Wiley & Sons. Chichester. (1996) 67. M.C. Smith. Multivariable root-locus behaviour and the relationship to transfer function pole-zero structure. Int. J. Control. 43. 497–515. (1986) 68. J.P. Thorp. The singular pencil of linear dynamical systems. Int. J. Control. 18. 577–596. (1973) 69. S.H. Wang and E.J. Davison. Canonical forms of linear multivariable systems. SIAM J. Control & Optimisation. 14. 236–250. (1976) 70. M.E. Warren and A.E. Eckberg. On the dimensions of controllability subspaces: A characterisation via polynomial matrices and Kronecker invariants. SIAM J. Control. 13l. 434– 445. (1975) 71. J.C. Willems. Almost Invariant Subspaces: An Approach to High Gain Feedback Design, Part I: Almost Controlled Subspaces. IEEE Trans. Aut. Control. 26. 235–252. (1981) 72. W.A. Wolovich. Linear multivariable systems. Applied Maths. Sciences. 11. SpringerVerlag. (1974) 73. W A . . Wolowich and P.L Falb. On the structure of multivariable systems. SIAM J. Control. Vol. 7. 437–451. (1969) 74. W.M Wonham. Linear Multivariable Control: A Geometric Approach. Sec. Ed. Springer Verlag. New York. (1979) 75. D. Vafiadis and N. Karcanias. Canonical Forms for Descriptor Systems Under Restricted Systems Equivalence. Automatica. 33. 955–958. (1997) 76. A I. G . Vardulakis, D.J.N. Limebeeer and N. Karcanias. Structure and Smith–McMillan form of a Rational Matrix at Infinity. Int. J. Control. 35. 701–725. (1982)

98

N. Karcanias and E. Milonidis

77. A.I.G. Vardulakis and N. Karcanias. Structure, Smith–McMillan form and coprime MFDs of a rational matrix inside a region PΩ ∪ {∞}. Int. J. Control. 38. 927–957. (1983) 78. A.I.G. Vardulakis and N. Karcanias. Relations between strict Equivalence Invariants and Structure at Infinity of Matrix Pencil. IEEE Trans. Aut. Control. AC-28. 514–516. (1983) 79. G .Verghese and T. Kailath. Rational matrix structure. IEEE Trans. Aut. Control. AC-26. 434–439. (1981) 80. M. Vidyasagar. Control System Synthesis: A Factorization Approach. MIT Press. Cambridge. Mass. (1985)

3 Modelling and Model Reduction—State-Space Truncation† David J.N. Limebeer

3.1 Introduction The approximation of high-order plant and controller models by models of lower-order is an integral part of control system design and analysis. Until relatively recently model reduction was often based on physical intuition. For example, chemical engineers often assume that mixing is instantaneous and that packed distillation columns may be modelled using discrete trays. Electrical engineers represent transmission lines and the eddy currents in the rotor cage of induction motors by lumped circuits. Mechanical engineers remove high-frequency vibration modes from models of aircraft wings, turbine shafts and flexible structures. It may also be possible to replace high-order controllers by low-order approximations with little sacrifice in performance. We will develop elementary procedures which seek to automate the model reduction process. Suppose a high-order, linear, time-invariant model G is given, then the prototype b of G such that kG − Gk b ∞ L∞ model reduction problem is to find a low-order approximation G is small. Truncation methods of model reduction seek to remove, or truncate, unimportant states from state-space models. If a state-space model has its A-matrix in Jordan canonical form, statespace truncation will amount to classical modal truncation. For example, one may truncate all those states that correspond to “fast modes”—eigenvalues with a large negative real part. One’s interpretation of “fast” will obviously depend on the application, but this could mean modes outside the control system bandwidth. Since the poles of the truncated model are a subset of the poles of the original high-order model, any low-order modal approximation of a stable high-order model is guaranteed to be stable. As we will show later, it is also possible to b ∞ for modal truncation. Because any transfer function can be realized get a bound on kG − Gk in terms of an infinite number of state-space models, there are, in principle, also an infinite number of candidate truncation schemes. For a truncation scheme to be useful, it must preserve stability and carry with it a guaranteed error bound. The aim of these notes is to develop the balanced truncation method of model reduction, which satisfies an infinity norm bound on the absolute approximation error. † This

chapter is based on Chapter 9 of “Linear Robust Control” [1]. This book, together with the solutions manual, can be obtained from the authors in .pdf form on request.

M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 99 -122, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

100

D.J.N. Limebeer

It is well known that the modes of a realization that are either uncontrollable or unobservable do not appear in the corresponding system transfer function matrix. It is therefore natural to conjecture that the modes that are almost uncontrollable and unobservable can be omitted from the realization with little effect on the input-output characteristics of the model. The balanced realization has the property that mode i is equally controllable and observable, with these properties measured in terms of a number σi ≥ 0. As σi increases, the corresponding level of controllability and observability increases; the σi ’s are the singular values of the Hankel operator associated with the transfer function matrix G. The model reduction method that applies the truncation operation to a balanced realization is known as balanced truncation. For this algorithm the absolute error is guaranteed to satisfy the twice-the-sum-of-the-tail bound b ∞ ≤ 2(σr+1 + . . . + σn ), kG − Gk

b in which n is the McMillan degree of G and r is the McMillan degree of G. Example 3.1. The transfer function

g=

1 (s + 1)(s + 2)

has modal realization g = Cm (sI − Am )−1 Bm , in which −1 0 1 Am = , Bm = , 0 −2 −1 Cm = 1 1 .

(3.1) (3.2)

If we truncate the fast mode we obtain the reduced-order system gm =

1 . s+1

The norm of the error is

1 1 k∞ = . s+2 2 The transfer function g also has realization g = Cb (sI − Ab )−1 Bb , in which −0.40859 −0.970143 0.492479 Ab = , Bb = , 0.970143 −2.59141 −0.492479 Cb = 0.492479 0.492479 .

kg − gm k∞ = k

(3.3) (3.4)

This realization is balanced because the controllability and observability gramians are both equal and diagonal: 0.296796 0 Σ= , 0 0.0467961

so that σ1 = 0.296796 and σ2 = 0.0467961. If we truncate this realization, we obtain the reduced-order system (0.492479)2 gb = , s + 0.40859 which is stable and the norm of the error is kg − gb k∞ = 0.0935921 = 2σ2 .

(3.5) (3.6)

Modelling and Model Reduction State-Space Truncation

101

The point to note is that, in this case, the error associated with balanced truncation is considerably smaller than that associated with modal truncation. The generally good performance of balanced truncation has led to its widespread popularity and enduring usefulness. Another well known model order reduction method is the singular perturbation approximation. This technique is usually associated with a fast-slow decomposition of the state space, with the approximation achieved by setting the “fast” states to their steady-state values. Since the singular perturbation approximation is related to state-space truncation by the frequency inversion transformation s → 1/s, it is also considered to be a truncation method. Our main interest is in balanced singular perturbation approximation, in which the singular perturbation method is applied to a balanced realization. In this method, states corresponding to small σi ’s are set to their steady-state values. Although the error bounds for balanced truncation and balanced singular perturbation approximation are identical, the resulting models have different high- and low-frequency characteristics. Direct truncation gives a good model match at high frequency, while singular perturbation methods have superior low-frequency properties. We will develop most of our theory for direct truncation because of its notational simplicity.

3.2 State-Space Truncation Consider a linear, time-invariant system with the realization x(t) ˙ = Ax(t) + Bu(t), y(t) = Cx(t) + Du(t)

x(0) = x0 ,

(3.7)

and divide the state vector x into components to be retained and components to be discarded: x (t) x(t) = 1 . (3.8) x2 (t) The r-vector x1 (t) contains the components to be retained, while the (n − r)-vector x2 contains the components to be discarded. Now partition the matrices A, B and C conformably with x to obtain A11 A12 B1 A= , B= , A21 A22 B2 (3.9) C = C1 C2 . By omitting the states and dynamics associated with x2 (t), we obtain the lower-order system p(t) ˙ = A11 p(t) + B1 u(t) q(t) = C1 p(t) + Du(t).

p(0) = p0 ,

The rth -order truncation of the realization (A, B,C, D) is given by Tr (A, B,C, D) = (A11 , B1 ,C1 , D).

(3.10)

In general, very little can be said about the relationship between x and p, y and q or the transfer b associated with (A, B,C, D) and (A11 , B1 ,C1 , D). In particular, the function matrices G and G

102

D. J. N. Limebeer

truncated system may be unstable even if the full-order system is stable, and the truncated system realization may be nonminimal even if the full-order realization is minimal. One thing that clearly does hold is b G(∞) = G(∞), which means that all reduced-order models obtained by truncation have perfect matching at infinite frequency.

Example 3.2. (Modal Truncation). The truncation of modal realizations is common in engineering practise, because it is often the case that high-frequency modes may be neglected on physical grounds, or because the phenomena resulting in such modes only play a secondary role in determining the model’s essential characteristics. Let G be a transfer function matrix with an asymptotically stable modal realization     λ1 0 0 B1     A =  0 ... 0 , B =  ...  , (3.11) 0 0 λm Bm C = C1 . . . Cm (3.12)

For simplicity, we assume that each of the eigenvalues λi of the A-matrix has a simple Jordan structure. If the modes with a fast decay rate may be omitted from the model, the eigenvalues should be ordered so that |Re (λi )| is nondecreasing with increasing i. Alternatively, if one is to omit the high-frequency modes, the eigenvalues should be ordered so that |Im (λi )| is nondecreasing with increasing i. As a combination of these two, one might order the modes so that |λi | is nondecreasing with increasing i, so that those modes with the highest “natural” frequency are deleted. The error incurred in modal truncation depends not only on the λi ’s, but also on the size of the b we have residues Ci Bi . If modes labeled r + 1 to n are omitted by truncation to obtain G, and therefore that

b= G−G

n

Ci Bi s i=r+1 − λi

b ∞≤ kG − Gk

∑

n

kCi Bi k . i=r+1 |Re λi |

∑

Since the error associated with deleting a mode depends on the ratio kCi Bi k/|Re λi | and not |Re λi | alone, the modal decay rate is not a reliable guide as to whether a particular mode should be included in the reduced-order model. The main features of modal truncation are: 1. It is conceptually simple. 2. The poles of the reduced-order model are a subset of the poles of the original model. In addition, the poles of the reduced-order model retain their physical interpretation, because one knows, for example, that certain vibration modes are being retained while others are being omitted. 3. It is computationally cheap, because the main calculation is an eigenvalue decomposition of A.

Modelling and Model Reduction State-Space Truncation

103

3.2.1 The Truncation Error To help us with later work, in which we develop bounds for the error incurred by model reduction based on truncation, we develop some of the basic properties of the error system. Lemma 3.1. Suppose (A, B,C, D) is a realization of order n partitioned as in (3.9) and let b = D +C1 (sI − A11 )−1 B1 . Define G = D +C(sI − A)−1 B and G ˜ A(s) = A22 + A21 (sI − A11 )−1 A12 ˜B(s) = B2 + A21 (sI − A11 )−1 B1 ˜ C(s) = C2 +C1 (sI − A11 )−1 A12 .

(3.13)

Then: ˜ 1. det(sI − A) = det(sI − A11 ) det(sI − A(s)). 2. The truncation error system satisfies −1 ˜ b = C(s)(sI ˜ ˜ G(s) − G(s) − A(s)) B(s). (3.14) P 0 3. If AP + PA′ + BB′ = 0 and P = 1 , in which the partitioning is conformable with 0 P2 (3.9), then

A11 P1 + P1 A′11 + B1 B′1 = 0 (3.15) ∼ ∼ ˜ ˜ ˜ ˜ A(s)P2 + P2 A (s) + B(s)B (s) = 0. (3.16) ˜ ˜ If (A, B) is controllable, then sI− A(s) B(s) has full row rank for all s. Q1 0 ′ ′ 4. If A Q + QA +C C = 0 and Q = , in which the partitioning is conformable with 0 Q2 (3.9), then A′11 Q1 + Q1 A11 +C1′ C1 = 0 ˜ + C˜ ∼ (s)C(s) ˜ A (s)Q2 + Q2 A(s) = 0. ˜ sI − A(s) If (A,C) is observable, then has full column rank for all s. ˜ C(s) ˜∼

Proof. Write Φ (s) = (sI − A11 )−1 and note that 0 I 0 sI − A11 sI − A = ˜ 0 sI − A(s) −A21 Φ (s) I I −Φ (s)A12 . × 0 I

(3.17) (3.18)

(3.19)

1. This follows directly from (3.19). 2. From (3.19), we have C(sI − A)−1 B I Φ (s)A12 Φ (s) 0 = C1 C2 −1 ˜ 0 I 0 (sI − A(s)) I 0 B1 × B2 A21 Φ (s) I −1 ˜ ˜ ˜ = C1 Φ (s)B1 + C(s)(sI − A(s)) B(s),

which proves (3.14).

(3.20) (3.21) (3.22) (3.23)

104

D. J.N. Limebeer

3. Equation (3.15) is immediate from the assumed block diagonal structure of P. To prove (3.16), write BB′ = (sI − A)P + P(−sI − A)′ . Now use (3.19) to obtain B1 ′ ˜ ′ B1 B (−s) ˜ B(s) sI − A11 −A12 P1 P1 Φ ′ (−s)A′21 = ˜ 0 P2 0 sI − A(s) −sI − A′11 0 P1 0 + . ′ A21 Φ (s)P1 P2 −A′12 −sI − A˜ (−s)

(3.24) (3.25) (3.26)

The (2, 2)-block of this equation is (3.16). We also note that if x is such that ˜ ˜ B(s) =0 x∗ sI − A(s)

for some s, then

I 0 sI − A B = 0 A21 Φ (s) I ˜ ˜ B(s) and we see that (A, B) controllable implies x = 0. Hence sI − A(s) has full row rank when (A, B) is controllable. 4. This follows from calculations which are dual to those given in Item 3.

0 x∗

3.2.2 Singular Perturbation Approximation The steady-state error associated with state-space truncation is given by −1 b G(0) − G(0) = C1 A−1 11 B1 −CA B.

In applications requiring good low-frequency models this may be unacceptably large. In these cases, it is appropriate to use a singular perturbation approximation in preference to state-space truncation because of its greatly improved low-frequency model reduction characteristics. Consider the full-order model given by x(t) ˙ = Ax(t) + Bu(t),

x(0) = x0 ,

y(t) = Cx(t) + Du(t),

(3.27) (3.28)

which is partitioned as in (3.8) and (3.9). If x2 (t) represents the fast dynamics of the system, we may approximate the low-frequency behavior by setting x˙2 (t) = 0. This gives 0 = A21 x1 (t) + A22 x2 (t) + B2 u(t), which yields the quasi-steady-state solution x2 (t) = −A−1 22 (A21 x1 (t) + B2 u(t))

(3.29)

provided A22 is nonsingular. Eliminating x2 from the remaining equations using (3.29) yields

Modelling and Model Reduction State-Space Truncation −1 p(t) ˙ = (A11 − A12 A−1 22 A21 )p(t) + (B1 − A12 A22 B2 )u(t)

q(t) =

−1 (C1 −C2 A−1 22 A21 )p(t) + (D −C2 A22 B2 )u(t).

105 (3.30) (3.31)

The rth -order singular perturbation approximation (SPA) is given by

in which

b11 , Bb1 , Cb1 , D), b Sr (A, B,C, D) = (A

(3.32)

b11 = A11 − A12 A−1 A21 , Bb1 = B1 − A12 A−1 B2 , A 22 22 −1 b Cb1 = C1 −C2 A−1 A , D = D −C A 2 22 B2 . 22 21

(3.33)

The following result shows that SPA is equivalent to: (a) setting H(w) = G(w−1 ); (b) performb ing a state-space truncation of H(w) to obtain H(w); and (c) defining the reduced-order model b = H(s b −1 ). as G(s)

Lemma 3.2. Let G(s) = D + C(sI − A)−1 B, in which A is nonsingular, and let H(w) = G(w−1 ). Then: 1. H(w) = D −CA−1 B −CA−1 (wI − A−1 )−1 A−1 B. 2. The realizations of G(s) and H(w) have the same controllability and observability gramians (when they exist). 3. Suppose that A22 is nonsingular, that Gr (s) is the rth -order SPA of G(s) and that H r (w) is the rth -order system obtained by truncation of the realization of H(w) defined in Item 1. Then Gr (s) = H r (s−1 ).

Proof. 1. This follows from the identity (w−1 I − A)−1 = −A−1 − A−1 (wI − A−1 )−1 A−1 . 2. Suppose P and Q are the controllability and observability gramians of the realization of G satisfying AP + PA′ + BB′ = 0

(3.34)

A′ Q + QA +C′C = 0.

(3.35)

Multiplying by A−1 and (A−1 )′ gives A−1 P + P(A−1 )′ + (A−1 B)(A−1 B)′ = 0 QA 3. Writing

−1

+ (A

−1 ′

) Q + (CA

I A12 A−1 22 A= 0 I gives A−1 =

I −A−1 22 A21

Now truncate the realization of H(w):

0 I

−1 ′

) (CA

b11 0 A 0 A22

b−1 0 A 11 0 A−1 22

−1

) = 0.

I 0 A−1 A I 21 22

(3.36)

I −A12 A−1 22 . 0 I

(3.37)

106

D.J.N. Limebeer

I b−1 I 0 A−1 =A 11 0 −1 b−1 Bb1 I0 A B=A 11 I b−1 −CA−1 = −Cb1 A 11 0

b−1 Bb1 . b − Cb1 A D −CA−1 B = D 11

(3.38) (3.39) (3.40) (3.41)

Applying Item 1 to the realization of Gr given in (3.32), H r (w) = Gr (w−1 ), which is equivalent to Gr (s) = H r (s−1 ). Since the singular perturbation and truncation operations are related in a straightforward way, it suffices to develop all our theoretical results for state-space truncation. When the lowfrequency fidelity of the approximation is important, the singular perturbation approximation is the method of choice. Conversely, direct truncation should be preferred when good highfrequency modelling is the central concern.

Main Points of the Section 1. State-space truncation is a simple but general procedure for generating reducedorder models. The properties of the reduced-order model will depend on the realization selected for truncation. For example, reduced-order models obtained from the truncation of balanced and modal realizations (of the same full-order system) will generally be quite different. 2. State-space truncation produces zero error at infinite frequency. 3. Since the singular perturbation method of model reduction is related to statespace truncation by the bilinear transform s → 1/s, singular perturbation approximations have zero steady-state error.

3.3 Balanced Realization The aim of this section is to introduce balanced realizations, which are of interest because they have good absolute-error truncation properties.

3.3.1 Model Reduction Motivation b∈ Suppose we are given a G ∈ RH ∞ and our aim is to produce a reduced-order model G RH ∞ that approximates G. A natural criterion with which to measure the absolute error is b ∞. kG − Gk b G with the same input u, we get If we drive G and y = Gu,

and therefore that

b yb = Gu

Modelling and Model Reduction State-Space Truncation

b ∞ = sup kG − Gk

u∈L2

If G = D +C(sI − A)−1 B, then

ky − ybk2 . kuk2

107

(3.42)

x˙ = Ax + Bu

(3.43)

y = Cx + Du.

(3.44)

b ∞ to be small, the We assume that A is asymptotically stable (i.e., Re λi (A) < 0). For kG − Gk identity given in (3.42) suggests we should delete those components of the state-vector x that are least involved in the energy transfer from the input u to the output y. This observation leads us to consider two closely related questions: 1. What is the output energy resulting from a given initial state x(0) = x0 ? 2. What is the minimum input energy required to bring the state from zero to the given initial state x(0) = x0 ? The solutions are well known and closely related to the theory of Hankel operators [1]: 1. Suppose x(0) = x0 is given and that u(t) = 0 for t ≥ 0. By standard theory [1], the L2 [0, ∞) norm of y is given by kyk22 = x0′ Qx0 , in which Q is the observability gramian. 2. Consider the LQ problem min

Z 0

u∈L2 (−∞,0] −∞

u′ (t)u(t)dt

subject to x˙ = Ax + Bu with x(0) = x0 . This is equivalent to min

Z ∞

v∈L2 [0,∞) 0

v′ (τ )v(τ ) d τ

subject to d p(τ ) = −Ap(τ ) − Bv(τ ), p(0) = x0 , dτ with τ = −t, p(τ ) = x(t) and v(τ ) = u(t). By standard LQ theory [1], the optimal control is v(τ ) = B′ X p(τ ) and Z ∞

min v

0

v′ (τ )v(τ ) d τ = x0′ Xx0 ,

in which X is the solution to

−XA − A′ X − XBB′ X = 0 such that −A − BB′ X is asymptotically stable. If (A, B) is controllable and P is the controllability gramian satisfying AP + PA′ + BB′ = 0, then P is invertible and −P−1 A − A′ P−1 − P−1 BB′ P−1 = 0. Furthermore −A − BB′ P−1 = PA′ P−1 , which is asymptotically stable. Hence X = P−1 and we conclude that the optimal control is u(t) = B′ P−1 x(t) and that min

Z 0

u∈L2 (−∞,0]:x(0)=x0 −∞

u(t)′ u(t) dt = x0′ P−1 x0 .

108

D. J. N. Limebeer

Combining the answers to our two questions we get R∞ ′ 0 y (t)y(t) dt = R0

max

u∈L2 (−∞,0]:x(0)=x0 −∞ u′ (t)u(t) dt

=

x0′ Qx0 ′ x0 P−1 x0 1

(3.45) 1

α ′ P 2 QP 2 α , α ′α

1

x0 = P 2 α .

b ∞ small, the state-space for the trunThese calculations suggest that in order to keep kG − Gk cated system should be the space spanned by the eigenvectors corresponding to the larger 1

1

1

1

eigenvalues of P 2 QP 2 . That is, we should truncate a realization in which P 2 QP 2 is diagonal, with the eigenvalues ordered in descending order.

3.3.2 Balanced Realization In the last section we argued that an appropriate realization for absolute-error model reduction 1 1 is one in which P 2 QP 2 is diagonal. We now show that these realizations, known as a balanced realizations, always exist for asymptotically stable minimal realizations. The following is the most commonly used definition of a balanced realization: Definition 3.1. A realization (A, B,C) is balanced if A is asymptotically stable and

in which



AΣ + Σ A′ + BB′ = 0

(3.46)

A′ Σ + Σ A +C′C = 0,

(3.47)

 0 σ1 Ir1 0   Σ =  0 ... 0 , 0 0 σm Irm

σi 6= σ j , i 6= j and σi > 0 ∀i.

(3.48)

Note that n = r1 +· · ·+rm is the McMillan degree of C(sI −A)−1 B and that ri is the multiplicity of σi . We say that the realization is an ordered balanced realization if, in addition, σ1 > σ2 > · · · > σm > 0. In a balanced realization, the basis for the state space is such that each basis vector is equally controllable and observable, with its “degree” of controllability and observability given by the corresponding diagonal entry of Σ . Suppose (A, B,C) is a balanced realization and the initial condition x0 is partitioned as x0 = ′ ′ ′ with x an r × 1 vector. It follows from (3.45) that x1 · · · xm i i max

R∞ ′ m 0 y (t)y(t) dt = ∑ σi2 xi′ xi . R0 ′

u∈L2 (−∞,0]:x(0)=x0 −∞ u

σi2

(t)u(t) dt

i=1

This shows that is a measure of the extent to which the corresponding ri dimensional subspace of the state space is involved in the transfer of energy from past inputs to future outputs. These ideas can be related to the norm of the Hankel operator (with symbol G) [1]. The next result is concerned with the existence and uniqueness of balanced realizations.

109

Modelling and Model Reduction State-Space Truncation

Lemma 3.3. A given realization (A, B,C) can be transformed by a state transformation to a balanced realization if and only if it is asymptotically stable and minimal. Furthermore, a balanced realization obtained from such an (A, B,C) is unique up to: (a) the ordering of the σi ’s and (b) an orthogonal matrix S satisfying SΣ = Σ S. When the σi have multiplicity one (i.e., ri = 1 for all i), S is a diagonal matrix with diagonal elements ±1. When (A, B,C)is asymptotically stable and minimal, (TAT −1 , T B,CT −1 ) is balanced if T = 1 Σ 2 U ′ R−1 . When defining T , P = RR′ is a Cholesky factorization of P and R′ QR = U Σ 2U ′ is a singular value decomposition of R′ QR, in which P and Q are the controllability and observability gramians, which satisfy AP + PA′ + BB′ = 0

(3.49)

A′ Q + QA +C′C = 0.

(3.50)

Proof. To begin we note that if P and Q satisfy (3.49) and (3.50), then for any nonsingular T , (TAT −1 )(T PT ′ ) + (T PT ′ )(TAT −1 )′ + (T B)(T B)′ = 0 and (T ′ )−1 QT −1 (TAT −1 ) + (TAT −1 )′ (T ′ )−1 QT −1 + (CT −1 )′ (CT −1 ) = 0.

If (A, B,C) is balanced, it is asymptotically stable by assumption and Σ > 0 implies minimality. If (A, B,C) is asymptotically stable and minimal is has positive definite controllability and ob1 servability gramians P and Q satisfying (3.49) and (3.50) respectively. Setting T = Σ 2 U ′ R−1 gives 1 1 T PT ′ = (Σ 2 U ′ R−1 )RR′ (R′ )−1U Σ 2 = Σ

and

1

1

(T ′ )−1 QT −1 = (Σ − 2 U ′ R′ )Q(RU Σ − 2 ) = Σ .

Clearly, we may re-label the state components in a balanced realization to obtain another balanced realization. To determine the nonuniqueness that is possible while maintaining the same Σ , let S be a transformation that preserves the ordering of the σi in some balanced realization. Under this assumption we have Σ = SΣ S′ and Σ = (S−1 )′ Σ S−1 . This gives SΣ 2 = Σ 2 S, which implies that SΣ = Σ S since σi > 0. It now follows that Σ = SΣ S′ = Σ SS′ , so that I = SS′ as required.

Main Points of the Section 1. A balanced realization is an asymptotically stable and minimal realization in which the controllability and observability gramians are equal and diagonal. 2. Any stable transfer function matrix has a balanced realization. The balanced realization is unique up to ordering of the numbers σi and an orthogonal transformation that commutes with Σ . 3. An analysis of the extent to which states are involved in energy transfer from past inputs to future outputs motivates the consideration of the balanced realization as an appropriate realization for absolute-error model reduction.

110

D.J.N. Limebeer

3.4 Balanced Truncation Model reduction by balanced truncation simply applies the truncation operation to a balanced realization (A, B,C, D) of a system G. Suppose (A, B,C) is a balanced realization as described in Definition 3.1 and partition Σ as Σ1 0 (3.51) Σ= 0 Σ2 with



 σ1 Ir1 0 0   Σ1 =  0 . . . 0  , 0 0 σl Irl

 σl+1 Irl+1 0 0   .. Σ2 =  . 0 . 0 0 0 σm Im 

(3.52)

We never “split” states corresponding to a σi with multiplicity greater that one. If (A, B,C) b with realization (A11 , B1 ,C1 , D), is partitioned as in (3.9) conformably with Σ , we obtain G which is a balanced truncation of G. b is We will show that (A11 , B1 ,C1, D) is itself a balanced realization, which implies that G stable, has McMillan degree r = r1 + . . . + rl and that the approximation error satisfies the twice-the-sum-of-the-tail infinity norm bound b ∞ ≤ 2(σl+1 + . . . + σm ). kG − Gk

3.4.1 Stability Lemma 3.4. Suppose (A, B,C) is a balanced realization as described in Definition 3.1 and that (A11 , B1 ,C1 , D) is a balanced truncation of (A, B,C, D). Then (A11 , B1 ,C1 ) is a balanced realization. In particular, A11 is asymptotically stable and (A11 , B1 ,C1 ) is minimal. Note that by a trivial re-ordering argument (A22 , B2 ,C2 , D) is also a balanced realization. Proof. From (3.46) and (3.47) we have A11 Σ1 + Σ1 A′11 + B1 B′1 = 0

(3.53)

A′11 Σ1 + Σ1 A11 +C1′ C1 = 0.

(3.54)

If we can show that Re λi (A11 ) < 0, then it is immediate that (A11 , B1 ,C1 ) is a balanced realization because Σ1 > 0. Since Σ1 > 0, we have Re λi (A11 ) ≤ 0, but we still need to show that there can be no imaginary axis eigenvalues. Suppose, to obtain a contradiction, that there is a real ω such that jω I − A11 is singular. Let V be a basis for the kernel of jω I − A11 : ( jω I − A11 )V = 0.

(3.55)

Multiplying (3.54) on the left by V ∗ and on the right by V , and then multiplying just on the right by V , we obtain C1V = 0, ( jω I + A′11 )Σ1V = 0.

Modelling and Model Reduction State-Space Truncation

111

Multiplying (3.53) on the left by V ∗ Σ1 and on the right by Σ1V , and then multiplying just on the right by Σ1V , we obtain B′1 Σ1V = 0,

( jω I − A11 )Σ12V = 0.

Therefore Σ12V is also a basis for the right nullspace of jω I − A11 and hence 2

Σ12V = V Σ 1 2

for some matrix Σ 1 , which will have eigenvalues which are a subset of the eigenvalues of Σ12 . 2

(In fact, since ( jω I − A11 )(V T ) = 0 and Σ12 (V T ) = (V T )(T −1 Σ 1 T ) for any nonsingular T ,

we can assume V is such that those in Σ12 .)

2 Σ1

is diagonal—it will have diagonal entries that are a subset of

Now consider A21 Σ1 + Σ2 A′12 + B2 B′1 = 0

(3.56)

A′12 Σ1 + Σ2 A21 +C2′ C1

(3.57)

= 0,

which come from the (2, 1)-blocks of (3.46) and (3.47). Multiplying (3.56) on the right by Σ1V and multiplying (3.57) on the left by Σ2 and on the right by V , we obtain A21 Σ12V + Σ2 A′12 Σ1V = 0

Σ22 A21V

+ Σ2 A′12 Σ1V

(3.58)

= 0.

(3.59)

Subtracting these gives 2

Σ22 A21V = A21 Σ12V = A21V Σ 1 , which we may write as

"

2

Σ1 0 0 Σ22

#

I

I

2

Σ 1. A21V ′ 2 Since Σ 1 and Σ22 have no eigenvalues in common, I (A21V )′ must be a basis for the " # 2 2 Σ1 0 eigenspace of corresponding to the eigenvalues Σ 1 . That is, we have 0 Σ22

A21V

I A21V

=

=

I , 0

which amounts to A21V = 0. Combining this with (3.55) we obtain V ( jω I − A) = 0, 0 which contradicts the asymptotic stability of A. Notice that Lemma 3.4 does not assume that the balanced realization is ord ered in any way. The only assumption is that the partitioning of Σ does not split the states associated with a multiple σi .

112

D.J.N. Limebeer

3.4.2 Error Bound for “one-step” Truncation Our next result determines the infinity norm of the error that occurs when deleting the state(s) associated with one of the σi ’s. ˜ ˜ ˜ Lemma 3.5. Let A(s), B(s) and C(s) be proper real rational transfer function matrices, with˜ jω ) − jω I is nonsingular for all real ω . Suppose out poles on the imaginary axis, such that A( that ˜ + A˜∼ (s)) + B(s) ˜ B˜ ∼ (s) = 0 σ (A(s) ∼ ˜ + A˜ (s)) + C˜ ∼ (s)C(s) ˜ σ (A(s) =0

(3.60) (3.61)

−1 B(s) ˜ ˜ ˜ ˜ for some σ > 0. Then E(s) = C(s)(sI − A(s)) satisfies kEk∞ = 2σ . Moreover, if A(s) has odd dimension, then σ E(0) = 2σ .

Proof. The proof is divided into two steps: first we show that kEk∞ ≤ 2σ and then that equal˜ ity holds. In establishing the equality, we observe that σ E(0) = 2σ when A(s) has odd

˜ dimension. In the proof, we shall assume that the number of rows of C(s)( jω ) is greater than ˜ or equal to the number of columns of B(s)( jω ); when this is not the case, we may consider a dual argument based on E ∗ rather than E. Choose any real ω . From (3.60) and (3.61), we have ˜ jω ), ˜ jω )B˜ ∗ ( jω ) = C˜ ∗ ( jω )C( B( ˜ jω ) such that so there exists a matrix U( ∗ ˜ jω ) = σ 2 I, U˜ ( jω )U(

∗

˜ jω ) = 0. ˜ jω ) + C˜ ( jω )U( σ B(

(3.62)

(see Problem 3.1.) Now note that ˜ jω ) + E( jω ) ∗ U( ˜ jω ) + E( jω ) U( ∗ ∗ ˜ −∗C˜ ∗ U˜ + C( ˜ −1 B˜ ˜ jω I − A) = U˜ + B˜ ( jω I − A)

∗ ˜ −∗ ( jω I − A) ˜ + ( jω I − A) ˜ ∗ = σ I − σ B˜ ( jω I − A) ∗ ˜ −1 B˜ +(A˜ + A˜ ) ( jω I − A) 2

= σ 2 I.

(3.63) (3.64) (3.65) (3.66) (3.67)

Hence kEk∞ = sup σ E( jω )

(3.68)

ω

˜ jω ) + E( jω ) − U( ˜ jω ) = sup σ U( ω ˜ jω ) + E( jω ) + sup σ U( ˜ jω ) ≤ sup σ U( ω

ω

= 2σ .

We now show that there is a frequency ω0 such that σ E( jω0 ) = 2σ . Define ˜ jω ) − jω I, Φ˜ ( jω ) = A(

∗

˜ ( jω ) − Φ ˜ ( jω ) X( jω ) = Φ

(3.69) (3.70) (3.71)

Modelling and Model Reduction State-Space Truncation

113

and note that X( jω ) is skew-Hermitian since X( jω ) + X ∗ ( jω ) = 0.

(3.72)

+ λ )x∗ x

If X( jω )x = λ x, x 6= 0, it follows from (3.72) that (λ = 0, so that λ + λ = 0. That is, every eigenvalue of X( jω ) is on the imaginary axis. Now X(0) is real, so it has an equal number of eigenvalues with positive and negative imaginary part. For sufficiently large ω , all the eigenvalues of X( jω ) have negative imaginary part, since X( jω ) → − j2ω I as ω → ∞. Since the eigenvalues of a matrix are continuous functions of the matrix entries, it follows that there is a frequency, ω0 , such that X( jω0 ) is singular (i.e., has a zero eigenvalue). If X has odd dimension, X(0) must be singular, since it is a real matrix of odd dimension with all its eigenvalues on the imaginary axis, so we may take ω0 = 0 in this case. Let x 6= 0 and ω0 be selected so that X( jω0 )x = 0. Then 0= = = = =

1 X( jω0 )x 2 1 ˜ ∗ Φ ( jω0 ) − Φ˜ ( jω0 ) x 2 1 ∗ Φ˜ ( jω0 )x − (Φ˜ ( jω0 ) + Φ˜ ( jω0 ))x 2 1 ˜ ∗ Φ˜ ( jω0 )x + by (3.60) B( jω0 )B˜ ( jω0 )x 2σ ˜ jω0 )u, Φ˜ ( jω0 )x + B(

(3.73) (3.74) (3.75) (3.76) (3.77)

1 ˜∗ 2σ B ( j ω0 )x.

˜ ( jω0 )x = 0, in which u = Note that u 6= 0, since otherwise we would have Φ which is banned by assumption. Now using (3.62) we have ˜ jω0 )x, ˜ jω0 )u = C( −2U( giving

Hence ˜∗

˜ jω0 Since U ( jω0 )U(

˜ jω0 ) Φ˜ ( jω0 ) B( ˜ 0 C( jω0 )

x 0 = ˜ jω0 )u . u −2U(

˜ jω0 )u, E( jω0 )u = −2U( u 6= 0. we see that σ E( jω0 ) = 2σ .

) = σ 2 I,

Consider the situation in which the state(s) associated with one σi are deleted by balanced ˜ ˜ ˜ truncation and define A(s), B(s) and C(s) as in Lemma 3.1. Then by Lemma 3.4 the assumptions of Lemma 3.5 are satisfied and we conclude that the infinity norm of the error associated with this special “one-step” balanced truncation is exactly 2σ . Moreover, if the associated multiplicity r of σ is odd, then the maximum error (as measured by the maximum singular value) occurs at ω = 0.

3.4.3 The Error Bound for Balanced Truncation Lemma 3.5 provides an infinity norm bound on the absolute error associated with the truncation of the state(s) associated with a single σi in a balanced realization. To determine a bound that is applicable when the states associated with several σi ’s are deleted, we simply remove the σi ’s one at a time and allow the error to accumulate. This procedure yields the twice-the-sum-of-the-tail error bound.

114

D.J.N. Limebeer

Theorem 3.1. Let G = D + C(sI − A)−1 B, in which (A, B,C, D) is a balanced realization b = partitioned as in (3.51). Let r = r1 + · · · + rl , (A11 , B1 ,C1 , D) = Tr (A, B,C, D) and G D +C1 (sI − A11 )−1 B1 . Then b ∞ ≤ 2(σl+1 + · · · + σm ). kG − Gk b In the case that l = m − 1 equality holds, and σ G(0) − G(0) = 2σm if rm is odd.

(3.78)

Proof. Truncating the states associated with σl+1 , . . . , σm may be achieved by a succession of one-step truncation operations. At the kth step we delete the states associated with σm−k+1 , k = 1 . . . m − l, to obtain Gk from Gk−1 where G0 = G. Each truncation step preserves the balanced realization and each step incurs an error of E k = Gk−1 − Gk with kE k k∞ = 2σm−k+1 . b Now write The last of these steps gives Gm−l = G. b = (G0 − G1 ) + · · · + (Gm−l−1 − Gm−l ) G−G = E 1 + · · · + E m−l

(3.79)

and observe that the triangle inequality gives b ∞ ≤ kE 1 ||∞ + · · · + kE m−l k∞ . kG − Gk

Since kE k k∞ = 2σm−k+1 , we obtain the error bound (3.78).

By considering the case when all the states are deleted, the bound (3.78) yields kG − G(∞)k∞ ≤ 2(σ1 + · · · + σm ). Hence, kGk∞ = kG − G(∞) + G(∞)k∞

≤ kG(∞)k + 2(σ1 + · · · + σm ).

(3.80) (3.81)

Tightness of the Bound We have already established that the infinity norm error bound for one-step truncation is tight. In the case of multi-step truncation the situation is less clear cut—one can find examples for which the error bound is close to the true error and there are also cases for which it is conservative. It is natural to suspect that because the twice-the-sum-of-the-tail bound arises by repeated application of the triangle inequality, it gets weaker and weaker as more and more states are deleted. Although this is usually the case, the bound may remain tight however many states are deleted. The next example illustrates this point. Example 3.3. Consider the transfer function n

gn =

i=1

which may be realized as

αi

∑ s + αi

α > 0,

α 6= 1

115

Modelling and Model Reduction State-Space Truncation



 −α 0 0   A =  0 ... 0  0 0 −α n

√  α  .  B =  ..  √ αn

C = B′ .

It is easily verified that the controllability and observability gramians are equal and the (i, j) entry of the controllability gramian is √ α i− j Pi j = i− j i, j = 1, . . . , n. α +1 Since P = Q, we have

σi = giving

p λi (PQ) = λi (P),

i = 1, . . . , n,

2(σ1 + · · · + σn ) = 2 × trace(P) n = 2× 2 = n.

(3.82) (3.83) (3.84)

Therefore (3.79) implies kgn k∞ ≤ n. Since g(0) = n, equality must hold.

Systems of the form given in the above example have interlaced poles and zeros on the negative real axis (Problem 3.4). The bound is tight for such systems. At the other extreme, for systems that have interlaced poles and zeros on the imaginary axis, the bound exceeds the true error by a factor that is approximately twice the number of states that are deleted—see Enns [2].

Frequency Dependence of the Error The error bounds given in Theorem 3.1 says nothing about the way the true error varies as a b match function of frequency. Since the full-order system G and the reduced-order system G each other exactly at infinite frequency, one would expect good high-frequency fidelity from the reduced order model. Apart from the guaranteed satisfaction of the error bound, little can be said about the variation of the error at low and intermediate frequencies. In most cases, the σi ’s will have unit multiplicity and the truncation of n − r states will be achieved via n − r truncations of one state. Since one is odd, each of these truncations incurs a maximum error at zero frequency. One would therefore expect the largest truncation error to occur at low frequency.

Main Points of the Section 1. Any truncation of a balanced realization that does not “split” the states associated with a single singular value is called balanced truncation. 2. Balanced truncation preserves stability and minimality, and the approximation error satisfies the twice-the-sum-of-the-tail infinity norm error bound. 3. The infinity norm of a strictly proper transfer function matrix is bounded above by twice the sum of the σi ’s in its balanced realization.

116

D.J.N. Limebeer

3.5 Balanced Singular Perturbation Approximation The fact that balanced truncation generally incurs the greatest approximation error in the lowfrequency region is undesirable in many applications. An algorithm which produces zero error at zero frequency may be obtained via an easy modification to the basic balanced truncation algorithm. The idea is to simply replace s with 1/s as follows: 1. Set H(w) = G(1/w). 2. Let H r (w) be an rth -order balanced truncation of H(w). 3. Set Gr (s) = H r (1/s). This algorithm will have exact matching at zero frequency, thereby leading to prefect steadystate performance; it is a singular perturbation version of balanced truncation, which we call balanced singular perturbation approximation (BSPA). Since singular perturbation and truncation are connected by the simple fre quency inversion s → 1/s, the infinity norm error bounds will still hold. Secondly, since frequency inversion preserves stability, both algorithms will produce reduced-order models with the same stability properties. Finally, since frequency inversion preserves the controllability and observability gramians and hence a balanced realization, we obtain the following result: b =D b11 )−1 Bb1 , b + Cb1 (sI − A Theorem 3.2. Assume the hypotheses of Theorem 3.1 and define G b11 , Bb1 , Cb1 , D) b = Sr (A, B,C, D). Then: in which (A b11 , Bb1 , Cb1 , D) b is balanced with its controllability and observability 1. The realization (A b11 is stable, (A b11 , Bb1 ) is gramians given by Σl = diag(σ1 Ir1 , . . . , σl Irl ). In particular, A b b controllable and (C1 , A11 ) is observable. b ∞ ≤ 2(σl+1 + · · · + σm ), and equality holds if l = m − 1. 2. kG − Gk b 3. G(0) = G(0). 4. kG − G(0)k∞ ≤ 2(σ1 + · · · + σm ).

Proof. Since (A, B,C) is balanced, A is stable and therefore nonsingular. By Lemma 3.4, A22 b11 , Bb1 , Cb1 , D) b given by (3.33) are well defined is also stable and hence is nonsingular. Thus (A and the assumptions of Lemma 3.2 hold. All the items now follow directly from their balanced truncation counterparts by using Lemma 3.2.

Main Point of the Section The stability and infinity norm properties of any model reduction procedure will be preserved by a change of variables that maps the left-half plane into itself. The balanced singular perturbation approximation, which is equivalent to balanced truncation of the system obtained by setting w = 1/s, has the same infinity norm error as balanced truncation, but has zero steady-state error. In general, its performance at low frequencies is superior to that of balanced truncation.

Modelling and Model Reduction State-Space Truncation

117

3.6 Example We illustrate the model reduction methods described above with an eighth-order model of a flexible structure [3]. The model is 4

g=

ω2

∑ ki s2 + 2ζi ωi i s + ω 2 ,

i=1

i

in which i 1 2 3 4

ωi 0.56806689746895 3.94093897440699 10.58229653714164 16.19234386986640

ζi 0.00096819582773 0.00100229920475 0.00100167293203 0.01000472824082

ki 0.01651378989774 0.00257034576009 0.00002188016252 0.00027927762861

The controllability and observability gramians of a balanced realization for this system are 4.26819 4.25994 0.641754 0.640469 Σ = diag . 0.069856 0.0697163 0.00546623 0.00545529 By making use of the error bound given in Theorem 3.1, we see that if two states are eliminated, the infinity norm of the error is less than 0.0218431. If four states are eliminated, the error bound increases to 0.300988 and so on. Eliminating four states by balanced truncation gives   −0.000550222 −0.568066 0.000993291 0.00134457 0.068534  0.568066 −0.000549776 0.00134607 0.000994008 −0.0684399   s  0.000993291 −0.00134607 −0.00394544 −3.94093 −0.0711619  b g=  .  −0.00134457 0.000994008 3.94093 −0.00395455 0.0711725  0.068534 0.0684399 −0.0711619 −0.0711725 0 Figure 3.1 shows the gain of the full-order model g, the fourth-order balanced truncation model b g and the fourth-order balanced singular perturbation model. The gains of the error incurred by each method, along with the error bound, are shown in Figure 3.2. The solid line is the error bound, the dashed line is the balanced truncation error and the dash-dot line is the balanced singular perturbation error. Notice that the actual error is an order of magnitude less than the error bound. This is not unexpected for this type of system because its poles and zeros are interlaced and close to the imaginary axis. It is also worth noting that the balanced truncation reduced-order model is virtually identical to the fourth-order model obtained by deleting the two high-frequency modes. The main difference is that the balanced truncation reduced-order model has an additional zero at −3980.19.

3.7 Notes and References Balanced realizations first appeared with the work of Mullis and Roberts [4], who were interested in realizations of digital filters that are optimal with respect to round-off errors in the state update calculations. These issues are developed extensively in the book by Williamson [5].

118

D.J.N. Limebeer

10 1 10 0

10 -1

gain

10 -2

10 -3

10 -4

10 -5

10 -6

10 -7

10 -8 10 -1

10 0

10 1

10 2

10 3

frequency (rad/s)

Fig. 3.1. Flexible structure models: full-order g (solid), balanced truncation b g (dashed) and balanced singular perturbation (dash-dot) 10 0 10 -1

10 -2

gain

10 -3

10 -4

10 -5

10 -6

10 -7

10 -8 10 -1

10 0

10 1

10 2

10 3

frequency (rad/s)

Fig. 3.2. Model reduction errors: error bound (solid), balanced truncation (dashed) and balanced singular perturbation (dash-dot)

The balanced truncation method of model reduction is due to Moore, who argued that the method is sensible on system theoretic grounds. He also proved a weak version of the stability result. The asymptotic stability result, Lemma 3.4, is due to Pernebo and Silverman [6]. The twice-the-sum-of-the-tail infinity norm error bound for balanced truncation is due to Enns [2],

Modelling and Model Reduction State-Space Truncation

119

who also proposed a frequency weighted generalization He introduced formula (3.14) for the truncation error system and proved that the error for “one-step” truncation is 2σ when the multiplicity of σ is one. Our new proof shows that no matter what the multiplicity of σ is, the infinity norm of the error is exactly 2σ . The twice-the-sum-of-the-tail error bound was also proved by Glover [7] using a different approach. The balanced singular perturbation approximation method was introduced by Fernando and Nicholson [8]. Liu and Anderson [9] showed that it was related to balanced truncation by the transformation s → 1/s. Numerical algorithms for computing balanced realizations have been considered by Laub, Heath, Page and Ward [10]. Tombs and Postlethwaite [11], and Safonov and Chiang [12] have developed various algorithms that compute balanced truncation approximations without computing balanced realizations. None of these procedures eliminate the need to compute the controllability and observability gramians, which is a serious problem when large-scale models need to be reduced. Jaimoukha, Kasenally and Limebeer [13] have developed algorithms for computing approximate solutions to Lyapunov equations using Krylov subpace methods that can be effective even for models containing many hundreds of states. A relative error based truncation method of model reduction is the balanc ed stochastic truncation method of Desai and Pal [14]. Green derived an infinity norm bound for the relative error associated with this method [15]; the bound was subsequently improved by Wang and Safonov [16]. Another feature of the algorithm is that it preserves the closed-right-half-plane zeros of the model and will therefore produce a minimum phase approximation of a minimum phase system [17]. Discrete-time versions of the results in these notes (see Problem 3.6) hav e been proved by Al-Saggaf and Franklin [18].

3.8 Problems Problem 3.1. Let B and C be complex matrices such that BB∗ = C∗C. 1. If the number of rows of C is greater than or equal to the number of columns of B, show that for any nonzero real number σ , there exists a matrix U such that σ B + C∗U = 0 and U ∗U = σ 2 I. (Hint: Write a singular value decomposition of C and infer the form of the singular value decomposition of B; see Lemma 3.5 of [7].) 2. Show that for any nonzero real number σ , there exists a matrix U such that σ B +C∗U = 0 and U ∗U ≤ σ 2 I. Show further that the nonzero singular values of U are all equal to σ . Problem 3.2. Suppose (Aii , Bi ,Ci ), i = 1, 2 are two balanced realizations with controllability/observability gramians Σ1 and Σ2 respectively. Suppose also that Σ1 and Σ2 have no eigenvalues in common. Construct the unique matrices A21 and A12 such that A11 A12 B1 A= , B= , (3.85) A21 A22 B2 C = C1 C2 (3.86) Σ1 0 is a balanced realization with controllability/observability gramian . 0 Σ2

120

D.J.N. Limebeer

Problem 3.3. Assume the hypotheses of Lemma 3.5. Using the bounded real lemma [1], prove that ˜ jω ))−1 B( ˜ jω )( jω I − A( ˜ jω ) ≤ 2σ . sup σ C( ω

Problem 3.4. Show that the poles and zeros of the system gn defined in Example 3.3 have an interlacing property—between any two poles there is exactly one zero. If gn is the impedance of a linear circuit, explain why this interlacing property implies that the circuit is passive. Problem 3.5. Let (A, B,C, D) be any realization such that A is asymptotically stable and let P and Q satisfy AP + PA′ + BB′ = 0

(3.87)

A′ Q + QA +C′C = 0.

(3.88)

Show that the following procedure computes the rth -order balanced truncation approximation to G = D +C(sI − A)−1 B: a. Let P = UP SPUP′ and Q = UQ SQUQ′ be ordered singular value decompositions of P and Q. 1

1

b. Let SQ2 UQ′ UP SP2 have (ordered) singular value decomposition V Σ U ′ . c. Partition the matrices V , Σ and U as Σ1 0 , U = U1 U2 , V = V1 V2 , Σ = 0 Σ2

in which Σ1 is r × r. d. Define

1

−1

L = UQ SQ2 V1 Σ1 2 ,

1

−1

M = UP SP2 U1 Σ1 2 .

e. Define Ar = L′ AM, Br = L′ B and Cr = CM. b = D +Cr (sI − Ar )−1 Br . f. Define G

Problem 3.6. [18] A realization (A, B,C, D) of a discrete-time system G(z) = D + C(zI − A)−1 B is balanced if |λi (A)| < 1 for all i and there exists a positive definite diagonal matrix of the form given in (3.48) such that AΣ A′ − Σ + BB′ = 0 ′

′

A Σ A − Σ +C C = 0.

(3.89) (3.90)

1. Show that a given realization of a discrete-time system can be transformed into a discretetime balanced realization if and only if it is stable and minimal. 2. Suppose that (A, B,C, D) is a discrete-time balanced realization which is partitioned as in (3.9) and (3.48). Show that A11 is asymptotically stable (|λi (A11 )| < 1 for all i). 3. Define ˜ θ ) = A22 + A21 (e jθ I − A11 )−1 A12 A( ˜ θ ) = B2 + A21 (e jθ I − A11 )−1 B1 B( ˜ C(θ ) = C2 +C1 (e jθ I − A11 )−1 A12 .

(3.91) (3.92) (3.93)

Modelling and Model Reduction State-Space Truncation

121

Show that ˜ θ )Σ2 A˜∗ (θ ) − Σ2 + B( ˜ θ )B˜ ∗ (θ ) = 0 A( ∗ ˜ θ ) = 0. ˜ θ ) − Σ2 + C˜ ∗ (θ )C( A˜ (θ )Σ2 A(

(3.94) (3.95)

4. Suppose that A is asymptotically stable |λi (A)| < 1 for all i and that AA∗ − I + BB∗ = 0

A∗ A − I +C∗C = 0.

(3.96) (3.97)

Show that σ [C(e jθ I − A)−1 B] < " 2. 1 # h i 2 1 5. Show that (A11 , A12 Σ 2 B1 , A21 Σ2 ) is a balanced realization. 2 C1

6. Suppose that G(z) = D + C(zI − A)−1 B, in which (A, B,C, D) is a discrete-time balanced realization. Let (A11 , B1 ,C1 , D) be an rth -order balanced truncation of (A, B,C, D) and let b = D +C1 (zI − A11 )−1 B1 . Show that G(z) b ∞ < 2(σr+1 + . . . + σn ). kG − Gk

Problem 3.7. [19, 20] Let (A, B,C, D) be a balanced realization of a continuous-time system and let α be a nonnegative number. Define b = A11 + A12 (α I − A22 )−1 A21 A Bb = B1 + A12 (α I − A22 )−1 B2 Cb = C1 +C2 (α I − A22 )−1 A21

b = D +C2 (α I − A22 )−1 B2 , D

(3.98) (3.99) (3.100) (3.101)

in which (A, B,C, D) are partioned as in (3.9) conformably with a partioning of Σ in (3.52). The generalized singular perturbation approximation (GSPA) is defined by G S r (A, B,C, D) = b B, b D). b C, b (A,

1. Show that the GSPA reduced-order model results from replacing the dynamics associated with x2 by the exponential system x˙2 = α x2 . 2. Show that the GSPA approximant has zero error at s = α . +s , which 3. Suppose that 0 < α < ∞ and define the linear fractional transformation z = αα −s maps the left-half plane to the unit circle. Suppose also that we consider the equivalent α (z−1) ˜ − A) ˜ −1 B, ˜ in discrete-time system defined by F(z) = G( z+1 ). Show that F = D˜ + C(zI which A˜ = (α I + A)(α I − A)−1 = (α I − A)−1 (α I + A) √ B˜ = 2α (α I − A)−1 B √ C˜ = 2α C(α I − A)−1

D˜ = D +C(α I − A)−1 B.

(3.102) (3.103) (3.104) (3.105)

Show that this realization of F is a discrete-time balanced realization. 4. Show that for 0 < α < ∞, GSPA is equivalent to: (a) mapping to discrete-time via +s ; (b) discrete-time balanced truncation (see Problem 3.6) and (c) mapping back z = αα −s α (z−1)

to continuous-time via s = z+1 . 5. For 0 ≤ α ≤ ∞, show that reduced-order models obtained by GSPA enjoy the same stability and infinity norm error bound properties as balanced truncation and SPA.

122

D.J.N. Limebeer

References 1. M. Green and D. Limebeer, Linear Robust Control. Englewood Cliffs, N.J.: PrenticeHall, 1995. 2. D. Enns, “Model reduction with balanced realizations: An error bound and frequency weighted generalization,” in Proceedings of the IEEE Conference on Decision and Control, Las Vegas, 1984, pp. 127–132. 3. H. V. et al., “Flexible structure control laboratory development and technology demonstration,” California Institute of Technology, JPL Publication 88-29, 1987. 4. C. Mullis and R. Roberts, “Synthesis of minimum roundoff noise fixed point digital filters,” IEEE Transactions of Circuits and Systems, vol. 23, pp. 551–562, 1976. 5. D. Williamson, Digital Control and Implementation : Finite Wordlength Considerations. Englewood Cliffs, N.J.: Prentice-Hall, 1990. 6. L. Pernebo and L. Silverman, “Model reduction by balanced state space representations,” IEEE Transactions on Automatic Control, vol. 27, pp. 382–387, 1982. 7. K. Glover, “All optimal hankel-norm approximations of linear multivariable systems and their L∞ -error bounds,” International Journal of Control, vol. 39, no. 6, pp. 1115–1193, 1984. 8. K. Fernando and H. Nicholson, “Singular perturbational model reduction of balanced systems,” IEEE Transactions on Automatic Control, vol. 27, pp. 466–468, 1982. 9. Y. Liu and B. Anderson, “Singular perturbation approximation of balanced systems,” International Journal of Control, vol. 50, pp. 1379–1405, 1989. 10. A. Laub, M. Heath, C. Page, and R. Ward, “Computation of balacing transformations and other applications of simultaneous diagonalization algorithms,” IEEE Transactions on Automatic Control, vol. 32, pp. 115–122, 1987. 11. M.S.Tombs and I. Postlethwaite, “Truncated balanced realizations of a stable nonminimal state-space system,” Rpt. No. OUEL 1647/86, Engineering Dept., Oxford., 1986. 12. M. Safonov and R. Chiang, “A Schur method for balanced-truncation model reduction,” IEEE Transactions on Automatic Control, vol. 43, no. 7, pp. 729–733, 1989. 13. I. Jaimoukha, E. Kasenally, and D. Limebeer, “Numerical solution of large scale lyapunov equations using krylov subspace methods,” in Proceedings of the IEEE Conference on Decision and Control, Tuscon, 1992, pp. 1927–1932. 14. U. Desai and D. Pal, “A transformation approach to stochastic model reduction,” IEEE Transactions on Automatic Control, vol. 29, no. 12, pp. 1097–1100, 1984. 15. M. Green, “A relative error bound for balanced stochastic truncation,” IEEE Transactions on Automatic Control, vol. 33, no. 10, pp. 961–965, 1988. 16. W. Wang and M. Safonov, “A tighter relative-error bound for balanced stochastic truncation,” Systems and Control Letters, vol. 14, no. 4, pp. 307–317, 1990. 17. M. Green, “Balanced stochastic realizations,” Linear Algebra and its Applications, vol. 98, pp. 211–247, 1988. 18. U. Al-Saggaf and G. Franklin, “An error bound for a discrete reduced order model of a linear multivariable system,” IEEE Transactions on Automatic Control, vol. 32, pp. 815– 819, 1987. 19. P. Heuberger, “A familily of reduced order models based on open-loop balancing,” in Selected Topics in Identification, Modelling and Control, 1990, vol. 1, pp. 1–10. 20. U. Al-Saggaf and G. Franklin, “Model reduction via balanced realization: An extension and frequency weighting techniques,” IEEE Transactions on Automatic Control, vol. 33, pp. 681–692, 1988.

4 Linear Matrix Inequalities in Control Guido Herrmann, Matthew C. Turner, and Ian Postlethwaite

Summary. This chapter gives an introduction to the use of linear matrix inequalities (LMIs) in control. LMI problems are defined and tools described for transforming matrix inequality problems into a suitable LMI-format for solution. Several examples explain the use of these fundamental tools.

4.1 Introduction to LMI Problems The design of a controller should in general satisfy four basic requirements: 1. Closed-Loop Stability 2. Robustness: The closed-loop controller has to remain stable despite uncertainty in the mathematical plant description or disturbances. 3. Performance: The controller has to have certain dynamical or steady state characteristics such as rise time, overshoot, controller bandwidth or steady state error etc. 4. Robust Performance: The controller has to remain well performing and certainly stable, although disturbances and uncertainties affect the plant. 1 These control design requirements are usually best encoded in the form of an optimization criterion subject to some constraints e.g., terminal constraints or practical constraints on plant variables. Control design optimization criteria were initially based on the idea of Linear Quadratic Gaussian Control [1, 1, 8] which was later generalized to the idea of H2 -controller design [20]. The development of small gain theory [10, 17] layed the foundations of robust H∞ -control [3, 20]. For linear systems and suitable optimization criteria, such as H2 - and H∞ , the solution to the optimization problem is readily found solving Riccati-equations [3, 19, 20]. Many of these optimal control problems can be stated in terms of linear matrix inequalities and their existence can be traced back over 100 years to the work of Lyapunov. Linear matrix inequalities are matrix inequalities which are linear (or affine) in a set of matrix variables. 1

It is obvious that robust performance implies robustness and performance but it is a stronger requirement than considering robustness and performance each as a single entity.

M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 123-142, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

124

G. Herrmann, M.C . Turner, and I. Postlethwaite

ItisinterestingtonotethattheequivalencebetweentheRiccati-equationandLMIformulations of the control problem was found at an early point in the seventies [19]. This created the possibility of solving optimal control problems using LMI-methods, as the numerical LMIoptimization problem was found to be convex. However, only during the past 10-15 years has the development of sophisticated numerical routines, i.e., semi-definite programming [11,18], made it possible to solve LMIs in a reasonably efficient manner. From a control engineering perspective, one of the main attractions of LMIs is that they can be used to solve problems which involve several matrix variables, and, moreover, different structures can be imposed on these matrix variables. Another attractive feature of LMI methods is that they are flexible, so it is often relatively straightforward to pose a variety of problems as LMI problems amenable to LMI methods. Furthermore, in many cases the use of LMIs can remove restrictions associated with conventional methods and aid their extension to more general scenarios. Often LMI methods can be applied in instances where conventional methods either fail, or struggle to find a solution [15]. In actual fact, the flexibility of LMIs has created a much wider scope for controller design. They allow the efficient consideration of H2 and H∞ -constraints for performance, robustness and robust performance in the design of one single controller [13, 14]. Thus an advantage of LMIs, also in a pedagogical sense, is that they are able to unite many previous results in a common framework. This can also enable one to obtain additional insight into established areas. Some of the contents of this chapter are available as a University of Leicester Technical report [16] which draws heavily on the material contained within the book in [2] and the R MATLAB LMI control toolbox in [6]. A similar treatment can be found in [12]. In order to convey the main points, the presentation is somewhat condensed and the interested reader should consult the work in [2] for a more complete exposure to LMIs.

4.1.1

Fundamental LMI Properties

A notion central to the understanding of matrix inequalities is definiteness. In particular, a matrix Q is defined to be positive definite if xT Qx > 0 ∀x 6= 0

(4.1)

Likewise, Q is said to be positive semi-definite if xT Qx ≥ 0

∀x

(4.2)

It is common practise to write Q > 0 (Q ≥ 0) to indicate that it is positive (semi-) definite. In particular, we are interested in positive definite matrices which are also symmetric, i.e., Q = QT . A symmetric, positive definite matrix has two key features: it is square and all of its eigenvalues are positive real. A symmetric, positive semi-definite matrix shares the first attribute, but the last is relaxed to the requirement that all of its eigenvalues are positive real or zero. A matrix P = −Q is said to be negative (semi) definite if Q is positive (semi) definite. To indicate negative (semi) definiteness we write P < 0 (P ≤ 0). In fact, once the notation Q > 0 (Q ≥ 0) or P > 0 (P ≥ 0) is used, we usually also require Q and P to be symmetric and we will do so in the rest of this chapter, although from a mathematical point of view this is not necessary. Nevertheless, numerical solution routines for LMIs have enforced this fact as this simplifies the computational process and it is therefore a common assumption for LMIs [2].

Linear Matrix Inequalities in Control

125

The basic structure of an LMI is m

F(x) = F0 + ∑ xi Fi > 0

(4.3)

i=1

where x ∈ Rm is a variable and F0 , Fi are given constant symmetric matrices. The basic LMI problem - the feasibility problem - is to find x such that inequality (4.3) holds. Note that F(x) > 0 describes an affine relationship in terms of the matrix x. Normally the variable x, which we are interested in, is composed of one or many matrices whose columns have been ‘stacked’ as a vector. That is, F(x) = F(X1 , X2 , . . . , Xn )

(4.4)

where Xi ∈ Rqi ×pi is a matrix, ∑ni=1 qi × pi = m, and the columns of all the matrix variables are stacked up to form a single vector variable. Hence, from now on, we will consider functions of the form

F(X1 , X2 , . . . , Xn ) = F0 + G1 X1 H1 + G2 X2 H2 + . . .

(4.5)

n

= F0 + ∑ Gi Xi Hi > 0

(4.6)

i=1

where F0 , Gi , Hi are given matrices and the Xi are the matrix variables which we seek.

4.1.2

Systems of LMIs

In general, we are frequently faced with LMI constraints of the form

F1 (X1 , . . . , Xn ) > 0 .. .>0

(4.7)

Fp (X1 , . . . , Xn ) > 0

(4.9)

(4.8)

where n

Fj (X1 , . . . , Xn ) = F0 j + ∑ Gi j Xi Hi j

(4.10)

i=1

However, it is easily seen that, by defining F˜0 , G˜ i , H˜ i , X˜i as F˜0 = diag(F01 , . . . , F0p ) G˜ i = diag(Gi1 , . . . , Gip ) H˜ i = diag(Hi1 , . . . , Hip )

(4.12)

X˜i = diag(Xi , . . . Xi )

(4.14)

(4.11) (4.13)

126

G. Herrmann, M.C . Turner, and I. Postlethwaite

we actually have the inequality n

Fbig (X1 , . . . , Xn ) := F˜0 + ∑ G˜ i X˜i H˜ i > 0

(4.15)

i=1

That is, we can represent a (big) system of LMIs as a single LMI. Therefore, we do not distinguish a single LMI from a system of LMIs; they are the same mathematical entity. We may also encounter systems of LMIs of the form:

F1 (X1 , . . . , Xn ) > 0

(4.16)

F2 (X1 , . . . , Xn ) > F3 (X1 , . . . , Xn ).

(4.17)

Again, it is easy to see that this can be written in the same form as inequality (4.15) above. For the remainder of the chapter, we do not distinguish between LMIs which can be written as above, or those which are in the more generic form of inequality (4.15).

4.1.3

Types of LMI Problems

The term ‘LMI problem’ is rather vague and in fact there are several sub-groups of LMI problems. These will be described below in the same way that they are separated in the R MATLAB LMI toolbox. Note that by ‘LMI problem’ we normally mean solving an optimization problem or an eigenvalue problem with LMI constraints.

LMI Feasibility Problems These are simply problems for which we seek a feasible solution {X1 , . . . , Xn } such that F(X1 , . . . , Xn ) > 0

(4.18)

We are not interested in the optimality of the solution, only in finding a solution, which may not be unique. Example 4.1. (Determining stability of a linear system). Consider an autonomous linear system x˙ = Ax

(4.19)

then the Lyapunov LMI problem for proving asymptotic stability of this system is to find a P > 0 such that AT P + PA < 0 (4.20) This is obviously an LMI feasibility problem in P > 0. However, given any P > 0 which satisfies this, it is obvious that any matrix from the set P = {β P : scalar β > 0}

(4.21)

also solves the problem. In fact, as will be seen later, the matrix P forms part of a Lyapunov function for the linear system.

Linear Matrix Inequalities in Control

127

Linear Objective Minimization Problems These problems are also called eigenvalue problems. They involve the minimization (or maximization) of some linear scalar function, α (.), of the matrix variables, subject to LMI constraints:

min α (X1 , . . . , Xn ) s.t.

(4.22)

F(X1 , . . . , Xn ) > 0

(4.23)

where we have used the abbreviation ‘s.t.’ to mean ‘such that’. In this case, we are therefore trying to optimize some quantity whilst ensuring some LMI constraints are satisfied. Example 4.2. (Calculating the H∞ norm of a linear system). Consider a linear system

x˙ = Ax + Bw

(4.24)

z = Cx + Dw

(4.25)

then the problem of finding the H∞ norm of the transfer function matrix Tzw from w to z is equivalent to the following optimization procedure (see, for example the work in [5]):

 

min γ AT P + PA BT P C

s.t. 

(4.26)

CT

PB −γ I DT  < 0 D −γ I P>0

(4.27)

Note that although γ > 0 is unique, the uniqueness of P > 0 is, in general, not guaranteed.

Generalized Eigenvalue Problems The generalized eigenvalue problem, or GEVP, is slightly different to the preceding problem in the sense that the objective of the optimization problem is not actually convex, but quasiconvex. However, the methods used to solve such problems are similar. Specifically a GEVP is formulated as

min λ

s.t.

(4.28)

F1 (X1 , . . . , Xn ) + λ F2 (X1 , . . . , Xn ) < 0

(4.29)

F2 (X1 , . . . , Xn ) < 0

(4.30)

F3 (X1 , . . . , Xn ) < 0

(4.31)

The first two lines are equivalent to minimizing the largest ‘generalized’ eigenvalue of the matrix pencil F1 (X1 , . . . , Xn ) + λ F2 (X1 , . . . , Xn ). In some cases, a GEVP problem can be reduced to a linear objective minimization problem, through an appropriate change of variables.

128

G. Herrmann, M.C . Turner, and I. Postlethwaite

Example 4.3. (Bounding the decay rate of a linear system). A good example of a GEVP is given in [2]. Given a stable linear system x˙ = Ax, the decay rate is the largest α such that kx(t)k ≤ exp(−α t)β kx(0)k ∀x(t)

(4.32)

where β is a constant. If we choose V (x) = xT Px > 0 as a Lyapunov function for the system and ensure that V˙ (x) ≤ −2α V (x) it is easily shown that the system will have a decay rate of at least α . Hence the problem of finding the decay rate could be posed as the optimization problem

T

min −α

s.t.

(4.33)

A P + PA + 2α P < 0

(4.34)

−P < 0

(4.35)

This problem is a GEVP with the functions F1 (P) := AT P + PA

(4.36)

F2 (P) := −2P

(4.37)

F3 (P) := −I

4.2

(4.38)

Tricks in LMI Problems

Although many control problems can be cast as LMI problems, a substantial number of these need to be manipulated before they are in a suitable LMI problem format. Fortunately, there are a number of common tools or ‘tricks’ which can be used to transform problems into suitable LMI forms. Some of the more useful ones are described below.

4.2.1

Change of Variables

Many control problems can be posed in the form of a set of nonlinear matrix inequalities; that is, the inequalities are nonlinear in the matrix variables we seek. However by defining new variables it is sometimes possible to ‘linearise’ the nonlinear inequalities, hence making them solvable by LMI methods. Example 4.4. (State feedback control synthesis problem). Consider the problem of finding a matrix F ∈ Rm×n such that the matrix A + BF ∈ Rn×n has all of its eigenvalues in the open left-half complex plane. By the theory of Lyapunov equations (see [20]), this is equivalent to finding a matrix F and a positive definite matrix P ∈ Rn×n such that the following inequality holds (A + BF)T P + P(A + BF) < 0

(4.39)

AT P + PA + F T BT P + PBF < 0

(4.40)

or

Linear Matrix Inequalities in Control

129

This problem is not in LMI form due to the terms which contain products of F and P these terms are ‘nonlinear’ and as there are products of two variables they are said to be ‘bilinear’. If we multiply on either side of equation (4.40) by Q := P−1 (which does not change the definiteness of the expression since rank(P) = rank(Q) = n) we obtain QAT + AQ + QF T BT + BFQ < 0

(4.41)

This is a new matrix inequality in the variables Q > 0 and F, but it is still nonlinear. To rectify this, we simply define a second new variable L = FQ giving QAT + AQ + LT BT + BL < 0

(4.42)

We now have an LMI feasibility problem in the new variables Q > 0 and L ∈ Rm×n . Once this LMI has been solved we can recover a suitable state-feedback matrix as F = LQ−1 and our Lyapunov function as P = Q−1 . Hence, by making a change of variables we have obtained an LMI from a nonlinear matrix inequality. The key facts to consider when making a change of variables is the assurance that the original variables can be recovered and that they are not over-determined.

4.2.2

Congruence Transformation

For a given positive definite matrix Q ∈ Rn×n , we know that, for another real matrix W ∈ Rn×n such that rank(W ) = n, the following inequality holds W QW T > 0

(4.43)

In other words, definiteness of a matrix is invariant under pre and post-multiplication by a full rank real matrix, and its transpose, respectively. The process of transforming Q > 0 into equation (4.43) using a real full rank matrix is called a ‘congruence transformation’. It is very useful for ‘removing’ bilinear terms in matrix inequalities and is often used, in conjunction with a change of variables, to make a bilinear matrix inequality linear. Often W is chosen to have a diagonal structure. Example 4.5. (Making a bilinear matrix inequality linear). Consider T A P + PA PBF +CT V Q= <0 ⋆ −2V

(4.44)

where the matrices P > 0,V > 0 and F (definiteness not specified) are the matrix variables and the remaining matrices are constant. The ⋆ in the bottom left entry of the matrix denotes the term required to make the expression symmetric and will be used frequently hereafter. Notice that this inequality is bilinear in the variables P and F which occur in the (1,2) and (2,1) elements of the matrix Q ∈ R(n+p)×(n+p) . However, if we choose the matrix −1 P 0 W= ∈ R(n+m)×(n+m) (4.45) 0 V −1 which is full rank (rank(W ) = n + m) by virtue of the inverses of P and V (which exist as the matrices are positive definite), then calculating W QW T gives

130

G. Herrmann, M.C . Turner, and I. Postlethwaite

W QW T =

P−1 AT + AP−1 BFV −1 + P−1CT ⋆ −2V −1

<0

(4.46)

Hence, in the new variables X = P−1 , U = V −1 and L = FV −1 we have a linear matrix inequality T XA + AX BL + XCT W QW T = (4.47) ⋆ −2U

Notice that the original variables can be recovered by inverting X and U and computing F = LU −1

4.2.3

Schur Complement

The main use of the Schur complement is to transform quadratic matrix inequalities into linear matrix inequalities, or at least as a step in this direction. Schur’s formula says that the following statements are equivalent: i.

Φ=

Φ11 Φ12 <0 T Φ Φ12 22

ii.

Φ22 < 0 −1 T Φ11 − Φ12 Φ22 Φ12 < 0

A non-strict form involving a Moore-Penrose pseudo inverse also exists if Φ is only negative semi-definite; see [2]. Example 4.6. (Making a quadratic inequality linear). Consider the LQR-type matrix inequality (Riccati inequality) AT P + PA − PBR−1 BT P + Q < 0 (4.48)

where P > 0 is the matrix variable and the other matrices are constant with Q, R > 0. This inequality can be used to minimize the cost function J=

Z ∞ 0

(xT Qx + uT Ru)dt

(4.49)

for the computation of the state feedback controller u = Kx controlling x˙ = Ax + Bu with an initial condition of x(0) = x0 . We know that the solution to this problem is ˜ K = −R−1 BT P, and J = min u

˜ − PBR ˜ −1 BT P˜ + Q = 0 AT P˜ + PA

Z ∞ 0

˜ 0 (xT Qx + uT Ru)dt = x0T Px

However, the solution can be also obtained from a linear matrix inequality problem. Using a congruence transformation for S = P−1 , it follows from (4.48)

Linear Matrix Inequalities in Control SAT + AS − BR−1 BT + SQS < 0

131 (4.50)

We can now define

Φ11 := SAT + AS − BR−1 BT

(4.51)

Φ12 := S

(4.52)

Φ22 := −Q−1

(4.53)

and use the Schur complement identities. Thus, we can transform our Riccati inequality into T SA + AS − BR−1 BT S <0 (4.54) ⋆ −Q−1 In other words, we have transformed a quadratic matrix inequality into a linear matrix inequality. The target is to minimize x0T Px0 subject to (4.54). Alternatively, it is possible to minimize σ subject to σ > x0T Px0 and (4.54). Using the Schur complement again for x0T Px0 − σ < 0, it follows for Φ11 := −σ , Φ12 := x0T and Φ22 := −S: −σ x0T <0 x0 −S Hence, the alternative solution to the optimization problem is given by the following LMIproblem: min σ s.t. −σ x0T <0 x0 −S T SA + AS − BR−1 BT S <0 ⋆ −Q−1 for which the optimal controller is given by K = −R−1 BT S−1 .

4.2.4

The S-Procedure

The S-procedure is essentially a method which enables one to combine several quadratic inequalities into one single inequality (generally with some conservatism). There are many instances in control engineering when we would like to ensure that a single quadratic function of x ∈ Rm is such that F0 (x) ≤ 0

F0 (x) := xT A0 x + 2b0 x + c0

(4.55)

whenever certain other quadratic functions are positive semi-definite i.e., when Fi (x) ≥ 0 Fi (x) := xT Ai x + 2b0 x + c0 ,

i ∈ {1, 2, . . . , q}

(4.56)

To illustrate the S-procedure, consider i = 1, for simplicity. That is, we would like to ensure F0 (x) ≤ 0 for all x such that F1 (x) ≥ 0. Now, if there exists a positive definite (or zero) scalar, τ , such that Faug (x) := F0 (x) + τ F1 (x) ≤ 0 ∀x s.t.F1 (x) ≥ 0 (4.57)

132

G. Herrmann, M.C . Turner, and I. Postlethwaite

it follows that our goal is achieved. To see this, note that Faug (x) ≤ 0 implies that F0 (x) ≤ 0 if τ F1 (x) ≥ 0 because F0 (x) ≤ Faug (x) if F1 (x) ≥ 0. Thus, extending this idea to q inequality constraints we have that F0 (x) ≤ 0 whenever Fi (x) ≥ 0 (4.58) holds if

q

F0 (x) + ∑ τi Fi (x) ≤ 0, i=1

τi ≥ 0

(4.59)

In general the S-procedure is conservative; inequality (4.59) implies inequality (4.58), but not vice versa.2 The usefullness of the S-procedure is in the possibility of including the τi ’s as variables in an LMI problem. Example 4.7. (Combining quadratic constraints to yield an LMI.) An instructive example, taken from [2], Chapter 2, involves finding a matrix variable P > 0 such that T T x A P + PA PB x <0 (4.60) z z BT P 0 whenever x 6= 0 and z satisfy the constraint zT z ≤ xT CT Cx

(4.61)

Note that inequality (4.61) is equivalent to (xT CT Cx − zT z) ≥ 0

(4.62)

T T x x C C 0 ≥0 z z 0 −I

(4.63)

or

The two quadratic constraints (4.60) and (4.63) can thus be combined with the S-procedure to yield the LMI T A P + PA + τ CT C PB <0 (4.64) T B P −τ I in the variables P > 0 and τ ≥ 0.

4.2.5

The Projection Lemma and Finsler’s Lemma

In some types of control problems, particularly those seeking dynamic controllers, we encounter inequalities of the form

Ψ (X) + G(X)Λ H T (X) + H(X)Λ T GT (X) < 0

(4.65)

where X and Λ are the matrix variables and Ψ (.), G(.), H(.) are (normally affine) functions of X but not of Λ . Λ is an unstructured matrix variable i.e., all element values of Λ are not constrained and they are to be considered part of the unstructured matrix variable.3 In [5], it is proved that inequality (4.65) is satisfied, for some X and Λ , if and only if 2 3

Equivalence is only guaranteed for i = 1 (see [2]). It is for instance not permitted to constrain Λ to be sparse, symmetric or diagonal.

Linear Matrix Inequalities in Control (

T Ψ (X)W WG(X) G(X) < 0 T WH(X)Ψ (X)WH(X) < 0

133

(4.66)

where WG(X) and WH(X) are matrices with columns which form bases for the null spaces of G(X) and H(X) respectively. Alternatively, WG(X) and WH(X) are sometimes called orthogonal complements of G(X) and H(X) respectively. Note that WG(X) G(X) = 0 WH(X) H(X) = 0.

(4.67)

The main point of this result (referred to as Gahinet and Apkarian’s projection lemma) is that it enables one to transform a matrix inequality which is a, not necessarily linear, function of two variables, X and Λ , into two inequalities which are just functions of one variable. This has two useful consequences (i) It can facilitate the derivation of an LMI. (ii) There are less variables for computation. In [4], it is also proved that inequality (4.65) was equivalent to two inequalities Ψ (X) − σ G(X)G(X)T < 0 Ψ (X) − σ H(X)H(X)T < 0

(4.68)

for some real σ . In other words, inequalities (4.66) and (4.68) are equivalent. This result is often referred to as Finsler’s Lemma. Example 4.8. (State feedback control synthesis problem revisited). Consider again the state feedback synthesis problem of finding P > 0 and F such that (A + BF)T P + P(A + BF) < 0

(4.69)

Using the change of variables described earlier in Section 4.2.1 we can change this problem into that of finding Q > 0 and L such that QAT + AQ + LT BT + BL < 0 If we choose to eliminate the variable L using the projection lemma we get T WB (AQ + QAT )WB < 0, Q > 0 WIT (AQ + QAT )WI < 0, Q > 0

(4.70)

(4.71)

However, as WI is a matrix whose columns span the null space of the identity matrix which is N (I) = {0} the above equation simply reduces to WBT (AQ + QAT )WB < 0,

Q>0

which is an LMI problem. Alternatively, using Finsler’s Lemma we get AQ + QAT − σ BBT < 0, Q > 0 AQ + QAT − σ I < 0, Q > 0

(4.72)

(4.73)

However, we can neglect the second inequality because if we can find a σ satisfying the first inequality, we can always find one which satisfies the second. Notice that the use of both the projection lemma and Finsler’s Lemma effectively reduces our original LMI problem into two separate ones: the first LMI problem involves the calculation of Q > 0; the second ivolves the back-substitution of Q into the original problem in order for us to find L (and then F). The reader is, however, cautioned against the possibility of illconditioning in this two-step approach. For some problems, normally those with large numbers of variables, X can be poorly conditioned, which can hinder the numerical determination of Λ from equation (4.65).

134

G. Herrmann, M.C . Turner, and I. Postlethwaite

4.3

Examples

4.3.1

Lyapunov Stability for Continuous-Time Systems

The stability of a nonlinear systems is generally more difficult to ascertain than that of a linear systems. A sufficient (but not necessary) condition was given by Lyapunov; see, for example, [7]. Theorem 4.1 (Lyapunov’s Theorem for continuous-time systems). Given an unbounded positive definite function V (x) > 0 ∀x 6= 0 and an autonomous system x˙ = f (x), then the system x˙ = f (x) is asymptotically stable if

∂V V˙ (x) = f (x) < 0 ∂x

∀x 6= 0

(4.74)

For linear systems described by x˙ = Ax this condition (4.74) is necessary and sufficient. A suitable Lyapunov function is V (x) = xT Px,

P > 0.

This implies that the linear system is asymptotically stable if and only if xT PAx + xT AT Px < 0 ∀x 6= 0 This is equivalent to what has was previously said in Example 1.

4.3.2 L2 Gain In linear systems, the H∞ norm is equivalent to the maximum RMS energy gain of the system and is also called the H∞ -gain of the linear system. The equivalent measure for nonlinear systems is the so-called L2 gain, which is a bound on the RMS energy gain. Specifically a nonlinear system with input w(t) and output z(t) (see Figure 4.1) is said to have an L2 gain of γ if (4.75) kzk2 < γ kwk2 + β where β is a positive and k(.)k2 denotes the standard 2-norm-in-time (L2 norm) of a pRconstant ∞ T vector, i.e. kxk2 = t=0 x (t)x(t)dt. Thus the L2 gain of a system can be taken as a measure of the size of its output relative to the size of its input. For a linear system

Fig. 4.1. Nonlinear System

Linear Matrix Inequalities in Control

135

x˙ = Ax + Bw

(4.76)

z = Cx + Dw

(4.77)

this fact can be derived from the LMIs (4.26)-(4.27) as this is equivalent to: " # AT P + PA + 1γ CT C PB + 1γ CT D <0 BT P + 1γ DT C −γ I + 1γ DT D

(4.78)

T for P > 0. Writing this matrix inequality in terms of the vector xT wT , it follows that we need to find the minimum of γ so that # T " T h iT A P + PA + 1γ CT C PB + 1γ CT D x x < 0, ∀ xT wT 6= 0 (4.79) 1 1 T T T w w B P + γ D C −γ I + γ D D or 1 1 xT AT Px + xT PAx + xT CT Cx + xT (PB + CT D)w γ γ 1 1 +wT (BT P + DT C)x + wT DT Dw − γ wT w γ γ 1 = xT AT Px + xT PAT x + 2xT PBw + zT z − γ wT w < 0 γ

(4.80)

Defining now V = xT Px, it is easily derived from V˙ = xT AT Px + xT PAx + 2xT PBw and (4.80) that: 1 V˙ + zT z − γ wT w < 0 γ Integration in the interval [0, ∞) implies V (t = ∞) −V (t = 0) +

Z ∞ Z ∞ 1 T z (s)z(s)ds − γ wT (s)w(s)ds < 0 t=0

γ

t=0

Shifting terms, taking the square root and using the triangle inequality, it is easily shown that γ is indeed the L2 -gain of the linear system (4.76) as (4.75) follows.

4.3.3

Lyapunov Stability for Discrete-Time Systems

Discrete-time systems are often treated in less detail than continuous-time systems. However, digital technology for controller implementation makes discrete controller design and analysis a pertinent subject of interest. As for Lyapunov stability of continuous-time systems a similar stability theorem can be stated. Theorem 4.2 (Lyapunov’s Theorem for discrete systems). Given an unbounded positive definite function V (x) > 0 ∀x 6= 0 and an autonomous system x(k + 1) = f (x(k)), then the system x(k + 1) = f (x(k)) is asymptotically stable if

∆ V (x(k + 1)) = V (x(k + 1)) −V (x(k)) < 0 ∀x 6= 0

(4.81)

136

G. Herrmann, M.C . Turner, and I. Postlethwaite

As before, for linear systems described by x(k + 1) = Ax(k) this condition (4.81) is necessary and sufficient and a Lyapunov function is given by V (x) = xT Px,

P > 0.

Thus, the linear system is asymptotically stable if and only if xT AT PAx − xT Px < 0 ∀x 6= 0 or AT PA − P < 0.

4.3.4 l2 Gain Similar to a continuous-time system, the computation of the l2 -gain can be very helpful to the design and analysis of discrete-time control systems. As for continuous time (4.75), a discrete-time system with input w and output z has a finite l2 -gain, γ , if kzk2 < γ kwk2 + β

(4.82)

The significant difference is the definition of the l2 -norm k(.)k2 which is now given by an q T infinite sum kxk2 = ∑∞ k=0 x (k)x(k). For linear systems x(k + 1) = Ax(k) + Bw(k)

(4.83)

y(k) = Cx(k) + Dw(k)

(4.84)

the value of the finite l2 -gain, γ , is the same as the H∞ norm or the maximum RMS energy gain of the system. The computation of the l2 -gain is easily achieved by considering the following matrix inequality problem: "

min γ AT PA − P + 1γ CT C BT PA + 1γ DT C

s.t. #

AT PB + 1γ CT D −γ I + BT PB + 1γ DT D

(4.85) <0

(4.86)

−P < 0

(4.87)

for P > 0. This readily follows from the analysis of 1 ∆ V (x(k + 1)) + yT (k)y(k) − γ wT (k)w(k) < 0 γ

(4.88)

for a Lyapunov function V = xT Px. To complete the analysis for the l2 -gain, the sum of the inequality of (4.88) over k in the interval [0, ∞) is considered (rather than the integral as in Section 4.3.2). The problem with matrix inequality (4.86) is that it is not linear in γ . This is easily amended by using the Schur complement so that the l2 -gain computation can be equivalently expressed in a convex formulation:

Linear Matrix Inequalities in Control

137

min γ s.t.  AT PA − P AT PB CT T T T  B PA −γ I + B PB D  < 0 C D −γ I −P < 0

(4.89)



(4.90) (4.91)

The matrix P > 0 and the scalar γ > 0 are variables of this LMI-problem. Another approach of creating a linear matrix inequality from (4.86) is using the change of variable approach from Section 4.2.1. The first step is to multiply all elements of (4.86) by γ so that

min γ s.t. γ AT PA − γ P +CT C γ AT PB +CT D <0 γ BT PA + DT C −γ 2 I + γ BT PB + DT D −P < 0

(4.92) (4.93) (4.94)

Defining now the new variables Q = Pγ and µ = γ 2 allows us to formulate the equivalent optimization problem:

min µ AT QA − Q +CT C BT QA + DT C

AT QB +CT D

s.t.

(4.95)

<0 −µ I + BT QB + DT D −Q < 0

(4.96) (4.97)

which is clearly convex. Now the matrix Q > 0 and the scalar γ > 0 are variables of the LMI√ problem and the l2 -gain is readily computed with γ = µ .

4.3.5

Sector Boundedness

The saturation function is defined as sat(u) = [sat1 (u1 ), . . . , sat2 (um )]T

(4.98)

and sati (ui ) = sign(ui ) × min {|ui |, u¯i } , u¯i > 0 ∀i ∈ {1, . . . , m}, where u¯i is the i’th saturation limit. From this, the deadzone function can be defined as Dz(u) = u − sat(u)

(4.99)

It is easy to verify that the saturation function, sati (ui ) satisfies the following inequality ui sati (ui ) ≥ sat2i (ui )

(4.100)

or sati (ui )[ui − sati (ui )]wi ≥ 0 for some wi > 0. Collecting this inequality for all i we can write sat(u)T W [u − sat(u)] ≥ 0

(4.101) (4.102)

for some diagonal W > 0. Similarly, it follows that Dz(u)T W [u − Dz(u)] ≥ 0

(4.103)

for some diagnal W > 0. We will make use of this inequality in the next section when computing the L2 -gain of a linear system with input and output constraints.

138

G. Herrmann, M.C . Turner, and I. Postlethwaite

4.3.6

A Slightly More Detailed Example

Suppose we consider the computation of the L2 -gain for the linear SISO system with saturated input signal u: x˙ = Ax + bsat(u), x ∈ Rn (4.104) which is subject to a limitation in the measurement range of the output y (see Figure 4.2): y = sat(cx + dsat(u)).

(4.105)

Fig. 4.2. Linear SISO System with input/output saturation

This is usually the case for many practically controlled systems. For instance actuator inputs u can be constrained due to mechanical limits (e.g., valves) or due to digital-to-analogue converter voltage signal limits, while sensor signals are possibly constrained due to sensor voltage range limits or simply by analogue-to-digital converter limits. Hence, the analysis of such systems is vital to practical control systems and will be pursued in greater detail in a later chapter. An upper bound of theL2 -gain of this system can be computed by considering sector constraints for the saturation functions. Furthermore, the Projection Lemma will be used to show that the L2 -gain of (4.104)–(4.105) is in actual fact not larger than the L2 -gain of the system without any of the saturation limits. For analysis, we may define s = sat(u). Hence, similar to (4.103), it follows that sw1 (u − s) ≥ 0,

(4.106)

where w1 > 0. In the same way, it can be shown that for the output signal y we have: yw2 (cx + ds − y) ≥ 0

(4.107)

for an arbitrary diagonal matrix w2 > 0. We know from Section 4.3.2 that from 1 V˙ + y2 − γ u2 ≤ 0 γ it follows that our system of interest (4.104-4.105) has the L2 -gain γ . However, in our case, this analysis is conducted under the condition (4.106–4.107). Hence, from the S-procedure of Section 4.2.4, it follows that 1 V˙ + y2 − γ u2 + 2sw1 (u − s) + 2yw2 (cx + ds − y) < 0 γ

(4.108)

Linear Matrix Inequalities in Control

139

for xT s u y 6= 0 implies similarly an L2 -gain of γ . This is also easily derived when integrating the inequality over the interval [0, ∞] above and considering a similar procedure as in Section 4.3.2 under the assumptions of (4.106–4.107). By computing the expression for V˙ , it follows that: xT AT Px + xPAx + xT Pbs + sbT Px 1 + y2 − γ u2 + 2sw1 (u − s) + 2yw2 (cx + ds − y) < 0 γ Rewriting this inequality, it is shown that:  T  T A P + PA Pb x  s   bT P −2w1    u  0 w1 w2 c w2 d y

for

(4.109)

  0 cT w2 x s  w1 dw2    < 0  u  0 −γ 1 0 −2w2 + γ y

6 0. xT s u y =

This is equivalent to the matrix inequality of  T A P + PA Pb  bT P −2w1   0 w1 w2 c w2 d

 0 cT w2 w1 dw2   < 0.  0 −γ 1 0 −2w2 + γ

(4.110)

The target is now to minimize the value of γ > 0 to find the L2 -gain of our system of interest (4.104–4.105). However, the variable γ appears twice in the matrix inequality, once in the denominator of 1γ . This results in a non-linear and non-convex matrix inequality and it is not easily possible to avoid this non-linearity. Furthermore, this matrix inequality looks rather complicated as the positive definite matrix P, γ , w1 and w2 are variables. The Projection Lemma allows us to derive a significantly simpler matrix inequality which delivers the L2 -gain of our system. The first step is to rewrite the matrix inequality of (4.110) in the same format as for (4.65). Hence,   T A P + PA Pb 0 cT w2 T  b P 0 0 dw2      0 0 0 −γ w2 c w2 d 0 −2w2 + 1γ     0 0 1  −1     + (4.111)  0  w1 0 −1 1 0 +  1  w1 0 1 0 0 < 0. 0 0

Defining the matrices  T     A P + PA 0 0  bT P 1  −1      g1 =   0  , h1 =  1  , Ψ1 =  0 w2 c 0 0

allows us to write

Pb 0 0 w2 d

 0 cT w2 0 dw2  ,  −γ 0 1 0 −2w2 + γ

140

G. Herrmann, M.C . Turner, and I. Postlethwaite

Ψ1 + g1 w1 hT1 + h1 wT1 gT1 < 0.

(4.112) T Thus, the null space matrices Wg1 and Wh1 for g1 and h1 need to be found so that Wg1 g1 and h i WhT1 h1 are full rank and Wg1 g1 = 0, Wh1 h1 = 0. This implies: 

   I 000 I 000 Wg1 =  0 0 1 0  , Wh1 =  0 1 1 0  0001 0001

Hence, it follows from (4.66) that both  T  A P + PA 0 cT w2 <0 0 −γ 0 Wg1 Ψ1WgT1 =  w2 c 0 −2w2 + 1γ

and

Wh1 Ψ1WhT1



 AT P + PA Pb cT w2 T −γ dw2  < 0 = b P w2 c w2 d −2w2 + 1γ

have to be satisfied for (4.110) to be true. An extra iteration using the Projection Lemma for w2 as the variable implies that if Wh1 Ψ1WhT1 < 0 then we also have Wg1 Ψ1WgT1 < 0. Hence, we can focus our investigation on Wh1 Ψ1WhT1 only: Wh1 Ψ1WhT1 = Ψ2 + g2 w2 hT2 + h2 w2 gT2 where

(4.113)

   T   T  A P + PA Pb 0 0 c g2 =  0  , h2 =  d  , Ψ2 =  bT P −γ 0  0 0 1γ 1 −1

This allows us to derive the null space matrices Wg2 and Wh2 for g2 and h2 I 00 I 0 cT Wg2 = , Wh2 = 010 01 d so that Wg2 Ψ2WgT2

" # T T AT P + PA + cγ c Pb + c γ d AT P + PA Pb T , Wh2 Ψ2Wh2 = = . 2 bT P −γ bT P + dc −γ + dγ γ

As before we can easily see that Wg2 Ψ2WgT2 < 0 is always satisfied if Wh2 Ψ2WhT2 < 0. Hence, the matrix inequality " # T T AT P + PA + cγ c Pb + c γ d <0 (4.114) 2 bT P + dc −γ + dγ γ is satisfied for P > 0 if and only if (4.110) holds. From (4.78), the L2 -gain of the non-linear system (4.104)-(4.105) is exactly the same as for the linear system defined by the system quadruple (A, b, c, d). The computation of the L2 -gain follows the approach shown in Example 2. This result is easily understood by the following argument. For small input signals of u, the non-linearities in the system of (4.104)–(4.105) are not active. The system acts as a linear

Linear Matrix Inequalities in Control

141

system for which the L2 -gain has been found in Section 4.3.2. When the signal u reaches the saturation limits, the linear system is non-linearly constrained and the output may also be nonlinearly limited. Hence, the level of the L2 -gain for large amplitude signals must decrease. Thus, the maximum achievable L2 -gain is indeed given only by the linear component of (4.104)-(4.105).

4.4

Summary

This chapter has introduced some important tools for the analysis of closed-loop control systems via linear matrix inequalities. It has been shown that quadratic performance constraints, such as L2 -gain computation, can be equivalently encoded via linear matrix inequalities. Furthermore, we have shown that non-linearities can be considered using sector bounds. By posing control problems in terms of linear matrix inequalities a highly flexible design and anaiysis tool is obtained which allows the combination of different control requirements. wellestablished numerical solvers for semi-definite programming allow the efficient solution of these convex linear matrix inequality problems.

References 1. M. Athans. Optimal control : an introduction to the theory and its applications. McGrawHill, New York, London, 1966. 2. S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear Matrix Inequalities in System and Control Theory. Society for Industrial and Applied Mathematics, 1994. 3. J. C. Doyle, K. Glover, P. P. Khargonekar, and B. A. Francis. State-space solutions to standard H2 and H∞ control problems. IEEE Transactions on Automatic Control, 34(8):831– 847, 1989. 4. P. Finsler. Uber das Vorkommen definiter und semi-definiter Formen in Scharen quadratischer Formen. Comentarii Mathematica Helvetici, 9:192–199, 1937. 5. P. Gahinet and P. Apkarian. A linear matrix inequality approach to H∞ control. International Journal of Robust and Nonlinear Control, 4:421–448, 1994. 6. P. Gahinet, A. Nemirovski, A.J. Laub, and M. Chilali. LMI Control Toolbox. The MathWorks Inc., 1995. 7. H.K. Khalil. Nonlinear Systems. Prentice Hall, New Jersey, 1996. 8. H. Kwakernaak and R. Sivan. Linear Optimal Control Systems. Wiley-Interscience, New York, 1972. 9. F. L. Lewis. Optimal control. John Wiley and Sons, New York, 1986. 10. P. Moylan and D. Hill. Stability criteria for large-scale systems. IEEE Transactions on Automatic Control, 23:143–149, 1978. 11. Y. Nesterov and A. Nemirovskii. Interior-Point Polynomial Algorithms in Convex Programming. Studies in Applied Mathematics. SIAM, Philadelphia, 1993. 12. S. Skogestad and I. Postlethwaite. Multivariable Feedback Control: Analysis and Design. Wiley, Chichester, UK, 2nd edition, 2005. 13. C. W. Scherer. Mixed H2 /H∞ control for time-varying and linear parametrically-varying systems. Intenational Journal of Robust and Nonlinear Control, 6:929–952, 1996.

142

G. Herrmann, M.C . Turner, and I. Postlethwaite

14. C. W. Scherer, P. Gahinet, and M. Chilali. Multi-objective output-feedback control via lmi optimization. IEEE Transactions on Automatic Control, 42:896–911, 1997. 15. M.C. Turner, G. Herrmann, and I. Postlethwaite. Accounting for uncertainty in antiwindup synthesis. submitted, 2003. 16. M.C. Turner, G. Herrmann, and I. Postlethwaite. An introduction to linear matrix inequalities in control. University of Leicester Department of Engineering Techincal Report no 02-04, 2004. 17. A. van der Schaft. L2-Gain and Passivity Techniques in Nonlinear Control. Communications and Control Engineering Series. Springer-Verlag, Berlin, 2nd edition, 2000. 18. L. Vandenberghe and S. Boyd. Semidefinite programming. SIAM Review, 38:49–95, 1996. 19. J. C. Willems. Least stationary optimal control and the algebraic riccati equation. IEEE Transactions on Automatic Control, AC-16(6):621–634, 1971. 20. K. Zhou, J.C. Doyle, and K. Glover. Robust and Optimal Control. Prentice Hall, New Jersey, 1996.

5 Anti-w indup Compensation and the Control of Input-Constrained Systems Matthew C. Turner, Guido Herrmann , and Ian Postlethwaite

Summary. This chapter introduces the problem of input saturation in systems which are otherwise linear. It is demonstrated how input saturation may cause such systems to exhibit undesirable behaviour. The notion of anti-windup compensators is introduced, followed by a detailed description of a typical anti-windup problem and an effective design procedure. The chapter also emphasizes the role of various nonlinear stability and performance concepts such as Lyapunov stability, sector-boundedness and L2 gain.

Key words: anti-windup, constrained control, linear matrix inequalities

5.1 Introduction 5.1.1

Input Constraints in Control Systems

Control engineers, where possible, like to work under the assumption of linearity. The mathematics associated with the field of linear systems is well developed and underpins much of that which undergraduates learn about control theory, and also much of the control theory which is applied in industry. Even nonlinear techniques often attempt to “generalise” linear concepts, and frequently nonlinear systems are linearised to obtain linear models which locally yield good engineering approximations. The problem with the assumption of linearity is that it is sometimes unrealistic and can lead to erroneous results. Although there are many instances in which this is the case, this chapter will examine one important such instance: the problem of input saturation. Input saturation is prevalent in one form or another in virtually all engineering systems. Effectively it arises due to physical limitations on a control system’s actuators. It is easy to cite many instances of actuator saturation in virtually all areas of applied control. For instance, in the control of electric motors the voltage input is strictly limited by the rail voltage, to something like ±5, ±12 or ±24 volts. Similarly in the control of aeroplanes, the elevator which largely determines the pitch of the aircraft can only be deflected by a certain amout, say between ±20 degrees M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 143-174, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

144

M.C. Turner, G. Herrmann , and I. Postlethwaite

from its centre. A particularly good example of a (non)-symmetric input constraint occurs in process control where valves can only vary between closed and fully open i.e., this restricts fluid flow to between zero and 100 %. The valve cannot allow negative fluid flow or a flow of greater than 100 %. One of the reasons why saturation is so difficult to handle is that although its small signal behaviour is effectively equal to that of the identity operator, its large signal behaviour is complicated is “difficult to get rid of” - it cannot be linearised or inverted. We will return to this subject in more detail, but it is essentially attributable to the non-smooth nature of the function and the lack of bijectivity in the saturation operator. So we now know that input saturation is a feature common to most control systems, we know it is nonlinear and we know that it is difficult to wave a magic wand and make it disappear. What are the implications of the above? Well, the main problem with saturation is that even if we consider a linear plant, G(s), with input saturation and we assume that a linear controller, K(s), has been designed such that in the absence of saturation the system behaves very well, it is often the case that when saturation becomes active, it can seriously degrade the system’s performance. This performance degradation can occur even if the nominal controller, K(s), was designed such that the nominal (without saturation) closed loop had large robustness margins. This degradation of performance due to saturation has been known for a long time. Teel traces the history of saturation control back to Lozier ( [17]), although it is probable that its effects were known, but probably not well-documented, before this. It is certainly the case that engineers involved in the design and commissioning of PI controllers became aware of phenomena called “wind-up” (or “windup”) many years ago. In a typical wind-up situation when control saturation occurs, an error persists for longer than would be expected in a purely linear system. This forces the integrator state to achieve a value in excess of that reached in linear operation, and then, when control saturation ceases, the “excess energy” in the integrator state must be dissipated into the system. This dissipation of energy often manifests itself as a large overshoot before steady state tracking has begun and sometimes leads to the system’s actuators staying at the saturation level for long periods of time - sometimes referred to as actuator “lock-up”. See Figure 5.1. However, although such a problem can be a nuisance in industrial applications and certainly lowers the performance from what was predicted by linear design, this classical “windup” effect is rather benign compared to other pernicious effects of input saturation. There are many examples of saturation problems but perhaps the most notorious are those associated with so-called pilot-induced-oscillations (PIO’s). These saturation-induced events have led to the crashes of several aircraft (the SAAB Grippen, the Boeing V22 Osprey) and several nearmisses with others. Also, saturation has been linked with the Chernobyl nuclear power plant meltdown [23]. Thus the presence of saturation can lead to performance degradation from the mild to the severe and can also lead to loss of stability. Although this is not always critical, it is clear that some way of predicting the effects of saturation is required and, moreover, that some method of limiting the degradation that occurs is warranted.

5.1.2

Constrained System Description

Let us re-visit our linear plant G(s) which we shall assume is a MIMO system relating a nominal control input u(t) ∈ Rm to a measured output y(t) ∈ R p in the usual manner. To keep

Anti-windup Compensation and the Control of Input-Constrained Systems

8

3.5

3

6

2.5

Output [non-dimensional]

Output [non-dimensional]

145

Large overshoot 2

1.5

1

Excessive rise time

0.5

4

Acuator "lock−up"

2

0

−2

0

−4 −0.5

−6

−1

−1.5

0

5

10

15

Time [sec]

20

−8

0

5

10

15

Time [sec]

20

Fig. 5.1. Illustration of the archetypal windup phenomena (dashed, linear; solid, saturated)

it simple we shall not distinguish between a system’s transfer function and its time-domain linear operator explicitly i.e. y(s) = G(s)u(s) (5.1) where y(s), u(s) are the Laplace transforms of y(t), u(t), is entirely equivalent to y(t) = Gu(t)

(5.2)

where G is the time-domain linear convolution operator. Now, let us consider the constrained case, that is when y(t) = Gum (t) (5.3) where um (t) ∈ Rm may not be equal to u(t) as it is fed through the saturation element um (t) = satu¯ (u(t)) where the saturation operator satu¯ (.) is described by   sat1,u¯1 (u1 )   .. satu¯ (u) =   . satm,u¯m (um )

where each individual saturation component is described by the equation   u¯i ui > u¯i u |ui | ≤ u¯i sati,u¯i (ui ) =  i −u¯i ui < −u¯i

(5.4)

(5.5)

(5.6)

146

M.C. Turner, G. Herrmann , and I. Postlethwaite

Note that each i’th saturation function depends on two parameters: the i’th control signal, ui (t), which will vary with time and the i’th saturation-limit, u¯i > 0, which determines the maximum magnitude of the output of the saturation, i.e., |sati,u¯i (ui )| ≤ u¯i ∀ui (t). A more succinct way of expressing equation (5.6) is sati,u¯i (ui ) = sign(ui ) min(|ui |, u¯i )

(5.7)

Note also that this description of the saturation function is symmetric; both upper and lower limits are the same. Non-symmetric saturation functions can be considered with little extra difficulty (in fact this is trivial) but they complicate the notation so we do not consider them here. In fact, to simplify things further, for the remainder of the chapter we drop the subscript u¯ and simply write sat(u) to indicate saturation of a vector u(t). An interesting feature of the saturation function defined as such is that if we construct a hyperbox defined as U = [−u¯1 , u¯1 ] × [−u¯2 , u¯2 ] × . . . × [−u¯m , u¯m ] ⊂ Rm (5.8) we can see that if the control signal stays within this box, that is u(t) ∈ U ∀t, then sat(u) = u i.e., the saturation function behaves as the identity. This is an obvious fact as it essentially says that providing that our control signal is sufficiently small, we do not have any saturation problems. However, it gives us a convenient mathematical way of capturing this feature.

5.1.3

Constrained Control and Anti-windup

r(t)

κ( r,y,t )

G(s)

y(t)

u(t)

Fig. 5.2. Architecture for “one-shot” constrained control design

Let us now consider the control problem. We want to control the system y = Gsat(u)

(5.9)

using the control signal u(t), such that the system is stable and performs well. For nonlinear systems such as the one above, the words “stable” and “performance” are not as precise as for linear sytems. However it suffices to say that we want non-oscillatory and non-divergent behaviour and we want to make sure that some performance objectives are met. Given these goals, one is very tempted to attempt to design a controller - be it linear or nonlinear - directly. That is we could design a controller u = K (r, y,t)

(5.10)

where K (., ., .) is a nonlinear operator which closes the loop using y(t) and also takes information from the reference signal, r(t) ∈ Rnr . This approach is shown in Figure 5.2 and for

Anti-windup Compensation and the Control of Input-Constrained Systems

147

want of a better name we call this the constrained control approach. This seems a very obvious approach to take and indeed in some cases it may be appropriate. However, in this chapter we are going to argue for a different approach: a two-step design procedure characterised by the following two steps. 1. Design a linear controller u = K[r′ y′ ]′ such that for all u(t) ∈ U our closed loop system is stable and behaves in a desirable manner, meeting linear design criteria. (Note that when u(t) ∈ U , we have linear behaviour). 2. Design an extra (normally linear) element, Θ , such that when u(t) ∈ / U , Θ becomes active and improves the performance and stability properties of the system. So what the above procedure is saying is (i) design a linear controller to yield good small signal behaviour and (ii) design an extra element to improve the large signal behaviour of the system when saturation occurs. At this stage, the second part of our two-step design procedure has been left deliberately vague; this will be fleshed out shortly. The two-step design procedure has become known as the anti-windup (AW) approach to the control of constrained systems. Although something of a misnomer, due to its links with integrator windup, the term “antiwindup” is now associated with a two-step design procedure for coping with input saturation. There is still some disagreement in the literature on the minutae of the definition, but generally the two-step approach outlined above can be considered an anti-windup approach. A schematic of the anti-windup approach is shown in Figure 5.3. Here Θ represents the AW compensator which becomes active when sat(u) − u 6= 0 (i.e., u ∈ / U ), and injects two signals, θ1 ∈ Rm into p the control signal and θ2 ∈ R into the input of the controller. We now have two approaches

d r K(s)

+ −

u

um

G(s)

y

θ1 + − + +

θ2

Θ1 Θ= Θ 2

Fig. 5.3. Architecture for anti-windup compensator approach

to controller design for constrained systems which can be “in-formalised” as follows.

INFORMAL CONSTRAINED CONTROL PROBLEM Given the linear plant G(s), design a controller K such that the system in Figure 1 is stable and yields desirable performance for all u(t).

M.C. Turner, G. Herrmann , and I. Postlethwaite

148

INFORMAL ANTI-WINDUP CONTROL PROBLEM Given the linear plant G(s) and the linear controller K(s), design an anti-windup compensator Θ such that (i) when u(t) ∈ U ∀t, Θ (s) is not active and when u(t) ∈ / U ∀t, the system in Figure 2 is stable and yields performance as close to nominal as possible. Again note that these are only informal problems; the precise nature of the stability and performance objectives we seek has not yet been characterised. The key observation to make is that the constrained control problem does not make any requirements on small signal performance, whereas the anti-windup problem imposes a structure on the small signal performance through the local controller K(s) (compare Figures 5.2 and 5.3). Therefore the anti-windup problem may be conceptually more difficult than the constrained problem as we require our overall control strategy to coincide with a local (linear) control strategy providing saturation does not occur. Despite this apparent difficulty there are a number of features which make the anti-windup approach more appealing than the constrained control approach. Some of these are listed below: •

•

•

•

Given a nominal linear controller K(s), small signal performance is not altered by the anti-windup compensator Θ . This is particularly important if nominal linear performance is desirable and saturation problems occur only occasionally. With a good anti-windup compensator, linear performance would then be disrupted rarely. Linear controller design can be done independently of the anti-windup compensator design. Thus there is no restriction in the linear controller design, making this step fairly easy to accomplish. Anti-windup compensators can be introduced into control loops once a saturation problem is identified. Thus they can be introduced into systems which have legacy controllers which are known to function well in most situations and only encounter difficulties during saturation. Thus anti-windup compensators can be useful in industry where much time and effort has been used to assess a given controller and the anti-windup compensator can be used to improve the system’s performance only where necessary without requiring a complete controller re-design. Conceptually, it is appealing to think of designing the nominal controller for linear performance and then an additional control element (the anti-windup compensator) for nonlinear stability and performance.

For these reasons, most industrial control systems which have saturation alleviating properties have an anti-windup like structure and indeed, the controller design is typically carried out in the two-step anti-windup procedure. It is worth mentioning that recent work on one-shot constrained controller design has made very good progress in recent years and, in particular, the constraint handling abilities of model predictive control (MPC) make it a very good candidate for the constrained control problem. The interested reader is urged to consult [18] for details of this. For the remainder of the chapter, we focus on the anti-windup approach.

5.2

Problems Due to Saturation

Before we look into how we might alleviate saturation problems in systems which are otherwise linear, it makes sense to have some idea of the types of instances in which problematic

Anti-windup Compensation and the Control of Input-Constrained Systems

149

behaviour might occur. Although the nonlinear behaviour of the saturation function makes this difficult, it turns out that some indication of performance degradation can be inferred by looking at the dynamics of the nominal linear system (both linear plant and linear controller). First we look at some hints from classical control and then we move on to more formal statements.

5.2.1 Clues from Classical Control As saturation may be responsible for some very peculiar behaviour we shall start with very simple systems and consider how saturation may contribute to performance degradation during saturation. It therefore makes sense to start with single-input-single-output systems of the form in equation (5.9) but where the control and measured signals are purely scalar, viz u(t) ∈ R and y(t) ∈ R. In this case we can simplify our saturation to um (t) = sat(u) ∈ R Note that if we consider the input to the saturation block and the output of the saturation block, we can describe a nonlinear “gain” as the ratio of the output to the input. This is defined as  u¯ u > u¯ sat(u)  u 1 |u| ≤ u¯ β (u) = = (5.11)  u¯ u u < −u¯ −u sat(u)

sat(u)

Note that u ∈ (0, 1) when u > u¯ and likewise − u ∈ (0, 1) when u < −u. ¯ Therefore the “gain” of the saturation has the very interesting property that

β (u) ∈ (0, 1]

(5.12)

So it can be thought of as a function which represents a loop gain in the system. When β (u) = 1, and thus the saturation behaves as the identity, nominal performance is preserved. When β (u) < 1 and saturation occurs, then the system loses gain and behaves in a more “open-loop” manner. Finally when the gain β (u) ≈ 0 we can expect behaviour similar to open-loop. This extra gain term is depicted in Figure 5.4. The classical root-locus approach to control system

r(t)

β( u)

K(s) u(t)

um(t)

G(s)

y(t)

Fig. 5.4. [Overly] Simple approach to modelling saturation design tells us how the closed-loop poles move as the loop gain is varied. We know that the closed-loop poles start off (at zero gain) on the open-loop poles and then migrate along the root-locus to the open-loop zeros as gain is increased. At infinite gain, the closed-loop poles would sit on the open-loop zeros. The root-locus can therefore also be used to explain, in an approximate fashion, how the closed-loop poles move for varying amounts of saturation. We have the following interpretation: first for no saturation the closed-loop poles of the saturated system sit on the closed-loop poles of the nominal linear system. When saturation starts and as it gets more severe the poles move off the poles of the nominal closed-loop and head back

150

M.C. Turner, G. Herrmann , and I. Postlethwaite

down the root-locus towards the open-loop poles. For very severe saturation, the poles of the saturated system effectively sit on the poles of the open-loop system. Some examples where the root-locus allows us to anticipate poor behaviour are shown overleaf. From the plots overleaf, it is plausible to suggest that we can compile several properties about single-loop saturated systems just from looking at the root locus. To aid us in this analysis we define the term spec to represent the poles of a transfer function (or equivalently the eigenvalues of a system A-matrix) and we define the open-left half complex plant, the imaginary axis and the open-right half complex plane respectively as C− , C0 and C+ . We define the open-loop transfer function L(s) = G(s)K(s) and draw the following conclusions (guidance): • •

•

If a system L(s) is such that at least one pole is in C +, then for sufficiently large state, reference or disturbance, saturation will always cause stability problems (divergence). If a system L(s) is such that spec(L(s)) ∈ C− but part of the root locus between the nominal closed-loop poles and the open-loop poles is in C+ , saturation will cause stability problems (limit cycling) for sufficiently large states, references or disturbances. If a system L(s) is such that spec(L(s)) ∈ C− but the part of the root locus between the nominal closed-loop poles and the open-loop poles is in a region corresponding to light damping, one can expect oscillatory responses for certain references or disturbances.

The root-locus gives us a guide to anticipating problems with saturation given a linear plant G(s) and a linear controller K(s). However it must be remembered that as we are dealing with a nonlinear problem, this only gives us a guide as we are essentially treating the saturation function as a linear gain, which is not mathematically correct. Also, as saturated systems are very “input-dependent”, predictions made on the basis of the root locus may only apply to certain classses of reference or disturbance inputs. Furthermore, it must be remembered that the root-locus only gives us information about SISO systems. If we now move to the case where again u(t) ∈ Rm and y(t) ∈ R p , it is more difficult to predict a system’s behaviour accurately. This is particularly true in systems which feature strong coupling. Recall that in multi-input multi-output (MIMO) systems the control signals are vectors and hence have directions associated with them. Saturation may not only limit the magnitude of this vector but can also alter the direction of the control vector causing closedloops which, in the absence of saturation were decoupled, to loose their decoupling. Figure 5.5 shows an example of this change in direction. u2 _ u2

un−saturated control vector saturated control vector

_ u1

_ −u 1

u1

_ −u 2

Fig. 5.5. Control direction corruption by saturation

Anti-windup Compensation and the Control of Input-Constrained Systems

ROOT-LOCUS AND WINDUP SUSCEPTABILITY

Conditionally stable system Root Locus

6

5

40

4

30

4

3 20

Output [non−dimensional]

Imaginary Axis

Closed−loop poles 1

Gain = 0.042

0

−1

Control [non−dimensional]

2

2

0

−2

−2

10

0

−10

"bang−bang" actuator switching

−4 −20

−3 High−amplitude limit cycle

−6

−30

−4

−5 −10

−8

−6

−4 Real Axis

−2

0

−8

2

0

5

10 Time [sec]

15

−40

20

0

5

10 Time [sec]

15

20

System root-locus and corresponding responses with (−) and without (−−) saturation. System with lightly damped plant poles Root Locus 2.5

4

25

20

3 2

15

2

Closed−loop poles

0

−1

Output [non−dimensional]

Output [non−dimensional]

1

Imaginary Axis

1.5

1

0.5

10

5

0

−5

−2 −10

0

−3

−15

−4 −6

−5

−4

−3 Real Axis

−2

−1

−0.5

0

0

5

10 Time [sec]

15

20

−20

0

5

10 Time [sec]

15

20

System root-locus and corresponding responses with (−) and without (−−) saturation. System with one unstable plant pole Root Locus

Output [non-dimensional]

0.8

3

Output [non-dimensional]

14000

1

12000

0.6

Imaginary Axis

0.4

0.2

0

−0.2

Critical gain point −0.4

−0.6

−0.8

−1 −8

Divergent instability

10000

−7

−6

−5

−4

−3 Real Axis

−2

−1

0

1

8000

6000

1

0

−1

4000

−2

2000

0

2

Actuator lock−up

−3

0

5

10

15

Time [sec]

20

−4

0

5

10

15

20

Time [sec]

System root-locus and corresponding responses with (−) and without (−−) saturation

151

152

M.C. Turner, G. Herrmann , and I. Postlethwaite

The conclusion of this section is that although classical control can give us a useful guide to predicting saturation problems, it does not tell us the full story and we need something more powerful to guarantee stability and performance. In fact some of the predictions made by classical control turn out to be true (e.g., if G(s) has unstable poles) but this cannot be verified by classical control and we must turn to more systematic methods.

5.3

Stability of Systems with Input Saturation

The very first thing which it is desirable to do for systems with input saturation is to establish whether they are stable. Again, remember that we are considering nonlinear systems so that this may not be a trivial task and that most stability conditions will be necessary or sufficient but not both. For instance it can be verfied that G(s) having no poles in C+ is a necessary condition for global asympotic stability of any closed-loop containing it and input saturation, but it is far from sufficient. Converesely, it is easy to prove that a system of the form M(s) =

α s+α

is always stable when put in a unity negative feedback loop with saturation1 , but this is hardly a necessary condition. This chapter will introduce some techniques for establishing stability of linear systems interconnected with saturation elements. The techniques will only provide sufficient conditions for stability and will be expressed as linear matrix inequalities.

5.3.1

Definitions of Stability

Throughout this work we try to keep things as simple as possible and so we will confine our attention to the minimum mathematical baggage possible. However, we shall introduce concepts of stability in a fairly general but brief manner in the interests of rigour and clarity. First we consider zero input stability of state-space systems i.e., the stability of autonomous systems with no forcing term. Definition 5.1. Consider the system x˙ = f (x)

(5.13)

then we say the origin of the system is i) Globally asymptotically stable (GAS) if limt→∞ x(t) = 0 for all x ∈ Rn ii) Globally exponentially stable (GES) if there exist scalars, η1 > 0 and η2 > 0 such that kx(t)k ≤ η1 exp[−η2t]kx(0)k for all x ∈ Rn iii) Locally asymptotically stable (LAS) if limt→∞ x(t) = 0 for all x(t = 0) ∈ X ⊂ Rn and 0 ∈ intX . 1

This follows as both the saturation and M(s) are passive.

Anti-windup Compensation and the Control of Input-Constrained Systems

153

iv) Locally exponentially stable (LES) if there exist scalars, η1 > 0 and η2 > 0 such that kx(t)k ≤ η1 exp[−η2t]kx(0)k for all x(t = 0) ∈ X

⊂ Rn

and 0 ∈ intX .

Note that exponential stability is a stronger form of stability than merely asymptotic as it stipulates that the state must decay at a certain rate. The local versions of stability will not be dealt with in this chapter, although they play a crucial role in the assessment of saturated unstable plants (see [1, 6, 12, 22]). For systems with a forcing term, w we take a slightly different approach. We normally use a relationship between the forcing (input) signal, w ∈ Rnw and an output signal z ∈ Rnz . It is common to assume that signals belong to certain vector spaces. A vector space is a space which is defined in terms of its norm. In control, the L p norm is a popular way of measuring the size of signals. It is defined as kxk p :=

Z

0

∞

kx(t)k p dt

1

p

The space L p is then defined as the space of all signals with finite L p norm, that is L p := x : kxk p < ∞

It follows that if a signal belongs to L p it is “well-behaved” in some sense. This allows us to define more concrete notions on input-output stability Definition 5.2. Consider the system S : w 7→ y then we say the system is

i) L p stable if w ∈ L p implies z ∈ L p ii) Finite gain L p stable if kzk p ≤ γ kwk p + β ,

β >0

Note that the finite gain form of input-output stability is stronger as it forces a linear relationship between the L p norm of the input and the L p norm of the output. The smallest such quantity γ is sometimes called the induced L p norm or the L p gain of the system S and is denoted kS ki,p . It might at first seem that the two definitions of stability are unrelated, but there are connections between them and some which we shall use later on. One connection is summarised in the following theorem which is effectively paraphrased from [15]. Theorem 5.1. Consider the state-space system x˙ = f (x, w) S ∼ z = h(x, w)

(5.14)

If the state equation x˙ = f (x, 0) is GES, f (x, w) is Lipschitz in w and kh(x, w)k ≤ η1 kxk + η2 kwk then the system S is finite gain L p stable for all integers p ∈ [1, ∞]. This says that, if our state equation is globally exponentially (and therefore globally asymptotically) stable, then providing our output equation z = h(x, w) is sufficiently “nice” we get L p stability for free. A local version of this theorem can also be stated. Hence, at least for the time being, it makes sense to try and prove exponential stability of a state-space system.

154

M.C. Turner, G. Herrmann , and I. Postlethwaite

How can we prove exponential stability of a nonlinear system? Well the approach used throughout the control community is to find a Lyapunov function. Although this is too large a topic to be considered here it roughly boils down to finding a positive definite scalar function of the state, V (x) > 0 and then proving that the time derivative of this is negative along the dV (x) ∂ V (x) trajectories of the closed loop systems, i.e., dt = ∂ x f (x) < 0. In particular, we have the following theorem 2 Theorem 5.2. If there exists a function V (x) such that k1 kxkc ≤ V (x) ≤ k2 kxkc ∂ V (x) f (x) ≤ −k3 kxkc ∂x then the origin of x˙ = f (x) is globally exponentially stable.

(5.15) (5.16)

It is this Lyapunov argument which will be used extensively in the remainder of the chapter.

5.3.2

Saturation Modelling

Although it is possible to use the saturation function directly in proving stability using a Lyaponov function, it may be difficult. This is certainly the case with MIMO systems and to a lesser extent with SISO systems. Despite the saturation function’s simple description, it is nonlinear, and this complicates things. To make the saturation function more tractable, we will do two things. The first is a simple transformation and the second is known as sector bounding.

An Equivalent Representation Given the system y = Gsat(u), where the control signal u(t) is fed directly into the saturation element, it is useful to consider instead y = G(u + ∆ (u)) The system is now driven by the nominal control signal plus a nonlinear perturbation term, ∆ (.). This allows us to analyse the stability of a perturbed system, which is often conceptually simpler. Clearly ∆ (u) = sat(u) − u, but even more remarkably, ∆ (u) = −Dz(u), where Dz(.) is the deadzone function defined as   Dz1 (u1 )   .. (5.17) Dz(u) =   . Dzm (um )

where Dzi (ui ) = sign(ui ) max {0, |ui | − u¯i }. Equivalently it can be defined through the identity sat(u) = u − Dz(u).

(5.18)

A graph of a scalar version of the deadzone is shown in Figure 5.6. Thus if we have a system M(s) which is in a feedback loop with a saturation element, as shown in Figure 5.7, this can equivalently be represented by a feedback loop with [M(s) − I]−1 M(s) and a deadzone element. Effectively this lets us represent any system containing saturation as the nominal linear closed-loop plus a perturbation term involving the deadzone. Note that for all u(t) ∈ U , Dz(u) = 0 and thus the system behaves in a nominal way. 2

Essentially a simple version of Corollary 3.4 in [15] .

Anti-windup Compensation and the Control of Input-Constrained Systems

155

Dz(u)

_ −u

_ u

u

Fig. 5.6. The scalar deadzone function −1

M(s) um

(M−I) M

u

~

u

u

"−∆ " Fig. 5.7. Representation of linear and nonlinear elements

Sector Bounding

N(v) slope =β slope =α

v

Fig. 5.8. Graphical representation of Sector[α , β ] Let us consider a static (memoryless) nonlinear z = N (v) which is continuous in its argument (for simplicity). Consider the graph in Figure 5.8. This shows a nonlinear function which is quite irregular, but notice that the graph of this function always lies between the gradient α and the gradient β lines. It is thus said that the function lies within the Sector[α , β ] and in fact it can be seen that the following inequality holds.

α v2 ≤ vN (v) ≤ β v2

β >α

156

M.C. Turner, G. Herrmann , and I. Postlethwaite

It is easy to see that this implies v(α v − N (v)) ≤ 0

(5.19)

v(N (v) − β v) ≤ 0

(5.20)

(N (v) − β v)(N (v) − α v) ≤ 0

(5.21)

Hence it follows that

Thus any nonlinearity N (.) ∈ Sector[α , β ] will satisfy the inequality (5.21). Note that there are an infinite number of nonlinearities which belong to this sector and hence “modelling” a particular nonlinearity in this way introduces some conservatism. If we let N (.) = sat(.), or N (.) =Dz(.) it follows that the lower bound on the sector is α = 0 and the upper bound, is β = 1 for both nonlinearities. Hence, strangely, both the saturation and the deadzone nonlinearities inhabit the same sector. This idea can be easily extended to multivariable static nonlinearities. For simplicity we will only consider so-called decentralised nonlinearities in which N (.) : Rm 7→ Rm is a nonlinearity, with each element satisfying a scalar sector bound. Formally, we assume that N can be described as   N1 (u1 )   .. N (u) =  (5.22)  . Nm (um )

Ni ∈ Sector[αi , βi ]

(5.23)

In this case, using a similar process to the above, we say that a decentralised static nonlinearity N is in the Sector[A , B] if the following inequality holds (N (v) − Bv)(N (v) − A v) ≤ 0

(5.24)

where A = diag(α1 , . . . , αm )

(5.25)

B = diag(β1 , . . . , βm )

(5.26) (5.27)

Now, if we assume that the saturation and deadzone nonlinearities are multivariable, it follows that as each i’th component is within the sector[0, 1], that the full multvariable nonlinearity is such that sat(.) ∈ sector[0, I] Dz(.) ∈ sector[0, I] (5.28)

In turn this means that for both nonlinearities, the following inequality holds N (u)′ (N (u) − u) ≤ 0

(5.29)

N (u)′ X(N (u) − u) ≤ 0

(5.30)

or, more generally, that

where X is a positive definite diagonal matrix and N (u) is either Dz(u) or sat(u).

Anti-windup Compensation and the Control of Input-Constrained Systems

157

Remark 5.1. A very interesting feature of bounding both the saturation and deadzone nonlinearities in this way is that the sector bounds are not functions of the limits u. ¯ That means that no matter what u¯ is -for either saturation or deadzone - it plays no role in determining the sector bound. Thus henceforth, the subscript u¯ will be dropped as it plays no further role in the development of the results. However, it should be noted that if we wish to develop local versions of these results, the limits u¯ actually play a crucial role in determining the local stability properties. ⊓ ⊔

5.3.3

The Multivariable Circle Criterion

z

w ~

T(s)

u

u

"−∆ " Fig. 5.9. The nonlinear stability problem We are now ready to state the main result of the section. This result is a very well-known result which effectively is going to be presented in a slightly different form to that in which it is often presented. However, this form will be of direct relevance to the results to be introduced later. We consider the system w u (5.31) = T (s) u˜ z

where T (s) is a linear system with state-space realisation   x˙ = Ax + Bw w + Bu˜ T (s) ∼ u = Cx  z = Cz x + Dzw w + Dz u˜

(5.32)

where x ∈ Rn is the system’s state, u˜ = Dz(u) ∈ Rm is the “deadzoned” control signal 3 , w ∈ Rnw is the exogenous input and z ∈ Rnz is the performance output. Remark 5.2. We have deliberately restricted the feedback term u = Cx to be strictly proper to avoid questions of well-posedness (i.e., issues of existence and uniqueness of the loop equations). It is actually possible to use the more general feedback u = Cx + Du, ˜ at the expense of extra analysis. This chapter will not consider this additional analysis in detail, although the topic of well-posedness will be touched upon later. ⊓ ⊔ 3

This actually does not need to be the case - it could be any nonlinearity in the Sector [0, I]. . .

158

M.C. Turner, G. Herrmann , and I. Postlethwaite

We wish to examine the conditions under which, the closed-loop system with w = 0 is globally exponentially stable. We proceed using a Lyapunov argument and attempt to establish some linear matrix inequalities. The following is the main result of the section. Theorem 5.3. The system described by equation (5.32) with u = Dz(u) satisfies the following properties • •

When w = 0 the origin is globally exponentially stable. There exist β , γ > 0 such that kzk2 ≤ γ kwk2 + β

if the following linear matrix inequality is satisfied ′ A P + PA PB +C′W M := <0 ⋆ −2W

(5.33)

Proof. First, global exponential stability of the origin is established. Consider the Lyapunov function V (x) = x′ Px > 0 ∀x 6= 0 (5.34)

where P > 0 is a postitive definite symmetric matrix. Note that V (x) is radially unbounded so that if we can prove V˙ (x) < 0 ∀x 6= 0, then global asymptotic stability will follow. Therefore, let us consider V˙ (x) V˙ (x) = 2x′ P(Ax + Bu) (recall w = 0) (5.35) Next note that as u˜ = Dz(u) satisfies the sector condition it follows that V˙ (x) ≤ x′ (A′ P + PA)x + 2x′ PBu + 2 u˜′W (u − u) ˜ | {z } sector bound ′ ′

= x′ (A′ P + PA)x + 2x′ (PB +C W )u˜ − 2u˜ W u˜ =: V˙1 (x)

(5.36) (5.37) (5.38)

where W > 0 is a positive definite diagonal matrix. Note that the adjoining of the sector bounds to the Lyapunov function to create the augmented Lyapunov function is a simple application of the S-procedure. Note further that it follows that if V˙1 (x) < 0, then obviously V˙ < 0, although the converse is not true (this has introduced conservatism). The expression above can be majorised to get ′ ′ x A P + PA PB +CW x V˙1 (x) = (5.39) u˜ ⋆ −2W u˜ | {z } =:−M

Thus if LMI (5.33) as stated in Theorem 3 is negative definite it follows that

2

x

V˙1 (x) ≤ −λmin (M)

u˜ ≤ −λmin (M)kxk2

(5.40) (5.41)

This proves that the origin is asymptotically stable. Note now that this implies V˙ (x) ≤ λmin (M)kxk2 . Furthermore as P > 0, it follows that

λmin (P)kxk2 ≤ V (x) ≤ λmax (P)kxk2

(5.42)

Anti-windup Compensation and the Control of Input-Constrained Systems

159

Thus, applying Theorem 2 we can see that the origin is not only asymptotically stable, but exponentially stable too. Our proof is almost complete. What remains is to prove that a finite L2 gain exists between w and z. To do this simply note that in our case we have f (x, w) = Ax + BDz(Cx) + Bw w

(5.43)

h(x, w) = Cx + DDz(Cx) + Dw w

(5.44)

where Dz(x) is Lipschitz in x. From this it follows that f (x, w) and h(x, w) are both certainly Lipschitz and hence that they satisfy the conditions of Theorem 1 and hence we have a finite L2 gain between w and z, which completes the proof. ⊓ ⊔⊓ ⊔ Remark 5.3. The so-called “Multivariable Circle Criterion” is the workhorse of many modern anti-windup strategies. Its main advantage is that through an appropriate change of variables it can be used for convex synthesis of anti-windup compensators. ⊓ ⊔

5.4

Anti-windup Problem Definition

The anti-windup problem is more than just ensuring stability of a constrained input system; as mentioned earlier anti-windup stipulates that large signal stability and performance goals must be achieved while preserving a local controller structure - i.e., that dictated by the linear controller. This objective is more complicated than a simple constrained control problem, where no constraints are placed on the structure and small-signal performance of the system. As mentioned before we would like the anti-windup compensator to become active only during periods of saturation and to interfere with the linear control loop as little as possible. Furthermore, when a saturation event has passed we would like linear behaviour to be recovered, at least asymptotically. In this section we formalise the anti-windup problem which we would like to solve 4 . Definition 5.3. Let ulin ∈ Rm denote the nominal linear control signal; and yd ∈ R p be the difference between the linear output ylin and the actual output y (i.e., yd = ylin − y) which results from actuator saturation.

• •

Then the anti-windup compensator Θ (s) is said to solve the anti-windup problem if the closed loop system in Figure 5.3 is internally stable, well-posed and the following hold: 1. If dist(ulin , U ) = 0, ∀t ≥ 0, then yd = 0, Θ (s)). 2. If dist(ulin , U ) ∈ L2 , then yd ∈ L2 .

∀t ≥ 0 (assuming zero initial conditions for

The anti-windup compensator Θ (s) is said to solve strongly the anti-windup problem if, in addition, the following condition is satisfied: 3. The operator T p : ulin 7→ yd is well-defined and finite gain L2 stable. 4

This definition is heavily influenced by that proposed in [24].

M.C. Turner, G. Herrmann , and I. Postlethwaite

160

Some comments on this problem definition are appropriate. •

•

•

•

Note that stability and well-posedness are part of the anti-windup problem but are not the limit of it. A system is well-posed if unique solutions exist to the closed-loop equations. In general, for saturated systems, it requires a detailed analysis; existence of solutions is relatively easy to prove, but uniqueness is somewhat harder and requires an argument similar to that found in [26] or the use of an implicit function theorem as advocated in [7] The first condition of the definition involves the use of the distance function. This function effectively gives a measure of the “distance” of a vector from a set. For example assume we are measuring the distance of a vector z from a set Z , then the distance function is the Euclidean norm from the vector z to the nearest point in Z i.e., dist(z, Z ) = infv∈Z kv − zk. If z ∈ Z , then dist(z, Z ) = 0. In this work we assume that our sets are connected and, moreover, convex. The first condition of the definition says that if the linear control signal ulin remains in U for all time, then the deviation from the linear output, ylin to the actual output, y is zero for all time; that is the anti-windup compensator does not impede linear behaviour if saturation is not encountered. The second item also makes use of the distance function, but it looks at the L2 norm of the distance function. It effectively says that if the signal which can be constructed by measuring the distance between ulin and the set U is square integrable (i.e., well behaved) then the deviation from linear behaviour will also be well behaved, and importantly, decay to zero asymptotically (this can be proved using Barbalat’s lemma). Note this ensures that linear behaviour will at least be recovered asymptotically, providing saturation does not occur indefinitely 5 . The third item is the “strong” version of the anti-windup problem and it requires there to be a finite L2 gain between our linear control signal, ulin and the deviation from linear behaviour, yd . Ideally this should be as small as possible to represent small (in an “energy gain” sense) deviation from linear behaviour. This means that if our control signal has a certain energy, then the signal representing the difference between linear and nonlinear behaviour has an energy bounded by a linear function of the energy in the linear control signal.

This definition has now “fleshed” out the problem we would like to solve. The next section offers a certain solution to that problem.

5.5

An Anti-windup Solution

There are a number of ways to solve the above problem. The one we describe here is a method which we think is particularly attractive as it gives a nice graphical interpretation of the problem (we now have words, maths and figures!). It also gives a convex formulation of the problem which can be solved using linear matrix inequalities which we shall describe shortly.

5.5.1

Architecture

In order to appreciate the strength of the method, it is fruitful to consider Figure 5.10, which shows the anti-windup problem in a reasonably generic form (compare to Figure 5.3), but 5

It is important to realise that the key point in establishing this is that there is no direct feedthrough term in the loop.

Anti-windup Compensation and the Control of Input-Constrained Systems

161

d r K

u lin + − ud

y lin

M−I + +

y

G

+ −

yd

G2 M

~ u

Fig. 5.10. A special parametrisation of the AW problem

M−I

ud

Disturbance Filter

Nonlinear Loop

− ~ u

+

r

d K

G

u lin

yd

G 2M

+

−

ylin

y

Nominal Linear Transfer Function Fig. 5.11. An equivalent “decoupled” representation of the AW problem

with a special choice of our AW compensator, Θ . This architecture allows the anti-windup compensator Θ (s) to inject signals, which modify the controller’s behaviour, at the controller output and also one of its inputs. Here G(s) = [G1 (s) G2 (s)] is the finite-dimensional linear time invariant (FDLTI) plant which we have assigned the following state-space realisation x˙ p = A p x p + B p um + B pd d (5.45) G(s) ∼ y = C p x p + D p um + D pd d where x p ∈ Rn p is the plant state, um ∈ Rm is the control input to the plant, y ∈ R p is the plant (measurement) output and d ∈ Rnd is the disturbance. The FDLTI controller K(s) is assumed to be implemented as K(s) = [K1 (s) K2 (s)] and is assigned the following state-space realisation

162

M.C. Turner, G. Herrmann , and I. Postlethwaite K(s) ∼

x˙c = Ac xc + Bc ylin + Bcr r ulin = Cc xc + Dc ylin + Dcr r

(5.46)

where xc ∈ Rnc is the controller state, ulin ∈ Rm is the linear controller output r ∈ Rnr is the disturbance on the controller, normally the reference. The overall controller output is given by u = ulin + ud ∈ Rm where ud is a signal produced by the anti-windup compensator; if the anti-windup compensator is inactive, u = ulin . The signal ylin ∈ R p is the linear input to the controller and is given by y + yd ∈ R p ; if the anti-windup compensator is inactive ylin = y. Note that in terms of Figure 5.3, the signals ud and yd play the roles of θ1 and θ2 respectively. As it stands, the scheme depicted in Figure 5.3 does not have any illuminating features. However, consider a particular case of the generic scheme, as shown in Figure 5.10 (introduced in [29]), where the anti-windup compensator Θ (s) has been parameterised in terms of a transfer function M(s) ∈ RH ∞ and a copy of the plant G2 (s). Using the identity (5.18), Figure 5.10 can be re-drawn as the decoupled scheme in Figure 5.11. Notice that this system exhibits an attractive decoupled structure: 1. Nominal linear system. This represents the linear system which would have resulted if no saturation was present. 2. Nonlinear loop. This contains the nonlinear stability problem and assuming that the linear loop is stable and the linear operator G2 M is also stable, the entire nonlinear stability problem is contained within this part of the system 3. Disturbance filter. This part of the system determines how the system recovers after a saturation event has stopped. This part is responsible for both the speed and manner of recovery after saturation has ceased. In [30], it was shown that most anti-windup schemes can be interpreted as certain choices of M(s) and therefore schemes such as the Hanus conditioning scheme ( [8]) and the high gain approach ( [4, 14]) can be analysed in terms of Figure 5.11. The advantages of viewing antiwindup in terms of Figure 5.11 is that the nominal linear performance is separated from the nonlinear part of the scheme. Moreover, the stability of the scheme is dependent on the stability of the nonlinear loop, assuming stability of the nominal linear closed loop and stability of the plant. This leads to the following assumption.

Assumption 5.5.1 • •

I −K2 (s) The poles of −G2 (s) I

plane. The limit lims→∞ (I − K2 (s)G2 (s))−1 exists.

−1

are in the open left-half complex

The first point ensures that all nominal closed loop transfer functions are asymptotically stable (in RH ∞ ) and if r = 0 and d = 0, then limt→∞ [x′p (t) xc′ (t)]′ = 0. The second assumption ensures that the nominal linear system is well-posed i.e., unique solutions exist to the feedback equations. These assumptions will be satisfied in most circumstances. From Figure 5.11, it can be seen that the performance of the overall closed loop system, including the anti-windup compensator, is closely related to the mapping T p : ulin 7→ yd . This mapping represents the deviation from nominal linear behaviour in response to a saturation

Anti -windup Compensation and the Control of Input-Constrained Systems

163

event and can be used as a measure of the anti-windup compensator’s performance. If some appropriate norm of this mapping is small, then the anti-windup compensator is successful at keeping performance close to linear (which we assume is the desired performance). In fact, as the AW compensator is parameterised by M(s), the choice of M(s) dictates the system’s stability properties under saturation. This chapter proposes choosing M(s) such that performance during saturation is improved. In [26] (see also [9, 25]), the L2 gain of T p was minimised using a system of linear matrix inequalities and, furthermore, M(s) was chosen such that it corresponded to static or low order anti-windup compensators. The result of [26] demonstrated, using suitable examples, that direct minimisation of T p was central to good anti-windup performance, and compensators designed according to the ideas in [26] seemed to perform at least as well, and often better, than most other anti-windup compensators. This idea was extended to more general cases in [10].

5.5.2

Full Order Compensators

Full-order anti-windup compensators are the class of AW compensators which are of order equal to the plant; that is they have n p states. They are attractive because they always exist for any stable plant, G2 (s), and can be designed in a relatively straightforward manner, through judicious choice of M(s). In fact, an appealing choice of M is as a part of a right comprime factorisation of G2 (s) = N(s)M(s)−1 where M, N ∈ RH ∞ . This means that the disturbance filter is given as yd (s) = N(s)u(s) ˜ and is hence a stable transfer function. The central reason for the appeal of this choice of M(s) is that this means that the anti-windup compensator M(s) − I (5.47) Θ (s) = N(s) has order equal to that of the plant, providing the coprime factors share the same state-space. A state-space realisation of the operator T p is then easily calculated (using [33], for example) as 6    Ap + BpF Bp x˙2 M(s) − I x ∼  ud   (5.48) F 0  2 N(s) u˜ yd Cp + D p F D p

where we have used the state x2 ∈ Rn p to represent the anti-windup compensator’s state and to distinguish it from the plant state. Another endearing property of this choice of AW compensator is that there is no direct-feedthrough from u˜ to ud which prevents algebraic loops (and thus does not require us to deal with well-posedness issues). The following is the main result of the section: Theorem 5.4. Under Assumption 5.5.1, there exists a full anti-windup compensator Θ (s), described by equations (5.47) and (5.48), which solves strongly the anti-windup problem if there exist matrices Q > 0, U = diag(µ1 , . . . , µm ) > 0,L ∈ R(m×n) and a positive real scalar γ such that the following LMI is satisfied 6

We could in fact use a more general coprime factorisation as discussed in [27] and [10], although it is not clear that this gives us much advantage over the simple factorisation described above.

164

M.C. Turner, G. Herrmann , and I. Postlethwaite



 QA′p + A p Q + L′ B′p + B p L B pU − L′ 0 QC′p + L′ D′p ′   ⋆ −2U I UD p  <0   ⋆ ⋆ −γ I 0 ⋆ ⋆ ⋆ −γ I

(5.49)

Furthermore, if this inequality is satisfied, a suitable F achieving kT p ki,2 < γ is given by F = LQ−1 . ⊓ ⊔ Proof. Well-posedness. Note from equation (5.48), all equations which comprise the operator T p are explicit; hence the sytem is, trivially, well-posed. Stability and L2 gain. If there exists a positive definite symmetric matrix P > 0 which satisfies J=

d x2 (t)′ Px2 (t) + kyd k2 − γ 2 kulin k2 < 0 ∀[x2′ dt

u˜′

u′lin ]′ 6= 0

(5.50)

it follows that 1. If ulin = 0, then there exists a P > 0 such that v(x2 ) = x2′ Px2 is a Lyapunov function for the system; hence the system is globally stable. 2. Integrating J from 0 to T and taking the limit as T → ∞ kyd k22 < γ 2 kulin k22 + (x2 (0)′ Px2 (0) − x2 (∞)′ Px2 (∞))

(5.51)

which implies L2 gain less than γ from ulin to yd (and thus Item 3). The remainder of this part of the proof involves simplifying the expression for J. First, note that as the deadzone function is such that Dz ∈ Sector[0, I] we have that u˜i ui ≥ u˜2i ,

∀i ∈ {1, . . . , m}

(5.52)

From this, it follows that for some matrix W = diag(w1 , . . . , wm ) > 0 u˜′W (u − u) ˜ ≥0

(5.53)

Thus it follows that a sufficient condition for J < 0 is if d ˜ < 0 ∀[x2′ J˜ = x2 (t)′ Px2 (t) + kyd k2 − γ 2 kulin k2 + 2u˜′W (u − u) dt

u˜′

u′lin ]′ 6= 0 (5.54)

where u = ulin − ud . Hence using the state-space realisation (5.48), this becomes  (A p + B p F)′ P + P(A p + B p F) + (C p + D p F)′ (C p + D p F)  ⋆ ⋆  PBP + (C p + D p F)′ D p − F ′W 0 ′ −2W + D p D p W <0 ⋆ −γ 2 I

(5.55)

The LMI in the theorem then follows using the standard Schur complement and congruency transformation arguments, as reported, for example, in [21] (see also [25, 26], although the details are not the same). Item 1 of Definition 1. This follows from the stability of the nonlinear loop above, assuming that x2 (0) = 0 and the nominal, un-saturated system is asymptotically stable. Item 2 of Definition 1. This is implied directly from Item 3 of Definition 1.

⊓ ⊔⊓ ⊔.

Anti-windup Compensation and the Control of Input-Constrained Systems

165

Remarks •

•

•

Notice that a full-order anti-windup compensator always exists for any stable plant G2 (s). Simply choosing F = 0 recovers the so-called internal model control (IMC) solution { ( [32]). This is a stabilising, explicit solution, provided G2 (s) is stable, although it may not yield good performance if the plant has slow or lightly damped modes. It can be proved using the elimination lemma that the matrix A p must be Hurwitz in order to permit a globally stabilising anti-windup compensator. This is both necessary and sufficient for the existence of a full order compensator, but only becomes necessary, not sufficient, for low order or static compensators. This produces an AW compensator of order n p i.e., the same as that of the plant which typically is bad. There has been some success with lower order AW compensators - see [26] for example or even static ones [21, 26]. However, in this case, their existence is not guaranteed and they are also not guaranteed to satisfy optimal performance qualities either.

5.6

Simple Examples

This section demonstrates the use of the full-order “optimal” anti-windup compensators advocated in the previous section. The examples used are not case studies per se because they are grossly simplified. However it is hoped that they allow the main ideas of anti-windup compensation to be conveyed.

5.6.1

Simple 2nd-Order Example

The Nominal System We consider a nominal single-loop plant described by the transfer function G2 (s) =

y(s) β (s + α ) = u(s) s2 + 2ζ ωn s + ωn2

(5.56)

where the denominator has two complex conjugate roots with postive real parts. Such a transfer function is representative of the short-period pitch dynamics of an aircraft, for example, but also of many other systems. For this example ωn = 1.5, ζ = 0.1. For simplicity we assume no disturbance is acting on the system and hence G2 (s) ≡ 0. To stabilise the system and to ensure zero steady state error to step inputs, th e following PI controller is used ks (k p + ki ) KPI (s) = (5.57) s and to prevent proportional kick, the reference is filtered through the low-pass filter. 1 τs + 1 In terms of our generic controller description we thus have KF (s) =

K(s) = [K1 (s) K2 (s)] = [KPI (s)KF (s)

(5.58)

− KPI (s)]

(5.59)

This plant is interesting because the nominal linear controller, K(s) is actually quite good for linear operation. Noteworthy features are:

166 • • •

M.C. Turner, G. Herrmann , and I. Postlethwaite The controller damps the closed-loop poles, bringing them close to critical damping. The controller yields good robustness margins with infinite gain margin, over 60 degrees of phase margin and also an infinite gain reduction margin Nominal tracking is relatively good. There is some overshoot and some “drop back” but the rise and settling times are short.

Figure 5.12 shows the nominal linear response. 2.5

Control input [non-dimensional]

6

Output [non-dimensional]

2

1.5

1

0.5

5

4

3

2

1

0

−1

0 −2

−0.5

0

5

10

15

Time [sec]

20

−3

0

5

10

15

Time [sec]

20

Fig. 5.12. Linear response of Example 1

The Constrained System If we now introduce control signal saturation at ±1unit, the response now degrades to that shown in Figure 5.13. Note that the presence of input saturation has indeed caused performance problems in the system with large oscillations occurring in the output, y(t), increasing the overshoot and settling time. Compared to the linear response, the saturated response looks quite bad, despite the large linear robustness margins.

The Constrained System and Anti-W indup We will make a quick attempt at fixing some of the problems caused by saturation with a full-order anti-windup compensator described in Section 5. A plot of the system’s response with AW is shown in Figure 5.14. Note that while complete recovery of linear behaviour is not possible, the presence of the AW compensator has led to improvements over the “no AW” case. Further points to note are:

1.8

1

1.6

0.9

Control input [non-dimensional]

Output [non-dimensional]

Anti-windup Compensation and the Control of Input-Constrained Systems

1.4

1.2

1

0.8

0.6

0.4

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.2

0

167

0.1

0

5

10

15

Time [sec]

20

0

0

5

10

15

Time [sec]

20

Fig. 5.13. Saturated response of Example 1 •

•

•

It is usually difficult to recover linear behaviour entirely because saturation still has an effect on the system (it cannot simply be avoided!). However, the aim of AW is to provide a graceful degradation of this performance. In practice this often means something like accepting a lower rise time but trying to reduce overshoot and settling time. An exception to this rule is the case of redundant systems in which careful AW design could cause a redistribution of control effort to enable linear performance to be effectively recovered. In this example, for the reference demanded, quite a lot of saturation occurs. In particular note that the “steady state” value of the control signal for the 2 unit pulse demand is above the saturation level. This effectively restricts the ability of the system to track the reference at this level and thus explains why the response with AW sits below the linear response there simply is not enough control power to attain the demand. Such references are often called infeasible references as it is not feasible to expect a constrained input system to track them. Note that there is quite a slow mode in the compensator which causes the AW compensated system to resume linear behaviour lower than one would like. This is due to the “optimal” LMI routine yielding a slow pole in the AW compensator. In fact, this can be removed by appropriate re-optimisation using a low order design which allows faster linear recovery. The interested reader should consult [26].

Hopefully this has demonstrated that a simple application of AW can improve a system’s responses during saturation. The AW design here is by no means perfect and would require considerable tuning for a real system. Nevertheless, even a solution like this can lead to big improvements in damping during saturation.

168

M.C. Turner, G. Herrmann , and I. Postlethwaite 1

Control input [non-dimensional]

Output [non-dimensional]

1.5

1

0.5

0

−0.5

0.8

0.6

0.4

0.2

0

−0.2

−0.4

−0.6

−0.8

−1

0

5

10

15

Time [sec]

20

−1

0

5

10

15

Time [sec]

20

Fig. 5.14. Saturated response of Example 1 (with AW)

5.6.2

Lockheed Martin F104 Example

The Nominal System The nominal plant we use is a 4-state linear model of the longitudinal dynamics of a Lockheed Martin F104 Starfighter aircraft, taken from [2]. We consider the transfer function from elevator position to pitch attitude, i.e., G(s) =

θ (s) δ (s)

(5.60)

No controller is provided in [2], so we fit the aircraft with an H∞ loop-shaping controller 7 which yields excellent robustness to normalised coprime factor uncertainty - it is robust to a ball of radius εmax ≈ 0.28. This loop-shaping controller was designed such that, again, steady state error to a step reference demand was zero. Figure 5.15 shows the response of the linear system to a reference puls e train of 30 degrees in magnitude. Note that the linear system is well damped and the output tracks the reference demand well.

5.6.3 The Constrained System Saturation limits of ±25 degrees are introduced to the control signal and for reference demands of ±30 degrees this causes significant periods of saturation. Figure 5.16 shows the 7

As the F104 is a 1950’s aircraft, of course this is unlikely, but the idea is to demonstrate the principle.

Anti-windup Compensation and the Control of Input-Constrained Systems

80

30

60

Control input (elevator) [deg]

40

Output [deg]

20

10

0

−10

40

20

0

−20

−20

−40

−30

−60

−40

0

5

10

15

Time [sec]

20

169

−80

0

5

10

15

Time [sec]

20

Fig. 5.15. Linear response of F104 Example response of the saturated system. Although the response has not degraded as much as the first example (perhaps due to the H∞ controller’s good robustness properties), saturation has still caused the output to become more oscillatory. A pilot would find this type of response most uncomfortable.

5.6.4

The Constrained System and Anti-windup

Again, a quick fix of the saturation problem is attempted using a full-order AW compensator described in Section 5. Figure 5.17 shows the systems response to the same pulse train as before. Similarly to the first example, notice that although the AW compensator has not enabled a complete recovery of linear performance, it still manages to reduce the damping significantly, providing better ride quality for the pilot.

5.7

Conclusion

This chapter has tried to introduce the anti-windup problem in a fairly clear manner, and has tried to explain some of the technical nuances present in the problem and its solution. A further chapter will be devoted to case studies which expose the methods to some real practical problems.

M.C. Turner, G. Herrmann , and I. Postlethwaite 50

25

40

20

Control input (elevator) [deg]

170

30

Output [deg]

20

10

0

−10

15

10

5

0

−5

−20

−10

−30

−15

−40

−20

−50

0

5

10

15

Time [sec]

20

−25

0

5

10

15

Time [sec]

20

Fig. 5.16. Saturated response of F104 Example

5.7.1

Further Reading

This chapter has dwelt upon a very special class of anti-windup compensators and although the problem solved has been exactly the type of “true anti-windup” problem which practitioners are interested in, there has obviously been a Leicester bias to the presentation of the results. The following are a few good journal papers which we think the interested reader would do well to read. •

• •

• • •

[16] A unifying scheme for AW compensators. Although design methods are not really given, this is a classic paper which was arguably the first to give a common framework for anti-windup designs [31] This was where the decoupled scheme used in this chapter was first introduced. Most of the subsequent Leicester work has been based on this. [19,21]. Some of the first papers to apply LMI methods to anti-windup compensators. Although the some of these papers are fairly “brute-force” AW methods, they are interesting nonetheless. [7]. The first paper to give a full LMI interpretation of the AW compensation problem for stable systems. A number of insights are given in this paper. [26]. One of the first papers to give a direct solution to the problem as posed in this paper. Treats low order and static compensators in the main. [11]. A fairly complete solution to the discrete-time anti-windup problem for stable systems.

This is not an exhaustive list and there are many papers (including one or two really good conference papers) which can be found in the references given within these papers. We also

Anti-windup Compensation and the Control of Input-Constrained Systems 40

171

25

20

Control input (elevator) [deg]

30

Output [deg]

20

10

0

15

10

5

0

−5

−10

−10

−15

−20 −20

−30

0

5

10

15

Time [sec]

20

−25

0

5

10

15

Time [sec]

20

Fig. 5.17. Saturated response of F104 Example (with AW) suggest consulting the papers [3, 5, 13, 19, 20, 28] for other good and related ideas on antiwindup. Finally, it is important to remark that AW is a developing topic and the solution given here is by no means perfect. It does however give a fast, easy-to-obtain prototype solution which can then be honed in subsequent designs. We anticipate that the subject will grow more mature in the coming years, hopefully with the application of these techniques being matched by progress in the theory.

Acknowledgements The authors would like to thank Dr. Murray Kerr of the University of Leicester for his comments on an earlier version of the chapter.

References 1. C.C. Cao, Z. Lin, and D. Ward. An anti-windup approach to enlarging domain of attraction for linear systems subject to actuator saturation. IEEE Transactions on Automatic Control, 47(1):140–145, 2002. 2. M.V. Cook. Flight dynamics principles. Arnold, 1997. 3. S. Crawshaw and G. Vinnicombe. Anti-windup synthesis for guaranteed L2 performance. Proc. IEEE Conference on Decision and Control, 2000.

172

M.C. Turner, G. Herrmann , and I. Postlethwaite

4. C. Edwards and I. Postlethwaite. Anti-windup and bumpless transfer schemes. Automatica, 34(2):199–210, 1998. 5. A. H. Glattfelder and W. Schaufelberger. Stability of discrete override and cascade-limiter single-loop control systems. IEEE Transactions on Automatic Control, 33(6):532–540, 1988. 6. J. M. Gomes da Silv Jr. and S. Tarbouriech. Anti-windup design with guaranteed regions of stability: an LMI based approach. IEEE Transactions on Automatic Control, 50(1):106–111, 2005. 7. G. Grimm, J. Hatfield, I. Postlethwaite, A.R. Teel, M.C. Turner, and L. Zaccarian. Antiwindup for stable linear systems with input saturation: an LMI based synthesis. IEEE Transactions on Automatic Control, 48(9):1509–1525, 2003. 8. R. Hanus, M. Kinnaert, and J.L. Henrotte. Conditioning technique, a general anti-windup and bumpless transfer method. Automatica, 23(6):729–39, 1987. 9. G. Herrmann, M.C. Turner, and I. Postlethwaite. Discrete time anti-windup - part 2: extension to sampled data case. Proceedings of the European Control Conference, 2003. 10. G. Herrmann, M.C. Turner, and I. Postlethwaite. Some new results on anti-windup conditioning using the Weston-Postlethwaite approach. Proc. Conference on Decision and Control, 2004. 11. G. Herrmann, M.C. Turner, and I. Postlethwaite. Discrete-time and sampled-data antiwindup synthesis: stability and performance. International Journal of Systems Science, 2006. 12. H. Hindi and S. Boyd. Analysis of linear systems with saturation using convex optimisation. Proc. IEEE Conference on Decision and Control, pages 903–907, 1998. 13. P. Hippe. Windup prevention for unstable systems. Automatica, 39(11):1967–1973, 2003. 14. R.A. Hyde. The Application of Robust Control to VSTOL Aircraft. PhD thesis, Department of Engineering, University of Cambridge, 1991. 15. H.K. Khalil. Nonlinear Systems. Prentice Hall, New Jersey, 1996. 16. M.V. Kothare, P.J. Campo, M. Morari, and C.N. Nett. A unified framework for the study of anti-windup designs. Automatica, 30(12):1869–1883, 1994. 17. J.C. Lozier. A steady state approach to the theory of saturable servo systems. IRE Transactions on Automatic Control, pages 19–39, 1956. 18. J. Maciejowski. Predictive control with constraints. Pearson, 2001. 19. V.R. Marcopoli and S.M. Phillips. Analysis and synthesis tools for a class of actuatorlimited multivariable control systems: a linear matrix inequality approach. International Journal of Robust and Nonlinear Control, 6(9-10):1045–1063, 1996. 20. S. Miyamoto and G. Vinnicombe. Robust control of plants with saturation nonlinearity based on comprime factor representations. Proc. IEEE Conference on Decision and Control, pages 2838–2840, 1996. 21. E.F. Mulder, M.V. Kothare, and M. Morari. Multivariable anti-windup controller synthesis using linear matrix inequalities. Automatica, 37:1407–1416, 2001. 22. C. Pittet, S. Tarbouriech, and C. Burgat. Stability regions for linear systems with saturating control via Circle and Popov criteria. Proc. IEEE Conference on Decision and Control, 5:4518–4523, 1997. 23. G. Stein. Respect the unstable. IEEE Control Systems Magazine, (8):12–25, 2003. 24. A.R. Teel and N. Kapoor. The L2 anti-windup problem: Its definition and solution. Proceedings of the European Control Conference, 1997. 25. M.C. Turner, G. Herrmann, and I. Postlethwaite. Discrete time anti-windup - part 1: stability and performance. Proceedings of the European Control Conference, 2003. 26. M.C. Turner and I. Postlethwaite. A new perspective on static and low order anti-windup synthesis. International Journal of Control, 77(1):27–44, 2004.

Anti-windup Compensation and the Control of Input-Constrained Systems

173

27. M.C. Turner, I. Postlethwaite, and G. Herrmann. Further results on full-order anti-windup synthesis: exploiting the stability multiplier. Proc. IFAC Nonlinear Controller Symposium, 2004. 28. N. Wada and M. Saeki. Design of a static anti-windup compensator which guarantees robust stability. Transactions of the Institute of Systems, Control and Instrumentation Engineers, 12(11):664–670, 1999. 29. P.F. Weston and I Postlethwaite. Analysis and design of linear conditioning schemes for systems containing saturating actuators. IFAC Nonlinear Control System Design Symposium, 1998. 30. P.F. Weston and I Postlethwaite. Analysis and design of linear conditioning schemes for systems with nonlinear actuators. Internal Report, Dept. of Engineering, Leicester Univ., 98-6, 1998. 31. P.F. Weston and I Postlethwaite. Linear conditioning for systems containing saturating actuators. Automatica, 36(9):1347–1354, 2000. 32. A. Zheng and M. Morari. Anti-windup using internal model control. Interntional Journal of Control, 1994. 33. K. Zhou, J.C. Doyle, and K. Glover. Robust and Optimal Control. Prentice Hall, 1996.

6 Output Feedback H∞ Loop-Shaping Controller Synthesis Emmanuel Prempain and Ian Postlethwaite

Summary. This chapter introduces a Linear Matrix Inequality formulation of H∞ output feedback controller synthesis. Within this framework, the design procedure of McFarlane and Glover is detailed. Existence conditions for full-order and static output loop-shaping controllers are given in terms of matrix inequalities. The approach is extended to the class of polytopic systems. The effectiveness of the various controller synthesis algorithms is demonstrated on two aerospace control design examples.

Key words: Coprime Factorizations, H∞ and H2 Optimizations, Linear Matrix Inequality (LMI), Polytopic Systems, H∞ loop-shaping Control.

6.1 Introduction This chapter is concerned with the problem of designing controllers to meet robust stability together with performance requirements. A well-known H∞ design method achieving such a goal is the procedure of McFarlane and Glover [10]. This controller design procedure, also known as H∞ loop-shaping is based on coprime factor robustness. Here we give a Linear Matrix Inequality (LMI) formulation of the the loop-shaping design procedure of McFarlane and Glover. This formulation allows the derivation of necessary and sufficient conditions for the existence of an H∞ loop-shaping controller. Also, in the LMI framework, sufficient conditions for the existence of static H∞ loop-shaping control can be derived. We will see that the existence of a static H∞ loop-shaping output feedback controller is equivalent to the existence of a positive definite matrix R simultaneously satisfying two inequalities, where one inequality is quadratic in R (Riccati) and the other is linear in R. Because of the very special mathematical structure of the H∞ loop-shaping problem, it is possible to derive simple sufficient linear conditions. An extension of the concept of left coprime factorizations to Linear Parameter Varying (LPV) systems is also given. This novel extension enables us to formulate the LPV H∞ loop-shaping M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 175-194, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

176

E. Prempain and I. Postlethwaite

synthesis problem in terms of polytopic left coprime factors of the shaped LPV plant. The formulation presented here is interesting since it applies to static and low-order H∞ loopshaping LPV controller syntheses as well. The chapter is structured as follows. Section 6.2 reviews some of the basics of LMI-based H2 and H∞ performance objectives. Section 6.3 gives existence conditions for the standard H∞ output feedback controller synthesis in terms of LMIs. Section 6.4 details the formulation of full-order and zero-order (static) H∞ loop-shaping controller design problems and gives explicit conditions under which the syntheses reduce to solving a system of LMIs. In addition, the synthesis algorithms are illustrated on a helicopter model. Section 6.5 generalizes the results obtained in Section 6.4 to robust gain scheduled control for LPV systems. The effectiveness of the low-order loop-shaping controller synthesis proposed here is illustrated on a missile model example. Conclusions are given in Section 6.7. The notation used in this chapter is standard: Rm×n denotes the set of real m × n matrices, I is 1 1 the identity matrix, A > 0 means that A is positive definite and A 2 = (A 2 )T is the square root of a positive semi-definite symmetric matrix A. For a square matrix A, trace(A) is the trace of A. If P and K are Linear Time-Invariant (LTI) systems, the notation F (P, K) denotes the closed-loop system resulting from the interconnection of systems P and K. kG(s)k ∞ is the A B H∞ norm of the LTI system G(s). kG(s)k 2 is the H2 norm of G(s). The notation CD represents a state space realization of G.

6.2 Preliminaries 6.2.1 LMI Formulation of Performance Specifications This section introduces the well-known H∞ and H2 performance measures for LTI systems. Algorithms for computing the H∞ and H2 norms of an LTI system are given in terms of LMIs. Throughout this section, we consider the LTI system G(s) with state space representation x˙ = Ax + Bw

(6.1)

z = Cx + Dw

(6.2)

where x(t) ∈ Rn is the state vector, w(t) ∈ Rnw is the system input, z(t) ∈ Rnz is the measurement output.

H∞ Performance Definition 6.1. Consider the LTI system G(s) with state space equations (6.1)-(6.2). Suppose G(s) is stable. The H∞ norm of G(s) can be defined as the largest singular value of its fre quency response across frequency, that is kG(s)k ∞ = sup σmax (G( jω )). w

The H∞ norm is useful to enforce the stability robustness of the closed loop with respect to model uncertainties . In addition , the H∞ norm facilitates frequency domain design

Output Feedback H ∞ Loop-Shaping Controller Synthesis

177

specifications such as bandwidth and high frequency roll-off in an H∞ controller synthesis.

The H∞ norm of an LTI system can be computed as the solution to a particular linear matrix inequality optimization problem. Lemma 6.1. (Bounded Real Lemma, see e.g., [6], [5]). Suppose the system G(s) given in (6.1)-(6.2) is stable. The H∞ norm of G(s) is smaller than γ if and only if there exists a symmetric matrix P such that  AT P + PA PB CT T T  B P −γ I D  < 0, C D −γ I 

P > 0.

(6.3)

H2 Performance Definition 6.2. Consider the LTI system (6.1)-(6.2) with D = 0. Assume w(t) ∈ Rnw is a zero mean white noise, with variance E {w(t)w(τ )T } = δ (t − τ )Inw , where δ (t) is the Dirac Delta function. The H2 norm of G is defined as ZT 1 z(t)T z(t)dt (6.4) kGk22 = lim E T →∞ T 0

where the supremum is taken over all nonzero trajectories starting from x(0) = 0. The H2 norm is useful when disturbances/noises acting on the system are well represented by stochastic signals. The H2 objective can be computed as the solution of the LMI optimization problem given in Lemma 6.2. Lemma 6.2. (H2 norm computation, see e.g., [15]). kGk 22 is the global minimum of the optimization problem mintrace(Q)

(6.5)

AP + PAT + BB < 0 Q CP <0 T PC P

6.2.2 Normalized Left Coprime Factorization for LTI Systems This section introduces the notion of coprime factorizations. Left coprime factor representations are introduced here because they will play a central role in the output feedback synthesis algorithms presented in the subsequent sections. ˜ M) ˜ where N, ˜ M˜ are stable is a normalized left Definition 6.3. ( [19], [10]). The pair (N, coprime factorization of G if and only if

178

E. Prempain and I. Postlethwaite

(i) M˜ is square and invertible (ii) G = M˜ −1 N˜ ˜ ˜ = I) (iii) M˜ and N˜ are left coprime(that is,there exist U and V ,both stable,such that MV− NU ˜ jω )N˜ T (− jω ) + M( ˜ jω )M˜ T (− jω ) = I, ω ∈ R (iv) N( Theorem 6.1. Let a realization of the LTI system G(s) be given by (A, B,C, D), where A ∈ Rn×n , B ∈ Rn×nu , C ∈ Rny ×n and D ∈ Rny ×nu . (a) Define R = I + DT D, and R˜ = I + DDT . Suppose (C, A) is detectable and (A, B) has no uncontrollable mode on the imaginary axis. Then there is a normalized left coprime factorization G = M˜ −1 N˜ ! A + LC L B + LD ˜ ˜ , MN = 1 1 1 R˜ − 2 C R˜ − 2 R˜ − 2 D

wher L = −(BDT +YCT )R˜ −1 ,

(6.6)

and Y ≥ 0 is the solution to the Riccati equation

(A − BDT R˜ −1C)Y +Y (A − BDT R˜ −1C)T −YCT R˜ −1CY + BR−1 BT = 0.

(6.7)

(b) L is the minimizer of kF (GOI , Ls )k 2 over the set of real matrices Ls ∈ Rn×ny where   A0B I GOI :=  I 0 0 0  . C I D0

The fact that normalized left coprime factors can be constructed from an H2 optimization problem is interesting because this principle generalizes to parameter-dependent systems. Proof. (a) is a standard result which can be found in [20].

(b) In order to apply the standard H2 results of [3], we need to normalize the term I D in GOI . Using the singular value decomposition, it can be shown that I D = R˜ 1/2 [0 I]U, where U is an orthonormal matrix. Performing this regularization leads to the regular plant G˜ OI

 [0 B]U T I 0 0 0 . :=  1 R˜ − 2 C 0 I 0 

A I

The Hamiltonian matrix corresponding to the H2 norm minimization problem based on G˜ OI T is CT A 0 J2 := − R˜ −1 DBT C . −BBT −A −BDT

Clearly, the Riccati equation corresponding to J2 is equation (6.7). Let Y ≥ 0 be the solution to (6.7). The optimal H2 observer gain is given by L2 = −(YCT + BDT )R˜ −1/2 (see [20]) and L = L2 R˜ −1/2 solves the original H2 minimization problem based on GOI . Now note that the optimal observer gain L is also the observer gain required to obtain the normalized left coprime factors of plant G. This completes the proof. ⊔⊓ ⊓ ⊔

Output Feedback H∞ Loop- Shaping Controller Synthesis

179

6.3 H∞ Synthesis This section is concerned with the standard H∞ control formulation. The existence conditions for an H∞ controller were first given in [3], [8] in terms of two Riccati equations and a spectral radius condition. The Riccati framework provides a very elegant solution to the H∞ control problem. However, it is somewhat less general than the LMI optimization framework in the sense that the plant must satisfy some assumptions (rank conditions) which can be restrictive in practice. Here we briefly present the general H∞ control problem in the LMI optimization framework. The results of this section are standard and can be found in [6], [5], [9]. Let us consider the nth -order linear time-invariant generalized plant P(s) with state space equations x˙ = Ax + Bw w + Bu u

(6.8)

z = Cz x + Dzw w + Dzu u

(6.9)

y = Cy x + Dyw w + Dyu u

(6.10)

where x ∈ Rn is the plant state vector, w ∈ Rnw is the exogenous input which might include disturbances, measurement noises, reference signals etc., z ∈ Rnz is the error signal, u ∈ Rnu is the control input and y ∈ Rny is the measured input. System (6.8)-(6.10) is assumed to be stabilizable and detectable. This assumption ensures the existence of an internally stabilizing output feedback control law u = K(s)y of order n. Also we will assume Dyu = 0. This assumption makes the closed-loop state space matrices linear in the control matrices. It is standard and can always be satisfied via loop loop transformation (see e.g., [20] for details). The γ -suboptimal H∞ synthesis consists of finding an internally stabilizing output control law u = K(s)y that makes the H∞ norm of the closed-loop transfer function Tzw less than γ , where Tzw relates the disturbance w to the error signal z. Given a realization K(s) = CK (sI − AK )−1 BK + DK ,

AK ∈ Rn×n

of the output controller, a realization of the closed loop transfer function from w to z is A + Bu KCy BuCK Bw + Bu DK Dyw , Bcl = (6.11) Acl = BK Cy AK BK Dzw (6.12) Ccl = Cz + Dzw DK Cy DzuCK , Dcl = Dzw + Dzu DK Dyw

The output feedback H∞ control problem is simply solved by applying the Bounded Real Lemma given in Section 6.2 to the closed-loop system (6.11)-(6.12). Thus, if there exist a 2n-by-2n symmetric matrix Xcl > 0 and γ > 0 such that  T  T Acl Xcl + Xcl Acl Xcl Bcl Ccl  (6.13) BTcl Xcl −γ DTcl  < 0 Ccl Dcl −γ I then kTzw k∞ < γ . Now define

180

E. Prempain and I. Postlethwaite

ΩK :=

AK BK DK DK

.

(6.14)

It is easy to show that inequality (6.13) can be rewritten as Z + PT ΩK Q + QT ΩKT P < 0

(6.15)

where the matrices Z, P and Q depend only on the system matrices and on the matrix variable Xcl . Using the Elimination Lemma [2], the existence conditions for an internally γ -suboptimal H∞ controller reduce to solving two simpler LMIs NPT ZNP < 0,

NQT ZNQ < 0

(6.16)

where NP and NQ are bases of the null spaces of P and Q respectively. The next theorem, (Theorem 6.2), gives explicit expressions of inequalities (6.16) in terms of the system plant matrices and two Lyapunov matrices S and R which are related to Xcl in the following way. The matrix S is the matrix formed by the first n rows and n columns of Xcl while R is the matrix formed by the first n rows and n columns of Xcl−1 . The reader is referred to [6], [5] and [9] for more details. Theorem 6.2. (Existence conditions for a γ -suboptimal H∞ controller [6]). Consider the system given in (6.8)-(6.10) and let Nz and Nw be bases of the null spaces of (BTu , DTzu ) and (Cy , Dyw ) respectively. There exists an internally stabilizing output feedback controller K(s) such that the H∞ norm of the closed-loop transfer function from w to z is less than γ if and only if there exists two symmetric matrices R ∈ Rn×n and S ∈ Rn×n satisfying the following system of LMIs:   T AR + RAT RCzT Bw Nz 0  Nz 0 Cz R −γ I Dzw  <0 (6.17) 0 I 0 I DTzw −γ I BTw   T SA + AT S SBw CzT Nw 0  N 0 T T <0 (6.18) Bw S −γ I Dzw  w 0 I 0 I Cz Dzw −γ I RI > 0. (6.19) I S Remark 6.1. The third inequality in Theorem 6.2 ensures that Xcl > 0.

⊔ ⊓

The optimization variables R, S and γ can be computed with efficient LMI solvers [7], [18]. Then it can be shown that matrix Xcl is uniquely determined as the solution of the following linear equation I S R I = (6.20) Xcl 0 NT MT 0 where the matrices M and N are both in Rn×n and satisfy MN T = I − RS.

(6.21)

Once the matrix Xcl is determined, the state space matrices ΩK , of a γ -suboptimal H∞ controller, can be numerically obtained by solving the linear matrix inequality (6.15).

Output Feedback H∞ Loop-Shaping Controller Synthesis

181

It is worth mentioning that a reduced-order controller (i.e of order k < n) can be synthesized if the matrices R and S satisfy the additional rank constraint rank(I − RS) ≤ k.

(6.22)

A static output feedback controller (e.g., a controller with no states) can be obtained if rank(I − RS) = 0, that is, when R = S−1 . The literature is rich in algorithms that try to incorporate rank constraints into an optimization problem. The cone complementary algorithm [4] is one of them. However, because (6.22) is highly non convex in the variables R and S, its incorporation into an optimization problem may lead to formidable numerical difficulties. For this reason, there is still no efficient algorithm for fixed-order control law synthesis.

6.4 H∞ Loop-Shaping One H∞ synthesis procedure that meets stability and performance requirements together is the approach of McFarlane and Glover [10], [11]. The approach is based on the normalized coprime factors of a certain open-loop plant, known as the shaped plant. The method focuses on coprime factor robustness. The Glover–McFarlane design procedure is probably the H∞ design method preferred by most practical engineers since it combines classical open loop-shaping ideas. The method has many advantages over other H∞ control structures (see e.g., [16]) and tends to produce naturally robust controllers. This is essentially because the implicit H∞ minimization objective includes four closed loop sensitivity functions each of which is representative of a certain type of stability robustness. The design procedure is outlined below: 1) Pre and post filters are used to shape the plant singular values to give a desired open loop shape. The augmentation of the plant to be controlled with the shaping filters forms the shaped plant. 2) An H∞ feedback controller which robustly stabilizes with respect to perturbations of the the normalized left coprime factors is synthesized. 3) The feedback controller and shaping filters are combined to form the final controller. The design method was originally proposed by McFarlane and Glover [10] in the Riccati framework. Here we present the robust controller synthesis in terms of LMI optimization. With the LMI framework, we will see that it is relatively easy to extend the H∞ loop-shaping design procedure to low-order controller designs for LTI and polytopic systems.

6.4.1 LMI Formulation of the H∞ Loop-Shaping Controller Synthesis Let Gs be a strictly proper plant of order n having a stabilizable and detectable state-space realization: AB (6.23) Gs := C0

with A ∈ Rn×n ,B ∈ Rn×nu , C ∈ Rny ×n . Gs represents the shaped plant in the Glover–McFarlane H∞ loop-shaping design procedure.

182

E. Prempain and I. Postlethwaite

˜ Such Suppose Gs has a minimal normalized left coprime factorization, that is Gs = M˜ −1 N. coprime factors can be computed with the formulae of Theorem 6.1 given in Section 6.2. With the normalized left coprime factorization of Gs we form the generalized plant depicted in Figure 6.1. A state space representation of the interconnection structure of Figure 6.1 is 

A 0 P :=  C C

−L 0 I I

 B I   0 0

(6.24)

where L = −ZCT and the matrix Z is the unique symmetric positive semi-definite solution to the algebraic Riccati equation AZ + ZAT − ZCT CZ + BBT = 0

(6.25)

w z 2 y

M˜ −1

? l

z1 6 N˜

u

Fig. 6.1. Open loop Glover–McFarlane H∞ loop-shaping interconnection Theorem 6.3. (Existence conditions for a full-order γ -suboptimal H∞ loop-shaping controller). Let L = −ZCT where Z ≥ 0 is the stabilizing solution to (6.25). There exists an output feedback controller K such that

K −1 ˜ −1

I (I + Gs K) M < γ ∞

(6.26)

if γ > 1 and if there exist positive definite matrices R and S solving the inequalities R(A + LC) + (A + LC)T R − γ CT C < 0   AR + RAT − γ BBT RCT −L  CR −γ I I  < 0 −LT I −γ I RI > 0. I S

(6.27) (6.28)

(6.29)

Proof. The proof follows from the application of Theorem 6.2 to the augmented system (6.24). ⊔⊓ ⊓ ⊔

Corollary 6.1. (Existence conditions for a static γ -suboptimal H∞ loop-shaping controller). Let L = −ZCT where Z ≥ 0 is the stabilizing solution to (6.25). There exists a static output feedback controller K such that

Output Feedback H∞ Loop-Shaping Controller Synthesis

K −1 ˜ −1

(I + G K) M s

<γ

I ∞

183 (6.30)

if γ > 1 and if there exists a positive definite matrix R solving the system of inequalities (A + LC)R + R(A + LC)T < 0  AR + RAT − γ BBT RCT −L  CR −γ I I  < 0. T −L I −γ I 

(6.31) (6.32)

Proof. The proof follows from Theorem 6.3 if we set S = R−1 and eliminate the quadratic ⊔⊓ ⊓ ⊔ term −γ CT C in (6.27). Remark 6.2. The LMI conditions of Theorem 6.3 are necessary and sufficient for the existence of a full-order loop-shaping controller while those of Corollary 6.1 are only sufficient for the existence of a static H∞ loop-shaping controller. Therefore, Corollary 6.1 may lead to conservative results. ⊔ ⊓

6.4.2 Controller Reconstruction Full-order H∞ loop-shaping controllers can be computed numerically by solving the inequality (6.15). But it is preferable to use the reconstruction algorithms proposed in [9] and [5], which take advantage of the particular structure of (6.15). For the full-order controller syntheR hin f lmi of the LMI Control toolbox [7] can be directly used sis, the MATLAB command with the generalized plant Gs given in (6.24). For the static controller construction we need to proceed as follows. Since u = −Ky, a statespace realization of the closed-loop is given by

Tzw :=

Acl Bcl Ccl Dcl

 A − BKC BK − L =  −KC −K  . I (C − KC) 

(6.33)

From the Bounded Real Lemma the closed-loop system is stable and the H∞ -norm of Tzw is smaller than γ if and only if there exists a symmetric matrix R > 0 such that   T B Acl R + RATcl RCcl cl Ccl R −γ I Dcl  < 0. (6.34) Ω :=  BTcl DTcl −γ I

In this case,



 where A˜ =  

AR + RAT 0 CR −LT

0 −γ I 0 0

˜ C˜ − C˜ T K T B˜ T < 0 Ω = A˜ − BK    B RCT −L I 0 0  , B˜ =   and C˜ = CR 0 0 I .    D −γ I I 0 I −γ I

(6.35)

For given R and γ , one can solve the LMI in K (6.35) with any LMI solver or, more simply, K can be calculated using the explicit algebraic formulas given in [9], [7]. Static loop-shaping LMI conditions and controller reconstruction formulas for a non strictly proper plant (i.e., D 6= 0) are given in [14].

184

E. Prempain and I. Postlethwaite

6.4.3 Design Procedure for a Static H∞ Loop-Shaping Controller • •

• • •

The nominal plant G and the shaping functions W1 and W2 are combined to form the shaped plant Gs = W2 GW1 . Let (A, B,C) be a realization of Gs . Classical loop-shaping: Select W1 and W2 to get a desired open loop shape and compute the full-order McFarlane–Glover solution. Ensure that the closed-loop H∞ attenuation level γopt is small enough. Next, these weights are used as initial weights in the static output version of the design problem. Compute the matrix R > 0 and the scalar variable γ solution to the LMI system given in Corollary 6.1. If the LMI system is feasible, then make use of the Lyapunov matrix R > 0 and γ to solve the LMI problem (6.35) in K. The final feedback controller KST is then constructed using the static output feedback controller K with the shaping functions W1 and W2 such that KST = W1 KW2 . The order of KST is equal to sum of the orders of the weighting functions.

6.4.4 Static H∞ Flight Control System Design for the Bell 205 Helicopter This section describes the design of a static H∞ regulator which has been implemented on board a Bell 205 helicopter and tested in flight.

Plant Description The controller design was based on a 32-state nonlinear flight mechanic model of the Bell 205, provided by QinetiQ, Bedford. The linearized model corresponds to the hover situation. To simplify the synthesis of the resulting controller this model was first truncated and then residualised to 12 states. The main states which were removed were those associated with the rotor, which could be replaced with their steady state values. For control purposes, the helicopter can be thought of, roughly, as a three axis vehicle, the axes to be controlled being longitudinal, lateral and directional. Theses axes are associated with the three actuator positions producing the blade angle deflections as ordered and described in table 6.1. Generally speaking the longitudinal and lateral cyclic give lateral and longitudinal motion and the tail rotor generates a torque to counteract the main rotor torque and provides yaw motion. A fourth actuator, the collective, produces lift but, in this study, the torque axis is left under direct control of the pilot. An Attitude-Command Attitude-Hold strategy combined with Rate-Command Position-Hold is considered. The primary variables to be controlled are the pitch attitude (θ ), the roll attitude (φ ) and the yaw rate (r). The plant to be controlled is non square since we have 3 control inputs, 3 outputs to be controlled along with two extra rate gyro measurements p and q (table 6.1) used by the controller. In the sequel, G denotes the 12-state linearised model of the Bell 205 about the hover condition.

Static H∞ Helicopter Controller Design For the hover operating point the weights were chosen as W1 = diag(2

s + 0.6 s + 0.4 s + 0.6 ) , 1.4 ,2 s s s

(6.36)

Output Feedback H∞ Loop-Shaping Controller Synthesis

185

Table 6.1. Plant inputs and measured outputs

Description Measured outputs (y) Description Input (u) lateral cyclic actuator θ1c roll attitude φ longitudinal cyclic actuator θ1s pitch attitude θ tail rotor actuator δr yaw rate r roll rate p pitch rate q

W2 = diag(1, 1.1, 0.9, 0.7, 0.9)

(6.37)

Each input channel was augmented with a simple PI pre-compensator. The integral part was used to boost low frequency gain and improve performance. The proportional term was introduced to rectify the phase-lag added by the integrators at cross-over. W2 emphasizes the outputs to be controlled over the other measurements. The reader is referred to [12], [17], for a description of weighting function selection for helicopter control. The synthesis procedure described in section 2 was implemented using the solver in the LMI toolbox [7]. In this case, we get the following static controller   0.8897 −0.1374 0.2178 1.0567 −0.3132 K = −0.1915 −0.7287 −0.0967 −0.2070 −1.5119 (6.38) 0.1433 −0.1182 −1.0402 0.2405 −0.1470

with γ = 2.9877 which represents an upper bound on the H∞ closed-loop performance. The actual real closed-loop gain attenuation achieved with this static controller is equal to 2.69. The controller which is tested and implemented is KST = W1 KW2 , a 3rd-order controller called the static-based H∞ loop-shaping controller.

re f

- KST + 6−

u

-

G

y -

Fig. 6.2. Feedback configuration Figure 6.2 shows the feedback interconnection. The controller is the simple one-degree-ofT freedom regulator KST . The reference vector, re f = φd θd rd 0 0 , includes the three pilot inputs corresponding to the roll, the pitch and the yaw rate demands. Its two last components are identically equal to zero as rate commands are not required here. Time domain responses of the static and full-order loop-shaping controllers are given in Figures 6.3 and 6.4 respectively. Clearly the static and full-order loop-shaping controllers present similar primary responses in pitch and roll. However, we can see, from Figure 6.4, that the full-order controller is less active in the lateral axis (roll and yaw responses are slower). Also, the full-order controller presents important roll to yaw and yaw to roll static cross couplings. The full-order controller time responses are not as good because the weighting functions were finely tuned for the static

186

E. Prempain and I. Postlethwaite

controller and these weighting functions do not correspond to the best choice for the full-order synthesis. It is relatively easy to select the weighting functions to obtain a full-order controller with good performance. However, in this application, it is worth noting that model reduction techniques do not lead to satisfactory low-order controllers. Residualization and Hankel norm model reduction techniques were applied on a set of good performance full-order loop-shaping controllers. The best reduced controller was obtained by Hankel norm model reduction and it was of order 7. Setting the order to 3 in the model reduction algorithm produces controllers with unacceptable robustness and performance properties. More details on this application, including flight test results, can be found in [14].

roll

rad & rad/s

1 0.5 0 −0.5

0

0.5

1

1.5

2

2.5 pitch

3

3.5

4

4.5

5

0

0.5

1

1.5

2

2.5 yaw

3

3.5

4

4.5

5

rad & rad/s

1 0.5 0 −0.5

rad & rad/s

1 0.5

Roll attidude pitch attitude yaw rate roll rate pitch rate

0 −0.5

0

0.5

1

1.5

2

2.5 sec

3

3.5

4

4.5

5

Fig. 6.3. Closed-loop time responses to unity step-demands in roll, pitch and yaw. Static-based output feedback controller

6.5 H∞ Loop-Shaping for Polytopic Systems This section generalizes the loop-shaping controller synthesis results to polytopic systems. Polytopic systems are parameter dependent systems whose state space representations depend affinely on some parameters. In this context, we will suppose that the plant parameters are available for measurement.

Output Feedback H∞ Loop-Shaping Controller Synthesis

187

roll

rad & rad/s

1 0.5 0 −0.5

0

0.5

1

1.5

2

2.5 pitch

3

3.5

4

4.5

5

0

0.5

1

1.5

2

2.5 yaw

3

3.5

4

4.5

5

rad & rad/s

1 0.5 0 −0.5

rad & rad/s

1 Roll attidude pitch attitude yaw rate roll rate pitch rate

0.5 0 −0.5

0

0.5

1

1.5

2

2.5 sec

3

3.5

4

4.5

5

Fig. 6.4. Closed-loop time responses to unity step-demands in roll, pitch and yaw. Full-order controller.

6.5.1 Left Coprime Factors for Polytopic Systems Let a realization of the parameter dependent system G(θ ) be given by A(θ ) B C D

(6.39)

It is assumed that the parameter trajectories θ (t) = (θ1 , . . . , θm )T are bounded, that is, each component of the time-varying parameter θ (t) satisfies, for all t ≥ 0,

θ i ≤ θi (t) ≤ θ¯i . This assumption means that the parameter vector θ belongs to the hypercube Θ characterized by its r = 2m vertices Πi . More precisely, the hypercube Θ is defined as

Θ := [θ 1 , θ¯1 ] × . . . × [θ m , θ¯m ]. and, in this case, the parameter vector can be written as r

θ (t) =

∑ αi (t)Πi ,

i=1

r

αi (t) ≥ 0,

∑ αi (t) = 1.

i=1

The above expression is called a convex decomposition and the αi are called polytopic coordinates. Similarly, the parameter dependent evolution matrix A(θ ) can be rewritten as

188

E. Prempain and I. Postlethwaite r

A(θ ) =

i=1

where Ai = A(Πi ),

r

∑ αi (t)Ai ,

αi (t) ≥ 0,

∑ αi (t) = 1

(6.40)

i=1

i = 1, . . . , r.

Throughout this section, it is assumed that the pairs (A(θ ), B) and (C, A(θ )) are quadratically stabilizable and detectable. Quadratic stabilizability is equivalent to the existence of a state feedback matrix F and a positive definite matrix Q such that (A(θ ) + BF)Q + Q(A(θ ) + BF)T < 0,

θ ∈ Θ.

Similarly, quadratic detectability is equivalent to the existence of an observer matrix L and a positive definite matrix P such that P(A(θ ) + LC) + (A(θ ) + LC)T P < 0,

θ ∈ Θ.

Theorem 6.4. (Left coprime factorization for polytopic systems, [13]). Define R˜ = I + DDT ,

R = I + DT D.

˜ θ) ˜ θ )−1 N( a) There is a left coprime factorization G(θ ) = M( # " A(θ ) + LC L B + LD ˜ ˜ M(θ ) N(θ ) = 1 1 1 R˜ − 2 C R˜ − 2 R˜ − 2 D

with L = −(BDT + P−1CT )R˜ −1

(6.41)

if there exist matrices P > 0 and Z > 0 which are solutions to the optimization problem

min trace(Z) PAi + ATi P −CT C PB −CT D < 0, −R BT P − DT C i = 1, . . . , r Z I > 0. I P

˜ θ )(s) is contractive, that is, at frozen values of parameter θ ˜ θ )(s) N( b) M( ˜ θ )(s)M˜ T (θ )(−s) + N( ˜ θ )(s)N˜ T (θ )(−s) ≤ I, M( s = jω ,

ω ∈R

and the equality holds (e.g., the factorization is normalized) when the polytope reduces to a single point. This result shows that the determination of a left coprime factorization for an LPV system reduces to solving a special H2 filtering problem. Moreover, it can be shown that if the size of the polytope vanishes to zero, the factorization given in (6.41) approaches the normalized left coprime factorization of the nominal plant; see [13] for more details.

Output Feedback H∞ Loop-Shaping Controller Synthesis

189

6.6 LMI Conditions The H∞ loop-shaping design procedure of McFarlane and Glover for LTI systems can now be extended to the class of linear parameter varying LPV systems. Two results are proposed. Theorem 6.5 gives existence conditions for a full-order gain-scheduled LPV controller, while Theorem 6.2 gives existence conditions for a static LPV controller. Theorem 6.5. Let us consider the LPV plant G(θ ) given in (6.39)-(6.40) and let L = −(BDT + P−1CT )R˜ −1 where P > 0 is the solution to the optimization problem of Theorem 6.4. There exists a dynamic output feedback LPV controller K(θ ) such that

K(θ ) −1 ˜ −1

M( θ )K( θ )) θ ) (6.42) (I + G(

<γ

I ∞ for all parameter trajectories in the polytope Θ , if γ > 1 and if there exist positive definite matrices Q and S solving the inequalities: S(Ai + LC) + (Ai + LC)T S − γ CT R˜ −1C < 0,

i = 1, . . . , r   1 T T T Ai Q + QAi − γ BB QC − γ BDT −LR˜ 2 1   CQ − γ DBT R˜ 2  < 0, −γ R˜  1 1 −γ Iny R˜ 2 −R˜ 2 LT

i = 1, . . . , r QI ≥ 0. I S

(6.43)

(6.44) (6.45)

Proof. The proof follows from the application of Theorem 6.4 and makes use of some results given in [1] and [14]. ⊔⊓ ⊓ ⊔ Corollary 6.2. Let us consider the LPV plant G(θ ) given in (6.39)-(6.40) and let L = −(BDT + P−1CT )R˜ −1 where P > 0 is the solution to the optimization problem of Theorem 6.4. There exists a static output feedback LPV controller K(θ ) such that

K(θ ) −1 ˜ −1

(I + G( M( θ )K( θ )) θ )

<γ

I ∞

for all parameter trajectories in the polytope Θ , if γ > 1 and if there exists a positive definite matrix Q solution to the inequality system (Ai + LC)Q + Q(Ai + LC)T < 0, i = 1, . . . , r   1 T T T Ai Q + QAi − γ BB QC − γ BDT −LR˜ 2 1   R˜ 2  < 0, CQ − γ DBT −γ R˜  1 1 −γ Iny R˜ 2 −R˜ 2 LT

i = 1, . . . , r.

Proof. The proof is similar to that of Corollary 6.1.

(6.46)

(6.47) ⊔⊓ ⊓ ⊔

Remark 6.3. In Theorems 6.5 and 6.2, the plant matrices B, C, and D are required to be parameter independent. As is suggested in [1], low pass filters can be appended at the plant input (resp. output) to move the parameter dependence of B (resp. C) to the evolution matrix A. ⊓ ⊔

190

E. Prempain and I. Postlethwaite

6.6.1 Illustrative Example This example illustrates the design of an H∞ loop-shaping controller in the robust gain scheduled framework. We consider the LPV design example given in the LMI control toolbox [7]. We will see that it is possible to obtain good responses in terms of speed, damping and settling time with a low-order LPV regulator. The linear parameter dependent plant G0 represents a simplified model of the dynamics of a missile. G0 is a second-order system with two outputs and one input governed by the following equations  α˙ = −Zα (t)α + q    q˙ = −Mα (t)α + δm G0 : a = −α    zv q=q

(6.48)

The outputs are (azv ) the normalized vertical acceleration and (q) the pitch rate. The control input is (δm ) the fin deflection. The coefficients Zα (t) and Mα (t) are assumed to vary in Zα (t) ∈ [0.5, 4] Mα (t) ∈ [0, 106]

(6.49)

and are available in real-time. G0 is an affine 2-parameter plant, thus its corresponding polytopic representation has four vertices. More precisely, G0 (t) admits the following state space representation (A(t), B,C, D), where A(t) = ∑4k=1 ck (t)Ak , −0.5 1 −4 1 A1 = , A2 = , 0 0 0 0 −0.5 1 −4 1 , A4 = , A3 = −106 0 −106 0 0 −1 0 0 B= ,C = ,D = (6.50) 1 0 1 0 with polytopic coordinates

Mα (t) Zα (t) − 0.5 ), )(1 − 106 3.5 Mα (t) Zα (t) − 0.5 ), (1 − c2 (t) = 106 3.5 Zα (t) − 0.5 Mα (t) , ) c3 (t) = (1 − 106 3.5 Zα (t) − 0.5 Mα (t) ) c4 (t) = ( 106 3.5

c1 (t) = (1 −

(6.51)

In this problem, the vertical acceleration azv must be controlled. For extra realism, the actuator is modelled as a second-order system with a bandwidth of 188rad/s and a damping of 0.7. Similarly, each sensor has a bandwidth of 628 rad/s and a damping of 0.7.

Output Feedback H∞ Loop-Shaping Controller Synthesis

191

Therefore, the plant to be controlled Gaug has 8 states, 1 input, 2 outputs and four vertices. Values for the weighting matrices were chosen as W1 =

W2 =

80 s + 80

1000s+4100 s

0

(6.52)

0 35

(6.53)

W1 is a low pass filter introduced to limit the control action of the polytopic controller. The first entry of W2 behaves like an integrator at DC. It is used to boost the gain of Gaug at low frequency. Now, define the loop-shaping plant as Gs = W2 GaugW1 .

(6.54)

In this case, the open loop shaped plant Gs is a polytopic system with 4 vertices, 10 states, 1 input and 2 outputs. The optimization problem of Theorem 6.4 is feasible and returns two matrices P > 0 and Z > 0 with Z such that mintraceZ = 0.633. Substituting the matrix P into (6.41) gives the following observer matrix   −6.2497e − 002 1.8757e − 003  −3.0161e − 005 1.1060e − 004     −2.7245e − 004 1.5691e − 004     1.2874e − 006 −1.8917e − 003     4.5150e − 003 −7.5989e − 003   (6.55) L=  2.7078e − 002 −1.6726e − 002  .    4.3663e − 001 −7.5981e − 001     −1.0303e − 001 5.9623e − 001     4.1795e − 002 −3.1349e − 001  7.9450e − 002 −6.0183e − 001

Substituting the values of the plant matrices of Gs and the matrix L into the LMI system of Theorem 6.2, leads to a feasible solution Q > 0 with the objective γ = 6.9. With the values of Q and γ a static polytopic controller K can be constructed using a procedure similar to the one given for LTI systems. The final loop-shaping controller KST = W1 KW2 , is a polytopic controller of order 2 with 4 vertices. In this case, it worth noting that the standard polytopic H∞ synthesis procedure of the LMI control toolbox [7] produces a polytopic controller of order 12. The closed-loop time-responses of the controlled system for a unit step demand in azv along the following spiral trajectories of the parameters are given in Figure 6.5. Zα = 2.25 + 1.7e−4t cos(100t) Mα = 50 + 49e−4t sin(100t) Figure 6.5 shows that the time response of the acceleration with the static based LPV controller KST (solid line) is very close to the specified reference. Note that the response with the static LPV controller is better damped than the response provided by the polytopic regulator designed with the procedure of the LMI toolbox [7] (dashed line). Figure 6.6 shows the corresponding control effort. We can see that the low-order LPV controller is a bit more active than the LPV controller of [7]. This is not surprising since KST provides a faster acceleration response.

192

E. Prempain and I. Postlethwaite

step response of the gain−scheduled autopilot 1.4

1.2

1

azv

0.8

0.6

0.4

0.2

0

0

0.1

0.2

0.3

0.4

0.5

time (s)

Fig. 6.5. Unit step closed-loop responses of the gain scheduled autopilots: Static based polyR toolbox [7] (dashed line) topic regulator KST (solid line); controller from the MATLAB LMI

6.7 Conclusions In this chapter, we have addressed the design of LTI and gain-scheduled H∞ loop-shaping controllers. The determination of a left coprime factorization for an LPV system is given as a special robust H2 filtering problem. This can be viewed as a natural extension of the wellknown normalized left coprime factors for LTI systems to LPV systems. Existence conditions for full-order and static LPV loop-shaping controllers are given in terms of linear matrix inequalities. The effectiveness of the proposed approaches are illustrated on a realistic helicopter model example and a polytopic missile model example.

References 1. P. Akparian and P. Gahinet. A Convex Characterzation of Gain-Scheduled H∞ Controllers. IEEE Transactions on Autmatic Control, 40(9):853–864, May 1995. 2. S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear Matrix Inequalities in System and Control Theory. SIAM. Studies in Applied Mathematics, 1994. 3. J Doyle, K. Glover, P. Khargonekar, and B. Francis. State-space solutions to standard H2 and H∞ control problems. IEEE Transactions on Automatic Control, 34(8):831–847, August 1989. 4. L. El Ghaoui, F. Oustry, and M. AitRami. A cone complementary linearization algorithm for static output-feedback and related problems. IEEE Transactions on Automatic Control, 42:1171–1176, August 1997.

Output Feedback H∞ Loop-Shaping Controller Synthesis

193

control action (fin deflection in degrees) 10

5

u

0

−5

−10

−15

0

0.1

0.2

0.3

0.4

0.5

time (s)

Fig. 6.6. Unit step closed-loop responses of the gain scheduled autopilots: Static based polyR toolbox [7] (dashed topic regulator KST (solid line); and controller from the MATLAB LMI line)

5. P. Gahinet. Explicit controller formulas for LMI-based H∞ synthesis. Automatica, 32(7):1007–1014, 1996. 6. P. Gahinet and P. Apkarian. A linear matrix inequality approach to H∞ control. International Journal of Robust and Nonlinear Control, 4:421–448, 1994. 7. P. Gahinet, A. Nemirovski, A. J. Laub, and M. Chilali. LMI control toolbox. The Math Works, May 1995. 8. K. Glover and J. C. Doyle. State-space formulae for alll stabilizing controllers that satisfy an H∞ norm bound and relations to risk sensitivity. Systems and Control Letters, 11:167– 172, 1988. 9. T. Iwasaki and R. E. Skelton. All controllers for the general H∞ control problem: LMI existence conditions and state space formulas. Automatica, 30(8):1307–1317, 1994. 10. D. McFarlane and K. Glover. Robust controller design using normalized coprime factor plant description, Lecture notes in control and information Science. Springer-Verlag, Berlin, 1989. 11. D. McFarlane and K. Glover. A loop-shaping design procedure using H∞ synthesis. IEEE Transactions on Automatic Control, 37(6):759–769, 1992. 12. I. Postlethwaite, A. Smerlas, D. J. Walker, A. W. Gubbels, S. W. Baillie, M. E. Strange, and J. Howitt. H∞ control of the NRC Bell 205 fly-by-wire helicopter. Journal of American Helicopter Society, 44(4):276–284, 1999. 13. E. Prempain. Coprime Factorizations for Parameter-Dependent Systems. ROCOND’06, Toulouse, July 2006. 14. E. Prempain and I. Postlethwaite. Static H∞ loop-shaping control of a fly-by-wire helicopter. Automatica, 41:1517–1528, September 2005. 15. C. Scherer, P. Gahinet, and M. Chilali. Multi-objective output-feedback control via LMI optimization. IEEE Trans. Aut. Cont., 30(8):1307–1317, July 1997.

194

E. Prempain and I. Postlethwaite

16. S. Skogestad and I. Postlethwaite. Multivariable Feedback Control; Analysis and Design. Wiley, 2nd edition, 2005. 17. A. Smerlas, I. Postlethwaite, D. J. Walker, M. E. Strange, J. Howitt, R. I. Horton, A. W. Gubbels, and S. W. Baillie. Design and flight testing of an H∞ controller for the ncr bell 205 experimental fly-by-wire helicopter. AIAA-98-4300. Proc AIAA Guidance Navig. Contr. Conf. Boston, pages 1023–1033, August 1998. R ∗ toolbox for optimization over symmetric cones. 18. J. F. Sturm. SeDuMi-a MATLAB Department of Econometrics, Tilburg University, The Netherlands, October 2001. 19. M. Vidyasagar. Control system synthesis. The MIT Press, 1985. 20. K. Zhou, J. Doyle, and K. Glover. Robust and Optimal Control. Prentice-Hall International, Inc, 1995.

7 Stability and Asymptotic Behaviour of Nonlinear Systems: An Introduction Hartmut Logemann and Eugene P. Ryan

7.1

Introduction

To motivate a study of asymptotic behaviour of nonlinear systems modelled by ordinary differential equations and differential inclusions, we indicate how such equations/inclusions arise naturally in control of dynamical process by feedback. The concept of control pertains to modifying the behaviour of the process, by manipulation of inputs to the process, in order to achieve some prescribed goal. Fundamental to this is the notion of feedback: a strategy in which the inputs to the process are determined on the basis of concurrent observations on (or outputs from) the process. Consider first a finite-dimensional, continuous-time dynamical process, the state of which

inputs

6

-

process strategy

outputs -

?

Fig. 7.1. Control loop evolves in RN and is governed by a controlled ordinary differential equation, with initial data (t0 , x0 ), of the general form x(t) ˙ = g(t, x(t), u(t)),

x(t0 ) = x0 ,

(7.1)

where the function u is the input or control and the output or observation y is generated via an output map c: y(t) = c(t, x(t)). (7.2) Under feedback, the input u(t) at time t is determined by the output y(t) via a feedback map h: u(t) = h(t, y(t)) = h(t, c(t, x(t))). (7.3)

M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 195-220, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

196

H. Logemann and E.P. Ryan

Introducing the function f given by f (t, ξ ) = g(t, ξ , h(t, c(t, ξ ))), we see that the conjunction of (7.1) and (7.3) gives rise to an initial-value problem of the following type. x(t) ˙ = f (t, x(t)),

x(t0 ) = x0 .

(7.4)

Clearly, in order to ensure that this problem is well posed, the function f is required to be sufficiently regular. In due course, we will make precise the requisite regularity conditions which imply continuity – in their second arguments – of g, c and h. Are these continuity conditions reasonable? Perhaps not in the case of the feedback map h – since many feedback strategies are inherently discontinuous e.g., “bang-bang” or “on-off” control actions. A prototype example is the signum function sgn, which can be embedded in the following set-valued map defined on R:  x>0  {+1}, [−1, +1], x = 0. x 7→  {−1}, x<0 The graph of this map is depicted below.

R

Fig. 7.2. Signum function

Extrapolating this prototype, we see that situations can arise in which one resorts to a discontinuous feedback which can be embedded in an appropriately-defined set-valued map H: u(t) ∈ H(t, y(t)) = H(t, c(t, x(t)).

(7.5)

Introducing the set-valued map F given by F(t, ξ ) = {g(t, ξ , u) : u ∈ H(t, c(t, ξ ))}, we see that the conjunction of (7.1) and (7.5) gives rise to an initial-value problem for a differential inclusion: x(t) ˙ ∈ F(t, x(t)), x(t0 ) = x0 . (7.6)

Again, in order to ensure well-posedness of this problem, F is required to be sufficiently regular (in a sense to be made precise in due course). The essence of the paper is therefore a study of existence and asymptotic properties of solutions of initial-value problems of the form (7.4) or (7.6). Qualitative results on the long-term behaviour of such dynamical processes are of great importance in the applications of differential equations, dynamical systems, and control theory to science and engineering. Although Lyapunov’s famous memoire on the stability of motion (published in 1892 in Russian) was translated into French in 1907 and reprinted in the USA in 1949 (it was eventually translated into English [13] by A T Fuller in 1992, a hundred years after the publication of the original), it was only at the end of the 1950s that scientists in the West began to appreciate, use, and further develop Lyapunov’s seminal contributions to stability theory. This contrasted

Stability and Asymptotic Behaviour of Nonlinear Systems

197

with the pre-eminence Lyapunov’s direct method had achieved in the former Soviet Union as a major mathematical tool in the context of linear and nonlinear stability problems. Today, Lyapunov’s direct method is a standard ingredient of the syllabuses of university courses on differential equations, dynamical systems, and control theory taught in mathematics and engineering departments worldwide. With Lyapunov’s direct method as exemplar, the paper provides a self-contained and elementary approach to the analysis of certain aspects of the asymptotic behaviour of solutions of ordinary differential equations and differential inclusions. A compendium of results pertaining to asymptotic behaviour of functions is developed. This is achieved by elementary arguments based on concepts of meagreness and weak meagreness which help to capture certain asymptotic properties of functions. This compendium then forms the basis for a unified approach to various results (including generalizations of LaSalle’s invariance principle) on asymptotic behaviour of solutions of (nonautonomous) ordinary differential equations and (autonomous) differential inclusions. The material presented in the paper is based on [20] and its precursors [11,19]. Before embarking on this presentation, we assemble some terminology, notation and background analytical concepts.

7.2

Terminology and Notation

Throughout, N denotes the set of positive integers, and R + := [0, ∞). The Euclidean inner product and induced norm on RN are denoted by h·, ·i and k · k, respectively. Let A be a nonempty subset of RN , and let h : A → RP . For a subset U of RP , h−1 (U) denotes the preimage of U under h, that is, h−1 (U) := {ξ ∈ A : h(ξ ) ∈ U}; for notational simplicity, if u ∈ RP , then we write h−1 (u) in place of the more cumbersome h−1 ({u}). We recall that h is continuous at a point ξ0 ∈ A if, for every ε > 0, there exists δ > 0 such that kh(ξ0 ) − h(ξ )k ≤ ε for all ξ in A with kξ0 − ξ k ≤ δ . If h is continuous at ξ for all ξ in a subset B of A, then h is said to be continuous on B; if B = A, then we simply say that h is continuous. The function h is uniformly continuous on a subset B of A if, for every ε > 0, there exists δ > 0 such that kh(ξ1 ) − h(ξ2 )k ≤ ε for all points ξ1 and ξ2 of B with kξ1 − ξ2 k ≤ δ ; if B = A, then we say that h is uniformly continuous. It is convenient to adopt the convention that h is uniformly continuous on the empty set 0. / If h is continuous and B ⊂ A is compact, then h is uniformly continuous on B. If h is scalar-valued (that is, if P = 1), then h is lower semicontinuous if lim infξ ′ →ξ h(ξ ′ ) ≥ h(ξ ) for all ξ in A, while h is upper semicontinuous if −h is lower semicontinuous; we remark that h is continuous if, and only if, it is both upper and lower semicontinuous. The Euclidean distance function for a nonempty subset A ⊂ RN is the function dA : RN → R + given by dA (v) = inf{kv − ak : a ∈ A}. The function dA is globally Lipschitz with Lipschitz constant 1, that is, kdA (v) − dA (w)k ≤ kv − wk for all v, w ∈ RN . A function x : R + → RN is said to approach the set A if dA (x(t)) → 0 as t → ∞. For ε > 0, Bε (A) := {ξ ∈ RN : dA (ξ ) < ε } (the ε -neighbourhood of A); for a in RN , we write Bε (a) in place of Bε ({a}). It is convenient to set Bε (0) / = 0. / The closure of A is denoted by cl(A).

7.3

Background Concepts in Analysis

In the context of the real numbers R, Lebesgue measure µ is a map from a set (the Lebesgue measurable sets) of subsets of R to the extended non-negative reals [0, ∞]. It is an extension of the classical notation of length of an interval in R to more complicated sets. Whilst there

198

H. Logemann and E.P. Ryan

exist subsets of R which are not measurable, the Lebesgue measurable sets include, for example, all open and closed sets and all sets obtained from these by taking countable unions and intersections. Lebesgue measure has a number of intuitively appealing properties. For example, (i) if A, B ⊂ R are measurable sets with A ⊂ B, then µ (A) ≤ µ (B), (ii) if A = ∪n∈N Bn is a disjoint union of countably many measurable sets, then µ (A) = ∑n∈N µ (Bn ), (iii) µ is translation invariant, that is, if A ⊂ R is measurable, then its translation by b ∈ R, given by B := {a + B : a ∈ A} is measurable and µ (A) = µ (B). The notion of a set of zero (Lebesgue) measure has a simple characterization: µ (A) = 0 if, for each ε > 0, there exists a countable collection of intervals In of length |In | such that

∑ |In | < ε

and

n∈N

A ⊂ ∪n∈N In .

If A is a closed and bounded interval [a, b], then its Lebesgue measure is the length of the interval µ (A) = b − a and, since each of the sets {a},{b} and {a, b} has measure zero, each of the intervals (a, b], [a, b) and (a, b) also has measure b − a.

Let I ⊂ R be an interval and X ⊂ RN an non-empty set. Two functions x, y : I → X are said to be equal almost everywhere (a.e.) if the subset (of I) on which they differ has zero measure, precisely, if {t ∈ I : x(t) 6= y(t)} is a set of zero measure. Let x : I → X and consider a sequence (xn ) of functions I → X. The sequence (xn ) is said to converge almost everywhere (a.e.) to x if the subset of points t ∈ I at which (xn (t)) fails to converge to x(t) has measure zero. A function x : I → RN is said to be a measurable function if there exists a sequence (xn ) of piecewise constant functions I → X converging almost everywhere to x. We remark that the composition f ◦x of a semicontinuous (upper or lower) function f and a measurable function x is a measurable function. A measurable function x : I → X is said to be essentially bounded if there exists K such that kx(t)k ≤ K for almost every (a.e.) t ∈ I (equivalently, the set of points t ∈ I at which kx(t)k > K has measure zero): the set of such functions x is denoted by L∞ (I; X). A measurable function x : I → X is said to be locally essentially bounded if the restriction of x to every compact (that is, closed and bounded) subset ∞ (I; X). The (Lebesgue) of I is essentially bounded: the set of such functions is denoted by Lloc integral Z x(t)dt

I

of a measurable function x : I → X may be defined via the limit of integrals of a suitably chosen sequence of piecewise constant approximants of x. The function x is said to be integrable if Z

I

kx(t)kdt < ∞.

The set of such integrable functions x is denoted1 by L1 (I; X). The function x is said to be locally integrable if the restriction x|J of x to every compact subinterval J of I is integrable R 1 (I; RN ). A function y (that is, J kx(t)kdt < ∞): the set of such functions x is denoted by Lloc that is the indefinite integral of a locally integrable function x is said to be locally absolutely continuous, that is, a function of the form 1

More generally, for 1 ≤ p < ∞, the set of measurable functions x : I → X with the property that Z kx(t)k p dt < ∞ I

is denoted by L p (I; RN ).

Stability and Asymptotic Behaviour of Nonlinear Systems

t 7→ y(t) =

Z t

199

x(s)ds

c

1 (I; X); moreover, y is differentiable almost everywhere (a.e.), that is, for some c ∈ I and x ∈ Lloc the set of points t ∈ I at which the derivative y(t) ˙ fails to exist has measure zero; furthermore y(t) ˙ = x(t) for almost all t ∈ I. Thus, one may identify locally absolutely continuous functions as those functions for which the Fundamental Theorem of Calculus holds in the context of Lebesgue integration.

7.4

Initial-Value Problems: Existence of Solutions

We proceed to develop a theory of existence of solutions of the initial-value problems (7.4) and (7.6): for clarity of exposition, we will assume t0 = 0 in each and, in the context of the latter problem, we restrict attention to the case of an autonomous differential inclusion x˙ ∈ F(x).

7.4.1

Ordinary Differential Equations

Consider the initial-value problem (7.4) (with t0 = 0), viz. x(0) = x0 ∈ RN .

x(t) ˙ = f (t, x(t)),

(7.7)

The ensuing results pertaining to problem (7.7) can be found in standard texts (see, for example, [1, 8, 28, 30]). In order to make sense of the notion solution of (7.7), it is necessary to impose some regularity on the function f : R+ × RN → RN . A classical result is: if f is continuous, then (7.7) has at least one solution, that is, a continuously differentiable function x : I → RN , on some non-trivial interval I containing 0, with x(0) = x0 and satisfying the differential equation in (7.7) for all t ∈ I. However, to insist on continuity of f in its t dependence is difficult to justify (for example, the t-dependence of the function f may arise from modelling extraneous disturbances impinging on a dynamical system - there is no reason to suppose that such disturbances are continuous). What can we say about existence of solutions in such cases: indeed, how do we even define the concept of solution? Given that we have decided against imposing, on f , continuity with respect to its first argument, we have to contend with the possibility of “solutions” of (7.7) which fail to be continuously differentiable. As a first attempt at arriving at a sensible notion of solution, consider the integrated version of (7.7): Z x(t) = x0 +

t

0

f (s, x(s))ds.

(7.8)

We might now consider a solution of (7.7) to be a function x : [0, ω ) → RN such that (7.8)) holds for all t ∈ [0, ω ). For this definition to have substance, the integral on the righthand side must make sense. As outlined in the previous section, the integral does indeed make sense (as a Lebesgue integral) if the integrand s 7→ f (s, x(s)) is a locally integrable function which implies, in particular, that its indefinite integral is locally absolutely continuous. Therefore, we define a (forward) solution of (7.7) to be a locally absolutely continuous function x : [0, ω ) → RN , 0 < ω ≤ ∞, such that (7.8) holds (or, equivalently, such that x(0) = x0 and the differential equation in (7.7) is satisfied for almost all t ∈ [0, ω )). Consequently, a basic requirement on f is sufficient regularity to ensure that, if x(·) is locally absolutely continuous,

200

H. Logemann and E.P. Ryan

then s 7→ f (s, x(s)) is locally integrable. The following hypotheses (usually referred to as the Carath´eodory conditions are sufficient for this to hold:  for each fixed ξ , f (·, ξ ) is a measurable function;    for each fixed t, f (t, ·) is a continuous function; (H): for each compact K ⊂ RN , there exists a locally integrable function    m such that k f (t, ξ )k ≤ m(t) for all (t, ξ ) ∈ R × K.

A solution x : [0, ω ) → RN of (7.7) is said to be maximal, and [0, ω ) is said to be a maximal interval of existence, if x does not have a right extension that is also a solution. We are now in a position to state the fundamental existence result for the initial-value problem (7.7).

Theorem 7.1. Let f satisfy hypothesis (H). Then, for each x0 ∈ RN , (7.7) has a solution and every solution can be extended to a maximal solution. If x : [0, ω ) → RN is a maximal solution and ω < ∞, then, for every τ ∈ [0, ω ) and every compact set K ⊂ RN , there exists σ ∈ [τ , ω ) such that x(σ ) 6∈ K. This theorem asserts that, under hypothesis (H) and for each x0 , the initial-value problem (7.7) has at least one solution (there may be multiple solutions) and every solution can be maximally extended. By imposing further regularity on f , we can infer the existence of precisely one maximal solution (the uniqueness property). Additional regularity sufficient for uniqueness is the following local Lipschitz condition: (L) :

1 (R ; R ) such that for each compact K ⊂ RN , there exists λ ∈ Lloc + + k f (t, ξ ) − f (t, ζ )k ≤ λ (t)kξ − ζ k for all t ∈ R+ and all ξ , ζ ∈ K.

Theorem 7.2. Let f satisfy hypotheses (H) and (L). Then, for each x0 ∈ RN , (7.7) has a unique maximal solution. Next, we consider the autonomous counterpart of (7.7), viz. x(t) ˙ = f (x(t)),

x(0) = x0 ∈ RN ,

(7.9)

where f : RN → RN is locally Lipschitz: (La) :

for each compact K ⊂ RN , there exists λ > 0 such that k f (ξ ) − f (ζ )k ≤ λ kξ − ζ k for all ξ , ζ ∈ K.

In this autonomous setting, we are interested in solutions of (7.9) in both forwards and backwards time: thus, we deem a continuously differentiable function x : (α , ω ) → Rn to be a solution if 0 ∈ (α , ω ), x(0) = x0 and the differential equation in (7.9) holds for all t ∈ (α , ω ). A solution is maximal if it has no proper left or right extension that is also a solution. Theorem 7.3. Let f satisfy hypothesis (La). Then, for each x0 ∈ RN , the autonomous system (7.9) has unique maximal solution x : (α , ω ) → RN . This theorem implies the existence of a map (t, x0 ) 7→ ϕ (t, x0 ) defined by the property that, for each x0 , ϕ (·, x0 ) is the unique maximal solution x of (7.9). The domain of ϕ is given by D = dom(ϕ ) = {(t, x0 ) ∈ R × RN : t ∈ I(x0 )}

Stability and Asymptotic Behaviour of Nonlinear Systems

201

where I(x0 ) = (α , ω ) denotes the maximal interval of existence of the maximal solution of (7.9). We refer to ϕ as the flow generated by f .

Proposition 7.1. D is an open set and ϕ is continuous. A set S ⊂ RN is said to be ϕ -invariant or invariant under the flow ϕ if, for every x0 ∈ S, the unique maximal solution of (7.9) has trajectory in S.

7.4.2

Autonomous Differential Inclusions

We now turn attention to the autonomous counterpart of the initial-value problem (7.6), viz. x(t) ˙ ∈ F(x(t)),

x(0) = x0 ∈ RN ,

(7.10)

where ξ 7→ F(ξ ) ⊂ RN is a set-valued map defined on RN . There is growing literature (see, for example, [2], [6], [7], [10], [12], [27]) pertaining to the study of differential inclusions. By a (forward) solution of (7.10), we mean a locally absolutely continuous function x : [0, ω ) → RN , 0 < ω ≤ ∞, with x(0) = x0 , such that the differential inclusion in (7.10) is satisfied for almost all t ∈ [0, ω ). A solution x : [0, ω ) → RN is maximal, and [0, ω ) is a maximal interval of existence, if x has no right extension that is also a solution. We proceed to address the issue of identifying regularity conditions on F sufficient to guarantee the existence of at least one solution of (7.10). To this end, let U denote the class of set-valued maps ξ 7→ F(ξ ) ⊂ RN , defined on RN , that (a) take nonempty convex compact values (that is, for each ξ ∈ RN , F(ξ ) is a non-empty, convex and compact subset of RN ) and (b) are upper semicontinuous at each ξ ∈ RN . A set-valued map F is upper semicontinuous at ξ ∈ RN if, for each ε > 0, there exists δ > 0 such that F(ξ ′ ) ⊂ Bε (F(ξ )) for all ξ ′ in Bδ (ξ ), as illustrated below.

ξ′

b

ξ

b

F

F(ξ ′ ) F(ξ )

Bδ (ξ ) Bε (F(ξ )) Fig. 7.3. Upper semi-continuity of set-valued map Theorem 7.4. Let F ∈ U . For each x0 ∈ RN , (7.10) has a solution and every solution can be extended to a maximal solution. If x : [0, ω ) → RN is a maximal solution with ω < ∞, then x is unbounded. A set S ⊂ RN is said to be weakly invariant with respect to the differential inclusion in (7.10) if, for each x0 ∈ S, there exists at least one maximal solution of (7.10) with trajectory in S.

202

H. Logemann and E.P. Ryan

7.4.3 ω -L imit Sets In his well-known book [4, pp. 197], Birkhoff introduced the notion of an ω -limit point in the context of trajectories of dynamical systems. For the purposes of this paper, it is useful to define the concept of an ω -limit point for arbitrary RN -valued functions defined on R+ . Let x : R + → RN . A point ξ ∈ RN is an ω -limit point of x if there exists an unbounded sequence (tn ) ⊂ R + such that x(tn ) → ξ as n → ∞; the (possibly empty) ω -limit set of x, denoted by Ω (x), is the set of all ω -limit points of x. The following two lemmas highlight well-known properties of ω -limit sets (see, for example, [1], [12], [14], and [30]). Lemma 7.1. The following hold for any function x : R + → RN : (a) Ω (x) is closed. (b) Ω (x) = 0/ if and only if kx(t)k → ∞ as t → ∞. (c) If x is continuous and bounded, then Ω (x) is nonempty, compact, and connected, is approached by x, and is the smallest closed set approached by x. (d) If x is continuous and Ω (x) is nonempty and bounded, then x is bounded and x approaches Ω (x). If x happens to be a maximal solution of (7.9) or of (7.10), then we can say more. Lemma 7.2 (a) Let x : R+ → RN be a bounded solution of (7.9). Then Ω (x) is nonempty, compact, connected, is approached by x, is the smallest closed set approached by x, and is invariant under the flow ϕ generated by f . (b) Let x : R+ → RN be a bounded solution of (7.10). Then Ω (x) is nonempty, compact, connected, is approached by x, is the smallest closed set approached by x, and is weakly invariant with respect to the differential inclusion in (7.10).

7.5

Barbalat’s Lemma, LaSalle’s Invariance Principle, and ˘ Lyapunov Stability

A function y : R+ → R is Riemann integrable (on R+ ) if the improper Riemann integral exists, that is, y is Riemann integrable on [0,t] for each t ≥ 0 and the limit 0 y(s)ds R limt→∞ 0t y(s)ds exists and is finite. If y belongs to L1 and is Riemann integrable on [0,t] for each t ≥ 0, then y is Riemann integrable on R+ . First, we highlight the following simple observation, due to Barb˘alat [3]. R∞

Lemma 7.3 (Barb˘alat’s lemma). If y : R + → R is uniformly continuous and Riemann integrable, then y(t) → 0 as t → ∞. Proof. Suppose to the contrary that y(t) 6→ 0 as t → ∞. Then there exist ε > 0 and a sequence (tn ) in R + such that tn+1 − tn > 1 and |y(tn )| ≥ ε for all n in N. By the uniform continuity of y, there exists δ in (0, 1) such that, for all n in N and all t in R + , |tn − t| ≤ δ

=⇒

|y(tn ) − y(t)| ≤ ε /2 .

Stability and Asymptotic Behaviour of Nonlinear Systems

203

Therefore, for all t in [tn ,tn + δ ] and all n in N, |y(t)| ≥ |y(tn )| − |y(tn ) − y(t)| ≥ ε /2, from which it follows that Z Z tn +δ tn +δ εδ = y(t)dt |y(t)|dt ≥ t 2 t n

n

for each n in N, contradicting the existence of the improper Riemann integral

R∞ 0

y(t)dt.

⊓ ⊔

Lemma 7.3 was originally derived in [3] to facilitate the analysis of the asymptotic behaviour of a class of systems of nonlinear second-order equations with forcing. Subsequently, Barb˘alat’s lemma has been widely used in mathematical control theory (see, for example, [9, p. 89], [23, p. 211], and [26, p. 205]). The following corollary is an immediate consequence of statement (c) of Lemma 7.1 and Lemma 7.3. Corollary 7.1. Let G be a nonempty closed subset of RN , and let g : G → R be continuous. Assume that x : R + → RN is bounded and uniformly continuous with x(R+ ) ⊂ G. If g ◦ x is Riemann integrable, then Ω (x) ⊂ g−1 (0) and x approaches g−1 (0). We will use Corollary 7.1 to derive LaSalle’s invariance principle. Let the vector field f : RN → RN be locally Lipschitz and consider the initial-value problem x˙ = f (x) ,

x(0) = x0 ∈ RN .

(7.11)

Let ϕ denote the flow generated by f , and so t 7→ ϕ (t, x0 ) is the unique solution of (7.11) defined on its maximal interval of existence I(x0 ). If R+ ⊂ I(x0 ) and ϕ (· , x0 ) is bounded on R+ , then, by assertion (a) of Lemma 7.2, Ω (ϕ (· , x0 )) is invariant with respect to the flow ϕ . The following integral-invariance principle provides an intermediate step towards LaSalle’s principle and is a consequence of Corollary 7.1. Theorem 7.5 (Integral-invariance principle). Let G be a nonempty closed subset of RN , let g : G → R be continuous, and let x0 be a point of G. Assume that R+ ⊂ I(x0 ), ϕ (· , x0 ) is bounded on R+ , and ϕ (R+ , x0 ) ⊂ G. If the function t 7→ g(ϕ (t, x0 )) is Riemann integrable on R+ , then ϕ (· , x0 ) approaches the largest invariant subset contained in g−1 (0). Proof. Since ϕ (· , x0 ) is bounded on R+ and satisfies the differential equation, it follows that the derivative of ϕ (· , x0 ) is bounded on R+ . Consequently, ϕ (· , x0 ) is uniformly continuous on R+ . An application of Corollary 7.1 together with the invariance property of Ω (ϕ (· , x0 )) establishes the claim. ⊓ ⊔. Before proceeding to derive LaSalle’s principle, we briefly digress to an example which illustrates that the above integral-invariance principle is of independent interest. Example 7.1. Theorem 7.5 is particularly useful in the context of observed systems. In applications, it is frequently impossible to observe or measure the complete state x(t) of (7.11) at time t. To illustrate the latter comment, consider the observed system given by (7.11) and the observation z = c(x) , (7.12) where c : RN → RP is continuous with c(0) = 0. The observation z (also called output or measurement) depends on the state and should be thought of as a quantity that can be observed

204

H. Logemann and E.P. Ryan

x˙ = f (x), x(0) = x0

x

c

z = c(x)

Fig. 7.4. Observed system

or measured: an important special case occurring when z is given by one component of the state. Observability concepts relate to the issue of precluding the possibility that different initial states generate the same observation: the initial state of an observable system can, in principle, be recovered from the observation. The system given by (7.11) and (7.12) is said to be zero-state observable if the following holds for each x0 in RN : z(·) = c(ϕ (· , x0 )) = 0

=⇒

ϕ (· , x0 ) = 0 ,

that is, the system is zero-state observable if x(·) = 0 is the only solution generating the zero observation z(·) = 0. The following corollary of Theorem 7.5 is contained in [5, Theorem 1.3] and essentially states that, for a zero-state observable system, every bounded trajectory with observation in L p necessarily converges to zero. Corollary 7.2. Assume that the observed system given by (7.11) and (7.12) is zero-state observable. For given x0 in RN assume that R+ ⊂ I(x0 ) and that ϕ (· , x0 ) is bounded on R+ . If R∞ 0 p 0 0 kc(ϕ (t, x ))k dt < ∞ for some p in (0, ∞), then limt→∞ ϕ (t, x ) = 0.

Proof. By the continuity and boundedness of ϕ (· , x0 ), it follows from Lemma 7.1 that ϕ (· , x0 ) approaches its ω -limit set Ω := Ω (ϕ (· , x0 )) and that Ω is the smallest closed set approached by ϕ (· , x0 ). An application of Proposition 7.5 with G = RN and g(·) = kc(·)k p shows that Ω ⊂ g−1 (0) = c−1 (0). Let ξ be a point of Ω . By the invariance property of Ω , ϕ (t, ξ ) lies in Ω for all t in R. Consequently, c(ϕ (· , ξ )) = 0. Zero-state observability ensures that ϕ (· , ξ ) = 0, showing that ξ = 0. Hence Ω = {0}, so limt→∞ ϕ (t, x0 ) = 0. ⊓ ⊔ Theorem 7.5 is essentially contained in [5, Theorem 1.2]: the proof given therein is not based on Barb˘alat’s lemma. The above proof of Theorem 7.5 is from [11]. LaSalle’s invariance principle (announced in [15], with proof in [16]) is now a straightforward consequence of Theorem 7.5. For a continuously differentiable function V : D ⊂ RN → R (where D is open), it is convenient to define the directional derivative V f : D → R of V in the direction of the vector field f by V f (ξ ) = h∇V (ξ ), f (ξ )i. Corollary 7.3 (LaSalle’s invariance principle). Let D be a nonempty open subset of RN , let V : D → R be continuously differentiable, and let x0 be a point of D. Assume that R+ ⊂ I(x0 ) and that there exists a compact subset G of RN such that ϕ (R+ , x0 ) ⊂ G ⊂ D. If V f (ξ ) ≤ 0 for all ξ in G, then ϕ (· , x0 ) approaches the largest invariant subset contained in V f−1 (0) ∩ G. Proof. By the compactness of G and the continuity of V on G, the function V is bounded on G. Combining this with Z t 0

V f (ϕ (s, x0 ))ds =

Z t 0

R

(d/ds)V (ϕ (s, x0 ))ds = V (ϕ (t, x0 )) −V (x0 ) ,

we conclude that the function t 7→ 0t V f (ϕ (s, x0 ))ds is bounded from below: but this function is also nonincreasing (because V f ≤ 0 on G) and hence must converge to a finite limit as

Stability and Asymptotic Behaviour of Nonlinear Systems

205

t → ∞. Therefore, the function t 7→ V f (ϕ (t, x0 )) is Riemann integrable on R+ . An application of Theorem 7.5 (with g = V f |G ) completes the proof. ⊓ ⊔ Assume that f (0) = 0, that is, 0 is an equilibrium of (7.11). The equilibrium 0 is said to be stable if for every ε > 0 there exists δ > 0 such that if kx0 k ≤ δ , then R+ ⊂ I(x0 ) and kϕ (t, x0 )k ≤ ε for all t in R+ . The equilibrium 0 is said to be asymptotically stable if it is stable and there exists δ > 0 such that kϕ (t, x0 )k → 0 as t → ∞ for every x0 satisfying kx0 k ≤ δ . Theorem 7.6 (Lyapunov’s stability theorem). Let D be a nonempty open subset of RN such that 0 ∈ D, and let V : D → R be continuously differentiable with V (0) = 0. If V (ξ ) > 0 for all ξ in D \ {0} and V f (ξ ) ≤ 0 for all ξ in D, then 0 is a stable equilibrium. Proof. Let ε > 0 be arbitrary. Without loss of generality, we may assume that the closed ball ¯ ε (0) is contained in D. Since the sphere Sε := {ξ ∈ RN : kξ k = ε } is compact and V is B continuous and positive-valued on Sε , we see that V achieves a minimum value m > 0 on Sε , that is, V (ξ ) ≥ m for all ξ ∈ Sε and V (ξ ) = m for some ξ ∈ Sε .

By continuity of the non-negative valued function V and since V (0) = 0, there exists δ ∈ (0, ε ), such that kξ k < δ =⇒ V (ξ ) < m. Let x0 be such that kx0 k < δ . Let x(·) = ϕ (·, x0 ) be the maximal solution of (7.11) with maximal interval of existence I(x0 ) = (α , ω ). Seeking a contradiction, suppose x(τ ) ∈ Sε for some τ ∈ (0, ω ) and assume τ is the first such time (and so kx(t)k < ε for all t ∈ [0, τ )). Then, d V (x(t)) = V f (x(t)) ≤ 0 dt

∀ t ∈ [0, τ ].

Therefore, t 7→ V (x(t)) is non-increasing on [0, τ ], whence the contradiction m ≤ V (x(τ )) ≤ V (x(0)) = V (x0 ) < m. Therefore, the positive trajectory x([0, ω )) is contained in the closed ball of radius ε centred ⊓ ⊔ at the origin in RN (and so ω = ∞). This completes the proof. Combining Corollary 7.3 and Theorem 7.6, we immediately obtain the following asymptotic stability theorem. Theorem 7.7 (Asymptotic stability theorem). Let D be a nonempty open subset of RN such that 0 ∈ D, and let V : D → R be continuously differentiable with V (0) = 0. If V (ξ ) > 0 for all ξ in D \ {0}, V f (ξ ) ≤ 0 for all ξ in D, and {0} is the largest invariant subset of V f−1 (0), then 0 is an asymptotically stable equilibrium. Example 7.2. In this example, which can also be found in [30], we describe a typical application of Theorem 7.7 in the context of a general class of nonlinear second-order systems. Consider the system y(t) ¨ + r(y(t), y(t)) ˙ = 0,

(y(0), y(0)) ˙ = (p0 , v0 ) ∈ R2 ,

(7.13)

206

H. Logemann and E. P. Ryan

where r : R2 → R is locally Lipschitz and differentiable with respect to its second argument. Furthermore, we assume that r(0, 0) = 0. Setting x(t) = (x1 (t), x2 (t)) = (y(t), y(t)), ˙ the second-order system (7.13) can be expressed in the equivalent form (7.11), where f : R + × R2 → R2 and x0 ∈ R2 are given by f (p, v) = (v, −r(p, v)) ,

x0 = (p0 , v0 ) .

(7.14)

Let ε > 0, set D = (−ε , ε ) × (−ε , ε ), and define V : D → R,

(p, v) 7→

Z p 0

r(s, 0)ds + v2 /2.

It follows from the mean-value theorem that, for each (p, v) in D, there exists a number θ = θ (p, v) in the interval (0, 1) such that V f (p, v) = −v(r(p, v) − r(p, 0)) = −v2

∂r (p, θ v) . ∂v

(7.15)

C LAIM . Consider (7.11) with f and x0 given by (7.14). If pr(p, 0) > 0 for all p in (−ε , ε ) \ {0} and (∂ r/∂ v)(p, v) > 0 for all (p, v) in D satisfying pv 6= 0, then the equilibrium 0 is asymptotically stable. We proceed to establish this claim. Using the hypotheses and (7.15), we infer that V (p, v)>0 for all (p, v) in D \ {0} and V f (p, v) ≤ 0 for all (p, v) in D. Writing ϕ (t, x0 ) = (x1 (t), x2 (t)), we see that for x0 = (p0 , 0) in D with p0 6= 0, x˙2 (0) = −r(p0 , 0) 6= 0. Similarly, for x0 = (0, v0 ) in D with v0 6= 0, x˙1 (0) = v0 6= 0. We conclude that solutions with these initial conditions do not remain in V f−1 (0), showing that {0} is the largest invariant subset of V f−1 (0). The claim now follows from Theorem 7.7. As a special case of (7.13), consider the nonlinear oscillator usually referred to as the Li´enard equation

y(t) ¨ + d(y(t))y(t) ˙ + k(y(t)) = 0 ,

(y(0), y(0)) ˙ = (p0 , v0 ) ∈ R2 ,

where d(y)y˙ represents a friction term that is linear in the velocity and k(y) models a restoring force. We assume that the functions d : R → R and k : R → R are locally Lipschitz and k(0) = 0. It follows from the foregoing discussion on the stability of (7.13) (with r now given by r(p, v) = d(p)v + k(p)) that 0 is an asymptotically stable equilibrium state of the Li´enard equation, provided that there exists ε > 0 such that pk(p) > 0 and d(p) > 0 for all p in (−ε , ε ) with p 6= 0.

7.6

Generalizations of Barbalat’s Lemma ˘

In Theorems 7.8 and 7.9 below, we present generalizations of Barb˘alat’s lemma and of Corollary 7.1, which will be exploited in subsequent analyses of the behaviour of solutions of non-autonomous differential equations and autonomous differential inclusions. To this end, we introduce the notion of (weak) meagreness that will replace the assumption of Riemann integrability in Barb˘alat’s lemma. The concept of meagreness is defined via the Lebesgue measure µ . However, for the purposes of this tutorial paper, we do not wish to assume familiarity with measure theoretic concepts. As an alternative, we also introduce the notion of weak meagreness (which does not require measure theory).

Stability and Asymptotic Behaviour of Nonlinear Systems

207

Definition 7.1 (a) A function y : R + → R is said to be meagre if y is Lebesgue measurable and µ ({t ∈ R + : |y(t)| ≥ λ }) < ∞ for all λ > 0. (b) A function y : R + → R is said to be weakly meagre if

lim ( inf |y(t)|) = 0

n→∞ t∈In

for every family {In : n ∈ N} of nonempty and pairwise disjoint closed intervals In in R + with infn∈N |In | > 0, where |In | denotes the length of the interval In . From Definition 7.1 it follows immediately that a meagre function is weakly meagre. The converse is not true, even in the restricted context of continuous functions. We remark that, if a function y : R+ → R is weakly meagre, then 0 belongs to Ω (y). The property of (weak) meagreness of a function implies that the function is “close to zero” in some sense: however, it is not the case that (weakly) meagre functions converge to zero as t → ∞. Indeed, there exist continuous and unbounded functions that are (weakly) meagre: the following is an example of one such function. Example 7.3. A continuous unbounded meagre function Consider the continuous unbounded function R + → R + , t 7→ x(t) =

∑ xn (t) n∈N

where, for each n ∈ N, xn is continuous and supported on In := [n, n + 1/n2 ], with graph as shown below. For every λ > 0, the total “length” (measure) of the set {t ∈ R + : |x(t)| ≥ λ } xn (t) n 0

b

b

n

n + 1/n2

t

Fig. 7.5. A continuous unbounded meagre function cannot exceed the sum of the lengths |In | = 1/n2 of the intervals In = [n , n + (1/n2 )], whence, µ {t ∈ R + : |x(t)| ≥ λ } ≤

1 < ∞ ∀ λ > 0. 2 n∈N n

∑

and so x is meagre (and, a fortiori, weakly meagre). The above definitions of (weak) meagreness are somewhat obscure: the following result gives more tangible sufficient conditions for meagreness and weak meagreness, respectively. Proposition 7.2. Let y : R+ → R be measurable. Then the following statements hold:

208

H. Logemann and E. P. Ryan

(a) If there exists a lower semicontinuous function α : R+ → R such that α −1 (0) = {0}, infs≥σ α (s) > 0 for all σ > 0, and α (|y(·)|) belongs to L1 , then y is meagre. R (b) If there exists τ > 0 such that limt→∞ tt+τ |y(s)|ds = 0, then y is weakly meagre. R (c) If y is continuous and for every δ > 0 there exists τ in (0, δ ) such that tt+τ y(s)ds converges to 0 as t → ∞, then y is weakly meagre. Proof. We prove only part (c) (the proofs of parts (a) and (b) are even more straightforward). Let y : R+ → R be continuous. We show that ifR y is not weakly meagre, then there exists δ > 0 such that for every τ in (0, δ ) the integral tt+τ y(s)ds does not converge to 0 as t → ∞. The claim follows then from contraposition. So assume that y is not weakly meagre. Then there exists a family {In : n ∈ N} of nonempty, pairwise disjoint closed intervals with δ = infn∈N µ (In ) > 0 and a number ε > 0 such that inft∈In |y(t)| ≥ ε for each n. Since y is continuous, the function y has the same sign on In for each n. Without loss of generality, we may assume that there are infinitely many intervals In on which y is positive. Then there exists a sequence (nk ) in N such that y has positive sign on Ink for all k. Denoting the left endpoint of Ink by tk , we obtain Z tk +τ tk

y(s)ds ≥ ετ > 0,

for each k in N and τ in (0, δ ), showing that the integral t → ∞.

R t+τ t

y(s)ds does not converge to 0 as ⊔ ⊓

It follows immediately from Proposition 7.2(a) that, if y belongs to L p for some p ∈ [1, ∞), then y is meagre (and, a fortiori weakly meagre). The following result will play a role in the subsequent derivation of generalized versions of Barb˘alat’s lemma. Lemma 7.4. Let A and B be nonempty subsets of RN such that cl(Bλ (B)) ⊂ A for some λ > 0. If x : R + → RN is uniformly continuous on x−1 (A), then there exists τ > 0 such that t ∈ R + , x(t) ∈ B

=⇒

x(s) ∈ Bλ (B) ∀ s ∈ [t − τ , t + τ ] ∩ R + .

(7.16)

Proof. Seeking a contradiction, suppose that property (7.16) does not hold. Then there exist sequences (sn ) and (tn ) in R + such that x(tn ) ∈ B and x(sn ) 6∈ Bλ (B) for all n, and sn −tn → 0 as n → ∞. Evidently, sn 6= tn for all n. Define In to be the closed interval with left endpoint min{sn ,tn } and right endpoint max{sn ,tn }, and write Tn = {s ∈ In : s 6∈ x−1 (Bλ (B))}. For each n, let τn in Tn (a compact set) be such that |τn − tn | = min |s − tn |. s∈Tn

Clearly, dB (x(τn )) = λ and dB (x(tn )) = 0 for each n. Combining this information with the facts that τn belongs to In and limn→∞ (sn − tn ) = 0, we conclude that (i) kx(tn ) − x(τn )k ≥ |dB (x(tn )) − dB (x(τn ))| = λ > 0,

(ii) tn , τn ∈ x−1 (A),

(iii) |tn − τn | → 0 as n → ∞, contradicting the hypothesis of the uniform continuity of x on x−1 (A). Therefore, property (7.16) holds. ⊔ ⊓

Stability and Asymptotic Behaviour of Nonlinear Systems

209

The following two theorems, the main results of this section, provide our generalizations of Barb˘alat’s lemma. Theorem 7.8. Let G be a nonempty closed subset of RN , let g : G → R be a function, and let x : R + → RN be continuous with x(R+ ) ⊂ G. Assume that each ξ in G for which g(ξ ) 6= 0 has a neighbourhood U such that inf{|g(w)| : w ∈ G ∩U} > 0

(7.17)

and x is uniformly continuous on x−1 (U). If g ◦ x is weakly meagre, then the following statements hold: (a) (b) (c) (d)

Ω (x) is contained in g−1 (0). If g−1 (0) is bounded and Ω (x) 6= 0, / then x is bounded and x approaches g−1 (0). If x is bounded, then g−1 (0) 6= 0/ and x approaches g−1 (0). If x is bounded and g−1 (0) is totally disconnected, then Ω (x) consists of a single point x∞ which lies in g−1 (0) (in particular, limt→∞ x(t) = x∞ ).

Proof. If Ω (x) = 0, / then statement (a) holds trivially. Now assume that Ω (x) 6= 0. / Let ξ be a point of Ω (x). Since G is closed and x(R+ ) ⊂ G, Ω (x) ⊂ G and thus ξ belongs to G. We show that g(ξ ) = 0. Seeking a contradiction, suppose that g(ξ ) 6= 0. By the hypotheses, there exists a neighbourhood U of ξ such that (7.17) holds and x is uniformly continuous on x−1 (U). Choose δ > 0 such that the closure of Bδ (ξ ) lies in U. Then

ε = inf{|g(w)| : w ∈ G ∩ Bδ (ξ )} > 0 .

(7.18)

Choose δ1 in (0, δ ). Since ξ is an element of Ω (x), there exists a sequence (tn ) in R+ with tn+1 − tn > 1 and x(tn ) in Bδ1 (ξ ) for all n. An application of Lemma 7.4 (with A = U, B = Bδ1 (ξ ) and λ = δ − δ1 ) shows that there exists τ in (0, 1) such that x(t) is in Bδ2 (ξ ) for all t in ∪n∈N [tn ,tn + τ ]. Therefore, by (7.18), |(g ◦ x)(t)| ≥ ε

(t ∈ [tn ,tn + τ ], n ∈ N) .

(7.19)

Finally, since tn+1 −tn > 1 for all n and τ belongs to (0, 1), the intervals [tn ,tn + τ ] are pairwise disjoint. Combined with (7.19) this contradicts the weak meagreness of g ◦ x and establishes (a). A combination of statement (a) and Lemma 7.1 yields statements (b)-(d). ⊔ ⊓ We remark that lower semicontinuity of the function ξ 7→ |g(ξ )| is sufficient to ensure that (7.17) holds for some neighbourhood U of any ξ in G with g(ξ ) 6= 0. Barb˘alat’s lemma follows immediately from an application of Theorem7.8(b) tothesituation wherein N = 1, G = R, g = idR , and x = y, in conjunction with the observation that a uniformly continuous and Riemann integrable function y : R+ → R is weakly meagre, implying that 0 is a member of Ω (y) and thus ensuring that Ω (y) 6= 0. / Corollary 7.1 is a simple consequence of statements (a) and (c) of Theorem 7.8. When compared with Theorem 7.8, the next result (Theorem 7.9) posits that x be uniformly continuous on x−1 (Bε (g−1 (0))) for some ε > 0. We remark that, in certain situations (for

210

H. Logemann and E.P. Ryan

example, if g−1 (0) is finite), this assumption is weaker than the uniform continuity assumption imposed on x in Theorem 7.8. On the other hand, the assumption imposed on g in Theorem 7.9 is stronger than that in its counterpart in Theorem 7.8. However, under these modified hypotheses, Theorem 7.9 guarantees that x approaches g−1 (0) 6= 0/ without assuming the nonemptiness of Ω (x) or the boundedness of x. Theorem 7.9. Let G be a nonempty closed subset of RN , and let g : G → R be such that g−1 (0) is closed and, for every nonempty closed subset K of G, K ∩ g−1 (0) = 0/

=⇒

inf |g(ξ )| > 0 .

ξ ∈K

(7.20)

Furthermore, let x : R + → RN be continuous with x(R+ ) ⊂ G. If (i) x is uniformly continuous on x−1 (Bε (g−1 (0))) for some ε > 0 and (ii) g ◦ x is weakly meagre, then the following statements hold: (a) g−1 (0) 6= 0, / x approaches g−1 (0), and Ω (x) is contained in g−1 (0). (b) If g−1 (0) is bounded, then x is bounded, x approaches g−1 (0), and Ω (x) is a nonempty subset of g−1 (0). (c) If g−1 (0) is bounded and totally disconnected, then Ω (x) is a singleton {x∞ }, where x∞ is a point of g−1 (0) (hence, limt→∞ x(t) = x∞ ). Proof. For convenience, we set Z = g−1 (0). It is clear that Z 6= 0/ (otherwise, by (7.20) and the closedness of G, γ = infξ ∈G |g(ξ )| > 0 and so |g(x(t))| ≥ γ for all t in R + , which contradicts the weak meagreness of g ◦ x). To prove statements (a) and (b), it now suffices to show that x approaches Z. From the closedness of Z it then follows immediately that Ω (x) ⊂ Z; moreover, if Z is bounded, then we can conclude that x is bounded and so Ω (x) 6= 0. / Since, by assumption, the trajectory of x is contained in G, it is immediate that, if G = Z, then x approaches Z. Consider the remaining case, wherein Z is a proper subset of G. By the closedness of Z, there / For θ in (0, δ ), define exists δ in (0, ε /3) such that G \ Bδ (Z) 6= 0.

ι (θ ) = inf{|g(ξ )| : ξ ∈ G \ Bθ (Z)} > 0 , wherein positivity is a consequence of (7.20) and the closedness of G \ Bθ (Z). Seeking a contradiction, we suppose that limt→∞ dZ (x(t)) 6= 0. Then there exist λ in (0, δ ) and a sequence (tn ) in R+ with tn → ∞ as n → ∞ and dZ (x(tn )) ≥ 3λ for all n. By the weak meagreness of g ◦ x, there exists a sequence (sn ) in R+ with sn → ∞ as n → ∞ and |g(x(sn ))| < ι (λ ) for all n, so dZ (x(sn )) ≤ λ for all n. Extracting subsequences of (tn ) and (sn ) (which we do not relabel), we may assume that sn is in (tn ,tn+1 ) for all n. We now have dZ (x(tn )) ≥ 3λ ,

dZ (x(sn )) ≤ λ ,

sn ∈ (tn ,tn+1 )

for all n. By the continuity of dZ ◦ x, there exists for each n a number σn in (tn , sn ) such that x(σn ) belongs to B := {ξ ∈ G : dZ (ξ ) = 2λ }. Extracting a subsequence (which, again, we do not relabel), we may assume that σn+1 − σn > 1 for all n. Noting that cl Bλ (B) ⊂ Bε (Z) and invoking Lemma 7.4 (with A = Bε (Z)), we conclude the existence of τ in (0, 1) such that dZ (x(t)) ≥ λ for all t in [σn , σn + τ ] and all n. Therefore, {t ∈ R + : |g(x(t))| ≥ ι (λ )} ⊃ ∪n∈N [σn , σn + τ ],

Stability and Asymptotic Behaviour of Nonlinear Systems

211

which (on noting that the intervals [σn , σn + τ ], are each of length τ > 0 and form a pairwise disjoint family) contradicts the weak meagreness of g ◦ x. Therefore, x approaches Z, implying that statements (a) and (b) hold. Finally, invoking the fact that the ω -limit set of a bounded continuous function is connected, we infer statement (c) from statement (b). ⊔ ⊓

7.7

Nonautonomous Ordinary Differential Equations

Consider again the initial-value problem (7.7) for a nonautonomous ordinary differential equation x(t) ˙ = f (t, x(t)),

x(0) = x0 ∈ RN ,

(7.21)

where f : R + ×RN → RN is a Carath´eodory function, that is, f satisfies hypothesis (H). Recall that, by Theorem 7.1, for each x0 ∈ RN , (7.21) has at least one solution and every solution can be extended to a maximal solution. With a view to highlighting a particular subclass of Carath´eodory functions f , we introduce the notion of uniform local integrability. 1 and Definition 7.2. A function m : R + → R is uniformly locally integrable if m belongs to Lloc if for each ε > 0 there exists τ > 0 such that

Z t+τ t

for all t in R + .

|m(s)|ds ≤ ε

Clearly, a locally integrable function m : R + → R is uniformly locally integrable if, and only R if, the function t 7→ 0t |m(s)|ds is uniformly continuous. It is readily verified that, if m belongs to L p for some p (1 ≤ p ≤ ∞), then m is uniformly locally integrable. We now introduce a particular subclass of Carath´eodory functions. Definition 7.3. For a nonempty subset A of RN , F(A) denotes the class of Carath´eodory functions f : R + × RN → RN with the property that there exists a uniformly locally integrable function m such that k f (t, ξ )k ≤ m(t) for all (t, ξ ) in R + × A. The next proposition shows that under suitable uniform local integrability assumptions relating to f , solutions of (7.21) satisfy the uniform continuity assumptions required for an application of Theorems 7.8 and 7.9. Proposition 7.3. Let A and B be nonempty subsets of RN with the property that Bε (A) ∩ B 6= 0/ for some ε > 0, and let f belong to F(Bε (A) ∩ B). If x : R + → RN is a global solution of (7.21) such that x(R+ ) ⊂ B, then x is uniformly continuous on x−1 (A). Proof. If x−1 (A) = 0, / then the claim holds trivially. Assume that x−1 (A) 6= 0. / Since f belongs to F(Bε (A) ∩ B), there exists a uniformly locally integrable function m such that k f (t, w)k ≤ m(t) for all (t, w) in R + × (Bε (A) ∩ B). Let δ in (0, ε ) be arbitrary. Choose τ > 0 such that R t+τ m ≤ δ for all t in R + . Let t1 and t2 be points of x−1 (A) with 0 ≤ t2 − t1 ≤ τ . We will t complete the proof by showing that kx(t2 ) − x(t1 )k ≤ δ . If we define J = {t > t1 : x(s) ∈ Bε (A) for all s ∈ [t1 ,t]}, it follows that kx(t) − x(t1 )k ≤

Z t t1

m(s)ds ≤

Z t1 +τ t1

m(s)ds ≤ δ

for all t in J with t ≤ t1 + τ . Since δ < ε , t1 + τ belongs to J, whence kx(t2 ) − x(t1 )k ≤ δ .

⊔ ⊓

212

H. Logemann and E.P. Ryan

In the following, we combine Proposition 7.3 with Theorems 7.8 and 7.9 to derive results on the asymptotic behaviour of solutions of (7.21).

Theorem 7.10. Let G be a nonempty closed subset of RN , and let g : G → R be a function. Assume that each ξ in G for which g(ξ ) 6= 0 has a neighbourhood U such that (7.17) holds and f belongs to F(U ∩ G). If x : R + → RN is a global solution of (7.21) with x(R+ ) ⊂ G and g ◦ x is weakly meagre, then statements (a)-(d) of Theorem 7.8 hold. Proof. Let ξ in G be such that g(ξ ) 6= 0. By the hypotheses, there exists a neighbourhood U of ξ such that (7.17) holds and f belongs to F(U ∩ G). Let ε > 0 be sufficiently small so that B2ε (ξ ) lies in U. Then, setting A = Bε (ξ ), we see that f is in the class F(Bε (A) ∩ G). By Proposition 7.3, it follows that x is uniformly continuous on x−1 (A). An application of Theorem 7.8 completes the proof. ⊓ ⊔ We remark that Theorem 7.10 contains a recent result by Teel [29, Theorem 1] as a special case. In the next theorem, it is assumed that f is a member of F(Bε (g−1 (0)) ∩ G) for some ε > 0. Under the additional assumption that g satisfies (7.20), it is then guaranteed that x approaches g−1 (0) (without positing the boundedness of x). Theorem 7.11. Let G be a nonempty closed subset of RN , and let g : G → R be such that g−1 (0) is closed and (7.20) holds for every nonempty closed subset K of G. Assume that f belongs to F(Bε (g−1 (0)) ∩ G)) for some ε > 0. If x : R + → RN is a global solution of (7.21) with x(R+ ) ⊂ G and g ◦ x is weakly meagre, then statements (a)-(c) of Theorem 7.9 hold. Proof. Fix δ in (0, ε ). By Proposition 7.3, x is uniformly continuous on the set x−1 (Bδ (g−1 (0))). An application of Theorem 7.9 completes the proof. ⊓ ⊔ In the following we use Theorem 7.10 to obtain a version of a well-known result on ω limit sets of solutions of nonautonomous ordinary differential equations. For a nonempty open subset D of RN and a continuously differentiable function V : R+ × D → R, we define V f : R+ ×D → R (the derivative of V with respect to (7.21) in the sense that (d/dt)V (t, x(t)) = V f (t, x(t)) along a solution x of (7.21)) by V f (t, ξ ) =

N ∂V ∂V (t, ξ ) + ∑ (t, ξ ) fi (t, ξ ) ∂t ∂ i=1 ξi

for all (t, ξ ) ∈ R+ × D, where f1 , . . . , fN denote the components of f . Corollary 7.4. Let D be a nonempty open subset of RN , and let V : R+ × D → R be continuously differentiable. Assume that V satisfies the following two conditions: (a) for each ξ in cl(D) there exists a neighbourhood U of ξ such that V is bounded from below on the set R+ × (U ∩ D); (b) there exists a lower semicontinuous continuous function W : cl(D) → R+ such that V f (t, ξ ) ≤ −W (ξ ) for all (t, ξ ) in R+ × D. Furthermore, assume that for every ξ in cl(D) there exists a neighbourhood U ′ of ξ such that f belongs to F(U ′ ∩ D). Under these assumptions, if x : R+ → RN is a global solution of (7.21) with x(R+ ) ⊂ D, then Ω (x) ⊂ W −1 (0).

Stability and Asymptotic Behaviour of Nonlinear Systems

213

Proof. If Ω (x) = 0/ there is nothing to prove, so we assume that Ω (x) 6= 0. / Since (d/dt)V (t x(t)) , = V f (t, x(t)) for all t in R+ , it follows from assumption (b) that the function t 7→ V (t, x(t)) is nonincreasing, showing that the limit l of V (t, x(t)) as t → ∞ exists, where possibly l = −∞. Let ξ ∈ Ω (x) ⊂ cl(D). Then there exists a nondecreasing unbounded sequence (tn ) in R+ such that limn→∞ x(tn ) = ξ . By assumption (a) there exists a neighbourhood U of ξ such that V is bounded from below on R+ × (U ∩ D). Now x(R+ ) ⊂ D, so there exists n0 such that x(tn ) ∈ U ∩ D whenever n ≥ n0 . Consequently, the nonincreasing sequence (V (tn , x(tn ))) is bounded from below, showing that l > −∞. Therefore 0≤

Z ∞ 0

(W ◦ x)(t)dt ≤ −

Z ∞ 0

V f (t, x(t))dt =−

Z ∞ 0

(d/dt)V (t, x(t))dt = V (0, x0 ) − l < ∞ ,

verifying that W ◦ x is in L1 , hence is weakly meagre. By assumption, for each ξ in cl(D) there exists an open neighbourhood U ′ of ξ such that f belongs to F(U ′ ∩ D), implying that f is in F(U ′ ∩ cl(D)). Therefore, an application of Theorem 7.10 with G = cl(D) and g = W establishes the claim. ⊓ ⊔ Corollary 7.4 is essentially due to LaSalle [17] (see also [14, Satz 6.2, p. 140]). However, we point out that the assumption imposed on f in Corollary 7.4 is weaker then that in [14] and [17], wherein it is required that, for every ξ in cl(D), there exists a neighbourhood U ′ of ξ such that f is bounded on the set R+ × (U ′ ∩ D). Furthermore, we impose only lower semicontinuity on the function W (in contrast to [14] and [17], wherein continuity of W is assumed). The next result is a consequence of Theorem 7.11. It shows, in particular, that under a mild assumption on f every global L p -solution of (7.21) converges to zero. Corollary 7.5. Assume that there exists ε > 0 such that f belongs to F(Bε (0)), and let x : R+ → RN be a global solution of (7.21). Then the following statements hold: (a) If kx(·)k is weakly meagre, then limt→∞ x(t) = 0. (b) If x belongs to L p for some p in (0, ∞), then limt→∞ x(t) = 0. Proof. If kx(·)k is weakly meagre, then an application of Theorem 7.11 with G = RN and g = k · k shows that limt→∞ x(t) = 0. This establishes statement (a). To prove statement (b), let x be in L p for some p in (0, ∞). Then, by Proposition 7.2(a), the function kx(·)k is meagre and hence is weakly meagre. By part (a) of the present result, limt→∞ x(t) = 0. ⊓ ⊔ Obviously, if (7.21) is autonomous (i.e., the differential equation in (7.21) has the form x(t) ˙ = f (x(t))), then the assumption that f belongs to F(Bε (0)) for some ε > 0 is trivially satisfied. Thus we may conclude that every weakly meagre global solution t 7→ x(t) of an autonomous ordinary differential equation converges to 0 as t → ∞.

7.8

Autonomous Differential Inclusions

In Section 5, we investigated the behaviour of systems within the framework of ordinary differential equations with Carath´eodory righthand sides. However, as already alluded to in the

214

H. Logemann and E.P. Ryan

Introduction, there are many meaningful situations wherein this framework is inadequate for purposes of analysis of dynamic behaviour. A prototypical example is that of a mechanical system with Coulomb friction which, formally, yields a differential equation with discontinuous right-hand side. Other examples permeate control theory and applications: a canonical case is a discontinuous feedback strategy associated with an on-off or switching device. Reiterating earlier comments, such discontinuous phenomena can be handled mathematically by embedding the discontinuities in set-valued maps, giving rise to the study of differential inclusions. The next goal is to extend our investigations on ordinary differential equations in this direction. Recall that U denotes the class of set-valued maps ξ 7→ F(ξ ) ⊂ RN , defined on RN , that are upper semicontinuous at each ξ in RN and take nonempty, convex, and compact values. The object of our study is the initial-value problem (7.10) for an autonomous differential inclusion, viz. x(t) ˙ ∈ F(x(t)), x(0) = x0 ∈ RN , F ∈ U . (7.22)

Recall that, by Theorem 7.4, for each x0 ∈ RN , (7.22) has at least one solution and every solution can be extended to a maximal solution. The following proposition shows that, under suitable local boundedness assumptions on F, the solutions of (7.22) satisfy the uniform continuity assumptions required for an application of Theorems 7.8 and 7.9. For a subset A of RN and for a member F of U we denote (in a slight abuse of notation) the set ∪a∈A F(a) by F(A). Proposition 7.4. Let A and B be subsets of RN . Assume that F(Bε (A) ∩ B) is bounded for some ε > 0 and that x : R + → RN is a solution of (7.22) with x(R+ ) ⊂ B. Then x is uniformly continuous on x−1 (A). / then the assertion holds trivially. Assume that x−1 (A) 6= 0, / and let δ Proof. If x−1 (A) = 0, in (0, ε ) be arbitrary. Define θ = sup{kvk : v ∈ F(Bε (A) ∩ B)}, and let τ > 0 be sufficiently small so that τθ ≤ δ . Adopting an argument similar to that used in the proof of Proposition 7.3, it can be shown that kx(t2 ) − x(t1 )k ≤ δ for all t1 and t2 in x−1 (A) with 0 ≤ t2 − t1 ≤ τ , proving that x is uniformly continuous on x−1 (A). ⊔ ⊓ We now invoke Theorems 7.8 and 7.9 to derive counterparts of Theorems 7.10 and 7.11 for differential inclusions. Theorem 7.12. Let G be a nonempty closed subset of RN , let g : G → R have the property that each ξ in G for which g(ξ ) 6= 0 has a neighbourhood U such that (7.17) holds. If x : R + → RN is a solution of (7.22) with x(R+ ) ⊂ G and g ◦ x is weakly meagre, then statements (a) and (d) of Theorem 7.8 hold. Moreover, the following statements are true: (b′ ) If g−1 (0) is bounded and Ω (x) 6= 0, / then x is bounded and x approaches the largest subset of g−1 (0) that is weakly invariant with respect to (7.22). (c′ ) If x is bounded, then g−1 (0) 6= 0/ and x approaches the largest subset of g−1 (0) that is weakly invariant with respect to (7.22). Proof. Let ξ in G be such that g(ξ ) 6= 0. By hypothesis, there exists ε > 0 such that (7.17) holds with U = Bε (ξ ). By the upper semicontinuity of F, together with the compactness of its values, F(Bε (U) ∩ G) is bounded (see [2, Proposition 3, p. 42]. By Proposition 7.4, x is uniformly continuous on x−1 (U). Therefore, the hypotheses of Theorem 7.8 are satisfied, so statements (a)-(d) thereof hold. Combining statements (b) and (c) of Theorem 7.8 with the ⊔ ⊓ weak invariance of Ω (x) yields statements (b′ ) and (c′ ).

Stability and Asymptotic Behaviour of Nonlinear Systems

215

Theorem 7.13. Let G be a nonempty closed subset of RN , let g : G → R be such that g−1 (0) is closed and (7.20) holds for every nonempty closed subset K of G. Assume that F(Bε (g−1 (0))∩ G) is bounded for some ε > 0. If x : R + → RN is a global solution of (7.21) with x(R+ ) ⊂ G and g ◦ x is weakly meagre, then statements (a) and (c) of Theorem 7.9 hold. Moreover, the following also holds: (b′ ) If g−1 (0) is bounded, then x is bounded and x approaches the largest subset of g−1 (0) that is weakly invariant with respect to (7.22). Proof. Fix δ in (0, ε ). By Proposition 7.4, x is uniformly continuous on the set x−1 (Bδ (g−1 (0))). It follows immediately from Theorem 7.9 that statements (a)-(c) thereof hold. Assuming that g−1 (0) is bounded, a combination of statements (b) of Theorem 7.9 with the weak invariance of Ω (x) yields statement (b′ ). ⊓ ⊔ If there exists a locally Lipschitz function f : RN → RN such that F(x) = { f (x)} (in this case, the differential inclusion (7.22) “collapses” to an autonomous differential equation which, for every x0 ∈ RN , has a unique solution satisfying x(0) = x0 ), then the conclusions of Theorems 7.12 and 7.13 remain true when every occurence of “weakly invariant” is replaced with “invariant”. We mention that precursors of Theorems 7.12 and 7.13 have appeared in [11] and [25]. Next, we exploit Theorem 7.13 to generalize LaSalle’s invariance principle(see Corollary 7.3) to differential inclusions. Corollary 7.6. Let D be a nonempty open subset of RN , let V : D → R be continuously differentiable, and set VF (ξ ) = maxy∈F(ξ ) h∇V (ξ ), yi for all ξ in D. Let x : R+ → RN be a solution of (7.22) and assume that there exists a compact subset G of RN such that x(R+ ) ⊂ G ⊂ D. If VF (ξ ) ≤ 0 for all ξ in G, then x approaches the largest subset of VF−1 (0) ∩ G that is weakly invariant with respect to (7.22). Proof. For later convenience, we first show that the function VF : D → R is upper semicontinuous. Let (ξn ) be a convergent sequence in D withlimit ξ in D. Define l = lim supn→∞ VF (ξn ). From (VF (ξn )) extract a subsequence VF (ξnk ) with VF (ξnk ) → l as k → ∞. For each k, let yk be a maximizer of the continuous function y 7→ h∇V (ξnk ), yi over the compact set F(ξnk ), so VF (ξnk ) = h∇V (ξnk ), yk i. Let ε > 0 be arbitrary. By upper semicontinuity of F, F(ξnk ) ⊂ Bε (F(ξ )) for all sufficiently large k. Since yk lies in F(ξnk ), F(ξ ) is compact and ε > 0 is arbitrary, we infer that (yk ) has a subsequence (which we do not relabel) converging to a point y∗ in F(ξ ). Therefore, lim sup VF (ξn ) = l = lim VF (ξnk ) = lim h∇V (ξnk ), yk i n→∞

k→∞

k→∞

= h∇V (ξ ), y∗ i ≤ VF (ξ ) , confirming that VF is upper semicontinuous. Evidently, d V (x(t)) = h∇V (x(t)), x(t)i ˙ ≤ VF (x(t)) ≤ 0 dt for almost every t in R + , which leads to

216

H. Logemann and E.P. Ryan

V (x(t)) −V (x(0)) ≤

Z t 0

VF (x(s))ds ≤ 0

(7.23) R

for all t in R+ . Since x is bounded, we conclude that the function t 7→ 0t VF (x(s))ds is bounded from below. But this function is also nonincreasing (because VF ≤ 0 on G), which ensures that R limt→∞ 0t VF (x(s))ds exists and is finite. Consequently, VF ◦ x is an L1 -function, showing that VF ◦ x is weakly meagre. Since VF is upper semicontinuous and VF ≤ 0 on G, the function G → R given by ξ 7→ |VF (ξ )| is lower semicontinuous. Therefore, each ξ in G with VF (ξ ) 6= 0 has a neighbourhood U such that inf{|VF (w)| : w ∈ G ∩U} > 0. By statement (c′ ) of Theorem 7.12 (with g = VF |G ) it follows that x approaches the largest subset of VF−1 (0) ∩ G that is weakly invariant with respect to (7.22). ⊓ ⊔ In Corollary 7.6, it is assumed that the solution x is global (that is, defined on R+ ) and has trajectory in some compact subset G of D. These assumptions may be removed at the expense of strengthening the conditions on V by assuming that its sublevel sets are bounded and that VF (ξ ) ≤ 0 for all ξ in D. Corollary 7.7. Let D, V , F, and VF be as in Corollary 7.6. Assume that the sublevel sets of V are bounded and that VF (ξ ) ≤ 0 for all ξ in D. If x : [0, ωx ) → RN is a maximal solution of (7.22) such that cl(x([0, ωx ))) ⊂ D, then x is bounded, ωx = ∞, and x approaches the largest subset of VF−1 (0) that is weakly invariant with respect to (7.22). Proof. Since (d/dt)V (x(t)) = VF (x(t)) ≤ 0 for almost all t in [0, ωx ), we have the counterpart of (7.23): V (x(t)) ≤ V (x(0)) for all t in [0, ωx ). Since the sublevel sets of V are bounded, it follows that x is bounded. By assertion (b) of Lemma 7.2, ωx = ∞. An application of Corollary 7.6, with G = cl(x(R + )), completes the proof. ⊓ ⊔ Example 7.4. In this example we describe a typical application of Corollary 7.7. In part (a) of the example we analyze a general class of second-order differential inclusions; in part (b) we discuss a special case, a mechanical system subject to friction of Coulomb type. (a) Let k : R → R be continuous with the property lim

Z p

|p|→∞ 0

k = ∞.

(7.24)

Let (p, v) 7→ C(p, v) ⊂ R be upper semicontinuous with nonempty, convex, compact values and with the property that, for all (p, v) in R2 , C∗ (p, v) := max{vw : w ∈ C(p, v)} ≤ 0 .

(7.25)

Consider the system y(t) ¨ + k(y(t)) ∈ C(y(t), y(t)), ˙

(y(0), y(0)) ˙ = (p0 , v0 ) ∈ R2 .

(7.26)

Setting x(t) = (y(t), y(t)), ˙ the second-order initial-value problem (7.26) can be expressed in the equivalent form x(t) ˙ ∈ F(x(t)),

x(0) = x0 = (p0 , v0 ) ∈ R2 ,

where the set-valued map F ∈ U is given by

(7.27)

Stability and Asymptotic Behaviour of Nonlinear Systems

F(p, v) = {v} × {−k(p) + w : w ∈ C(p, v)}.

217

(7.28)

By Theorem 7.4, (7.27) has a solution and every solution can be extended to a maximal solution; moreover, every bounded maximal solution has interval of existence R + . C LAIM A. For each x0 = (p0 , v0 ) in R2 , every maximal solution x = (y, y) ˙ of (7.27) is bounded (hence, has interval of existence R + ) and approaches the largest subset E of C∗−1 (0) that is weakly invariant with respect to (7.27). To establish this claim, we define V : R2 → R by V (p, v) =

Z p 0

k(s)ds + v2 /2 .

Observe that, by property (7.24) of k, V is such that, for every sequence (ξn ) in R2 , V (ξn ) → ∞ as n → ∞ and, as a result, every sublevel set of V is bounded. Moreover, VF (p, v) = max h∇V (p, v), θ i = C∗ (p, v) ≤ 0 θ ∈F(p,v)

for all (p, v) ∈ R2 .

˙ be a maximal solution of (7.27). An Let x0 = (p0 , v0 ) be a point in R2 and let x = (y, y) application of Corollary 7.7, with D = R2 , completes the proof of Claim A. (b) As a particular example, consider a mechanical system wherein a mass is subject to a restoring force k friction force of Coulomb type: the system can be written formally as y(t) ¨ + sgn(y(t)) ˙ + k(y(t)) = 0. Again, we assume that k is continuous with property (7.24). This system may be embedded in the differential inclusion (7.26) with the set-valued map C given by   {−1}, v > 0 C(p, v) := [−1, 1], v = 0  {+1}, v < 0.

(7.29)

C LAIM B. For each x0 = (p0 , v0 ) in R2 , every maximal solution x = (y, y) ˙ of (7.27) (with F and C given by (7.28) and (7.29)) is bounded and approaches the set k−1 ([−1, 1]) × {0}. To prove this claim, we first note that in this case the function C∗ (defined in (7.25)) is given by C∗ (p, v) = −|v| ≤ 0. Therefore, C∗−1 (0) = R × {0}. By Claim A, for each x0 = (p0 , v0 ) in R2 , every maximal solution x = (y, y) ˙ of (7.27) is bounded, is defined on R + , and approaches the largest subset E of R×{0} that is weakly invariant with respect to (7.27) (equivalently, (7.26)). Clearly (0, 0) ∈ E and so E is non-empty. To conclude Claim B, it suffices to show that E ⊂ k−1 ([−1, 1])×{0} . Let (p1 , 0) ∈ E be arbitrary. By weak invariance of E, there exists a solution (z, z˙) : R + → R2 of z¨ + k(z(t)) ∈ C(z(t), z˙(t)), with (z(0), z˙(0)) = (p1 , 0), such that (z(t), z˙(t)) ∈ E ⊂ R × {0} for all t ∈ R + . Therefore, for all t ∈ R + , z(t) = p1 and z˙(t) = 0 = z¨(t). By the differential inclusion, it follows that k(p1 ) ∈ C(p1 , 0) = [−1, 1] and so p1 ∈ k−1 ([−1, 1]). This completes the proof of Claim B.

218

H. Logemann and E. P. Ryan

References 1. H. Amann, Ordinary Differential Equations: An Introduction to Nonlinear Analysis, de Gruyter, Berlin, 1990 2. J. P. Aubin and A. Cellina, Differential Inclusions, Springer-Verlag, Berlin, 1984. 3. I. Barb˘alat, Syst`emes d’´equations diff´erentielles d’oscillations non lin´eaires, Revue de Math´ematiques Pures et Appliqu´ees IV (1959) 267-270. 4. G. D. Birkhoff, Dynamical Systems, American Mathematical Society, Colloquium Publications Vol 9, Providence, 1927. 5. C. I. Byrnes and C. F. Martin, An integral-invariance principle for nonlinear systems, IEEE Trans. Automatic Control AC-40 (1995) 983-994. 6. F. H. Clarke, Optimization and Nonsmooth Analysis, Wiley, New York, 1983. 7. F. H. Clarke, Yu. S. Ledyaev, R. J. Stern and P. R. Wolenski, Nonsmooth Analysis and Control Theory, Springer-Verlag, New York, 1998. 8. E. A. Coddington and N. Levinson, Theory of Ordinary Differential Equations, McGrawHill, New York, 1955. 9. C. Corduneanu, Integral Equations and Stability of Feedback Systems, Academic Press, New York, 1973. 10. K. Deimling, Multivalued Differential Equations, Walter de Gruyter, Berlin, 1992. 11. W. Desch, H. Logemann, E. P. Ryan, and E.D. Sontag, Meagre functions and asymptotic behaviour of dynamical systems, Nonlinear Analysis: Theory, Methods & Applications 44 (2001) 1087-1109. 12. A. F. Filippov, Differential Equations with Discontinuous Righthand Sides, Kluwer, Dordrecht, 1988. 13. A. T. Fuller, The General Problem of the Stability of Motion (A.M. Lyapunov) (translation), Int. J. Control, bf 55 (1992), 531-773. 14. H. W. Knobloch and F. Kappel, Gew¨ohnliche Differentialgleichungen, B.G. Teubner, Stuttgart, 1974. 15. J. P. LaSalle, The extent of asymptotic stability, Proc. Nat. Acad. Sci. USA 46 (1960) 363-365. 16. , Some extensions of Liapunov’s Second Method, IRE Trans. Circuit Theory CT-7 (1960) 520-527. , Stability theory for ordinary differential equations, J. Differential Equations 4 17. (1968) 57-65. 18. , The Stability of Dynamical Systems, SIAM, Philadelphia, 1976. 19. H. Logemann and E. P. Ryan, Non-autonomous systems: asymptotic behaviour and weak invariance principles, J. Differential Equations 189 (2003) 440-460. 20. H. Logemann and E. P. Ryan, Asymptotic Behaviour of Nonlinear Systems, American Mathematical Monthly, 111 (2004), 864-889. 21. A. M. Lyapunov, Probl`eme g´en´eral de la stabilit´e du mouvement, Ann. Fac. Sci. Toulouse 9 (1907), 203-474. Reprinted in Ann. Math. Study No. 17, 1949, Princeton University Press. 22. A. M. Lyapunov, The general problem of the stability of motion (translator: A. T. Fuller), Int. J. Control 55 (1992) 531-773. 23. V. M. Popov, Hyperstability of Control Systems, Springer-Verlag, Berlin, 1973. 24. E. P. Ryan, Discontinuous feedback and universal adaptive stabilization, in Control of Uncertain Systems, (D. Hinrichsen and B. M˚artensson, eds.), Birkh¨auser, Boston, 1990, pp. 245-258.

Stability and Asymptotic Behaviour of Nonlinear Systems

25. 26. 27. 28. 29. 30.

219

, An integral invariance principle for differential inclusions with application in adaptive control, SIAM J. Control & Optim. 36 (1998) 960-980. S. Sastry, Nonlinear Systems: Analysis, Stability and Control, Springer-Verlag, New York, 1999. G. V. Smirnov, Introduction to the Theory of Differential Inclusions, American Mathematical Society, Providence, 2002. E. D. Sontag, Mathematical Control Theory, 2nd Edition, Springer, New York, 1998. A. R. Teel, Asymptotic convergence from L p stability, IEEE Trans. Auto. Control AC-44 (1999) 2169-2170. W. Walter, Ordinary Differential Equations, Springer-Verlag, New York, 1998.

8 Sliding-Mode Observers † Christopher Edwards, Sarah K. Spurgeon, Chee P. Tan, and Nitin Patel

8.1

Introduction

The focus of much of the research in the area of control systems theory during the seventies and eighties addressed the issue of robustness – i.e., designing controllers with the ability to maintain stability and performance in the presence of discrepancies between the plant and model. One nonlinear approach to robust controller design which emerged during this period is the Variable Structure Control Systems methodology. Variable Structure Control Systems evolved from the pioneering work in Russia of Emel’yanov and Barbashin in the early 1960’s. The ideas did not appear outside the Soviet Union until the mid 1970’s when a book by Itkis [20] and a survey paper by Utkin [29] were published in English. Variable structure systems concepts have subsequently been utilized in the design of robust regulators, model-reference systems, adaptive schemes, tracking systems and state observers. The ideas have successfully been applied to problems as diverse as automatic flight control, control of electrical motors, chemical processes, helicopter stability augmentation, space systems and robotics [12, 23, 28, 30]. Variable structure control systems comprise a collection of different, usually quite simple, feedback control laws and a decision rule. Depending on the status of the system, a decision rule, often termed the switching function, determines which of the control laws is ‘on-line’ at any one time. Unlike, for example, a gain-scheduling methodology, the decision rule is designed to force the system states to reach, and subsequently remain on, a pre-defined surface within the state-space. The dynamical behaviour of the system when confined to the surface is described as the ideal sliding motion. The advantages of obtaining such a motion are twofold: firstly there is a reduction in order and secondly the sliding motion is insensitive to parameter variations implicit in the input channels. The latter property of invariance towards so-called matched uncertainty makes the methodology an attractive one for designing robust controllers for uncertain systems [12, 23, 28, 30]. This chapter considers the use of these ideas for robust estimation of states (and parameters). It will provide a perspective on the development of sliding-mode observers for continuous † This

is based in part on C. Edwards, S.K. Spurgeon and C.P. Tan, ‘On the Development and Application of Sliding-mode Observers’, in ‘Variable Structure Systems: Towards the 21st Century’, X.Yu and J.X. Xu, Springer-Verlag, 2002.

M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 221-242, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

222

C. Edwards et al.

time systems – primarily those which can be well represented as linear systems subject to bounded nonlinearities/uncertainties. This encompasses a wide range of real engineering systems. Sliding-mode observers for systems modelled by nonlinear systems, and whose design methodology explicitly exploits the nonlinear system structure, have recently been considered in [2]; sliding-mode observers for discrete time systems is addressed elsewhere in this volume. Consider initially a nominal linear system x(t) ˙ = Ax(t) + Bu(t)

(8.1)

y(t) = Cx(t)

(8.2)

where A ∈ Rn×n , B ∈ Rn×m and C ∈ R p×n . Without loss of generality assume that C has full row rank; i.e., there is no redundancy amongst the measurements. The objective here is to obtain an estimate of the state x(t) by measuring the quantities y(t) and u(t). The design of a linear time invariant dynamical system to estimate x(t) from the inputs and outputs was extensively studied in the 1960’s by Luenberger. An algebraic condition on the matrix pair (A,C) – the notion of observability – was proposed as a necessary and sufficient condition for state reconstruction [22]. (For simplicity, throughout this chapter, the observability condition will be assumed to hold, although technically some of the developments only require the weaker restriction of detectability on the pair (A,C). See for example [4].) One way of viewing the approach of Luenberger is to regard the observer as a model of the system in which the discrepancy between its output, and the output of the system, is fedback through a designer specified gain. Generally speaking, in sliding-mode observers, instead of feeding back the output error between the observer and the system linearly, the output error is fedback via a discontinuous switched signal.

8.2

A Discontinuous Observer

To be specific, consider the dynamical system z˙(t) = Az(t) + Bu(t) + Gn ν

(8.3)

where Gn ∈ Rn×p is to be specified, and ν is a discontinuous injection signal depending on the output error ey (t) = Cz(t) − y(t). As is common in sliding-mode approaches, the analysis is greatly simplified if an appropriate change of coordinates is introduced. In the case of observer design, the coordinate transformation associated with the invertible matrix T Nc Tc = (8.4) C where Nc ∈ Rn×(n−p) spans the null-space of C, is appropriate. As a result of this transformation, the triple (A, B,C) has the form A11 A12 B1 A= B= (8.5) C = 0 Ip A21 A22 B2

where the partitions of A and B are conformable with respect to C. Define the state estimation error e(t) = z(t) − x(t) and suppose (in the new coordinate system) that the output error injection gain

Sliding-Mode Observers

Gn =

L Ip

223

(8.6)

where L ∈ R(n−p)×p is a gain matrix to be specified.

Suppose the state estimation error e(t) in the coordinates of (8.5) is partitioned as (e1 , ey ) then e˙1 (t) = A11 e1 (t) + A12 ey (t) + Lν

(8.7)

e˙y (t) = A21 e1 (t) + A22 ey (t) + ν

(8.8)

As argued in [30], if the components of the discontinuous term

ν = −ρ

ey key k

if ey 6= 0

(8.9)

then for a large enough scalar ρ an ideal sliding motion is induced in finite time on the surface S = {e ∈ Rn : ey = Ce = 0}

(8.10)

During the sliding motion ey = e˙y = 0 and equation (8.8) can be written as 0 = A21 e1 + νeq

(8.11)

where νeq represents the equivalent output error injection term necessary to maintain a sliding motion on S . This is the natural analogue of the so-called equivalent control occurring in the design of sliding-mode controllers [30]. Rearranging (8.11) and substituting back into equation (8.7) it follows that the reduced order sliding motion is governed by e˙1 (t) = (A11 − LA21 )e1 (t)

(8.12)

In order to be able to sustain a sliding motion, and for (asymptotic) state estimation error decay, the eigenvalues of (A11 − LA21 ) must be stable. However if (A,C) is observable then so is (A11 , A21 ); see [30] for example. Therefore there exists L for which (A11 − LA21 ) is stable. Consider the example 0 1 0 A= B= C= 1 1 (8.13) −1 0 1

An appropriate coordinate change to generate the canonical form in (8.5) is given by x 7→ Tc x where 1 −1 Tc = (8.14) 1 1 This results in (with abuse of notation) 0 1 A= −1 0

−1 B= 1

C= 0 1

(8.15)

In the following the gain Gn,1 = −1 which gives (A11 − Gn,1 A21 ) = −1 and the nonlinear gain ρ = 2. The sliding-mode observer from (8.3) can then be written as z˙1 (t) = z2 (t) − u(t) + 2sign ey (t)

z˙2 (t) = −z1 (t) + u(t) − 2sign ey (t)

(8.16) (8.17)

224

C. Edwards et al. .

Output Estimation Error

0.4

.

0.2

0

−0.2

.

.

.

.

.

.

−0.4

−0.6

−0.8

−1

0

1

2

3

4

5 Time

6

7

8

9

10

Fig. 8.1. Switching function

For simplicity assume that u(t) ≡ 0. In the simulations which follow the initial condition for the observer is [ 0 0 ]T and the initial condition of the system is [ 1 0 ]T . Figure 8.1 is a plot of the output estimation error ey against time. This may also be viewed as the switching function, and as expected, it is forced to zero in finite time. Figure 8.2 shows a comparison of the states of the system and the states of the plant. It is apparent that although ey = 0 in finite time (after approximately 0.5 seconds) the state estimation error e(t) → 0 exponentially and is different from zero for several seconds. 1.5 Plant Observer

1st State

1

0.5

0

−0.5

−1

−1.5

0

1

2

3

4

5

6

7

8

9

10

0

1

2

3

4

5 Time

6

7

8

9

10

1.5

2nd State

1

0.5

0

−0.5

−1

−1.5

Fig. 8.2. A comparison of the observer and system states

Sliding-Mode Observers

225

The description thus far describes the first published1 work on sliding-mode observers by Utkin [30]. An important consequence of inducing a sliding motion is that robust state estimation can be obtained. In the same way that sliding-mode controllers exhibit complete rejection of a class of uncertainty, so will the observer just described. Suppose that the real system is governed by x(t) ¨ = − sin x(t)

(8.18)

and that the measured output corresponds to y(t) = x(t) + x(t) ˙

(8.19)

Physically this may be viewed as a pendulum system where x(t) represents angular displacement from the vertical. The system in (8.13) may be viewed as a linearization of the system above at the equilibrium point (0, 0). Using the observer given in (8.16)–(8.17), the error system is governed by e˙1 (t) = ey (t) − sin x1 (t) − ν

e˙y (t) = −e1 (t) + sin x1 (t) + ν

(8.20) (8.21)

where ν = −2 sign ey . During the sliding motion ey = e˙y = 0 and

νeq = e1 (t) − sin x1 (t)

(8.22)

Substituting into equation (8.20) the sine terms cancel, leaving e˙1 (t) = −e1 (t)

(8.23)

The sliding motion is therefore independent of the nonlinearity/uncertainty resulting from the linearization, and the observer states track the real states asymptotically. Once again, in the simulations which follow the initial condition for the observer is given by [ 0 1 ]T and the initial condition of the (nonlinear) system is [ 1 0 ]T . Figure 8.3 is a plot of the output estimation error ey against time. Again the switching function is forced to zero in finite time. Figure 8.4 shows a comparison of the states of the system and the states of the plant. Once again perfect asymptotic tracking is obtained despite the mismatch between the system and the linear model about which the observer was designed.

8.3

Observers with Linear and Discontinuous Injection

The formulation just described generally requires large values of ρ in order to ensure sliding for a broad range of initial state estimation errors – particularly if the underlying system is unstable. As discussed in [7], from a practical viewpoint, this may cause difficulties. A tradeoff is usually necessary between the requirement of a large ρ to ensure a sliding-mode occurs and its subsequent reduction to prevent excessive chattering (whilst still ensuring sliding). For 1

It is more accurate to say the first published work in English on sliding observers is [30]. Earlier published work on this theme appears in Russian; see for example the references in [8].

226

C. Edwards et al.

Switching function

1

0.5

0

0

0.5

1

1.5

2

2.5 Time, sec

3

3.5

4

4.5

5

Fig. 8.3. Switching function 1.5

1

1st State

0.5

0

−0.5

−1 0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

0.5

1

1.5

2

2.5 Time, sec

3

3.5

4

4.5

5

1

2nd state

0.5

0

−0.5

−1

−1.5 0

Fig. 8.4. A comparison of the observer and system states

this reason it may be preferable to include a linear output error injection term. Rather than the formulation described in (8.3) consider the dynamical system z˙(t) = Az(t) + Bu(t) − Gl ey (t) + Gn ν

(8.24)

where Gl , Gn ∈ Rn×p and ey (t) = Cz(t) − y(t). This is effectively the observer structure of Slotine et.al. [25], where it is argued that the linear gain should be chosen to enhance the size of the so-called sliding patch i.e., the domain in the state estimation error space in which sliding occurs. With a well designed linear gain Gl this observer enjoys the same robustness

Sliding-Mode Observers

227

prop erties. Indeed in certain situations (which will be discussed in more detail later) global state estimation error convergence properties can be proven. To demonstrate this, consider once more the nonlinear system (8.18)–(8.19) and its linearization (8.13). Again change coordinates according to (8.14) to obtain the realization in (8.15). The observer design and analysis will be performed in this coordinate system. Choose as the output error injection feedback gains 0 −1 Gl = Gn = (8.25) 2 1

It can be shown that λ (A − Gl C) = {−1, −1}. Furthermore the state estimation error system is governed by e˙1 (t) = ey (t) − sin x1 (t) − ν

e˙y (t) = −e1 (t) − 2ey (t) + sin x1 (t) + ν

(8.26) (8.27)

Consider now the quadratic form V (e1 , ey ) = e21 + 2e1 ey + 2e2y = (e1 + ey )2 + e2y

(8.28)

as a candidate Lyapunov function for the error system above. Clearly the expression in (8.28) is positive definite. Furthermore, after some algebra, it can be shown that V˙ = −2e21 − 6ey e1 − 6e2y + 2ey sin x1 + 2ey ν

≤ −2e21 − 6ey e1 − 6e2y + 2|ey || sin x1 | − 2ρ |ey |

≤ −2e21 − 6ey e1 − 6e2y − 2|ey |

≤ −2(e1 + 23 ey )2 − 23 e2y − 2|ey | ≤ 0 and therefore (global) asymptotic stability of the error system has been proven.

8.4

The Walcott and˙ Zak Observer

The problem of robust state estimation for systems with bounded matched uncertainty was ˙ [32]. They sought to design a sliding-mode observer for the first explored by Walcott & Zak system x(t) ˙ = Ax(t) + Bu(t) + B f (t, y, u)

(8.29)

y(t) = Cx(t)

(8.30)

where f : R+ × R p × Rm 7→ Rm represents lumped uncertainty or nonlinearities. The function is assumed to be unknown but bounded so that k f (t, y, u)k ≤ ρ (t, y, u)

(8.31)

where ρ (·) is known. The observer proposed in [32] has the form z˙(t) = Az(t) + Bu(t) − GCe(t) − P−1CT F T ν where the discontinuous scaled unit vector term

(8.32)

C. Edwards et al.

228

FCe(t) ν = ρ (t, y, u) kFCe(t)k

(8.33)

and e(t) = z(t) − x(t). The symmetric positive definite matrix P ∈ Rn×n and the gain matrix G are assumed to satisfy (8.34) PA0 + AT0 P < 0 where A0 := A − GC, and the structural constraint PB = (FC)T

(8.35)

for some F ∈ Rm×p . Under these circumstances the quadratic form given by V (e) = eT Pe can be shown to guarantee quadratic stability [32]. Furthermore an ideal sliding motion takes place on SF = {e ∈ Rn : FCe = 0}

in finite time. Remarks • •

If p > m then sliding on SF is not the same as sliding on Ce(t) = 0 and so the observer structure in (8.32) is different from the observer in (8.24). A system theoretic interpretation of (8.34)–(8.35) by Steinberg & Corless [27] is that the transfer function matrix G(s) = FC(sI − A0 )−1 B is strictly positive real.

8.4.1

Synthesizing the Gains

The problem of synthesizing P, G and F (and incorporating some sort of design element) is non-trivial. The original work in [32] postulated the use of symbolic manipulation tools to solve a sequence of constraints arising from ensuring that the principal minors of both P and the right hand-side of (8.34) are positive and negative respectively. In the original work, the class of systems for which the design problem has a solution is not identified. For low order systems the synthesis problem is quite tractable: the observer designed for the pendulum ˙ observer. The structural requirements (8.34)–(8.35) described earlier is in fact a Walcott & Zak ˙ were shown in [14] to be solvable if and only if of Walcott & Zak • •

rank(CB) = m any invariant zeros of (A, B,C) ∈ C−

Details of the constructive design algorithms are described in [14] and follow on from earlier work described in [9]. Under these circumstances a (semi) analytic expression for the solution to the observer design problem is given in terms of a gain matrix L ∈ R(n−p)×(p−m) and a stable matrix As22 ∈ R p×p .

8.5

A Convex Parameterization

˙ and proposes a This section considers a more general problem than that of Walcott & Zak numerically based solution methodology which exploits all the degrees of freedom available in the design.

Sliding-Mode Observers

229

Consider the dynamical system x(t) ˙ = Ax(t) + Bu(t) + Dξ (t, x, u)

(8.36)

y(t) = Cx(t)

(8.37)

where A ∈ Rn×n , B ∈ Rn×m ,C ∈ R p×n and D ∈ Rn×q where p ≥ q. Assume C and D are full rank and the function ξ : R+ × Rn × Rm → Rq is unknown but bounded so that kξ (t, x, u)k ≤ r1 kuk + α (t, y)

(8.38)

where r1 is a known scalar and α : R+ × R p → R+ is a known function. An observer of the form (as described earlier in §8.3) z˙(t) = Az(t) + Bu(t) − Gl ey (t) + Gn ν

(8.39)

will be considered where Gl ∈ Rn×p and Gn ∈ Rn×p and ey (t) = Cz(t) − y(t). The discontinuous vector ν is defined by ( e −ρ (t, y, u)kPo D2 k key k if ey 6= 0 y ν= (8.40) 0 otherwise where Po ∈ R p×p is symmetric positive definite. The matrices Po and D2 will be defined formally later. The function ρ : R+ × R p × Rm → R+ satisfies

ρ (t, y, u) ≥ r1 kuk + α (t, y) + γ0

(8.41)

where γ0 is a positive scalar. If the state estimation error e(t) := z(t) − x(t), then it is straightforward to show from equations (8.36)–(8.37) and (8.39) that e(t) ˙ = A0 e(t) + Gn ν − Dξ (t, x, u)

(8.42)

where A0 := A − Gl C. In [14] it is argued that necessary and sufficient conditions for the existence of a stable sliding motion on S = {e ∈ Rn : ey = 0} that is independent of ξ are 1. rank (CD)=q 2. any invariant zeros of (A, D,C) lie in the left half plane ˙ observer described in §8.4.1. These of course are analogous to those for the Walcott & Zak Again a (semi) analytic expression for the gains Gl and Gn in terms of a matrix L ∈ R(n−p)×(p−q) and a stable matrix As22 ∈ R p×p can be demonstrated [14]. This parameterization however only represents a specific subclass of possible solutions. Additional degrees of freedom are available which are not exploited because the simplicity of solution is lost. The next section considers the design of the matrices Gl , Gn and Po so that a sliding motion takes place on S . Assuming that conditions 1 and 2 are satisfied, a new parameterization is given which seeks to exploit all of the design freedom which is available. A canonical form from [9] will constitute a useful starting point. If conditions 1 and 2 above are satisfied there exists a coordinate transformation T0 in which the system (A, D,C) can be written as

230

C. Edwards et al.



 A¯ 11 A¯ 12 A¯ =  A¯ 211 ¯  A A¯ 212 22

D¯ =

0 D¯ 2

C¯ = 0 T

(8.43)

where the sub-matrices A¯ 11 ∈ R(n−p)×(n−p) , A¯ 211 ∈ R(p−q)×(p−n) represent a detectable pair, D¯ 2 ∈ Rq×q is nonsingular, and T ∈ R p×p is orthogonal. Further the unobservable modes of (A¯ 11 , A¯ 211 ) are the invariant zeros of (A, D,C) [12]. In order to make the partitions in (8.43) conformable it is convenient to introduce the term 0 D2 := (p−q)×q (8.44) D¯ 2 Let G¯ l and G¯ n represent the observer gain matrices in the new coordinate system and define ¯ The gain matrix G¯ l is to be determined but assume A¯ 0 = A¯ − G¯ l C. ¯ T −1 −LT (8.45) Po G¯ n = TT where L¯ ∈ R(n−p)×p and L¯ = L 0 with L ∈ R(n−p)×(p−q) , and the orthogonal matrix T is part of the output distribution matrix C¯ from (8.43). ¯ that satisfies P¯ A¯ 0 + Proposition 8.1. If there exists a positive definite Lyapunov matrix P, A¯ T0 P¯ < 0, with the structure P¯ P¯ L¯ P¯ = ¯ T1¯ ¯ 1¯ T ¯ ¯ > 0 (8.46) L P1 P2 + L P1 L where P¯1 ∈ R(n−p)×(n−p) and P¯2 ∈ R p×p , then the error system in equation (8.42) is quadratically stable. Proof. Consider the quadratic form given by V (e) ¯ = e¯T P¯ e¯

(8.47)

as a candidate Lyapunov function where e¯ := T0 e. Notice that if P¯1 , P¯2 > 0 then P¯ > 0 from the Schur expansion. From (8.42) the derivative along the system trajectory V˙ = e¯T (A¯ T0 P¯ + P¯ A¯ 0 )e¯ + 2e¯T P¯ G¯ n ν − 2e¯T P¯ D¯ ξ From the definitions in (8.43), (8.45) and (8.46) 0 ¯ ¯ PGn = ¯ T Po−1 = C¯ T P2 T

(8.48)

(8.49)

if the symmetric positive definite matrix Po := T P¯2 T T ¯ 2 = 0 and therefore Using the special structures of L¯ and D2 , LD 0 P¯ D¯ = ¯ = C¯ T Po D2 P2 D2 if the matrix

(8.50)

(8.51)

Sliding-Mode Observers D2 := T D2

231 (8.52)

Consequently, (8.48) becomes V˙ = e¯T (A¯ T0 P¯ + P¯ A¯ 0 )e¯ + 2ey T ν − 2ey T Po D2 ξ ≤ e¯T (A¯ T0 P¯ + P¯ A¯ 0 )e¯ − 2ρ kPo D2 kkey k − 2ey T Po D2 ξ Using the uncertainty bounds for ξ from equations (8.38) and (8.41) V˙ ≤ e¯T (A¯ T0 P¯ + P¯ A¯ 0 )e¯ − 2ρ kPo D2 kkey k + 2kPo D2 k [r1 kuk + α (y)] key k ≤ e¯T (A¯ T0 P¯ + P¯ A¯ 0 )e¯ − 2γ0 kPo D2 kkey k Since (A¯ T0 P¯ + P¯ A¯ 0 ) < 0 it follows that V˙ < 0 for all e¯ 6= 0.

⊓ ⊔

Corollary 8.1. An ideal sliding motion takes place on S in finite time. Furthermore the sliding dynamics are given by the system matrix A¯ 11 + LA¯ 211 . Proof. Using Proposition 8.1, a modification to Corollary 6.1 in [12] shows that sliding takes place on S in finite time. Using the concept of equivalent output error injection, the sliding motion is governed by ¯ ¯ ¯ ¯¯ ¯ A¯ 0 = A11 + LA211 A12 + LA22 (I − G¯ n (C¯ G¯ n )−1C) 0 0 Hence the sliding motion is governed by A¯ 11 + LA¯ 211 as claimed.

⊓ ⊔

Remarks: Since (A¯ 11 , A¯ 211 ) is detectable by construction, there exists a family of matrices L ∈ R(n−p)×(p−q) such that A¯ 11 + LA¯ 211 is stable. If a further linear change of co-ordinates In−p L¯ TL = (8.53) 0 T ¯ D, ¯ and its Lyapunov matrix P, ¯ C) ¯ the system matrix, disturbance is applied to the triple (A, distribution matrix and the output distribution matrix will be in the form A11 A12 0 A = C = 0 Ip (8.54) D= D2 A21 A22

where A11 = A¯ 11 + LA¯ 211 . In the new co-ordinate system, the Lyapunov matrix will be P¯1 0 −1 T ¯ −1 P = (TL ) P(TL ) = (8.55) 0 Po where Po is defined in (8.50). The nonlinear output error injection gain matrix 0 Gn = Po−1

(8.56)

As argued in [9], the fact that P is block diagonal Lyapunov matrix for A0 = A − Gl C implies that A11 is stable and hence the sliding motion is stable. From a design and synthesis perspective the problem of determining Gl , Gn and P can be posed in such a way that Linear Matrix Inequalities (LMIs) [3] can be used to numerically synthesize the required matrices.

232

8.6

C. Edwards et al.

A Case Study: Road Tyre Friction Estimation

Increasingly, commercial vehicles are being fitted with micro-processor based systems to enhance the safety, improve driving comfort, increase traffic circulation, and reduce environmental pollution associated with vehicles. Examples of such products are anti-lock brake systems (ABS), traction control systems (TCS), adaptive cruise control (ACC), active yaw control, active suspension systems, and engine management systems (EMS). Many of these systems rely on the physical parameters of the vehicle and the conditions in which it is required to operate. Some of these vehicle related parameters are fixed (or are at least subject to negligible variation); some depend on specific scenarios relating to the way in which the vehicle is being used (e.g., loading) and may be thought of as constant unknown parameters (which may be estimated or accounted for by the robustness of the control system); but others – particularly the tyre/road friction coefficient – are subject to severe and short term variations. Various methods have been developed to predict tyre/road friction [1, 6, 15, 21, 24, 33]. Most are passive, in the sense that they make use of available data from existing sensors to compute an estimate of the tyre/road friction and assume the parameters related to the vehicle and the tyres are constant. The methods from [1, 6, 15, 21, 24, 33] use observers designed around mathematical models of friction and simple vehicle models. Ray [21, 24] primarily uses a Kalman filter but [21] also examines a least squares approach. Extensive work by Canudas de Wit et. al [6] and Horowitz et. al [33] has used a LuGre friction model [5] and investigates nonlinear adaptive observers based on measurements of wheel speed. The work described here uses the same friction model as in [1, 6, 33] but employs a slidingmode observer to estimate the road/tyre friction coefficient. During sliding, an equivalent output estimation error is used to estimate the road surface parameter rather than the adaptive scheme in [33].

8.6.1

Tyre/Road Friction and Vehicle Modelling

The tyre/road coefficient of friction is defined as

µ=

Friction force Fx = Fn Normal force

(8.57)

The quantity µ is a nonlinear function of many physical variables including the velocity of the vehicle, the road surface conditions and so-called slip. Longitudinal slip, s, can be defined for a braking scenario as v − rω where v > rω and v 6= 0 (8.58) s= v where r is the effective rolling radius of the tyre, v is the linear speed of the tyre centre and ω is the angular speed of the tyre. Slip is an indirect measure of the fraction of the contact patch on the road surface and it creates braking and accelerating forces. The plot in Figure 8.5 is a typical µ -slip curve associated with a dry asphalt road surface. The figure has been obtained from using an expression known as the pseudo-steady state LuGre friction model2 [6]. When s = 0, free rolling of the wheel takes place, whilst s = 1 represents 2

Specifically equation (8.40) from [6] has been used together with the parameter values in Table 8.1 from the appendix. In [6] slip during a braking manoeuvre is defined as s = (rω − v)/v, i.e., with opposite polarity to the definition used here – accordingly s has been replaced by −s to obtain the plots in Figures 8.5-8.6.

Sliding-Mode Observers

233

Relationship between µ and s

1

0.8

µ

0.6

0.4

0.2

0

0

0.1

0.2

0.3

0.4 0.5 0.6 longitudinal slip s

0.7

0.8

0.9

1

Fig. 8.5. A typical µ -slip curve

a locked wheel condition. It can be seen from Figure 8.5 that there is an unique value, s, ˆ associated with the maximum value of µ . During a braking manoeuvre, maintaining the slip at s = sˆ provides an optimal stopping distance. Less favourable (and potentially more dangerous) road surface conditions tend to decrease the peak value of the mu/slip curve (by up to 25% in icy conditions) and alter the value of s. ˆ In many friction models this effect is taken into account by means of a ‘road surface condition’ parameter. Figure 8.6 captures this effect by means of a parameter θ which forms part of the pseudo-steady state LuGre friction model [34]. Figure 8.6 shows curves representing the relationship between µ and s for different values of θ . This has been obtained keeping the vehicle velocity v at 13.41 m/s (30 mph) and using different values for θ . For the purpose of developing a friction estimator, consider the dynamic LuGre friction model from [6] together with the vehicle dynamics: z˙ = vr − θ σ0 |vr |z/h(vr )

(8.59)

J ω˙ = −rFx − σω ω − kb Pb

(8.61)

mv˙ = 4Fx − Fav

(8.60)

where vr = rω − v is the relative velocity and z is an internal frictional state. The friction force Fx produced by the tyre/road contact is given by Fx = Fn (σ0 z + σ1 z˙ + σ2 vr )

(8.62)

where σ0 is the stiffness coefficient, σ1 is the damping coefficient, and σ2 is the viscous relative damping coefficient. The scalar function

234

C. Edwards et al. Relationship between µ and s when θ varies from 1 to 4 θ=1 1

0.8

θ = 1.5

µ

0.6

θ=2

θ = 2.5 0.4

θ=4 0.2

0

0

0.1

0.2

0.3

0.4 0.5 0.6 longitudinal slip s

0.7

0.8

0.9

1

Fig. 8.6. Pseudo static model at v = 13.41 m/sec and θ between 1 and 4 vr 1/2

h(vr ) := µc + (µs − µc )e−| vs |

(8.63)

where vs is the Stribeck relative velocity, µs is the normalized static friction coefficient, µc is the normalized Coulomb friction. The parameter θ in (8.59) captures the changes in the road characteristics: typically, θ =1 represents dry, θ = 2.5 represents wet and θ = 4 represents icy road conditions. In (8.61) and (8.60), which represent a simple model of the vehicle, J is the moment of inertia of the wheel, σω represents the viscous rotational friction coefficient, m is the total mass of the vehicle, Fav represents the aerodynamic force, kb the brake system gain and Pb is the actual applied braking pressure (the control variable). The values of these quantities used in this paper are given in Table 8.1. In what follows the state ω has been replaced by vr = rω − v. This clearly just represents a linear change of coordinates. Assuming the vehicle is on a flat road and the load is equally distributed on all four tyres then Fn = mg 4 . This removes the appearance of the normal force Fn from the state-space model that will be developed.

8.6.2

Observer Design

In this section a sliding-mode observer will be designed based on the assumption that only the angular velocity ω is available (which can be measured easily). The states of the system have been chosen as x = (z, v, vr ). From equations (8.59)– (8.62) it can be verified that the state-space equations can be written as x(t) ˙ = Ax(t) + Bu(t) + Dθ x1 (t) f (x3 )

(8.64)

Sliding-Mode Observers where the control signal u(t) = Pb (t) and   0 0 1 g(σ1 + σ2 )  A =  gσ0 −gσv qσ0 gσv − σJw q(σ1 + σ2 ) − σJω

where the static nonlinearity

 0 B= 0  − rkJb





 −1 D =  −gσ1  −qσ1

235

(8.65)

σ0 |x3 | (8.66) h(x3 ) Since it is assumed in this section that only angular wheel speed ω is measured, the output distribution matrix C = 0 1r 1r (8.67) f (x3 ) :=

The following nonlinear observer is proposed

˙ˆ = Ax(t) x(t) ˆ + Bu(t) + Gl ey + Dν

(8.68)

where Gl = col(g1 , g2 , g3 ) and g1 , g2 and g3 are scalar gains; and the nonlinear injection ν = −ksgn(Fey ) where F and k are scalar gains. The main objective is to synthesize an observer to generate an estimated angular velocity ωˆ = Cxˆ such that ey = ω − ωˆ ≡ 0 in finite time despite the nonlinear friction terms in (8.59) – or equivalently (8.64) – which have been ignored. Define e = x − xˆ as the state estimation error. The dynamics for the error system can be obtained from (8.64) and (8.68) as e˙ = (A − Gl C)e + D θ f (x3 )x1 − k sgn(Fey ) (8.69)

Suppose e = col(e1 , e2 , e3 ), then in the coordinates e1 , e2 and ey it can be shown that ¯ D, ¯ where ¯ C) (A, D,C) 7→ (A,   0 −1 r ¯ 11 A¯ 12 A  g(σ1 + σ2 )r A¯ = ¯ ¯ =  gσ0 −g(σv + σ1 + σ2 ) A21 A22 σω Fn r Fn r Fn r2 − J σ0 − J (σ1 + σ2 ) − J J (σ1 + σ2 )

and

  −1 ¯1 D D¯ = ¯ =  −gσ1  D2 Fn r J σ1

(8.70)

(8.71)

The output distribution matrix is now

C¯ = 0 0 1

A useful choice for the linear gain of the observer is A¯ L¯ + A¯ 12 − L¯ α G¯ l = ¯11 A22 + A¯ 21 L¯ − α

(8.72)

(8.73)

¯ where L¯ := D¯ −1 2 D1 and α is a negative scalar. After some algebra it can be shown that     r + (g + σα1 ) FJn r g¯1   G¯ l =  g¯2  =  g(α + g(σv + σ1 + σ2 ) − σσ10 ) FJn r + gr(σ1 + σ2 )  (8.74) σ σ w 0 g¯3 (σ1 + σ2 )q − J + σ1 − α

236

C. Edwards et al.

In the original coordinates e1 , e2 and e3 the linear output error injection gain g1 = g¯1 , g2 = g¯2 and g3 = rg¯3 − g¯2 . Proposition 8.2. The error system (8.69) is quadratically stable. Proof. Consider as a potential Lyapunov candidate V (e) = eT Pe where     1 0 0 10 0 ¯ ¯1 L¯ P − P 1 1 0 1 0  P = 0 1 r  −L¯ T P¯1 P¯2 + L¯ T P¯1 L¯ 0 1r 1r 0 0 1r

(8.75)

where P¯1 ∈ R2×2 is an s.p.d. matrix which is a Lyapunov matrix for σ0 σ2 − σ1 σ1 A¯ 11 − L¯ A¯ 21 = 0 −σv g

(8.76)

and P¯2 is a sufficiently large positive scalar such that P(A − Gl C) + (A − Gl C)T P =: −Q < 0

(8.77)

It is clear from (8.76) that A¯ 11 − L¯ A¯ 21 is stable with eigenvalues at {− σσ10 , −σv g} and so the existence of P1 is guaranteed. Also the existence of a sufficiently large P2 to ensure (8.77) is also guaranteed [9]. It can be verified by direct substitution that PD = FCT for some scalar F. Then V˙ = −eT Qe − 2eT PD k sgn(Fey ) − θ f (x3 )x1 ≤ −eT Qe − 2|Fey |(k − |θ f (x3 )x1 |)

≤0

for large enough k and so the state estimation error e is quadratically stable as claimed. As argued in [9], in a domain of the origin, a sliding motion takes place on S = {e : Ce = 0}. ♯ Now, from first principles, the reduced order motion whilst sliding will be investigated. The output error derivative satisfies re˙y = (g + q)(σ0 e1 + σ1 (e3 − θ f (x3 )x1 ) + σ2 e3 + σ1 ν ) − (k2 + k3 +

σω )ey J

and so if a sliding-mode is enforced in finite time, ey = e˙y = 0 [30] and the above equation becomes 0 = (g + q)(σ0 e1 + σ1 (e3 − θ f (x3 )x1 ) + σ2 e3 + σ1 νeq ) (8.78) where νeq represents the equivalent output error injection signal necessary to maintain a sliding motion in the state estimation error space. Therefore from (8.78) an expression for νeq is given by 1 (8.79) νeq = − (σ0 e1 + σ1 (e3 − θ f (x3 )x1 ) + σ2 e3 ) σ1

2 During the sliding motion ey := e3 +e = 0 which implies e3 = −e2 . Substituting this value r of νeq into the first two equations of the components of (8.69), and using the fact that the

Sliding-Mode Observers

237

e3 = −e2 , it can be shown that the sliding motion is governed by the reduced order linear system σ0 σ2 − σ1 σ1 e˙1 e1 = (8.80) e˙2 0 −σv g e2 and so the sliding motion is stable and e1 → 0 and e2 → 0. When e1 , e2 = 0 the expression for the equivalent output error injection signal (8.79) becomes

νeq = θ f (x3 )x1

(8.81)

Equation (8.81) indicates that the expression

θˆ =

νeq f (xˆ3 )xˆ1

(8.82)

may be used as an estimate for the parameter θ during the sliding motion. Since xˆ1 → x1 and xˆ3 → x3 it follows f (xˆ3 )xˆ1 → f (x3 )x1 and so θˆ → θ . Notice (8.81) is only valid when xˆ1 6= 0 and xˆ3 6= 0 since f (xˆ3 ) > 0 if x3 6= 0. The following tables give the friction model parameters and the vehicle parameters used in the subsequent simulations. Table 8.1. Vehicle and LuGre friction model parameters parameter vs σ0 σ1 σ2 µc µs

value 6.57 181.54 0.70 0.0018 0.8 1.55

unit m/s 1/m s/m s/m

parameter r J m σv σw kb

value unit 0.323 m 2.603 kg/m2 1701 kg 0.005 1 0.9

In the simulations which follow the initial conditions for the plant represented by z, v and vr are 0, 40 and 0, whilst 0, 40.1 and 0.1 are used for the observer. This represents a deliberate mis-match for the purpose of demonstration. The discontinuous term in the observer has been replaced by a sigmoidal approximation. In the following simulation a brake pressure signal u(t) has been selected to generate an ‘emergency braking’ (Figure 8.7). In the following simulation the gain on the nonlinear injection term k = 50 whilst g1 = 0.3392, g2 = −2.57864 and g3 = 46.4196. These have been calculated according to the formula in (8.73) and guarantee a sliding motion will take place in finite time. Figure 8.8 shows the evolution of the plant states (as dotted lines) and the estimated states (as solid lines). Figure 8.9 shows the state estimation errors and the normalized estimation error for θ . As expected from Figure 8.8, the error in the friction state estimate given in Figure 8.9 is small and quickly becomes zero. The error in the velocity estimate and the relative velocity estimate show an asymptotic decay but at a slow rate. This is to be expected since one of the poles associated with the system matrix in (8.80) is −gσv = −0.049 and is therefore slow. The fact that the state estimation errors have not decayed, results in a small error for the estimate of θ . Figure 8.10 shows the switching function associated with the observer i.e., the output estimation error. It shows sliding occurs almost instantly in the state estimation error space and so the estimated angular velocity ωˆ from the observer exactly tracks the plant output ω throughout the simulation.

238

C. Edwards et al.

Brake input P (Pa) b

2500

2000

P

b

1500

1000

500

0

0

1

2

3 time(sec.)

4

5

6

Fig. 8.7. Brake input signal −3

0

x 10

Internal friction state z

Velocity v m/s 40

−2 30

−4

20

−6

−8

10

0

2

4

6

0

0

2

time(sec.)

4

6

4

6

time(sec.)

Relative velocity v m/s

θ

r

10

4.5

4

8

3.5

6

3

4

2.5

2

2

1.5

0

−2

1

0

2

4

time(sec.)

6

0.5

0

2

time(sec.)

Fig. 8.8. Evaluation of plant and estimated states, road condition parameter θ

8.7

Summary

This chapter has discussed the development of sliding-mode observers for a class of uncertain systems. Recent work has been reported which formulates the design problem in convex terms so that efficient numerical methods can be employed to synthesize the observer gains. A case

239

Sliding-Mode Observers

Error in friction state z

−4

5

Error in velocity v m/s

x 10

−0.08

−0.085

0

−0.09

−0.095

−5 −0.1

−0.105

−10

−0.11

−15

0

2

4

6

−0.115

0

2

time(sec.)

4

6

time(sec.)

Error in relative velocity v m/s

Normalised error in estimated θ

r

0.15

0.1

0.1 0.05

0.05

0 0

−0.05

−0.1

−0.05

−0.15

−0.2

0

2

4

−0.1

6

0

2

time(sec.)

4

6

time(sec.)

Fig. 8.9. Estimation errors ey

0.1

0

−0.1

−0.2

−0.3

−0.4

−0.5

−0.6

0

0.005

0.01

0.015

0.02

0.025

0.03

0.035

0.04

0.045

0.05

time(sec.)

Fig. 8.10. Output estimation error study has also been presented which demonstrates how sliding-mode observers can be used for parameter estimation as well as state estimation.

8.8

Notes and References

The pendulum example is effectively the case study considered in [31]. Interestingly in this paper a sliding-mode observer was compared to other direct nonlinear observer design

240

C. Edwards et al.

methodologies for the pendulum example. It was argued that the sliding-mode observer had the best performance and was the least involved in terms of design. ˙ observer problem is given in [11] An almost completely analytic solution to the Walcott & Zak for square systems. In [10] an analysis is given of the closed loop system arising from using a state feedback sliding-mode controller implemented using the state estimates generated from a sliding-mode observer. For a class of systems with matched uncertainty the closed loop retains the invariance properties of its full state feedback counterpart. An overview of much of this work appears in [12]. In Herrmann et. al [19] the minimum phase condition is alleviated to some extent for systems with many outputs. In essence, rather than seeking a closed-loop observer based controller which is robust to matched uncertainty, robustness with respect to a strict subset of matched uncertainty is achieved by effectively generating an uncertainty distribution matrix of lower dimension which is contained within the range space of the input distribution matrix. For details see [19]. The use of sliding-mode observers for disturbance estimation is discussed in [17]. In fact [17] describes a design framework for a broader class of systems than the one described in this chapter – namely linear time varying systems. Sliding-mode ideas have previously been used in the literature for fault detection. Sreedhar et. al [26] consider a model-based sliding-mode observer approach although in their design procedure it is assumed that the states of the system are available; a different approach is adopted by Hermans & Zarrop [18] who attempt to design an observer in such a way that in the presence of a fault the sliding motion is destroyed. The approach in [14] seeks to design the sliding-mode observer to maintain sliding even in the presence of faults. More detailed simulations and analysis of a (slightly different) sliding-mode based FDI scheme is described in [13]. An interesting use of the concept of equivalent output error injection is described in [16]. A recursive design procedure based on successive evaluations of appropriate output error injection signals is shown to provide estimation of all unknown states of observable linear systems in finite time.

Acknowledgements The first author would like to thank the Royal Academy of Engineering for the provision of a Luverhulme Senior Research Fellowship.

References 1. L. Alvarez, R. Horowitz J. Yi, and L. Olmos. Dynamic friction model-based tyre/road friction estimation and emergency braking control. Journal of Dynamic Systems, Measurement and Control, 2005. 2. J.P. Barbot. Sliding-mode observers. In W. Perruquetti and J.P. Barbot, editors, Slidingmodes in Automatic Control. Marcel Dekker, 2001. 3. S.P. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear Matrix Inequalities in Systems and Control Theory. SIAM: Philadelphia, 1994.

Sliding-Mode Observers

241

4. W.L. Brogan. Modern Control Theory. Prentice Hall, Englewood Cliffs NJ., 1991. ´ om, and P. Lischinsky. A new model for control 5. C. Canudas de Wit, H. Olsson, K.J. Astr¨ of systems with friction. IEEE Transactions on Automatic Control, pages 419–425, 1995. 6. C. Canudas de Wit, P. Tsiotras, X. Claeys, J. Yi, and R.Horowitz. Friction tyre/road modeling, estimation and optimal braking control. In R. Johansson and A.Rantzer, editors, Nonlinear and Hybrid Systems in Automotive Control, pages 147–210. Springer-Verlag, London, UK, 2003. 7. C.M. Dorling and A.S.I. Zinober. A comparative study of the sensitivity of observers. In Proceedings of the IASTED Symposium on Applied Control and Identification– Copenhagen, pages 6.32–6.38, 1983. 8. S. Drakunov and V.I. Utkin. Sliding-mode observers: tutorial. In Proceedings of the 34th IEEE Conference of Decision and Control, pages 3376–3378, 1995. 9. C. Edwards and S.K. Spurgeon. On the development of discontinuous observers. International Journal of Control, 59:1211–1229, 1994. 10. C. Edwards and S.K. Spurgeon. Robust output tracking using a sliding-mode controller/observer scheme. International Journal of Control, 64:967–983, 1996. 11. C. Edwards and S.K. Spurgeon. Sliding-mode output tracking with application to a multivariable high temperature furnace problem. International Journal of Robust and Nonlinear Control, 7:337–351, 1997. 12. C. Edwards and S.K. Spurgeon. Sliding Mode Control: Theory and Applications. Taylor & Francis, 1998. 13. C. Edwards and S.K. Spurgeon. A sliding-mode control observer based FDI scheme for the ship benchmark. European Journal of Control, 6:341–356, 2000. 14. C. Edwards, S.K. Spurgeon, and R.J. Patton. Sliding-mode observers for fault detection. Automatica, 36:541–553, 2000. 15. F. Gustafsson. Slip-based tyre-road friction estimation. Automatica, pages 1087–1099, 1997. ¨ uner, and V.I. Utkin. On sliding-mode observers via equivalent control 16. I. Hasakara, U. Ozg¨ approach. International Journal of Control, 71:1051–1067, 1998. 17. H. Hashimoto, V.I. Utkin, J.X. Xu, H.Suzuki, and F. Harashima. Vss observers for linear time varying systems. In Proceedings of the 16th Annual Conference of the IEEE Industrial Electronic Society, pages 34–39, 1990. 18. F.J.J. Hermans and M.B. Zarrop. Sliding-mode observers for robust sensor monitoring. In Proceedings of the 13th IFAC World Congress, pages 211–216, 1996. 19. G. Herrmann, S.K. Spurgeon, and C. Edwards. A robust sliding-mode output tracking control for a class of relative degree zero and non-minimum phase plants: a chemical process application. International Journal of Control, 72:1194–1209, 2001. 20. U. Itkis. Control Systems of Variable Structure. Wiley, New York, 1976. 21. U. Kiencke. Real-time estimation of adhesion characteristic between tyre and road. In Proceedings of the IFAC World Congress, 1993. 22. D.G. Luenberger. An introduction to observers. IEEE Transactions on Automatic Control, 16:596–602, 1971. 23. W. Perruquetti and J.P. Barbot (Eds). Sliding Modes in Automatic Control. Marcel Dekker, 2001. 24. L.R. Ray. Nonlinear tyre force estimation and road friction identification - simulation and experiments. Automatica, 33:1819–1833, 1997. 25. J.J.E. Slotine, J.K. Hedrick, and E.A. Misawa. On sliding observers for nonlinear systems. Transactions of the ASME: Journal of Dynamic Systems, Measurement and Control, 109:245–252, September 1987.

242

C. Edwards et al.

26. R. Sreedhar, B. Fern´andez, and G.Y. Masada. Robust fault detection in nonlinear systems using sliding-mode observers. In Proceedings of the IEEE Conference on Control Applications, pages 715–721, 1993. 27. A. Steinberg and M.J. Corless. Output feedback stabilisation of uncertain dynamical systems. IEEE Transactions on Automatic Control, 30:1025–1027, 1985. 28. V. Utkin, J. Guldner, and J. Shi. Sliding Mode Control in Electromechanical Systems. Taylor & Francis, 1999. 29. V.I. Utkin. Variable structure systems with sliding-modes. IEEE Transactions on Automatic Control, 22:212–222, 1977. 30. V.I. Utkin. Sliding Modes in Control Optimization. Springer-Verlag, Berlin, 1992. ˙ 31. B.L. Walcott, M.J. Corless, and S.H. Zak. Comparative study of nonlinear state observation techniques. International Journal of Control, 45:2109–2132, 1987. ˙ 32. B.L. Walcott and S.H. Zak. State observation of nonlinear uncertain dynamical systems. IEEE Transaction on Automatic Control, 32:166–170, 1987. 33. J. Yi, L. Alvarez, X. Claeys, and R. Horowitz. Emergency braking control with an observer-based dynamic tyre/road friction model and wheel angular velocity measurement. Vehicle System Dynamics, 39:81–97, 2003. 34. J. Yi, L. Alvarez, R. Horowitz, and C. Canudas de Wit. Adaptive emergency braking control using a dynamic tyre/road friction model. In Proc. of 39th IEEE CDC, Sydney, Australia, pages 456–461, 2000.

9 Sliding-M ode Control in Systems with Output Time Delay Alan S.I. Zinober, G. Liu , and Yuri B. Shtessel

9.1

Introduction

The main features of sliding-mode control (SMC) and the associated feedback control law of Variable Structure Control (VSC) systems will be summarized in this chapter. SMC is a well-known solution to the problem of the deterministic control of uncertain systems, since it yields invariance to a class of parameter variations [4,7,33–36,43]. The characterizing feature is sliding motion, which occurs when the system state repeatedly crosses certain subspaces, or sliding hyperplanes, in the state space. A sliding controller may comprise nonlinear and linear parts, and has been well studied in the literature. Numerous practical applications of SMC have been reported in the literature. These include aircraft flight control [32] ,helicopter flight control, spacecraft flight control, ship steering, turbogenerators, electrical drives, overhead cranes, industrial furnace control, electrical power systems, robot manipulators, automobile fuel injection and magnetic levitation. There has been research on determining sliding hyperplanes by various approaches including complete and partial eigenstructure assignment, and reduction of the sensitivity to unmatched parameter variations [5, 6]. After presenting the underlying theory of the sliding-mode, we shall describe some of the techniques relating to the design of the sliding hyperplanes. For completeness we also present a suitable control law to ensure the attainment of the slidingmode. A straightforward scalar illustrative example is presented. Sliding-mode observers have also been designed and there have been important extensions of SMC to higher order sliding-modes (see, for example, [9, 10, 20, 21]. We then consider an interesting application of SMC (ully reported in [22] to the tracking of a nonminimum phase control system which is difficult to control [28, 29]. Output tracking in SISO fully linearizable nonlinear systems with an output time delay is considered using sliding-mode control and the system centre approach. Using Pad´e approximations for the delay, the problem is reduced to the tracking of a nonminimum phase control system. This system is transformed into a corresponding state tracking problem, where the state tracking profiles are generated by the equations of the stable system centre. The sliding-mode control approach is developed and numerical examples are presented. Also the describing function is used to evaluate the limit cycle generated by the feedback of the actual delayed output and these results agree very closely with the system simulations. M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 243-264, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

244

A. S. I. Zinober, G. Liu, and Y.B. Shtessel

The paper is organized as follows: Section 7.7 is dedicated to the output tracking problem formulation. Section 7.10 uses Pad´e approximation to approximate the delayed system. Section 4 presents the system centre and sliding-mode control to solve the output tracking problem. A numerical example demonstrating various aspects of different Pad´e approximations is presented. We use the describing function to evaluate the limit cycle generated by the actual delayed output. Conclusions are summarized in Section 9.4.

9.2 9.2.1

Sliding-Mode Control Regulator System

As our basic control system we consider the uncertain linear regulator x(t) ˙ = [A + ∆ A(t)]x(t) + [B + ∆ B(t)]u(t) + f (x, u,t)

(9.1)

where x is an n-vector of states and u is an m-vector of controls. It is assumed that n > m, that B is of full rank m and that the pair (A, B) is controllable. The matrices ∆ A and ∆ B represent the variations and uncertainties in the plant parameters and the control interface respectively. f represents uncertain time-varying additive terms. It is assumed further that the parameter uncertainties and disturbances are matched, occurring only on the control channels, i.e., R(B) = R([B, ∆ B]) (where R(·) denotes the range space); and that rank [B + ∆ B(t)] = m for all t ≥ 0. This implies that for suitable choice of limiting values of the control, one can achieve total invariance to parameter variations and disturbances [7]. The overall aim is to drive the system state from an initial condition x(0) to the state space origin as t → ∞. The jth component u j ( j = 1, . . . , m) of the state feedback control vector u(x) has a discontinuity on the jth switching surface which is a hyperplane M j passing through the state origin. Defining the hyperplanes by M j = {x : c j x = 0}, ( j = 1, 2, . . . , m)

(9.2)

(where c j is a row n-vector), the sliding-mode occurs when the state lies in M j for all j, i.e., in the sliding subspace M=

m \

Mj

(9.3)

j=1

In practice the control discontinuity may be replaced by a soft nonlinearity to reduce chattering [3].

9.2.2

Model-Following Control System

Model-following control systems are very widely used in practice and VSC can be designed in a manner very similar to the regulator system [18, 32, 44]. Consider the system x(t) ˙ = [A + ∆ A(t)]x(t) + [B + ∆ B(t)]u(t) + f (x, u) x˙m (t) = Am xm (t) + Bm r(t)

(9.4)

Sliding-Mode Control in Systems with Output Time Delay

245

where the first equation (as in (9.1)) describes the actual plant, and the second equation is the model plant with xm an n-vector of model states and r a vector of reference inputs. It is desired that the actual plant states follow the model states. The error e(t) = xm (t) − x(t)

(9.5)

should be forced to zero as time t → ∞ by suitable choice of the control u. Subject to the matrices A, B, ∆ A, ∆ B, Am and Bm satisfying certain structural and matching properties [18], we can achieve the desired objective with suitable control. The error model satisfies e(t) ˙ = Am e(t) + [(Am − A)x(t) − f + Bm r(t)] − Bu(t) (9.6)

and, subject to certain matching conditions [32], the model equations with suitable linear control components, reduce to e(t) ˙ = Am e(t) − Bu(t) (9.7)

The VSC of this error system may be readily designed, using the techniques previously described, by associating e with x in earlier sections. The sliding hyperplanes are now in the error state space. Further details and examples of practical time-varying and nonlinear avionics systems are given in [32]. Tracking problems can also be controlled and in the second half of this chapter a tracking problem is studied.

9.2.3

Sliding-Mode

When considering the synthesis of the sliding hyperplanes, it is sufficient to study the ideal regulator system, without uncertainties and disturbances, given by x(t) ˙ = Ax(t) + Bu(t)

(9.8)

Matched uncertainties are handled by suitable choice of the control function. From (9.2) the sliding-mode satisfies s = Cx(t) = 0 , t ≥ ts (9.9) where ts is the time when the sliding subspace is reached, and C is an m × n matrix. Differentiating equation (9.9) with respect to time, and substituting for x(t) ˙ from (9.8) gives Cx(t) ˙ = CAx(t) +CBu(t) = 0 ,

t ≥ ts

(9.10)

Equation (9.10) may be rearranged to give CBu(t) = −CAx(t)

(9.11)

The hyperplane matrix C is selected so that |CB| 6= 0, and therefore the product CB is invertible. Hence (9.11) may be rearranged to give the following expression for the equivalent control [33] ueq (t) = −(CB)−1CAx(t) = −Kx(t) (9.12) where ueq (t) is the linear open-loop control which would force the trajectory to remain in the null space of C, during sliding. Substituting for ueq (t) from equation (9.12) into (9.8) gives x(t) ˙ = {I − B(CB)−1C }Ax(t) , = (A − BK)x(t)

t ≥ ts

which is the system equation for the closed-loop system dynamics during sliding.

(9.13) (9.14)

246

A.S.I. Zinober, G. Liu, and Y.B. Shtessel

This motion is independent of the actual nonlinear control and depends only on the choice of C, which determines the matrix K. The purpose of the control u is to drive the state into the sliding subspace M , and thereafter to maintain it within the subspace M . The convergence of the state vector to the origin is ensured by suitable choice of the feedback matrix K. The determination of the matrix K or alternatively, the determination of the matrix C defining the subspace M , may be achieved without prior knowledge of the form of the control vector u. (The reverse is not true). The null space of C, N (C), and the range space of B, R(B), are, under the hypotheses given earlier, complementary subspaces, so N (C) ∩ R(B) = {0}. / Since motion lies entirely within N (C) during the ideal sliding-mode, the dynamic behaviour of the system during sliding is unaffected by the controls, as they act only within R(B). The development of the theory and design principles is simplified by using a particular canonical form for the system, which is closely related to the Kalman canonical form for a multivariable linear system. By assumption the matrix B has full rank m; so there exists an orthogonal n × n transformation matrix T such that 0 T B= (9.15) B2 where B2 is m × m and nonsingular [5]. The orthogonality restriction is imposed on T for reasons of numerical stability, and to remove the problem of inverting T when transforming back to the original system. The transformed state is y = T x, and the state equation becomes y(t) ˙ = T AT T y(t) + T Bu(t)

(9.16)

The sliding condition is CT T y(t) = 0, ∀ t ≥ ts . If the transformed state y is now partitioned as yT = (yT1 yT2 ); y1 ∈ R n−m , y2 ∈ R m (9.17)

and the matrices T AT T , T B and CT T are partitioned accordingly, then equation (9.16) may be written as the following pair of equations y˙1 (t) = A11 y1 (t) + A12 y2 (t) y˙2 (t) = A21 y1 (t) + A22 y2 (t) + B2 u(t)

(9.18)

The sliding condition becomes C1 y1 (t) +C2 y2 (t) = 0 , where T AT T =

A11 A12 A21 A22

,

t ≥ ts

CT T = (C1

(9.19)

C2 )

(9.20)

and C2 is nonsingular because CB is nonsingular. This canonical form is central to hyperplane design methods and it plays a significant role in the solution of the reachability problem i.e., the determination of the control form ensuring the attainment of the sliding-mode in M [5, 6, 36, 38–42]. Equation (9.19) defining the sliding-mode is equivalent to y2 (t) = −Fy1 (t) where the m × (n − m) matrix F is defined by

(9.21)

Sliding-Mode Control in Systems with Output Time Delay

F = C2−1C1

247 (9.22)

so that in the sliding-mode y2 is related linearly to y1 . The sliding-mode satisfies equation (9.21) and y˙1 = A11 y1 (t) + A12 y2 (t) (9.23) This represents an (n − m) th order system in which y2 plays the role of a state feedback control. So we get y˙1 (t) = (A11 − A12 F)y1 (t) (9.24)

which is known as the reduced-order equivalent system. The design of a stable sliding-mode such that y → 0 as t → ∞, requires the determination of the gain matrix F such that A11 − A12 F has n − m left-hand half-plane eigenvalues.

9.2.4

Feedback Control

Once the sliding hyperplanes have been selected, attention must be turned to solving the reachability problem. This involves the selection of a state feedback control function u : R n → R m which will drive the state x into N (C) and thereafter maintain it within this subspace. There is a virtually unlimited number of possible forms for this control function, the only essential features of the form chosen being discountinuity on one or more subspaces containing N (C). In general the variable structure control law consists of two additive parts; a linear control law uℓ and a nonlinear part un , which are added to form u. The linear control is merely a state feedback controller uℓ (x) = Lx (9.25) while the nonlinear feedback controller un incorporates the discountinuous elements of the control law. Consider here the unit vector control un (x) = ρ

s Cx =ρ , ksk kCxk

ρ >0

(9.26)

in the form [23] u(x) = Lx + ρ

Nx k Mx k

(9.27)

where the null spaces of N, M and C are coincident : N (N) = N (M) = N (C). Starting from the transformed state y, we form a second transformation T2 : R n → R n such that z = T2 y where T2 =

In−m 0 F Im

(9.28)

The matrix T2 is clearly nonsingular, with inverse I 0 T2 −1 = n−m −F Im Partitioning zT = (zT1

(9.29)

(9.30)

zT2 ) with z1 ∈ R n−m and z2 ∈ R m z1 = y1 ;

z2 = Fy1 + y2

(9.31)

248

A.S.I. Zinober, G. Liu, and Y.B. Shtessel

from which it is clear that the conditions s ≡ 0 and z2 ≡ 0 are equivalent (in the sense that the points of the state space at which s = 0 are precisely the points at which z2 = 0). The transformed system equations become z˙1 = Σ z1 + A12 z2

(9.32)

z˙2 = Θ z1 + Φ z2 + B2 u

(9.33)

where

Σ = A11 − A12 F

Θ = F Σ − A22 F + A21 Φ = FA12 + A22

(9.34)

In order to attain the sliding-mode it is necessary to force z2 and z˙2 to become identically zero. Define the linear part of the control to be uℓ (z) = −B−1 2 {Θ z1 + (Φ − Φ∗ )z2 }

(9.35)

where Φ∗ is any m × m matrix with left-hand half-plane eigenvalues. Transforming back into the original state space (x-space) gives L = −B−1 2 (Θ

Φ − Φ∗ )T2 T

(9.36)

The linear control law uℓ drives the state component z2 to zero asymptotically; to attain N (C) in finite time, the nonlinear control component un is required. This nonlinear control must be discontinuous whenever z2 = 0, and continuous elsewhere. Letting P2 denote the positive definite unique solution of the Lyapunov equation P2 Φ∗ + Φ∗ T P2 = −Im

(9.37)

then P2 z2 = 0 if and only if z2 = 0, and we may take un (z) = −ρ

B−1 2 P2 z2 , k P2 z2 k

z2 6= 0

(9.38)

where ρ > 0 is a scalar parameter to be selected by the designer. When z2 = 0, un may be arbitrarily defined as any function satistying k un k ≤ ρ . Expressing the control in x-space, we have N = −B−1 2 (0 P2 )T2 T M = (0 P2 )T2 T

(9.39) (9.40)

For the more general system (9.1) in which disturbances and uncertainties are present, a similar control structure may be employed. However, in this case the scalar ρ of (9.38) is replaced by a time-varying state-dependent function incorporating two design parameters γ1 , γ2 , upon which the time to reach N (C) also depends [23]. Discontinuous control produces chatter motion in the neighbourhood of the sliding surface. In may practical applications this cannot be tolerated. There are numerous techniques to “smooth” the control function. Perhaps the most straightforward smoothed continuous nonlinear control, which eliminates the chatter motion, is (see, for example, [3, 43]

Sliding-Mode Control in Systems with Output Time Delay

u(x) = Lx + ρ

9.2.5

Nx , kMxk + δ

δ > 0+

249

(9.41)

Second-Order Example

To illustrate some of the main ideas consider the simple scalar double-integrator plant x(t) ¨ = bu(t) + f (t) with the positive parameter b uncertain but taking known bounded maximum and minimum values, and f (t) a disturbance. Here the sliding subspace will be a one-dimensional space, a straight line through the state origin s = gx + x˙ = 0 During sliding for t > ts we require lim s˙ > 0

and

s=0

and

s→0−

lim s˙ < 0

s→0+

and then s˙ = 0

i.e., the state remains on the sliding surface. So x˙ = −gx with eigenvalue −g i.e., the dynamics of a reduced first-order system (with order n − 1). x(t) = x(ts )e−gt One obtains exactly the closed-loop eigenvalue −g by specifying the sliding line (9.2.5). The dynamics in the sliding-mode are independent of the parameter b. The discontinuous control s u=ρ |s| can maintain sliding motion on s = 0 within a bounded region of the state origin, for a range of values of b with the precise value of b not required to be known [33]. The equivalent control which theoretically can maintain the state on the sliding line, is the linear control ueq = −

gx b

To achieve the sliding-mode with this linear control would require exact knowledge of b; unlike for the case of nonlinear control. Smooth nonlinear control has the form s u=ρ |s| + δ Simulation results are presented in Figures. 9.2.5 and 9.2.5 for discontinuous (δ = 0) and smooth control (δ = 0.01) with ρ = 1. The state trajectories are very similar. During the sliding-mode the smooth control is equal to the equivalent control ueq , which is included in the control graph of Figure 9.2.5. Note the elimination of chatter when using the smoothing control. The system is invariant to a matched disturbance function f (t).

A.S.I. Zinober, G. Liu, and Y. B. Shtessel

250

states; relay)

control

0.5

0.5

0

0

u

1

x1 x2

1

−0.5

−1

−0.5

0

2

4

−1

6

0

2

4

t

6

t

s(t)

phase plane 0

1.2

1

−0.2

0.8

s

x2

0.6

−0.4

0.4

0.2

−0.6

0

−0.2

0

2

4

t

6

−0.8

0

0.2

0.4

0.6

0.8

1

x1

Fig. 9.1. Double integrator with discontinuous relay control

9.3

Application Problem

We now consider an interesting application of SMC to the tracking of a nonminimum phase control system which is difficult to control [28, 29]. This was fully reported in [22]. Output tracking in SISO fully linearizable nonlinear systems with an output time delay is considered using sliding-mode control and the system centre approach. Using Pad´e approximations for the delay, the problem is reduced to the tracking of a nonminimum phase control system. This system is transformed into a corresponding state tracking problem, where the state tracking profiles are generated by the equations of the stable system centre. The sliding-mode control approach is developed and numerical examples are presented. Also the describing function is used to evaluate the limit cycle generated by the feedback of the actual delayed output and these results agree very closely with the system simulations. Output time delay is a common feature in many systems and must be taken into account when designing a controller. The output tracking of a real-time reference profile in nonlinear systems with output delay by sliding-mode control was addressed in the paper [SHT 03a]. In addition to the first-order Pad´e approximation, the more precise second and third-order Pad´e approximations have been used to replace the output delay element so that we can transfer the problem into the output tracking of a nonminimum phase system. For the output reference profile given by a linear exogenous system, the nonminimum phase output tracking problem is transformed to the corresponding state tracking problem. Bounded state tracking profiles are generated by equations of the stable system centre [27]. The sliding-mode control algorithm is designed for the approximate nonminimum phase model of the system. In simulations many comparisons are made between different order Pad´e approximations and different delay time.

Sliding-Mode Control in Systems with Output Time Delay

251

Also the describing function is used to evaluate the limit cycle generated by the feedback of the actual delayed output and these results agree very closely with the system simulations. control 1

0.5

0.5

0

0

u

x1 x2

states; smoothed (d=0.01) 1

−0.5

−1

−0.5

0

2

4

−1

6

0

2

4

t

6

t

s(t)

phase plane 0

1.2

1

−0.2

0.8

s

x2

0.6

−0.4

0.4

0.2

−0.6

0

−0.2

0

2

4

6

−0.8

0

0.2

0.4

t

0.6

0.8

1

x1

Fig. 9.2. Double integrator with smoothed control

9.3.1

Problem Formulation

Consider a controllable fully feedback linearizable nonlinear SISO dynamic system without time delay x˙ = f (x,t) + g(x,t)u, y = h(x) (9.42) where x(t) ∈ Rn is a state vector, y(t) ∈ R1 a controlled output and u(t) ∈ R1 is a control input. As a fully linearizable relative degree n system, it can be transformed [ISI 95] to y(n) = ϕ (ξ ,t) + b(ξ ,t)u

(9.43)

where ξ = [y, y, ˙ · · · , y(n−1) ]T ∈ Rn . Following the approach developed in [12], we define a coordinate transformation   1 0 ··· ··· 0  0 1 ··· ··· 0  z  =  ··· ··· ··· ··· 0 ξ q a0 a1 · · · an−2 1

with ai > 0, where

q = yn−1 + an−2 yn−2 + · · · + a1 y˙ + a0 y

is a new output to get a relative degree 1 system,

z(t) ∈ Rn−1

and

q(t) ∈ R1 .

(9.44)

(9.45)

252

A.S. I. Zinober, G. Liu, and Y. B. Shtessel

The system (9.42) is rewritten in the new basis (9.44) as  z˙1 = z2 ,      z˙2 = z3 , ··· ···   z˙ = q − an−2 zn−1 − · · · − a1 z2 − a0 z1    n−1 ˆ q,t)u q˙ = ϕˆ (z, q,t) + b(z, where

and

(9.46)

ϕˆ (z, q,t) = −a0 an−2 z1 − (a1 an−2 − a0 )z2 − · · · − (a2n−2 − an−3 )zn−1 +an−2 q + ϕ (z, q − a0 z1 − · · · − an−2 zn−1 ,t) bˆ = b(q − a0 z1 − · · · − an−2 zn−1 ,t).

Note that the internal dynamics of the system (9.46) are stable and can be disregarded when solving the output tracking problem as time increases. So output tracking in system (9.42) can be transformed to output tracking in the scalar system ˆ q,t)u q˙ = ϕˆ (z, q,t) + b(z,

(9.47)

using the transformation (9.44). Assume here the desired command output profile is given in the form yc = A¯ + B¯ sin ωn t

(9.48)

Then we can assume the output tracking ( command ) profile qc (corresponding to q) as follows: qc = A + B sin ωn t +C cos ωn t (9.49) where A, B,C and ωn are piecewise constants and A, B,C are unknown. The signal (9.49) can be described by a linear system of exogenous differential equations with the following characteristic equation: (9.50) P3 (λ ) = λ 3 + 0λ 2 + ωn2 λ + 0 ˆ = bˆ 0 (·)(1 + δ (·)), |δ (·)| ≤ β1 < 1. The function Now we restrict ϕˆ (·) < α1 , α1 > 0 and b(·) bˆ 0 (·) is assumed known. Also assume the output of (9.47) is accessible with time delay y(t) ˆ = q(t − τ ).

(9.51)

The problem is to design sliding-mode control u(t) that forces the output variable y(t) ˆ to track asymptotically the command profile.

9.3.2

Pade´ Approximations and Time Delay Systems

In this section, we consider the Pad´e approximation to approximate the time delay. The Pad´e approximation [2] uses the quotient of two polynomials to estimate a power series. We will use the direct solutions of the Pad´e equations [L/M] = PL (x)/QM (x), where PL (x) is a polynomial of degree L and QM (x) is a polynomial of degree M. When we approximate a formal power series f (x), the explicit equation is

Sliding-Mode Control in Systems with Output Time Delay

lim

x→∞

253

QM (x) f (x) − PL (x) =0 xL+M+1

Letting f (x) = ex , we can get the first, second and third-order Pad´e approximations x 2 x 1− 2 x x2 1+ + 2 12 ex ≈ x x2 1− + 2 12 x x2 x3 1+ + + 2 10 120 ex ≈ x x2 x3 1− + − 2 10 120 ex ≈

1+

y(s) ˆ

The corresponding Laplace transform of equation (9.51) q(s) = e−sτ can always be approximated by first-order, second-order and third-order Pad´e approximation as e−sτ

e−sτ

e−sτ

sτ 2 ≈ sτ 1+ 2 sτ 1− 2 ≈ sτ 1+ 2 sτ 1− 2 ≈ sτ 1+ 2 1−

s2 τ 2 12 s2 τ 2 + 12 s2 τ 2 s3 τ 3 + − 10 120 2 2 s τ s3 τ 3 + + 10 120 +

(9.52)

where s is the Laplace variable. We introduce a new output variable y˜ to make the approximation exact and the system (9.51) can be always approximated by

For the first-order

for the second-order

and for the third-order

ζ˙ = Q˜1 ζ + Q˜2 y˜ ˜ ζ ,t) + ψ (y, ˜ ζ )u y˜˙ = ϕ¯ (y,

  Q˜1 = τ2 , Q˜2 = − τ4 ϕ¯ (y, ˜ ζ ,t) = −ϕˆ (z, q,t) + τ2 ζ − τ4 y˜  ˆ q,t) ψ (y, ˜ ζ ) = −b(z,

 0 1  12 1 ˜   Q˜1 = 12 6 , Q2 = − τ 6 − τ2 τ τ ˜1 ζ + Q˜2 y˜ ¯ ˆ ϕ ( y, ˜ ζ ,t) = ϕ (z, q,t) + Q    ˆ q,t) ψ (y, ˜ ζ ) = b(z,

(9.53)

(9.54)

(9.55)

254

A.S.I. Zinober, G. Liu, and Y.B. Shtessel

  0      Q˜1 =  0

  24  1 0 τ  0 1  , Q˜2 = −  288 τ2 60 12 120 2256 − τ3 τ2 τ τ3    ϕ¯ (y, ˜ ζ ,t) = −ϕˆ (z, q,t) + Q˜1 ζ + Q˜2 y˜   ˆ q,t) ψ (y, ˜ ζ ) = −b(z,

(9.56)

The system (9.53) is nonminimum phase because Q˜ 1 has at least one non-Hurwitz eigenvalue. Note that the output delay system tracking problem has been transformed into a nonminimum phase system output tracking problem (with no delay) by the Pad´e approximation.

9.3.3

System Centre Method and Sliding-Mode Control

The nonminimum phase system tracking problem is solved in this section by the system centre method and sliding mode-control. The equation of the system centre (ideal internal dynamics) [27] that defines a command (tracking) profile ζc (t) for the internal state vector for the system with delay approximated by the nonminimum phase system (9.53) is

ζ˙c = Q˜1 ζc + Q˜2 qc

(9.57)

which is unstable. According to the theorem of system centre in [27], the stable system centre ζ˜c (t) that converges asymptotically to the bounded particular solution ζc (t) of the unstable equation of system centre [16], i.e. ζ˜c (t) → ζc (t), is given by (3) (2) (1) (2) (1) ζ˜c + c2 ζ˜c + c1 ζ˜c + c0 ζ˜c = −(P2 θc + P1 θc + P0 θc )

where θc = Q˜2 qc , and the coefficients of P0 , P1 , P2 are computed as follows:  −1 P = c0 Q˜1    0 −1 −2  (c −ω 2 )Q˜ +(c −c ω 2 )Q˜ P1 = 1 n 1 2 ˜0 −2 2 n 1 Q I+ ω n 1  −1 −2 −3   c Q˜ +(c1 −ωn2 )Q˜1 +c0 Q˜1  P2 = 2 1 2 ˜ −2

(9.58)

(9.59)

I+ωn Q1

The coefficients c0 , c1 , c2 are chosen to provide the specified eigenvalues of the homogeneous differential equation. (3) (2) (1) (9.60) ζ˜c + c2 ζ˜c + c1 ζ˜c + c0 ζ˜c = 0. The sliding-mode control is introduced with the sliding function defined as

σ = eq +C1 e˜ζ

(9.61)

where eq = qc − y, ˜ and e˜ζ = ζ˜c − ζ .

Assume that the sliding-mode exists on the sliding surface σ = 0, then eq = −C1 e˜ζ . Therefore e˙˜ζ = (Q˜1 − Q˜2C1 )e˜ζ + (ζ˙˜c − Q˜1 ζ˜ + Q˜2 qc ), eq = −C1 e˜ζ

(9.62)

Sliding-Mode Control in Systems with Output Time Delay

255

Since ζ˜c (t) → ζc (t) with increasing time, then e˜ζ (t) → eζ (t) and ζ˙˜c − Q˜1 ζ˜ + Q˜2 qc → 0. So the sliding-mode dynamics (9.62) asymptotically approaches the homogeneous differential equation e˙˜ζ = (Q˜1 − Q˜2C1 )e˜ζ (9.63) The coefficient C1 is selected to provide asymptotic tracking error dynamics (9.63) to zero. The sliding-mode control is designed as follows: u = bˆ −1 [q˙c − ϕˆ (·) +C1 (Q1 eζ + Q2 eq ) − ρ sign(σ )]

(9.64)

where ρ is a sufficiently large positive gain. Note that the standard sliding-mode controller design can be found in the classical book [8].

9.3.4

Numerical Example and Simulations

Consider the relative degree 2 second-order system: x˙1 = x2 x˙2 = −x2 + u y = x1

(9.65)

Transforming the plant to the form (9.46) with relative degree equal to one, introduce a new state vector in accordance with (9.44) z1 1 0 x = · 1 (9.66) q a0 1 x2 The system (9.65) is rewritten in the new basis (9.66) as z˙1 = −a0 z1 + q q˙ = −(a0 − 1)a0 z1 + (a0 − 1)q + u

(9.67)

The desired command output profile has the characteristic equation (9.50). Now assume that the system output (9.67) is accessible with a time delay yˆ = q(t − τ ). The problem is to design the SMC that provides asymptotic tracking yˆ → qc as time increases. Replacing the time-delay function by the respective first-order, second-order and third-order Pad´e approximation to the system output y, ˆ the system (9.67) is approximately represented by the nonminimum phase systems without delay z˙1 = −a0 z1 − y˜ + ζ ζ˙ = τ2 ζ − τ4 y˜ y˙˜ = ( τ2 − a0 + 1)ζ + (a0 − 1 − τ4 )y˜ + (a0 − 1)a0 z1 − u

(9.68)

z˙1 = −a0 z1 + y˜ − ζ1 ζ˙1 = ζ2 − 12 τ y˜ ζ˙2 = τ6 ζ2 − 12 − 72 y˜ τ2 τ2 ˙y˜ = (− 12 + a0 − 1)y˜ + ζ2 − (a0 − 1)ζ1 − a0 (a0 − 1)z1 + u τ

(9.69)

256

A.S.I. Zinober, G. Liu, and Y. B. Shtessel

z˙1 = −a0 z1 − y˜ + ζ1 ζ˙1 = ζ2 − 24 τ y˜ ζ˙2 = ζ3 − 288 y˜ τ2 60 − ζ˙3 = 12 ζ ζ + 120 y˜ ζ − 2256 3 τ τ2 2 τ3 1 τ3 ˙y˜ = (− 24 + a0 − 1)y˜ + ζ2 − (a0 − 1)ζ1 + a0 (a0 − 1)z1 − u τ

(9.70)

where the output y˜ is an approximation to the original output y. ˆ Now we select parameters as follows: c0 = 1000, c1 = 300, c2 = 30, ωn = 2, a0 = 20 and A¯ = 1, B¯ = 2. Also we choose C1 = −0.75 for the first-order Pad´e approximation, [−0.75, 0] for the secondorder Pad´e approximation and [−0.75, 0, 0] for the third-order Pad´e approximation. Notice that the sign of control in the second-order Pad´e approximation is positive and those in the first and third-order Pad´e approximations are negative. The control is u = 25sign(σ ) for the second-order Pad´e approximation and u = −25sign(σ ) for the first-order and the third-order Pad´e approximations. R R We have used MATLAB Simulink

and Scilab to get the tracking results for different τ (see Figures. 1–9 ). Note that all the errors mentioned are the maximum magnitude of absolute errors after reaching steady state. Comparing Figures 1–3 we see that for τ = 0.2 the first-order Pad´e approximation has good tracking results (0.0068) although the errors for the second-order (0.0005) and the third-order (0.0045) Pad´e are smaller. However when τ increases to 0.5 (see Figures 4–6), the first-order Pad´e results is bad since the output error is about 0.068. So the advantages of second-order (0.00078) and third-order Pad´e (0.0048) become obvious. For τ = 1.0 the first-order Pad´e approximation is quite clearly bad (see Figure 7), so we should not use the first-order Pad´e approximation to estimate the delayed output. Also the second-order Pad´e error (0.035) is bigger than the third-order Pad´e error (0.004). In conclusion higher-order Pad´e approximations will generally have better results for larger τ and the first-order Pad´e can only be used in smaller time delays.

pade1 0.06 0.05 0.04 0.03

error

0.02 0.01 0.00 −0.01 −0.02 −0.03 −0.04 1

2

3

4

5

6

7

8

9

time

Fig. 9.3. 1st-order Pad´e τ = 0.2: error

10

11

Sliding-Mode Control in Systems with Output Time Delay

257

pade2 2.0e−003

1.5e−003

1.0e−003

error

5.0e−004

0.0e+000

−5.0e−004

−1.0e−003

−1.5e−003

−2.0e−003 3

11

10

9

8

7

6

5

4

time

Fig. 9.4. 2nd-order Pad´e τ = 0.2: error pade3 0.006

0.004

0.002

error

0.000

−0.002

−0.004

−0.006

−0.008

−0.010 0

1

2

3

4

5

6

7

8

9

10

time

Fig. 9.5. 3rd-order Pad´e τ = 0.2: error Also the effects of variations in the parameters (C1 , a0 , ρ and the system centre formulation) have been tested. Detailed results will appear in a future paper.

9.3.5

Feedback by yˆ and Describing Function

In a real-life situation the feedback should be the actual delayed output yˆ and not y, ˜ i.e., yˆ should be included in the sliding-mode surface design. Let us consider the first-order case to establish the effect of feedback by yˆ by both simulation and also using the Describing Function approach. Using the Scilab program for the whole system with the actual delayed output, we get a limit cycle as in Figure 10. In fact, it can be shown that the frequency and the magnitude of the limit cycle are not altered by A and B, so we can assume both of them zero, i.e., qc = 0. The Describing Function method (DF) is briefly introduced and used to analyse the limit cycle above. DF is a well-known method that can be used to detect oscillations (limit cycles). The basic assumptions for our system are as follows:

A.S.I. Zinober, G. Liu, and Y. B. Shtessel

258

pade1 0.15

0.10

error

0.05

0.00

−0.05

−0.10

−0.15 1

9

8

7

6

5

4

3

2

10

time

Fig. 9.6. 1st-order Pad´e τ = 0.5: error pade2 2.0e−003

1.5e−003

1.0e−003

error

5.0e−004

0.0e+000

−5.0e−004

−1.0e−003

−1.5e−003

−2.0e−003 2

4

3

5

6

7

8

9

10

time

Fig. 9.7. 2nd-order Pad´e τ = 0.5: error No input, i.e. qc = 0. Nonlinearity is symmetric. Nonlinearity does not depend on frequency. Assume at the input point of the nonlinearity, e(t) = E sin(ω t)), i.e., first-order sinusoidal approximation. 5. Linear part (Laplace transform for the whole system) acts as a low-pass filter, i.e., higherorder harmonic components are damped.

1. 2. 3. 4.

In general, the describing function at the output point of the nonlinearity can be defined as n(t) = N(e(t)) = N(E sin(ω t)) = n(ω t) where n(ω t) is defined by the Fourier series n(ω t) = From the assumptions above

∞ ∞ A0 + ∑ Ak cos(kω t) + ∑ Bk sin(kω t) 2 k=1 k=1

Sliding-Mode Control in Systems with Output Time Delay

pade3 0.020

0.015

0.010

error

0.005

0.000

−0.005

−0.010

−0.015

−0.020

−0.025 1

7

6

5

4

3

2

8

9

10

time

Fig. 9.8. 3rd-order Pad´e τ = 0.5: error pade1 3

2.5

2

1

c

y and y

1.5

0.5

0

−0.5

−1

−1.5

0

5

10

15

20

time t

Fig. 9.9. 1st-order Pad´e τ = 1.0: yc and y n(ω t) = A1 cos(ω t) + B1 sin(ω t) where A1 =

2 T

Z T /2

n(ω t) cos(ω t)dt =

1 π

Z π

n(α ) cos(α )d α

B1 =

2 T

Z T /2

n(ω t) sin(ω t)dt =

1 π

Z π

n(α ) sin(α )d α

and

−T /2

−T /2

The nonlinearity is the sign function

−π

−π

  M t >0 0 t =0 N(t) =  −M t < 0

So N(t) is an odd function and therefore A1 = 0, B1 =

1 π

Z π

−π

n(α ) sin(α )d α =

4M π

25

259

260

A.S.I. Zinober, G. Liu, and Y.B. Shtessel pade2 0.06

0.04

0.02

error

0.00

−0.02

−0.04

−0.06

−0.08 2

5

4

3

6

9

8

7

10

time

Fig. 9.10. 2nd-order Pad´e τ = 1.0: error pade3 0.020 0.015 0.010 0.005

error

0.000 −0.005 −0.010 −0.015 −0.020 −0.025 −0.030 4

6

5

8

7

9

10

time

Fig. 9.11. 3rd-order Pad´e τ = 1.0: error

The describing function is 4M πE It remains to calculate the total Laplace transform function for the whole system.The Laplace transform ( from u to z1 ) for the system with feedback yˆ is N( jω ) =

1 Z1 (s) = U(s) s(s + 1)

(9.71)

4 (s + a0 ) τ τ −s −s τ ˆ For qc = 0, σ = −yˆ + 0.75ζ . Also Y (s) = Q(s)e = (s + a0 )Z(s)e and ζ (s) = 2 s+ τ

σ (s) = (s + a0 )Z1 (s)( Therefore the total transfer function

3 τ

s + τ2

− e−sτ )

(9.72)

Sliding-Mode Control in Systems with Output Time Delay

261

Solution 0.10

0.05

y

0.00

−0.05

−0.10

−0.15 0

2

1

3

4

5

6

7

8

9

10

time

Fig. 9.12. Feedback by delayed output yˆ for 1st-order Pad´e τ = 0.2: y

G(s) = −

3 4M s + a0 ( τ 2 − e−sτ ) π E s(s + 1) s + τ

(9.73)

For a limit cycle to exist G( jω )N( jω ) = −1 + 0 j.

(9.74)

Many ‘stable’ solutions exist and the values of E and T (corresponding to ω ) are listed for largest E value in the tables. From the Tables 1–3, we can see that the period T and the Table 9.1. τ = 0.1: magnitudes of E and yˆ and period T Program for τ = 0.1 E T yˆ Simulink DF 1.40 0.148 0.949 Simulink for whole system 1.405 0.148 0.945

Table 9.2. τ = 0.2: magnitudes of E and yˆ and period T Program for τ = 0.2 E T yˆ Simulink DF 4.10 0.36 2.4815 Simulink for whole system 4.035 0.361 2.47

Table 9.3. τ = 0.3: magnitudes of E and yˆ and period T Program for τ = 0.3 E T yˆ Simulink DF 8.9 0.569 5.452 Simulink for whole system 8.896 0.57 5.427 magnitudes of E and yˆ are consistent. This indicates that the describing function method can be used to evaluate the limit cycle.

262

9.4

A.S.I. Zinober, G. Liu, and Y. B. Shtessel

Conclusions

Sliding-mode control has been described in detail. Output tracking in causal nonlinear systems with an output delay via sliding-mode control is considered. The problem is reduced to the tracking of a real-time output reference profile given by an exogenous system in a nonminimum phase system without delay via the first, second and third-order Pad´e approximations. The nonminimum phase output tracking problem is transformed to a corresponding state tracking problem and the system centre method is used to deal with the unstable internal dynamics. We use sliding-mode control to solve the problem and present many numerical results. The describing function is used to evaluate the limit cycle generated when the actual delayed output feedback is used in the feedback and these results agree very closely with system simulations.

References 1. D.P. Atherton (1982) Nonlinear Control Engineering, Van Nostrand Reinhold Company Ltd 2. G.A. Baker Jr. (1973) Recursive Calculation of Pad´e Approximation, Pad´e Approximants and Their Applications. ed Graves-Morris PR 83–91 3. J.A. Burton and A.S.I. Zinober (1986) Continuous approximation of variable structure control. Int J Systems Science 17: 876–885 ˙ and G.P. Matthews (1988) Variable structure control of nonlinear 4. R.A. DeCarlo, S.H. Zak multivariable systems: a tutorial. Proc IEEE 26: 1139–1144 5. C.M. Dorling and A.S.I. Zinober (1986) Two Approaches to Hyperplane Design in Multivariable Variable Structure Control Systems. Int J Control44: 65–82 6. C.M. Dorling and A.S.I. Zinober (1988) Robust hyperplane design in multivariable variable structure control systems. Int J Control48: 2043–2054 7. B. Dra˘zenovi´c (1969) The invariance conditions in variable structure systems. Automatica 5: 287–295 8. C. Edwards C and S.K. Spurgeon (1998) Sliding Control. Taylor and Francis, Bristol 9. L. Fridman and A. Levant (1996) Higher order sliding modes as the natural phenomena of control theory. in: Robust Control Variable Structure and Lyapunov Techniques, eds Garafalo F, Glielmo G, Lecture Notes in Control and Information Science, 217, Springer Verlag, London 107–133 10. L. Fridman and A. Levant (2002) Higher order sliding modes in Sliding Mode Control in Engineering, eds Barbot JP, Perruguetti W, Marcel Dekker, New York 3–101 11. K. Furuta and S.B. Kim (1987) Pole Assignment in a Specified Disc. IEEE Trans Aut ContrAC-32: 423–427 12. S. Gopalswamy and J.K. Hedrick (1993) Tracking Nonlinear Nonminimum Phase Systems Using Sliding Control. International Journal of Control 57: 1141–4458 13. A. Isidori (1995) Nonlinear control systems. Springer-Verlag, London 14. Y.-T. Juang Z.-C. Hong and Y.-T. Wang (1989) Robustness of Pole Assignment in a Specified Region. IEEE Trans Aut ContrAC-34: 758–760 15. J. Kautsky, N.K. Nichols and P. Van Dooren (1985) Robust Pole Assignment in Linear State Feedback. Int J Control41: 1129–1155 16. G. Klein and B.C. Moore (1977) Eigenvalue generalised eigenvector assignment with state feedback. IEEE Trans Aut ContrAC–22: 140–141

Sliding-Mode Control in Systems with Output Time Delay

263

17. A. Koshkouei, K.J. Burnham KJ and A.S.I. Zinober A (2004) Dynamic Sliding Mode Control Design. Proceeding of IEE Conference, Bath 18. I.D. Landau (1979) Adaptive control: The model reference approach, M. Dekker, New York 19. A. Levant (2003) Higher-order sliding modes, differentiation and output-feedback control. Int J Control 76: 924-=941, Special Issue on Sliding-Mode Control 20. A. Levant (2004) Finite Differences in Homogeneous Discontinuous Control.Proc. of the 44th IEEE Conference on Decision and Control, Atlantis, Paradise Island, Bahamas 14–17 21. A. Levant (2005) Homogeneity Approach to High-order Sliding Mode Design. Automatica41: 823–830 22. G. Liu, A.S.I. Zinober and Y.B. Shtessel (2006) Output Delay Systems Tracking Using System Centre Approach and Sliding Mode Control. Proc CTS-HYCON Workshop, Paris, July 2006 23. E.P. Ryan and M. Corless (1984) Ultimate boundedness and asymptotic stability of a class of uncertain dynamical systems via continuous and discontinuous feedback control. IMA J Math Control Information 1: 223–242 24. M.G. Safonov and R.Y. Chiang RY (1988) CACSD using the state-space L∞ theory – a design example. IEEE Trans Aut ContrAC-33: 477–479 25. M.G. Safonov, A.J. Laub and C.L. Hartmann (1981) Feedback properties of multivariable systems: the role and use of the return difference matrix. IEEE Trans Aut ContrAC-26: 47–65 26. S.L. Shah, D.G. Fisher and D.E. Seborg (1975) Eigenvalue/eigenvector assignment for multivariable systems and further results output feedback control. Electron Letters 11: 388–389 27. I. Shkolnikov and Y.B. Shtessel (2002) Tracking in a Class of Nonmimumum Phase Systems with Nonlinear Internal Dynamics via Sliding Mode Control using Method of System Centre. Automatica: 38: 837–842 28. Y.B. Shtessel, A.S.I. Zinober and I. Shkolnikov (2003) Sliding Mode Control for Nonlinear Systems with Output delay via Method of Stable System Centre. Journal of Dynamic Systems, Measurement, and Control 125: 29. Y.B. Shtessel, A.S.I. Zinober and I. Shkolnikov (2003) Sliding Mode Control of Boost and Buck-boost Power Converters using Method of Stable System Centre. Automatica 39: 1061–1067 30. Y.B. Shtessel and I. Shkolnikov (2003) Aeronautical and Space Vehicle Control in Dynamic Sliding Manifolds. International Journal of Control 25: 253–257 31. V. Sinswat and F. Fallside (1977) Eigenvalue/eigenvector assignment by state-feedback. Int J Control26: 389–403 32. S.K. Spurgeon, M.K. Yew, A.S.I. Zinober and R.J. Patton (1990) Model-following control of time-varying and nonlinear avionics systems, in Deterministic control of uncertain systems, ed. Zinober ASI, Peter Peregrinus Press, 96–114 33. V.I. Utkin (1977) Variable structure systems with sliding mode. IEEE Trans Aut ContrAC-22: 212–222 34. V.I. Utkin (1978) Sliding Modes and Their Application in Variable Structure Systems, MIR, Moscow 35. V.I. Utkin (1992) Sliding Modes in Control Optimization, Springer-Verlag, Berlin 36. V.I. Utkin and K.D. Yang (1978) Methods for Constructing Discontinuity Planes in Multidimensional Variable Structure Systems. Autom. Remote Control 39: 1466–1470

264

A.S.I. Zinober, G. Liu, and Y. B. Shtessel

37. V. Utkin, J. Guldner, Shi J (1999) Sliding Modes in Electromechanical Systems, Taylor and Francis, London, 38. C.A. Woodham (1991) Eigenvalue Placement for Variable Structure Control Systems, PhD Thesis, University of Sheffield 39. C.A. Woodham, A.S.I. Zinober (1990) New Design Techniques for the Sliding Mode. Proc IEEE International Workshop on VSS and their Applications, Sarajevo, 220–231 40. C.A. Woodham, A.S.I. Zinober (1991) Eigenvalue Assignment for the Sliding Hyperplanes. Proc IEE Control Conference, Edinburgh, 982–988 41. C.A. Woodham, A.S.I. Zinober (1991) Robust Eigenvalue Assignment Techniques for the Sliding Mode. IFAC Symposium on Control System Design, Zurich, 529–533 42. C.A. Woodham, A.S.I. Zinober(1993) Eigenvalue placement in a specified sector for variable structure control systems. Int J Control57: 1021–1037 43. A.S.I. Zinober (1990), Deterministic control of uncertain systems ed, Peter Peregrinus Press, London 44. A.S.I. Zinober, O.M.E. El-Ghezawi and S.A. Billings (1982) Multivariable variablestructure adaptive model-following control systems. Proc IEE129D: 6–12

10 Control Engineering and Systems Biology Burton W. Andrews and Pablo A. Iglesias

Summary. Engineers use feedback, both positive and negative, to perform a wide array of signaling functions. Biological systems are also faced with many of the same requirements In this tutorial we examine examples from different cellular signaling systems to show how biology also uses feedback paths to perform many of the same tasks.

10.1 Introduction Though the components differ greatly, there are striking similarities between cellular organisms and engineered, man-made objects. In both classes of systems, most functions are performed by operational modules consisting of a wide range of interacting subsystems [1]. In biology, these modules consist of many classes of molecules. In fact, as argued in [1], much of twentieth-century biology attempted to reduce biological phenomena to the behavior of individual molecules. Much effort has been spent in determining the identity and specific function of molecules responsible for different phenomelogical processes. This is a necessary first step to reverse engineering biology, and is one that has produced great results [2]. However, just as the individual components of an electronic circuit do not uniquely identify the system’s function, neither do the molecules of a biological entity solely determine the cellular function. Understanding how these biological modules achieve their function will require both knowledge of the identities of the molecules as well as how these components are arranged. Recently, the study of biology at this system level has received considerable attention, giving rise to the field of “systems biology” [3, 4]. Interest in understanding biology in terms of a system-theoretic viewpoint is not new — Wiener’s cybenetics provides an early viewpoint [5] — yet the interest in the field from biologists is. Inherent in all systems biology approaches is the notion that biological function does not arise out of individual genes or proteins, but that it is the concerted interaction of different components that gives rise to the observed phenomena [3]. Thus, to understand a biological system, it is necessary not only to grasp the role played by each of the individual components, but also how these individual biochemical species interact to form the signaling network. M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 267-288, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

268

B.W. Andrews and P. A. Iglesias

To elucidate these networks, systems biology relies on several complementary approaches. One is the so-called high-throughput model of biological investigation. High-throughput techniques allow the automated perturbation of large numbers of components (eg. genes) thus probing the cellular response to different changes in a systematic way. A second arrow in the systems biology quiver is the use of computational methods which allow biologists to test conceptual models that involve more than a few variables. More importantly, they provide predictive power by generating hypotheses that can be tested experimentally. The new focus on biology at the systems level affords an immense opportunity for control and systems engineers. One common aspect of systems and control analysis is that it works with systems at an abstract mathematical level that does not rely on specific knowledge of the components. As such, subsystems can represent electronic, mechanical, chemical or biochemical components — or combinations thereof. In this paper we consider mathematical models of several well-known signaling pathways. We show how feedback connections are crucial in achieving the desired function. In doing so we highlight how control engineering tools may prove to be fruitful tools for reverse engineering biology.

10.2 Negative Feedback Negative feedback loops are used in engineering systems to provide stability, to regulate systems and provide a level of robustness. As we demonstrate in this section, these same requirements give rise to negative feedback loops throughout biology.

10.2.1 Negative Feedback: Regulation Many problems in control can be framed as regulation problems, where the goal is to design a controller that can drive the output of a system that has been perturbed by possibly nonmeasurable disturbances to a constant set point [6]. Output regulation is also crucial for the functioning of all biological systems, where it is known as homeostasis, from the simplest bacteria to humans. Not surprisingly, feedback control is essential in achieving homeostasis. For example, the production of tryptophan, an essential amino acid, is regulated in bacteria by a series of three feedback loops [7]. Mammals rely on negative feedback loops to maintain blood calcium and glucose levels within tight bounds for their survival [8, 9].

Feedback Inhibition In biology, the term feedback inhibition is used when the catalytic function of an enzyme acting early in a reaction pathway is inhibited by a downstream product of that pathway [10]. This is illustrated by the system in Figure 10.1. This functional cascade appears in many cellular systems. One of the better known is the MAPK (Mitogen-Activated Protein Kinase) class of pathways that are found in a diverse array of signaling sytems1 . In this “module,” the 1

A kinase is an enzyme that transfers a phosphate group to substrates. This phosphorylation process is one of the fundamental switches in biology. The enzyme that removes the phosphate group is known as a phosphatase.

Control Engineering and Systems Biology

269

Fig. 10.1. Feedback inhibition in a cascade of biochemical reactions. In this system, several proteins stimulate downstream signals in a linear fashion (S stimulates X, which stimulates Y , which stimulates Z.) Feedback inhibition is present when one of the downstream molecules inhibits an earlier process. In this case, Z reduces the production rate of X. Examples in biology include the large class of mitogen-activated protein kinase (MAPK) systems that are found throughout biology.

MAPK (pronounced MAP kinase, and corresponding to element Z in Figure 10.1) is activated by the MAPKK (MAP kinase kinase, element Y ) which is activated by the MAPKKK (MAP kinase kinase kinase, element X). Comprehensive analyses of MAPK systems can be found elsewhere [11, 12]. Here, we assume that the MAPK pathway can be described by the system of differential equations2 d[X] = −k−1 [X] + k1 [S], dt d[Y ] = −k−2 [Y ] + k2 [X], dt d[Z] = −k−3 [Z] + k3 [Y ]. dt

(10.1) (10.2) (10.3)

This is, of course, the cascade of three linear first order systems and its transfer function, from signal S to output Z is: k1 k2 k3 . (10.4) (s + k−1 )(s + k−2 )(s + k−3 ) If we assume that a negative feedback exists between Z and the signaling molecule S, we replace (10.1) by d[X] = −k−1 [X] + k1 [S] − k[Z], (10.5) dt and obtain the new closed-loop transfer function: k1 k2 k3 . (s + k−1 )(s + k−2 )(s + k−3 ) + kk1 k2 k3

(10.6)

What difference can this feedback make? Clearly, at steady-state the gain of the system k1 k2 k3 kk1 k2 k3 + k−1 k−2 k−3 is a decreasing function of k. However, note that for small t (corresponding to |s| ↑ ∞ in (10.4) and (10.6)), the gain of the system is not affected. In effect, negative feedback leaves the 2

In computational biology, it is customary that square brackets around the name of a biochemical species denote its concentration.

270

B. W. Andrews and P.A. Iglesias

no feedback

160

Response

120

k = 0.005 80

k = 0.01 k = 0.02

40

0 0

10

20

30

40

50

60

Time (sec)

Fig. 10.2. Step response for the cascade system in Figure 10.1 and described by the equations (10.1–10.5). Parameters used are k1 = k2 = k3 = 1, k−1 = 0.1, k−1 = 0.2, k−1 = 0.3 and variable feedback gain k.

initial response unchanged, but decreases the steady-state. This inhibition has the effect of qualitatively changing the response of the system; see Figure 10.2. Without feedback, the system response rises to a persistent signal; with feedback, the response of the system is more transient.

Growth Factor Signaling We have just demonstrated how negative feedback can give rise to a transient signal. It is interesting to consider a specific biological example where this observation is of importance, in particular, in resolving a perplexing observation in the response in a cell line that is used extensively for studying nerve growth factor3 (NGF) signaling [13]. In these cells, two different growth factors, nerve growth factor (NGF) and epidermal growth factor (EGF), stimulate the same MAPK cascade. However, whereas EGF stimulation induces a rapid transient activation of the MAPK cascade, NGF stimulus results in sustained activation. Though these two behaviors had been observed experimentally, the reason behind these contrasting responses was unknown. Using a computer simulation of growth factor-induced MAPK cascade activation, Brightman and Fell determined that it is differences in the feedback regulation of the signals initiated by EGF and NGF that determines the duration of cascade activation [13]. Their model, of course, is considerably more complicated than that described above, as the system includes many more elements than the system of Figure 10.1 and the model uses more realistic dynamics than those used here. Nevertheless, the basic systems analysis is the same: negative feedback can lead to transient responses and attenuation of signal responses at steady-state.

3

Growth factors are proteins that bind to transmembrane receptors and are used to signal cells to activate differentiation or proliferation.

Control Engineering and Systems Biology

L + R GGDPαβγ

271

RL GGTPα + Gβγ

downstream effectors

RGS GGDPα

Fig. 10.3. Reactions describing the GPCR response. Ligand (L) binds to the GPCR (R) leading to an occupied receptor (RL). This induces the dissociation of the G-protein into its Gα - and Gβ γ subunits. The latter stimulates the synthesis of the RGS proteins. RGS proteins increase the rate of GGTP hydrolysis. Once GGDP is formed, the subunits reunite into a trimer. α α

G-protein Signaling We consider another ubiquitous cellular signaling mechanism that relies on negative feedback to generate transient responses to external stimuli. G-protein4 coupled receptors (GPCRs) are found in the signaling networks of eukaryotic organisms from single-celled amoebae and yeast to humans, where it is used in the eye (rhodopsin) as well as in the nose. GPCRs form the most common type of receptors in mammalian cells — it estimated that about 3–5% of genes in the genome are used to encode such receptors. Incredibly, about half of all known drugs target GPCRs [10]. GPCRs consist of a transmembrane component that traverses the cell membrane seven times. The intracellular component is coupled to G-proteins consisting of three subunits: Gα , Gβ and Gγ . The Gα subunit is bound to either GTP or GDP. In its inactive state, GGDP αβ γ , all three subunits are coupled, and Gα is bound to GDP. Binding of a stimulus to the transmembrane receptor causes Gα to release GDP and bind to GTP, and the Gα and Gβ γ subunits to separate. The Gβ γ subunit then signals downstream processes. The dissociation of the G-protein subunits is reversible. Hydrolysis5 of GTP leads to the formation of GGDP which reassociates α with free Gβ γ . This process can be accelerated by proteins known as regulators of G-protein signaling (RGS). Hao and coworkers developed a computational model for the effect of negative feedback on GPCR signaling that highlights the effect of negative feedback [14]. In this model, described in Figure 10.3, the dynamics of Gβ γ can be described by the ordinary differential equation GDP d[Gβ γ ] GTP GTP = k1 + k1′ [RL] [Gαβ γ ] − k2 [Gα ] − k3 [RGS][Gα ] dt = k1 + k1′ [RL] [Gβ γ ]T − [Gβ γ ] − k2 [GαGTP ] − k3 [RGS][GGTP α ], 4

5

(10.7)

G-protein is short for guanine nucleotide binding proteins. These are signaling proteins that are activated through the exchange of guanosine diphosphate (GDP) for guanosine triphosphate (GTP). In this case, hydrolysis refers to the exchange of GTP with GDP.

272

B. W. Andrews and P. A. Iglesias

where we have assumed that GTP hydrolysis is the rate limiting step in reformation of GGDP αβ γ and the total concentration of Gβ γ is constant: GDP [Gβ γ ]T = [Gβ γ ] + [Gαβ γ ]. GDP in the absence of stimulus, We have also assumed that some Gβ γ is produced from Gαβ γ indicated by the rate constant k1 . We assume the total Gα concentration is constant and given by GTP GDP [Gα ]T := [GGDP αβ γ ] + [Gα ] + [Gα ].

Moreover, by the assumption that GTP hydrolysis is the rate limiting step in reformation of GDP rapidly combines with G , we can assume that [GGDP ] ≈ 0 so that GGDP βγ α αβ γ , so that Gα GDP GDP GDP [GαGTP ] = [Gα ]T − [Gαβ γ ] − [Gα ] ≈ [Gα ]T − [Gαβ γ ].

Thus, [GGTP α ] ≈ [Gα ]T − [Gβ γ ]T + [Gβ γ ].

Substituting this into (10.7) leads to:

d[Gβ γ ] = k1 + k1′ [RL] [Gβ γ ]T − [Gβ γ ] − (k2 + k3 [RGS]) [Gβ γ ], dt

where, in the last line, we assumed that [Gβ γ ]T ≈ [Gα ]T . We rewrite this last equation as: d[Gβ γ ] = k1 + k1′ [RL] [Gβ γ ]T − k7 + k1′ [RL] + k3 [RGS] [Gβ γ ], dt

(10.8)

where k7 := k1 + k2 . We now focus on the feedback path. Assume that

d[RGS] = k4 + k5 [Gβ γ ] − k6 [RGS], dt

(10.9)

where k4 is a constant rate of RGS production in the absence of stimulus, k5 is the stimulusinduced rate of RGS production, and k6 is the degradation rate. The complete system of equations can be rewritten as: d[Gβ γ ] = k1 + k1′ [RL] [Gβ γ ]T − k7 + k1′ [RL] + k3 [RGS] [Gβ γ ], dt d[RGS] = k4 + k5 [Gβ γ ] − k6 [RGS]. dt Making the substitutions x :=

[Gβ γ ] , [Gβ γ ]T

the system becomes

τ := k7 t,

u :=

k1′ [RL], k7

and

y :=

k7 [RGS], k5 [Gβ γ ]T

Control Engineering and Systems Biology

1.0

A

273

B

1.0

g(x,y,u)=0 0.8

0.6

0.6

f(x,y,u)=0, k=1

x

no feedback

0.8

f(x,y,u)=0, k=0

x

k=1 0.4

0.4

0.2

0.2

k = 10

f(x,y,u)=0, k=10 0

1.0

1.5

2.0

0

0

1

2

3

4

5

Time (sec)

y

Fig. 10.4. Nullclines (A) and time response (B) for the GPCR signaling system, described by (10.10) and (10.11). Parameters chosen are u = 1, α = 0.5, and β = γ = 1. dx = f (x, y, u) := (α + u) − (1 + u + ky) x, dτ dy = g(x, y, u) := β + x − γ y, dτ where

α :=

k1 , k7

β :=

k4 , k5 [Gβ γ ]T

γ :=

k6 k7

and

k :=

(10.10) (10.11)

k3 k5 [Gβ γ ]T k72

.

The equilibrium is determined by the intersection of the two nullclines 0 = f (x, ¯ y, ¯ u) ¯ and 0 = g(x, ¯ y, ¯ u); ¯ or α + u¯ x¯ = (10.12) = γ y¯ − β , 1 + u¯ + ky¯ as shown in Figure 10.4, along with the time response, for varying k. As the feedback strength increases, the equilibrium value decreases demonstrating the regulating role of RGS on the system response.

10.2.2 Negative Feedback: Sensitivity and Robustness A common use of negative feedback in engineered systems is to reduce the effect of parameter variation on the system response.

Gene regulation We consider a simple example, due to Vinnicombe [15], of feedback reduction of sensitivity to parameter variation during gene regulation. The system can be described by the reactions: c

α

0/ −→ mRNA −−→ 0/ where [mRNA] represents the concentration of a particular messenger RNA (mRNA) being translated by DNA at a constant average rate c. Degradation is assumed at constant rate α . The mean rate of change of [mRNA] can be described by the differential equation

274

B. W. Andrews and P.A. Iglesias

d[mRNA] = c − α [mRNA]. dt

(10.13)

At steady-state, the production and degradation rates are equal, and the average number of molecules of mRNA is [mRNA]0 = c/α . Small changes, δα , to the parameter α lead to a change δ[mRNA] = [mRNA] − [mRNA]0 in the number of mRNA molecules. Linearizing (10.13) about α and [mRNA]0 yields dδ[mRNA] dt

= −αδ[mRNA] − [mRNA]0 δα ,

which has a steady-state value of

δ[mRNA] δα ≈− . [mRNA]0 α Thus, a small percentage change in the degradation parameter α leads to approximately the same percentage change in the steady-state number of mRNA molecules. If a consistent response is needed by the cell, this effect is undesirable because variations in parameters within an individual cell or across a population lead to response variations of the same order. A solution to the above sensitivity problem is to decrease these fluctuations through negative feedback. If we assume that the mRNA codes for a DNA-binding protein that promotes further transcription of the mRNA, then the system becomes f ([mRNA])

mRNA α

∅

(10.14)

where the transcription rate f ([mRNA]) is a function of the number of mRNA molecules. In this case, small perturbations in α yield dδ[mRNA] dt

= ( f ′ ([mRNA]0 ) − α )δ[mRNA] − [mRNA]0 δα ,

where the linearization is about [mRNA]0 and α . With feedback, the relative steady-state number of mRNA molecules is δ[mRNA] 1 δα ≈− . [mRNA]0 1 − f ′ ([mRNA]0 )/α α If f ′ ([mRNA]0 ) < 0, then decreasing the slope (more negative) reduces the sensitivity of the system output to changes in the parameter α . While perhaps overly simplistic, this example provides a useful insight into sensitivity reduction in biological networks through negative feedback. Other, more realistic examples are described in [16–19].

Bacterial Chemotaxis Robust systems maintain consistent, stable behavior in the face of uncertainties; both internal (eg. component variability) and external (eg. disturbances). Robustness has been found to exist in many biological systems [20–24].

Control Engineering and Systems Biology

I

k1

D+L

R+L k-1

k-r

E

kr

k-d

I

RL

275

k2 k-2

kd

DL E

Fig. 10.5. General four-state receptor model of receptor-mediated adaptation in E. coli. The receptors can be found bound (RL and DL) or not (R and D). Methylation by CheR, labeled I, inactivates the receptors. Similarly, demethylation, by CheB, labeled E, activates the system.

To illustrate the connection between biological robustness and control engineering, we consider the chemotactic response of Escherichia coli. E. coli is a rod-shaped bacteria, aproximately one micron in diameter and two microns long, that is propelled by rotation of 6–10 flagella. These bacteria move towards high concentrations of chemical attractants (a process known as chemotaxis) by monitoring and responding to temporal changes in attractant concentration. The signaling network governing this response is one of the best-understood signaling networks in all of biology. The response to a step change in chemoattractant consists of an initial transient followed by a return to the prestimulus level, a property known to biologists as adaptation. Adaptation allows cells to detect changes in the level of attractant and hence assess the values of the direction the cell is moving. This adaptation has been shown to be robust, in the sense that the bacteria will return to their prestimulus state even when internal conditions are varied greatly [20, 25]. To illustrate the control-theoretic aspects of robust adaptation, we consider a general framework for adaptation that is representative of the signaling network found in E. coli [26]. Chemotactic receptors can exist in one of two configurations, which we term either modified or unmodified. Biologically, these correspond to whether receptors are methylated or not6 . Methylation and demethylation are mediated by two enzymes: CheR and CheB, respectively. Because methylation “inactivates” the receptor, we will refer to CheR and CheB as the inactivator (I) and excitation (E) processes, respectively. Additionally, these receptors can be occupied by the chemoattractant molecules (so-called, ligand-bound) or not. Thus, as illustrated in Figure 10.5, receptors can be found in one of four states: R, RL, D and DL. To develop a numerical model of this system, we makes some assumptions about how the enzymes act on their substrates. An enzymatic reaction takes a substrate, S, and an enzyme, E, and creates a product, P, without consuming the enzyme. A classical model of enzyme catalysis, due to Michaelis–Menten, states that the product of an enzymatic reaction appears at the rate: [S][E] d[P] ∝ . dt kM + [S] 6

Methylation refers to the replacement of a hydrogen atom (H) with a methyl group (CH3 ).

276

B. W. Andrews and P. A. Iglesias

The constant, kM , is known as the Michaelis–Menten constant [27]. Note that there are two distinct regimes in this equation. If the amount of substrate is small relative to kM then: d[P] ∝ [S][E]. dt This is known as the linear regime, as the rate of production is linear in both substrate and enzyme concentrations. If, however, the amount of substrate is large relative to kM , then d[P] ∝ [E]. dt This is known as the saturated regime, in that extra substrate will not increase the rate of production. In our model of receptor-mediated adapation in bacterial chemotaxis, we make contrasting assumptions on the enzymatic production of CheB and CheR: that excitation (CheB) acts at saturation but that inhibition (CheR) acts in the linear regime. We make two more assumptions. We assume that fractions of the unmodified receptors, α1 R and α2 RL are active, and that the total activity of the system is given by the sum of these two fractions. Moreover, the inactivator CheB acts only on this fraction of receptors. Finally, we assume that the rate at which CheB acts on modified and unmodified receptors is equal: k−1 = k−2 in Figure 10.5. Based on these assumptions, we can describe the system dynamics with the following set of differential equations: d[R] = k−1 [E] − k1 [I]α1 [R] − kr [R][L] + k−r [RL], dt d[RL] = k−2 [E] − k1 [I]α2 [RL] + kr [R][L] − k−r [RL], dt d[D] = −k−1 [E] + k1 [I]α1 [R] − kd [D][L] + k−d [DL], dt d[DL] = −k−2 [E] + k2 [I]α2 [RL] + kd [D][L] − k−d [DL]. dt Note that the first two equations are decoupled from the second two. Making the variable transformation: [A] := α1 [R] + α2 [RL], which corresponds to the system’s activity, and [B] := leads to:

where

d dt

[R] + [RL] , k1 [I]

[A] −a1 ([L], [I]) a0 ([L], [I]) [A] b([I]) = + [A]0 , [B] −1 0 [B] 1 a0 ([L], [I]) := (α1 k−r + α2 kr [L] + α1 α2 k1 [I])k1 [I], a1 ([L], [I]) := kr [L] + k−r + (α1 + α2 )k1 [I], b1 ([I]) :=

(α1 k−1 + α2 k−2 )k1 [I] , k−1 + k−2

[A]0 = [E](k−1 + k−2 )/k1 [I].

(10.15)

Control Engineering and Systems Biology

x

277

y

Fig. 10.6. A. Autocrine loop of the epidermal growth factor receptor (EGFR) system. Epidermal growth factor (EGF) binding to EGFRs trigger the MAPK response. This, in turn, leads to the secretion of EGF completing a positive feedback loop. B. Linear system incorporating a positive feedback loop.

The system has one stable steady-state at [A] = [A]0 , which is independent of [L]. Thus, adaptation is a robust feature of the signaling network, and the steady-state response is robust to changes in ligand. It is worth highlighting how the system achieves adapation. If we consider the second row of (10.15): Z d[B] = −([A] − [A]0 ) ⇐⇒ [B](t) = − e(t)dt, dt | {z } e(t), error

we see that an integral feedback loop is being implemented where the activity [A](t) is being compared to the “set-point” [A]0 . The error [A] − [A]0 is then integrated. This is an example of the internal model principle: step changes in the external “disturbance” — which in this case is the chemoattractant concentration — can be rejected in a robust manner because the system has an integrator inside the loop [23]. The experimental verification of this robustness has been confirmed [20].

10.3 Positive Feedback 10.3.1 Positive Feedback: Amplification In engineering systems, the use of positive feedback for signal amplification dates back to the early 1900’s when Edwin Armstrong discovered the gain benefits of positive feedback for a triode amplifier [28]. Amplification through positive feedback is also common in biological systems. We consider one example: the epidermal growth factor receptor (EGFR) system where positive feedback from an autocrine loop — a loop where a cell secretes some of the very signal that stimulates its signaling network — leads to an amplified MAPK response [29]. Here we present an illustrative model that provides insight into the means by which positive feedback can induce gain. A detailed model of the EGFR system can be found in [29].

278

B. W. Andrews and P. A. Iglesias

Response

0.4

0.3

Increasing k 0.2

0.1

0

0

5

10

15

20

25

Time (sec)

Fig. 10.7. Response of the linear system (10.16) for varying values of the positive feedback gain k to a one-second pulse of unit amplitude. Parameter values used are α = 1 and k = 0.2, 0.4, 0.6 and 0.8.

Consider the system shown in Figure 10.6A. A primary stimulus (eg., an exogenous growth factor or some intracellular stimulant) and membrane receptor-bound ligands (eg. transforming growth factor [TGF]) serve as inputs to an intracellular signaling pathway (the extracellular signal-regulated MAPK cascade in the EGFR system). The output of the signaling pathway activates ligand-releasing proteases7 . The released ligand can then rebind to surface receptors creating a positive feedback loop. To illustrate the effect of such a feedback, we describe the signaling pathway by the transfer function 1 H(s) = , (s + α )2 which is a reasonable linear model of a two-level MAPK cascade [30]. The autocrine loop is modeled as a simple positive feedback gain k as in Figure 10.6B. The transfer function from stimulus x to output y is Y (s) 1 . (10.16) = X(s) (s + α )2 − k Figure 10.7 shows the response of this system to a pulse of duration one second, a perturbation similar to that studied in [29], for increasing values of k. In the EGFR system, increases in the effect of positive feedback may arise from factors such as an increase in protease activity (increased ligand release rate) or an increase in the number of receptors [29]. Figure 10.7 shows that positive feedback yields amplification and an increased duration of the response of the system. This effect is representative of observations both experimentally and in theoretical models of the EGFR system [29, 31] and demonstrates the use of positive feedback as an amplifier in biological systems.

10.3.2 Positive Feedback: Switching and Memory Positive feedback is often used in engineering applications to induce switching behavior, usually characterized by an all-or-nothing response to an input. The response remains relatively constant for some range of input values but changes abruptly once the input reaches some threshold value; see Figure 10.8. Monostable switches have a single stable (input-dependent) 7

A protease is an enzyme that breaks bonds between the amino acids of proteins.

Control Engineering and Systems Biology

B

Output

Output

A

279

stable “high” equilibrium

unstable stable “low” equilibrium

Input

Input

Fig. 10.8. Monostable (A) and bistable (B) switches

steady-state response at all input values (Figure 10.8A) while bistable switches exhibit hysteresis (Figure 10.8B) and are often said to have memory due to the delayed response to input transitions. An example is the Schmitt trigger, which uses an operational amplifier and positive feedback to implement a bistable circuit [32]. cooperative feedback

S

A

A*

Fig. 10.9. Model representative of the positive feedback found in the oocyte maturation signaling network

Many cells rely on switch-like behavior for response regulation [33–38], enabling the discretization of responses to graded inputs and, in some cases, making the response irreversible. For example, positive feedback loops have been shown to yield irreversible maturation8 of Xenopus oocytes (frog egg cells) when stimulated with sufficient amounts of progesterone. In particular, a MAPK, known as p42, and the protein kinase Cdc2, both of which are activated during oocyte maturation, are further activated through positive feedback. Active p42 MAPK increases the activity of both p42 MAPK and Cdc2 via intermediate proteins, and active Cdc2 increases the activity of both kinases as well [35]. A simplified model of this system, taken from [35], is shown in Figure 10.9. A signaling protein A is activated both by an external stimulus S (progesterone in the Xenopus oocyte system) and by positive feedback from the active form A∗ of A. If the feedback is cooperative9 with Hill coefficient n, the system response can be described by the differential equation 8

9

Maturation refers to the process in which egg cells reduce the number of chromosomes in half. Thus leaving them with a single copy of each chromosome. This enables them to be fertilized. An enzymatic reaction is said to be cooperative if the rate of production is proportional to n + [S]n ). The exponent n is known as the Hill coefficient. the term [S]n [E]/(kM

280

B. W. Andrews and P.A. Iglesias

0.02

A

s3

B

s3

Synthesis rate Degradation rate 0.01

s2

s2

s1

s1

0 0.02

C

s3

D

s3

0.01

s2

s2

s1

s1

0 0

0.5

1

0

Active response (A*)

0.5

1

Active response (A*)

Fig. 10.10. The strength of the positive feedback loop gain can change the number of equilibria. These panels plot the degradation (dotted line) and synthesis rates (solid lines) for (10.17) as a function of the active response [A∗ ]. Equilibria for (10.17) are characterized by points of intersection. The synthesis rate is plotted for three different constant signaling strengths: s1 = 0.001, s2 = 0.005 and s3 = 0.02. Each panel represents a different feedback strength: k1 = 0, 0.05, 0.15 and 0.25 in panels A–D respectively. For no or low feedback strengths, each of the external signal values leads to a single equilibrium (panels A and B). For higher values of the feedback gain, multiple equilibria can exist — note that for s2 in panel C, three equilibria appear. The two black circles correspond to stable equilibria, the open circle is an unstable equilibrium. Smaller input signals lead to a low response; higher input signals to a high output. Finally, for even greater feedback gains (panel D), these multiple equilibria exist only for small input values. The location of the equilibria can be plotted as a function of the signaling strength s as in Figure 10.11. [A∗ ]n d[A∗ ] = s([Atot ] − [A∗ ]) + k1 n ([Atot ] − [A∗ ]) − k−1 [A∗ ] , dt K + [A∗ ]n | {z } | {z } degradation synthesis

(10.17)

where k−1 is the degradation rate of A∗ , K is the Michaelis–Menten constant of the feedback reaction, k1 is the strength of the feedback and [Atot ] := [A] + [A∗ ]. As shown in Figure 10.10, as the strength of the feedback gain is increased, multiple equilibria can arise. These are shown in Figure 10.11 as a function of stimulus concentration for varying feedback gains. As the feedback strength increases, the system progresses from a monostable (panels a–c) to a bistable, but reversible switch (panels d–e). Systems in this regime will show hysteretic responses to the external stimulus s. For greater feedback strengths (panel f), the

Control Engineering and Systems Biology

281

1

Response (A*)

0.5

A

k1=0.04

B

k1=0.07

C

k1=0.11

D

k1=0.15

E

k1=0.18

F

0.02

0.04

k1=0

0 1

0.5

0 0

0

0.02

0.04

0

0.02

0.04

Stimulus (s) Fig. 10.11. Bistability and irreversibility of the system of Figure 10.9, representative of the behavior of Xenopus oocyte maturation. Parameters used are [Atot ] = 1, K = 1, k−1 = 0.01, and n = 5 with s ranging from 0 to 0.04.

switch becomes irreversible. Once the input signal increases beyond the low-to-high transition point, the system is in the high state. Because the high-to-low transition is not admissible (it is negative, which is not biologically plausible) the system can not return to the low state. Thus, this is appropriate for cell fate decisions that cannot be reversed, such as differentiation and maturation.

10.3.3 Positive Feedback: Oscillations One of the uses of positive feedback in engineering is the generation of oscillatory behavior. Originally discovered around 1915 [28], positive feedback-induced oscillators are found in numerous engineering applications. Many biological systems also rely on oscillatory behavior for various cellular functions [39]. A system which has drawn considerable interest is the oscillatory signaling network controlling cAMP in the amoebae Dictyostelium discoideum [40–46]. When nutrients are abundant, Dictyostelium are single-celled amoebae that live in soil and feed on bacteria. However, when the food supply is exhausted, they begin a developmental process that lasts approximately 24 hours. In this time they develop a mechanism that enables them to secrete cAMP10 periodically. These oscillatory signals synchronize and generate propagating cAMP waves that can be observed under a microscope [47]. Cells react to these waves by chemotaxing towards the source of cAMP. There, up to 100 000 cells aggregate and form a multicellular organism. This aggregate forms a slug with the ability to migrate towards light (phototaxis). Eventually, a fruiting body consisting of a stalk and spore are formed. The approximately 5 000 cells in the 10

Cyclic Adenosine MonoPhosphate; pronounced “cyclic AMP,” is a common second messenger in biology.

282

B. W. Andrews and P.A. Iglesias

Fig. 10.12. Oscillation-inducing autocrine loop found in Dictyostelium

spore are inactive but remain viable and can revert to their single-cell existence in the presence of new nutrients [48]. The circuit underlying oscillatory cAMP signaling in Dictyostelium is based on competing positive and negative feedback loops, as shown in Figure 10.12, which is adapted from [46,49]. Cells sense cAMP from the external using GPCRs. This sensed cAMP is used by cells to chemotax toward the sensed signal. Though the exact chemotaxis pathway is different, relying on a spatial, rather than temporal signal, the cell uses an adaptation mechanism similar to that in the bacterial chemotaxis pathway to process the sensed signal [50–53]. Signals downstream of this receptor are used to stimulate synthesis of intracellular cAMP which is then secreted by the cell, thereby completing an autocrine loop. We now show how this interplay of positive and negative feedback in the chemotactic signaling network of Dictyostelium leads to these oscillations [46]. As with the bacterial chemotaxis pathway, the signaling network of Dictyostelium can be modeled as a response regulator R that is excited and inhibited by processes E and I, respectively [51, 54, 55]. The dynamics of the active form of R are described by the differential equation d[R∗ ] = −k−r [I][R∗ ] + kr [E][R] dt = −(k−r [I] + kr [E])[R∗ ] + kr [E][R]T ,

(10.18)

where [R]T := [R] + [R∗ ]. The excitation and inhibition processes are each regulated by the external stimulus S (cAMP in the case of Dictyostelium): d[E] = −k−e [E] + ke [S], dt d[I] = −k−i [I] + ki [S]. dt

(10.19) (10.20)

As shown in [46], the system described by (10.18), (10.19) and (10.20) forms an integral control mechanism with negative feedback that yields a steady-state [R∗ ] that is independent of [S]. The concentration of cAMP within the cell, [C], is governed by

Control Engineering and Systems Biology

283

4

3

y g(x,y)=0 2

limit cycle 1

f(x,y)=0 0

0

1

2

x

3

4

Fig. 10.13. Phase-plane analysis for the oscillator of (10.25) and (10.26) following [46]. The parameters a = 1/2, b = 1/6 and ε = 0.01 were chosen so as to make the equilibrium point at x = y = 4/9 unstable. The nullcline for f (x, y) increases sharply as x ↓ 0 allowing for the application of the Poincar´e–Bendixson theorem. d[C] = −k1 [C] − k2 [C] + k3 [R∗ ]2 + k4 , dt

(10.21)

where the first term represents degradation, the second secretion, the third stimulus-induced synthesis, and the last constitutive synthesis. The quadratic dependence on [R∗ ] accounts for amplification in the signaling pathway [52, 56–58]. The cAMP concentration outside the cell, S (also the stimulus in (10.19) and (10.20)) is governed by d[S] = −k5 [S] + k6 [C]. dt

(10.22)

For simplicity, we assume fast dynamics of [R∗ ], [E] and [C] and replace these variables with their corresponding steady-states values to obtain d[I] = −a1 [I] + a2 [S], dt d[S] a4 [S]2 + a6 , = −a3 [S] + dt (a5 [S] + [I])2 where a1 := k−i , a2 := ki , a3 := k5 , a4 :=

k3 k6 ke2 kr2 [R]2T , 2 k2 ) (k1 + k2 )(k−e −r

a5 := kr ke /(k−r k−e ),

and a6 := k4 k6 /(k1 + k2 ). The change of variables x := and τ := a3 t, yields:

a22 a3 [S], a21 a4

y :=

a2 a3 [I], a1 a4

(10.23) (10.24)

284

B. W. Andrews and P. A. Iglesias

dx x2 = f (x, y) := −x + + ε, dt (ax + y)2 dy = g(x, y) := −b(y − x), dt

(10.25) (10.26)

where a = a1 a5 /a2 , b = a1 /a3 , and ε = a22 a6 /(a21 a4 ). It can be shown that the steady-state x = y ≈ 1/(1+a)2 is unstable iff b ≤ (1−a)/(1+a), and by the Poincar´e–Bendixson theorem, a stable limit cycle exists under this condition [46].

10.4 Discussion Feedback is used throughout engineering to accomplish a wide range of objectives. In this paper we have tried to demonstrate how biology, faced by many of the same functional requirements, has also relied on feedback regulation. Our goal has been to highlight how understanding how engineering systems are synthesized can help in understanding biological systems. We believe that this is one way in which control engineering can make a significant contribution to systems biology. As expected, we are starting to see this effect [52, 59–63]. Control engineers will also have an impact through the development of new theoretical and computational tools to study models. For example, by applying computational algorithms from robust control theory, it has been possible to evaluate and select between proposed models [24, 64–68]. Finally, it is our expectation that by having control engineers actively working in biology, the field of control systems will be enriched. In particular, these new systems will give rise to interesting control theory, thus closing the intellectual loop between biology and control systems. Again, we are beginning to see this effect [69, 70].

References 1. L. H. Hartwell, J. J. Hopfield, S. Leibler, and A. W. Murray, “From molecular to modular cell biology,” Nature, vol. 402, pp. 47–52, 1999. 2. T. D. Pollard, “The cytoskeleton, cellular motility and the reductionist agenda,” Nature, vol. 422, pp. 741–745, 2003. 3. T. Ideker, T. Galitski, and L. Hood, “A new approach to decoding life: Systems biology,” Annu. Rev. Genomics Hum. Genet., vol. 2, pp. 343–372, 2001. 4. H. Kitano, “Systems biology: A brief overview,” Science, vol. 295, pp. 1662–1664, 2002. 5. N. Wiener, Cybernetics or Control and Communication in the Animal and the Machine, 2nd ed. New York: MIT Press, 1961. 6. W. M. Wonham, Linear Multivariable Control: A Geometric Approach, 3rd ed. New York: Springer, 1985. 7. K. V. Venkatesh, S. Bhartiya, and A. Ruhela, “Mulitple feedback loops are key to a robust dynamic performance of tryptophan regulation in Escherichia coli,” FEBS Letters, vol. 563, pp. 234–240, 2004. 8. H. El-Samad, J. P. Goff, and M. Khammash, “Calcium homeostasis and parturient hypocalcemia: An integral feedback perspective,” J. Theor. Biol., vol. 214, pp. 17–29, 2002.

Control Engineering and Systems Biology

285

9. P. T. Saunders, J. H. Koeslag, and J. A. Wessels, “Integral rein control in physiology,” J. Theor. Biol., vol. 194, pp. 163–173, Sep 1998. 10. B. Alberts, A. Johnson, J. Lewis, M. Raff, K. Roberts, and P. Walter, Molecular Biology of the Cell, 4th ed. New York: Garland Science, 2002. 11. A. Levchenko, J. Bruck, and P. W. Sternberg, “Scaffold proteins may biphasically affect the levels of mitogen-activated protein kinase signaling and reduce its threshold properties,” Proc. Natl. Acad. Sci. USA, vol. 97, pp. 5818–5823, 2000. 12. C. Huang and J. E. Ferrell Jr, “Ultrasensitivity in the mitogen-activated protein kinase cascade,” Proc. Natl. Acad. Sci. USA, vol. 93, pp. 10 078–10 083, 1996. 13. F. A. Brightman and D. A. Fell, “Differential feedback regulation of the MAPK cascade underlies the quantitative differences in EGF and NGF signalling in PC12 cells,” FEBS Lett, vol. 482, pp. 169–174, 2000. 14. N. Hao, N. Yildirim, Y. Wang, T. C. Elston, and H. G. Dohlman, “Regulators of G protein signaling and transient activation of signaling: Experimental and computational analysis reveals negative and positive feedback controls on G protein activity,” J. Biol. Chem., vol. 278, pp. 46 506–46 515, 2003. 15. G. Vinnicombe, “Feedback networks,” in Contol of Uncertain Systems: Modelling, Approximation, and Design, ser. LNCIS, B. A. Francis, M. C. Smith, and J. C. Willems, Eds. Heidelberg: Springer, 2006, pp. 371–387. 16. M. Thattai and A. van Oudenaarden, “Intrinsic noise in gene regulatory networks,” Proc. Natl. Acad. Sci. USA, vol. 98, pp. 8614–8619, 2001. 17. J. Paulsson, “Summing up the noise in gene networks,” Nature, vol. 427, pp. 415–418, 2004. 18. Y. Morishita, T. J. Kobayashi, and K. Aihara, “Evaluation of the performance of mechanisms for noise attenuation in a single-gene expression,” J. Theor. Biol., vol. 235, pp. 241–264, 2005. 19. H. El-Samad and M. Khammash, “Regulated degradation is a mechanism for suppressing stochastic fluctuations in gene regulatory networks,” Biophys. J, vol. 90, pp. 3749–3761, 2006. 20. U. Alon, M. G. Surette, N. Barkai, and S. Leibler, “Robustness in bacterial chemotaxis,” Nature, vol. 397, pp. 168–171, 1999. 21. G. von Dassow, E. Meir, E. M. Munro, and G. M. Odell, “The segment polarity network is a robust developmental module,” Nature, vol. 406, pp. 188–192, 2000. 22. A. Eldar, R. Dorfman, D. Weiss, H. Ashe, B.-Z. Shilo, and N. Barkai, “Robustness of the BMP morphogen gradient in Drosophila embryonic patterning,” Nature, vol. 419, pp. 304–308, 2002. 23. T. M. Yi, Y. Huang, M. I. Simon, and J. Doyle, “Robust perfect adaptation in bacterial chemotaxis through integral feedback control,” Proc. Natl. Acad. Sci. USA, vol. 97, pp. 4649–4653, 2000. 24. L. Yang and P. A. Iglesias, “Positive feedback may cause the biphasic response observed in the chemoattractant-induced response of Dictyostelium cells,” Systems & Control Letters, vol. 55, pp. 329–337, 2006. 25. N. Barkai and S. Leibler, “Robustness in simple biochemical networks,” Nature, vol. 387, pp. 913–917, 1997. 26. P. A. Iglesias and A. Levchenko, “A general framework for achieving integral control in chemotactic biological signaling mechanisms,” in Proc. Conf. Decision and Control, Orlando, FL, 2001, pp. 843–848. 27. J. P. Keener and J. Sneyd, Mathematical Physiology. New York: Springer, 1998. 28. S. Bennett, A History of Control Engineering: 1800–1930. Stevenage, UK: Peter Peregrinus, 1979.

286

B. W. Andrews and P.A. Iglesias

29. S. Y. Shvartsman, M. P. Hagan, A. Yacoub, P. Dent, H. S. Wiley, and D. A. Lauffenburger, “Autocrine loops with positive feedback enable context-dependent cell signaling,” Amer. J. Phys. Cell Phys., vol. 282, pp. 545–559, 2002. 30. M. Chaves, E. Sontag, and R. J. Dinerstein, “Optimal length and signal amplification in weakly activated signal transduction cascades,” J. Phys. Chem. B, vol. 108, pp. 15 311– 15 320, 2004. 31. K. Mayawala, C. A. Gelmi, and J. S. Edwards, “MAPK cascade possesses decoupled controllability of signal amplification and duration,” Biophys. J., vol. 87, pp. L01–2, 2004. 32. A. Sedra and K. Smith, Microelectronic Circuits, 5th ed. New York: Oxford University Press, 2004. 33. D. Angeli, J. E. Ferrell Jr., and E. D. Sontag, “Detection of multistability, bifurcations, and hysteresis in a large class of biological positive-feedback systems,” Proc. Natl. Acad. Sci. USA, vol. 101, pp. 1822–1827, 2004. 34. J. R. Pomerening, E. D. Sontag, and J. E. Ferrell Jr., “Building a cell cycle oscillator: Hysteresis and bistability in the activation of Cdc2,” Nature Cell Biology, vol. 5, pp. 346– 351, 2003. 35. W. Xiong and J. E. Ferrell Jr., “A positive-feedback-based bistable ‘memory module’ that governs a cell fate decision,” Nature, vol. 426, pp. 460–465, 2003. 36. H. Meinhardt, “Orientation of chemotactic cells and growth cones: Models and mechanisms,” J. Cell Science, vol. 112, pp. 2867–2874, 1999. 37. A. Narang, K. K. Subramanian, and D. A. Lauffenburger, “A mathematical model for chemoattractant gradient sensing based on receptor-regulated membrane phospholipid signaling dynamics,” Ann. Biomedical Eng, vol. 29, pp. 677–691, 2001. 38. M. Postma and P. J. Van Haastert, “A diffusion-translocation model for gradient sensing by chemotactic cells,” Biophys. J., vol. 81, pp. 1314–1323, 2001. 39. A. Goldbeter, “Computational approaches to cellular rhythms,” Nature, vol. 420, pp. 238– 245, 2002. 40. J. L. Martiel and A. Goldbeter, “Autonomous chaotic behaviour of the slime mould Dictyostelium discoideum predicted by a model for cyclic AMP signalling,” Nature, vol. 313, pp. 590–592, 1985. 41. A. Goldbeter, “Mechanism for oscillatory synthesis of cyclic AMP in Dictyostelium discoideum,” Nature, vol. 253, pp. 540–542, 1975. 42. T. Hofer, J. A. Sherratt, and P. K. Maini, “Dictyostelium discoideum: Cellular selforganization in an excitable biological medium,” Proc. Biol. Sci., vol. 259, pp. 249–257, 1995. 43. H. Levine, I. Aranson, L. Tsimring, and T. V. Truong, “Positive genetic feedback governs cAMP spiral wave formation in Dictyostelium,” Proc. Natl. Acad. Sci. USA, vol. 93, pp. 6382–6386, 1996. 44. J. Halloy, J. Lauzeral, and A. Goldbeter, “Modeling oscillations and waves of cAMP in Dictyostelium discoideum cells,” Biophys. Chem, vol. 72, pp. 9–19, 1998. 45. M. T. Laub and W. F. Loomis, “A molecular network that produces spontaneous oscillations in excitable cells of Dictyostelium,” Mol. Biol. Cell, vol. 9, pp. 3521–3532, 1998. 46. P. A. Iglesias, “Feedback control in intracellular signaling pathways: Regulating chemotaxis in Dictyostelium discoideum,” Eur. J. Control, vol. 9, pp. 216–225, 2003. 47. K. J. Tomchik and P. N. Devreotes, “Adenosine 3’,5’-monophosphate waves in Dictyostelium discoideum: A demonstration by isotope dilution-fluorography,” Science, pp. 443–446, 1981. 48. R. H. Kessin, Dictyostelium: Evolution, Cell Biology, and the Development of Multicellularity. Cambridge: Cambridge University Press, 2001.

Control Engineering and Systems Biology

287

49. B. P. Ingalls, T.-M. Yi, and P. A. Iglesias, “Using control theory to study biology,” in System Modeling in Cellular Biology, Z. Szallasi, J. Stelling, and V. Periwal, Eds. Cambridge, MA: MIT Press, 2006, pp. 243–267. 50. C. A. Parent and P. N. Devreotes, “A cell’s sense of direction,” Science, pp. 765–770, 1999. 51. A. Levchenko and P. A. Iglesias, “Models of eukaryotic gradient sensing: Application to chemotaxis of amoebae and neutrophils,” Biophys. J, vol. 82, pp. 50–63, 2002. 52. L. Ma, C. Janetopoulos, L. Yang, P. N. Devreotes, and P. A. Iglesias, “Two complementary, local excitation, global inhibition mechanisms acting in parallel can explain the chemoattractant-induced regulation of PI(3,4,5)P3 response in Dictyostelium cells,” Biophys. J, vol. 87, pp. 3764–3774, 2004. 53. C. L. Manahan, P. A. Iglesias, Y. Long, and P. N. Devreotes, “Chemoattractant signaling in Dictyostelium discoideum,” Annu. Rev. Cell Dev. Biol., vol. 20, pp. 223–253, 2004. 54. J. Krishnan and P. A. Iglesias, “Analysis of the signal transduction properties of a module of spatial sensing in eukaryotic chemotaxis,” Bull. Math. Biol., vol. 65, pp. 95–128, 2003. 55. B. Kutscher, P. Devreotes, and P. A. Iglesias, “Local excitation, global inhibition mechanism for gradient sensing: An interactive applet,” Sci. STKE, p. PL3, 2004. 56. C. Janetopoulos, L. Ma, P. N. Devreotes, and P. A. Iglesias, “Chemoattractant-induced phosphatidylinositol 3,4,5-trisphosphate accumulation is spatially amplified and adapts, independent of the actin cytoskeleton,” Proc. Natl. Acad. Sci. USA, vol. 101, pp. 8951– 8956, 2004. 57. J. Krishnan and P. A. Iglesias, “A modeling framework describing the enzyme regulation of membrane lipids underlying gradient perception in Dictyostelium cells,” J. Theor. Biol., vol. 229, pp. 85–99, 2004. 58. ——, “A modelling framework describing the enzyme regulation of membrane lipids underlying gradient perception in Dictyostelium cells II: Input-output analysis,” J. Theor. Biol., vol. 235, pp. 504–520, Aug 2005. 59. K. Amonlirdviman, N. A. Khare, D. R. P. Tree, W.-S. Chen, J. D. Axelrod, and C. J. Tomlin, “Mathematical modeling of planar cell polarity to understand domineering nonautonomy,” Science, pp. 423–426, 2005. 60. H. El-Samad, H. Kurata, J. C. Doyle, C. A. Gross, and M. Khammash, “Surviving heat shock: Control strategies for robustness and performance,” Proc. Natl. Acad. Sci. USA, vol. 102, pp. 2736–2741, 2005. 61. P. A. Iglesias and A. Levchenko, “Modeling the cell’s guidance system,” Sci. STKE, p. RE12, 2002. 62. J. C. Effler, P. A. Iglesias, and D. N. Robinson, “Regulating cell shape during cytokinesis,” in Contol of Uncertain Systems: Modelling, Approximation, and Design, ser. LNCIS, B. A. Francis, M. C. Smith, and J. C. Willems, Eds. Heidelberg: Springer, 2006, pp. 203–224. 63. J. C. Effler, Y.-S. Kee, J. M. Berk, M. N. Tran, P. A. Iglesias, and D. N. Robinson, “Mitosis-sspecific mechanosensing and contractile-protein redistribution control cell shape,” Current Biology, vol. 16, no. 19, October 10 2006. 64. M. Morohashi, A. E. Winn, M. T. Borisuk, H. Bolouri, J. Doyle, and H. Kitano, “Robustness as a measure of plausibility in models of biochemical networks,” J. Theor. Biol., vol. 216, pp. 19–30, 2002. 65. L. Ma and P. A. Iglesias, “Quantifying robustness of biochemical network models,” BMC Bioinformatics, p. 38, 2002. 66. J. Stelling, E. D. Gilles, and F. J. Doyle 3rd, “Robustness properties of circadian clock architectures,” Proc. Natl. Acad. Sci. USA, vol. 101, pp. 13 210–13 215, 2004.

288

B. W. Andrews and P. A. Iglesias

67. R. J. Prill, P. A. Iglesias, and A. Levchenko, “Dynamic properties of network motifs contribute to biological network organization,” PLoS Biol., vol. 3, p. e343, 2005. 68. J. Kim, D. G. Bates, I. Postlethwaite, L. Ma, and P. A. Iglesias, “Robustness analysis of biochemical network models,” IEE Proc. Systems Biol., vol. 152, pp. 96–104, 2006. 69. D. Angeli and E. D. Sontag, “Monotone control systems,” IEEE Trans. Automat. Control, vol. 48, pp. 1684–1698, 2003. 70. D. Angeli, J. E. Ferrell Jr, and E. D. Sontag, “Detection of multistability, bifurcations, and hysteresis in a large class of biological positive-feedback systems,” Proc. Natl. Acad. Sci. USA, vol. 101, pp. 1822–1827, 2004.

11 Robust Control of a Distillation Column Da-Wei Gu

Summary. In this Chapter we present the design of a robust control system for a high-purity distillation column. The original nonlinear model of the column is of high order and it includes parametric gain and time delay uncertainty. A low order linearized distillation column model is used in design of a two-degree-of freedom (2DOF) H∞ loop shaping controller and a µ controller. Both controllers ensure robust stability of the closed loop system and fulfillment of a mixture of time domain and frequency domain specifications. A reduced order µ -controller is then found which preserves the robust stability and robust performance of the closed loop system. The simulation of the closed loop system with the nonlinear distillation column model shows very good performance for different reference and disturbance signals as well as for different values of the uncertain parameters.

11.1 Introduction Distillation is an important process in the separation and purification of chemicals. The process exploits the difference at boiling points of multi-component liquids. The control of distillation columns is difficult, because the distillation process is highly nonlinear and the corresponding linearized models are often ill-conditioned around the operating point. The aim of the design, presented in this chapter, is to find a controller which achieves robust stability and robust performance of the closed loop control system of a high-purity distillation column. The original nonlinear model of the column is of 82nd order and it includes uncertainties in the form of parametric gains and time delay. The uncertainty model is considered in the form of an input multiplicative complex uncertainty. In our design exercises, we found that it is difficult to achieve the desired performance of the closed loop system using one-degree-of-freedom controllers. Hence we turned to H∞ two-degree-of-freedom loop shaping design procedure and µ -synthesis/analysis method. The designs are based on a 6th order linearized distillation column model. Both designed controllers ensure robust stability of the closed loop system and achieve a mixed set of time domain and frequency domain † This

is based in part on Chapter 11 of Robust Control Design with Matlab by D.W. Gu, M. Konstatinov and P. Petkov, Springer, 2005.

M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 289-328, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

290

D.-W. Gu

specifications. We present several time domain and frequency domain characteristics of the corresponding closed loop systems which makes possible the comparison of controllers efficiency. An 11th order reduced order µ -controller is found which preserves the stability and performance of the closed loop system in the presence of uncertainties. The simulation of the closed loop system with this µ controller and with the nonlinear distillation column model is R conducted in Simulink and shows very good performance for different reference and disturbance signals as well as for different values of the uncertain parameters.

11.2 Dynamic Model of the Distillation Column

Overhead vapour VT

Condenser

MD

P

Condenser holdup

Reflux N N-1

Distillate D, yD

L

Feed F, zF

3 2 1 Reboiler holdup

Boilup V Reboiler

Bottom product B, xB

Fig. 11.1. The distillation column system A typical two-product distillation column is shown in Figure 11.1. The objective of the distillation column is to split the feed F, which is a mixture of a light and a heavy component with composition zF , into a distillate product D with composition yD , which contains most of the light component, and a bottom product B with composition zB , which contains most of the heavy component. For this aim the column contains a series of trays that are located along its height. The liquid in the columns flows through the trays from top to bottom, while the vapor in the column rises from bottom to top. The constant contact between the vapor and liquid leads to increasing the concentration of the more volatile component in the vapor, while

Robust Control of a Distillation Column

291

simultaneously increasing the concentration of the less volatile component in the liquid. The operation of the column requires that some of the bottom product is reboiled at a rate of V to ensure the continuity of the vapor flow and some of the distillate is refluxed to the top tray at a rate of L to ensure the continuity of the liquid flow. The notations used in the derivation of the column model are summarized in Table 11.1 and the column data are given in Table 11.2. Table 11.1. Column nomenclature Symbol F zF qF D and B yD and xB

Description Feed rate [kmol/min] feed composition [mole fraction] fraction of liquid in feed distillate (top) and bottom product flowrate [kmol/min] distillate and bottom product composition (usually of light component) [mole fraction] L reflux flow [kmol/min] V boilup flow [kmol/min ] N number of stages (including reboiler) Ntot = N + 1 total number of stages (including condenser) i stage number (1 - bottom, NF - feed stage, NT - total condenser) Li and Vi liquid and vapor flow from stage i [kmol/min] xi and yi liquid and vapor composition of light component on stage i Mi liquid holdup on stage i [kmol] (MB - reboiler, MD -condenser holdup) α relative volatility between light and heavy component τL time constant for liquid flow dynamics on each stage [min]

Table 11.2. Column Data N Ntot NF 40 41 21 B L V 0.5 2.70629 3.20629

F 1 yD 0.99

zF 0.5 xB 0.01

qF D 1 0.5 Mi τL 0.5 0.063

The index i denotes the stages numbered from the bottom (i = 1) to the top (i = Ntot ) of the column. Index B denotes the bottom product and D the distillate product. A particular highpurity distillation column with 40 stages (39 trays and a reboiler) plus a total condenser is considered. The nonlinear model equations are: 1. Total material balance on stage i

292

D.-W. Gu dMi /dt = Li+1 − Li +Vi−1 −Vi .

2. Material balance for light component on each stage i d(Mi xi )/dt = Li+1 xi+1 +Vi−1 yi−1 − Li xi −Vi yi . This equation leads to the following expression for the derivative of the liquid mole fraction dxi /dt = (d(Mi xi )/dt − xi (dMi /dt))/Mi .

3. Algebraic equations The vapor composition yi is related to the liquid composition xi on the same stage through the algebraic vapor-liquid equilibrium yi = α xi /(1 + (α − 1)xi ).

From the assumption of constant molar flows and no vapor dynamics, one obtains the following expression for the vapor flows Vi = Vi−1 . The liquid flows depend on the liquid holdup on the stage above and the vapor flow as follows Li = L0i + (Mi − M0i )/τL + λ (Vi−1 −V 0i−1 )

where L0i [kmol/min] and M0i [kmol] are the nominal values for the liquid flow and holdup on stage i and V 0i is the nominal boilup flow. If the vapor flow into the stage effects the holdup then the parameter λ is different from zero. For the column under investigation λ = 0.

The above equations apply at all stages except in the top (condenser), feed stage and bottom (reboiler). 1. For the feed stage, i = NF (it is assumed that the feed is mixed directly into the liquid at this stage) dMi /dt = Li+1 − Li +Vi−1 −Vi + F, d(Mi xi )/dt = Li+1 xi+1 +Vi−1 yi−1 − Li xi −Vi yi + FzF .

2. For the total condenser, i = Ntot (MNtot = MD , LNtot = LT ) dMi /dt = Vi−1 − Li − D,

d(Mi xi )/dt = Vi−1 − Li xi − Dxi .

3. For the reboiler, i = 1(Mi = MB ,Vi = VB = V )

d(Mi xi )/dt = Li+1 xi+1 −Vi yi − Bxi . As a result we obtain a nonlinear model of the distillation column of 82nd order. There are two states per tray, one representing the liquid composition and the other one representing the liquid holdup. The model has four manipulated inputs (LT ,VB , D and B) and three disturbances (F, zF and qF ). In order to find a linear model of the distillation column it is necessary to have a steady state operating point around which the column dynamics is to be linearized. However the model

Robust Control of a Distillation Column

293

contains two integrators, because the condenser and reboiler levels are not under control. To stabilize the column, we make use of the so called LV-configuration of the distillation column where we use D to control MD and B to control MB . This is done by two proportional controllers with both gains equal to 10. The nonlinear model is linearized at the operating point given in Table 11.2 (the values of F, L,V, D, B, yD , xB and zF ). These steady state values correspond to an initial state where all liquid compositions are equal to 0.5 and the tray holdups are also equal to 0.5 [kmol]. The steady state vector is obtained for t = 5000 min by numerical integration of the nonlinear model equations of the LV-configuration given in the M-file cola lv.m. The linearization is carried out by implementing the M-file cola lin which makes use of the equations given in the file cola lv lin.m. The 82nd order, linear model is stored in the variable G4 u and has four inputs (the latter two are actually disturbances) [LT VB F zF ] and two outputs [yD xB ]. Before reducing the model order, the model G4 u is scaled in order to make all inputs/disturbances and all outputs at about the same magnitude. This is done by dividing each variable by its maximum change, i.e., u = U/Umax ; y = Y /Ymax where U, Y are the input and output of the model G4 u in original units, Umax , Ymax are the corresponding maximum values allowed, and u, y are the scaled variables. The scaling is achieved by using the input scaling matrix   10 0 0 0 1 0 0   Si =   0 0 0.2 0  0 0 0 0.1 and output scaling matrix

So =

100 0 . 0 100

The scaled model is then found as G4 = SoG4 uSi . The final stage in selecting the column model is the order reduction of the scaled model G4 . This is done by using the commands sysbal and hankmr. As a result we obtain a 6th order model saved in the variable G. All commands for finding the 6th order liner model of the distillation column are contained in the file mod col.m. The frequency responses of the singular values of G are compared with the singular values of the 82nd order linearized model G4 in Figure 11.2. It is seen that the behaviour of both models is close until the frequency 2 rad/min.

11.3 Uncertainty Modelling The uncertainties considered in the distillation column control systems are a gain uncertainty of ±20% and a time delay of up to 1 min in each input channel. Thus the uncertainty may be represented by the transfer matrix

294

D.-W. Gu Frequency responses of the original plant and shaped plant

6

10

5

10

4

10

3

Magnitude

10

2

10

1

10

0

10

−1

solid line: singular values of G

10

dashed line: singular values of G4

−2

10

−3

10

−4

10

−3

10

−2

−1

10

10 Frequency (rad/min)

0

1

10

2

10

10

Fig. 11.2. Singular values of G and G4 Gd

W∆

∆1

1

u1

+

+ y1

∆2

W∆ 2 u2

G +

+

y2

Fig. 11.3. Distillation column with input multiplicative uncertainty

Wu =

0 k1 e−Θ1 s 0 k2 e−Θ2 s

where ki ∈ [0.8 1.2]; Θi ∈ [0.0 1.0]; i = 1, 2. It is convenient to represent this uncertainty by an input multiplicative uncertainty as shown in Figure 11.3 with ∆1 0 ∆= 0 ∆2 where |∆1 | ≤ 1, |∆2 | ≤ 1. The uncertainty weighting function W∆1 0 W∆ = 0 W∆2 is determined in the following way. Denote by W ui = 1 the nominal transfer function in the ith channel for ki = 1 and Θi = 0; i =1, 2.

Robust Control of a Distillation Column

295

According to Figure 11.3 we have that Wui = (1 +W∆i ∆i )W ui , i = 1, 2. Taking into account that |∆i | ≤ 1 it follows that the relative uncertainty should satisfy Wu ( jω ) − W u ( jω ) i i ≤ W∆i ( jω ) , i = 1, 2 W u ( jω ) i

where Wui ( jω ) = ki e jωΘi = ki (cos(ωΘi ) + j sin(ωΘi )). In this way, to choose the uncertainty weight W∆i is equivalent to determine an upper bound of the frequency response of the relative uncertainty Wu ( jω ) − W u ( jω ) q i i = (ki cos(ωΘi ) − 1)2 + (ki sin(ωΘi ))2 . W u ( jω ) i

The frequency responses of of the relative uncertainty

Approximation of uncertain time delay by multiplicative perturbation 2.5

2

Magnitude

1.5

1

0.5

0 −2 10

−1

10

0

10 Frequency (rad/min)

1

10

2

10

Fig. 11.4. Approximation of the uncertain time delay

Wu ( jω ) − W u ( jω ) i i W u ( jω ) i

are computed by the file unc col.m and shown in Figure 11.4. These responses are then approximated by 3rd order transfer functions by using the file wfit.m. As a result one obtains W∆i =

2.2138s3 + 15.9537s2 + 27.6702s + 4.9050 , i = 1, 2. 1.0000s3 + 8.3412s2 + 21.2393s + 22.6705

296

D.-W. Gu

11.4 Closed-Loop System Performance Specifications The aim of the distillation column control system design is to determine a controller which meets robust stability and robust performance specifications for the LV configuration. Since these specifications are difficult to be satisfied with a one-degree-of-freedom controller, we present the design of two-degree-of-freedom controllers which ensure robust stability and robust performance of the closed loop system. In the given case, the robust stability means guaranteed closed loop stability for all 0.8 ≤ k1 , k2 ≤ 1.2 and 0 ≤ Θ1 , Θ2 ≤ 1 minute. The time domain specifications are given in terms of step response requirements, which must be met for all values of k1 , k2 , Θ1 and Θ2 . Specifically, for a unit step command to the first input channel at t = 0, the scaled plant outputs y1 (tracking) and y2 (interaction) should satisfy: • • • • •

y1 (t) ≥ 0.9 for all t ≥ 30 min; y1 (t) ≤ 1.1 for all t; 0.99 ≤ y1 (∞) ≤ 1.01; y2 (t) ≤ 0.5 for all t; −0.01 ≤ y2 (∞) ≤ 0.01.

Correspondingly, similar requirements should be met for a unit step command at the second input channel. In addition, the following frequency domain specification should be met: •

•

ˆ jω ) < 316, for each ω , where Kˆ y denotes the feedback part of the unscaled conσ (Kˆ y S)( troller. (Here and further on, a variable with hat refers to the case of unscaled plant.) This specification is included mainly to avoid saturation of the plant inputs. ˆ jω ) ≤ 1, for ω ≥ 150. σ (Gˆ Kˆ y )( jω ) < 1, for ω ≥ 150; or σ (Kˆ y S)(

In the above, σ denotes the largest singular value, and Sˆ = (I + Gˆ Kˆ y ) < 1 is the sensitivity ˆ function for G. Wu

W∆

r

K

u

eu

∆

y+

G

Wp

ey

Wn

n

M

Fig. 11.5. Closed-loop interconnection structure of the Distillation Column system The block diagram of the closed-loop system incorporating the design requirements consideration represented by weights is shown in Figure 11.5. The plant enclosed by the dashed

Robust Control of a Distillation Column

297

rectangle consists of the nominal scaled model G plus the input multiplicative uncertainty. The controller K implements a feedback from outputs yD and xB and a feedforward from the reference signal r. The measurement of the distillate and bottom products composition is corrupted by the noise n. The desired dynamics of the closed loop system is sought by implementation of a suitably chosen model M. The model M represents the desired dynamic behavior of the closed-loop system from the reference signal to the outputs. The usage of a model of the desired dynamics allows to take easily into account the design specifications. The transfer function matrix of the model M is selected as # " 1 0 T s2 +2ξ T s+1 . M= 1 0 T s2 +2ξ T s+1

The coefficients of the transfer functions (T = 6, ξ = 0.8) in both channels of the model are chosen such that to ensure an overdamped response with the settling time of about 30 min. The off-diagonal elements of the transfer matrix are set as zeros in order to minimize the interaction between the channels. Model frequency response

0

10

−1

10

−2

Magnitude

10

−3

10

−4

10

−5

10

−6

10

−4

10

−3

10

−2

10

−1

10 Frequency (rad/sec)

0

10

1

10

2

10

Fig. 11.6. Model frequency response

The frequency response of the model M is shown in Figure 11.6. Let the scaled, two-degree-of-freedom controller be partitioned as K(s) = Ky (s) Kr (s) .

where Ky is the feedback part of the controller and Kr is the pre-filter part. It is easy to show that ˜ r − M) −Wp TWn Wp (SGK r ep = ˜ −1 Kr −Wu Ky SWn n eu Wu (I + Ky G)

298

D.-W. Gu

˜ y )−1 is the sensitivity function for the scaled plant, T = (I + GK ˜ y )−1 GK ˜ y where S = (I + GK ˜ is the complementary sensitivity function and G = G(I +W∆ ∆ ) is the uncertain, scaled plant model. The performance objective is to satisfy

˜ r − M) −Wp TWn

Wp (SGK

<1

−1

Wu (I + Ky G) ˜ Kr −Wu Ky SWn ∞

(11.1)

˜ for each uncertain G. The performance and control action weighting functions are chosen as # " 9.5s+3 s+1 0.3 0.55 9.5s+10 0.87 0.01s+1 0 −4 , W = Wp = . u s+1 9.5s+3 0 0.87 0.01s+1 0.3 0.55 9.5s+10 −4 The implementation of the performance weighting function Wp aims to ensure closeness of the system dynamics to the model over the low frequency range. Note that this function contains nonzero off-diagonal elements which make it easier to meet the time domain specifications. A small constant equal to 10−4 is added in the denominator in each channel to make the design problem regular. The usage of the control weighting function uW allows to limit the magnitude of control actions over the specified frequency range (ω ≥ 150). Inverse of Performance Weighting Function

1

10

0

10

−1

Magnitude

10

−2

10

−3

10

−4

10

−5

10

−6

10

−5

10

−4

10

−3

10

−2

−1

10 10 Frequency (rad/min)

0

10

1

10

2

10

Fig. 11.7. Inverse of performance weighting function

The magnitude plot of the inverse of the performance weighting function Wp is shown in Figure 11.7 and the magnitude plot of the control weighting function is shown in Figure 11.8. The noise shaping filter −2 s 10 s+1 0 Wn = s 0 10−2 s+1

Robust Control of a Distillation Column

299

Control Action Weighting Function

2

10

1

Magnitude

10

0

10

−1

10

−4

10

−3

−2

10

−1

10

10

0

1

2

10 10 Frequency (rad/min)

3

10

10

4

10

Fig. 11.8. Control action weighting function

is determined according to the spectral contents of the sensor noises accompanying the measurement of the distillate and bottom product composition.

Sensor Noise Weight

−2

10

−3

Magnitude

10

−4

10

−5

10

−6

10

−4

10

−3

10

−2

10

−1

10 Frequency (rad/min)

0

10

1

10

2

10

Fig. 11.9. Noise weighting function

The magnitude plot of the noise shaping filter is shown in Figure 11.9. The model transfer function, the performance and control weighting functions as well as the noise shaping filter are all set in the file wts col.m.

300

D.-W. Gu

11.5 Open-Loop and Closed-Loop System Interconnections

pertout {1-2}

pertin {1-2}

W∆ −

y

e_y

Wp

G +

control

Wu −

−

Wn

noise

e_u

M

ref

Fig. 11.10. Open-loop interconnection structure of the distillation column system The open-loop system interconnection is obtained by the M-file olp col. The internal structure of the eight input, ten output open-loop system, which is saved as the variable sys ic, is shown in Figure 11.10. The inputs and outputs of the uncertainties are saved as the variables pertin and pertout, the references and the noises – as the variables ref and noise, respectively, and the controls - as the variable control. All variables have two elements (i.e., 2-dimensional vectors).

pertin{1} pertin{2} ref{1} ref{2} noise{1} noise{2}

control{1} control{2}

1 2 3 4 5 6 7 8

sys_ic

1 2 3 4 5 6 7 8 9 10

pertout{1} pertout{2} e_y{1} e_y{2} e_u{1} e_u{2} -y{1}-noise{1} -y{2}-noise{2}

ref{1} ref{2}

Fig. 11.11. Schematic diagram of the open-loop interconnection

The schematic diagram showing the specific input/output ordering for the variable sys ic is given in Figure 11.11. The block-diagram used in the simulation of the closed-loop system is shown in Figure 11.12. The corresponding closed-loop system interconnection, which is saved as the variable sim ic, is obtained by the M-file sim col.m.

Robust Control of a Distillation Column pertout {1-2}

301

pertin {1-2}

W∆

y Gd

control −

−

Wn

noise

ref

Fig. 11.12. Closed-loop interconnection structure of the Distillation Column system

pertin{1} pertin{2} ref{1} ref{2} noise{1} noise{2}

control{1} control{2}

1 2 3 4 5 6 7 8

sim_ic

1 2 3 4 5 6 7 8 9 10

pertout{1} pertout{2} y{1} y{2} control{1} control{2} -y{1}-noise{1} -y{2}-noise{2}

ref{1} ref{2}

Fig. 11.13. Schematic diagram of the closed-loop interconnection The schematic diagram showing the specific input/output ordering for the variable sim ic is shown in Figure 11.13.

11.6 Controller Design Successful design of the Distillation Column control system may be obtained by using the H∞ loop shaping design procedure (LSDP) and the µ -synthesis. Note that in the case of LSDP we do not use the performance specifications implemented in the case of µ -synthesis. Instead of these specifications we use a pre-filter W1 and a post-filter W2 in order to shape appropriately the open-loop transfer function W1 GW2 .

11.6.1 Loop Shaping Design In the present case we choose a pre-filter with transfer function

302

D.-W. Gu

Frequency responses of the plant (−) and shaped plant (−−)

6

10

5

10

4

10

3

Magnitude

10

2

10

1

10

0

10

−1

10

−2

10

−3

10

−4

10

−3

−2

10

10

−1

10 Frequency (rad/min)

0

10

1

10

2

10

Fig. 11.14. Singular values of the original system and shaped system Robust stability

0

10

−1

mu

10

−2

10

−3

10

−4

10

−3

10

−2

−1

10

10

0

10 Frequency (rad/min)

1

10

2

10

3

10

Fig. 11.15. Robust stability for loop shaping controller

W1 =

0 1.7 1.1s+1 10s 1.1s+1 . 0 1.7 10s

The choice of the gain equal to 1.7 is done to ensure a sufficiently small steady-state error. Larger gain leads to smaller steady-state errors but worse transient response. The post-filter is taken simply as W2 = I2 .

Robust Control of a Distillation Column

303

The singular value plots of the original and shaped systems are shown in Figure 11.14. The design of the two-degree-of-freedom LSDP controller is done by using the M-file lsh col.m which implements the function ncfsyn. The controller obtained is of order 10. The robust stability analysis of the closed-loop system is done by the file mu col the frequency response plot of the structured value µ being shown in Figure 11.15. According to this plot the closed loop system preserves stability for all perturbations with norm less than 1/0.6814. As usual, the requirements for nominal performance and robust performance are not fulfilled with this controller. The closed loop frequency responses are obtained by using the file frs col.m.

Singular Value Plot of the Closed−loop Transfer Function Matrix

1

10

0

10

−1

10

−2

10

−3

10

−2

10

−1

0

10

10

1

10

Frequency (rad/min)

Fig. 11.16. Frequency response of the closed loop system with loop shaping controller

The singular value plot of the unscaled closed loop system transfer function is shown in Figure 11.16. Both low frequency gains are equal to 1 which ensures zero steady state errors in both channels. The singular value plots of the transfer function matrix in respect to the noises (Figure 11.17) show that the noises are attenuated at least 104 times at the system output. The singular value plots of the transfer function matrices Gˆ Kˆ y and Kˆ y Sˆ are shown in Figures 11.18 and 11.19, respectively. The maximum of the largest singular value of Gˆ Kˆ y is less than 1 for ω ≥ 150 and the maximum of the largest singular value of Kˆ y Sˆ is less than 200 so that the corresponding frequency domain specification is met. In Figures 11.20 we show the transient responses of the scaled closed-loop system obtained by the file prt col.m for different values of the uncertain gain and time delay. The time domain specification is met and the closed loop system transient response has a small settling time. The control action in the closed-loop system for the same variations of the uncertain parameters is shown in Figure 11.21.

304

D.-W. Gu

Singular Value Plot of the Noise Transfer Function

−3

10

−4

10

−5

10

−6

10

−7

10

−8

10

−9

10

−10

10

−3

10

−2

10

−1

10

0

10 Frequency (rad/min)

1

10

2

10

3

10

Fig. 11.17. Frequency response to the noises

Singular Value Plot of GKy

6

10

4

10

2

10

0

10

−2

10

−4

10

−6

10

−3

10

−2

10

−1

10

0

10 Frequency (rad/min)

1

10

Fig. 11.18. Singular value plot of Gˆ Kˆ y

2

10

3

10

Robust Control of a Distillation Column

305

Singular Value Plot of KyS

3

10

2

10

1

10

0

10

−1

10

−2

10

−1

−2

0

10

10

1

2

10 10 Frequency (rad/min)

3

10

10

Fig. 11.19. Singular value plot of Kˆ y Sˆ

Transient responses of the perturbed systems 1.5

1

1

y12

y11

Transient responses of the perturbed systems 1.5

0.5

0

−0.5

0.5

0

0

10

20

30

40

50 Time (min)

60

70

80

90

−0.5

100

0

10

20

1.5

1.5

1

1

0.5

0

−0.5

30

40

50 Time (min)

60

70

80

90

100

80

90

100

Transient responses of the perturbed systems

y22

y21

Transient responses of the perturbed systems

0.5

0

0

10

20

30

40

50 Time (min)

60

70

80

90

100

−0.5

0

10

20

30

40

50 Time (min)

60

70

Fig. 11.20. Perturbed transient responses for loop shaping controller

D.-W. Gu

306

Control action in the perturbed systems 1.5

1

1

u12

u11

Control action in the perturbed systems 1.5

0.5

0

−0.5

0.5

0

0

10

20

30

40

50 Time (min)

60

70

80

90

−0.5

100

0

10

20

1.5

1.5

1

1

0.5

0

−0.5

30

40

50 Time (min)

60

70

80

90

100

80

90

100

Control action in the perturbed systems

u22

u21

Control action in the perturbed systems

0.5

0

0

10

20

30

40

50 Time (min)

60

70

80

90

100

−0.5

0

10

20

30

40

50 Time (min)

60

70

Fig. 11.21. Perturbed control action for loop shaping controller

11.6.2 µ -Synthesis Let us denote by P(s) the transfer function matrix of the eight input, ten output open-loop system consisting of the distillation column model plus the weighting functions and let the block structure ∆P is defined as ∆ 0 : ∆ ∈ C2×2 , ∆F ∈ C4×4 . ∆P := 0 ∆F The first block of this matrix corresponds to the uncertainty block ∆ , used in modelling the uncertainty of the distillation column. The second block ∆F is a fictitious uncertainty 4 × 4 block, introduced to include the performance objectives in the framework of the µ -approach. The inputs to this block are the weighted error signals e p and eu the outputs being the exogenous inputs r and n. To meet the design objectives a stabilizing controller K is to be found such that, at each frequency ω ∈ [0, ∞], the structured singular value satisfies the condition

µ∆P [FL (P, K)( jω )] < 1.

Robust Control of a Distillation Column

307

The fulfillment of this condition guarantees robust performance of the closed-loop system, i.e.,

˜ r − M) −Wp TWn

Wp (SGK

(11.2)

Wu (I + Ky G) ˜ −1 Kr −Wu Ky SWn < 1 ∞

The µ -synthesis is done by using the M-file ms col.m. The uncertainty structure and other parameters used in the D-K iteration are set in the auxiliary file dk col.m.

Table 11.3. Results of the µ -synthesis Iteration Controller order Maximum value of µ 1 22 1.072 2 28 0.980 3 30 0.984 4 28 0.975 The progress of the D-K iteration is shown in Table 11.3. In the given case an appropriate controller is obtained after the fourth D-K iteration. The controller is stable and its order is equal to 28. It can be seen from the Table 11.3 that after the fourth iteration the maximum value of µ is equal to 0.975. The µ -analysis of the closed-loop system is done by the file mu col. The frequency response plot of the structured singular value for the case of robust stability is shown in Figure 11.22. The maximum value of µ is 0.709 which means that the stability of 1 the system is preserved under perturbations which satisfy k∆ k∞ < 0.709 . The frequency response of µ for the case of robust performance analysis is shown in Figure 11.23. The closed-loop system achieves robust performance the maximum value of µ being equal to 0.977. Robust stability

0

10

−1

10

−2

mu

10

−3

10

−4

10

−5

10

−3

10

−2

10

−1

10

0

10 Frequency (rad/min)

1

10

2

10

Fig. 11.22. Robust stability for µ controller

3

10

D.-W. Gu

Robust performance 1

0.9

0.8

0.7

0.6

mu

308

0.5

0.4

0.3

0.2

0.1

0 −3 10

−2

10

−1

10

0

10 Frequency (rad/min)

1

2

10

10

3

10

Fig. 11.23. Robust performance for µ controller

Singular Value Plot of the Closed−loop Transfer Function Matrix

0

10

−1

10

−2

10

−3

10

−4

10

−5

10

−2

10

−1

0

10

10

Frequency (rad/min)

Fig. 11.24. Closed-loop singular value plots

1

10

Robust Control of a Distillation Column

309

Singular Value Plot of the Noise Transfer Function

−3

10

−4

10

−5

10

−6

10

−7

10

−8

10

−9

10

−10

10

−3

10

−2

−1

10

10

0

10 Frequency (rad/min)

1

2

10

10

3

10

Fig. 11.25. Frequency responses in respect to noises The unscaled closed-loop system singular value plot is shown in Figure 11.24. The closed-loop bandwidth is about 0.1 rad/min. The frequency responses in respect to the noise are shown in Figure 11.25. It is seen from the figure that the noises in measuring the distillate and bottom product composition have a relatively small effect on the system output.

Singular Value Plot of the Sensitivity Function

1

10

0

10

−1

10

−2

10

−3

10

−4

10

−5

10

−3

10

−2

10

−1

10 Frequency (rad/min)

0

10

Fig. 11.26. Frequency responses of the sensitivity function

1

10

310

D.-W. Gu

Singular Value Plot of the Controller

5

10

4

10

3

10

2

10

1

10 −3 10

−2

−1

10

0

10 Frequency (rad/min)

1

10

10

Fig. 11.27. Singular values of the controller

ˆ The In Figure 11.26 we show the singular value plot of the unscaled sensitivity function S. singular value plots of the unscaled µ controller are shown in Figure 11.27. The singular value plots of Gˆ Kˆ y and Kˆ y Sˆ are shown in Figures 11.28 and 11.29, respectively. The maximum of

Singular Value Plot of GKy

6

10

4

10

2

10

0

10

−2

10

−4

10

−6

10

−3

10

−2

10

−1

10

0

10 Frequency (rad/min)

1

10

Fig. 11.28. Frequency responses of Gˆ Kˆ y

2

10

3

10

Robust Control of a Distillation Column

311

Singular Value Plot of KyS

3

10

2

10

1

10

0

10

−1

10

−2

10

−2

10

−1

10

0

1

10 10 Frequency (rad/min)

2

10

3

10

Fig. 11.29. Frequency responses of of Kˆ y Sˆ the largest singular value of Gˆ Kˆ y is less than 1 for ω ≥ 150 and the maximum of the largest singular value of Kˆ y Sˆ is less than 300 thus the frequency domain specification being met. Consider now the effect of variations of uncertain parameters on the system dynamics.

Singular Value Plot of the Sensitivity Function

1

10

0

10

−1

10

−2

10

−3

10

−4

10

−2

10

−1

10

0

1

10 10 Frequency (rad/min)

2

10

3

10

Fig. 11.30. Frequency responses of the perturbed sensitivity function

312

D.-W. Gu

Singular Value Plot of KyS

3

10

2

10

1

10

0

10

−1

10

−2

10

−1

−2

0

10

10

1

2

10 10 Frequency (rad/min)

3

10

10

Fig. 11.31. Perturbed frequency responses of Kˆ y Sˆ Transient responses of the perturbed systems

1

1

12

1.5

0.5

y

y

11

Transient responses of the perturbed systems 1.5

0

−0.5

0.5

0

0

10

20

30

40

50 Time (min)

60

70

80

90

−0.5

100

0

10

20

1

1

22

1.5

0.5

0

−0.5

40

50 Time (min)

60

70

80

90

100

80

90

100

Transient responses of the perturbed systems

1.5

y

y

21

Transient responses of the perturbed systems

30

0.5

0

0

10

20

30

40

50 Time (min)

60

70

80

90

100

−0.5

0

10

20

30

40

50 Time (min)

Fig. 11.32. Perturbed transient responses for µ controller

60

70

Robust Control of a Distillation Column

313

The frequency responses of the perturbed sensitivity function Sˆ obtained by the file pfr col.m are shown in Figure 11.30. The frequency responses of the perturbed transfer function matrix Kˆ y Sˆ are shown in Figure 11.30. The maximum of the largest singular value of this matrix does not exceed 300 for all values of the uncertain parameters. The perturbed transient responses of the scaled closed loop system with µ -controller are shown in Figure 11.32. The responses to the corresponding references have no overshoots and the interaction of channels is weaker than in the case of using Loop Shaping controller. The control action in the case of perturbed system with the µ -controller is shown in Figure 11.33. Consider now the reduction of controller order. For this aim we implement the M-file red col.m. After balancing of the controller and neglecting the small Hankel singular values its order is reduced to 11. In Figure 11.34 we compare the frequency responses of the maximum singular values of the scaled full order and reduced order controllers. The frequency responses of both full-order and reduced-order controllers coincide up to 23 rad/min which is much more than the closed loop

Control action in the perturbed systems

1

1

12

1.5

0.5

u

u

11

Control action in the perturbed systems 1.5

0

−0.5

0.5

0

0

10

20

30

40

50 Time (min)

60

70

80

90

−0.5

100

0

10

20

1.5

1

1

22

1.5

0.5

0

−0.5

30

40

50 Time (min)

60

70

80

90

100

80

90

100

Control action in the perturbed systems

u

u

21

Control action in the perturbed systems

0.5

0

0

10

20

30

40

50 Time (min)

60

70

80

90

100

−0.5

0

10

20

30

40

50 Time (min)

Fig. 11.33. Perturbed control action for µ controller

60

70

314

D.-W. Gu

bandwidth of the system. That is why the transient responses of the closed-loop system with full order and with reduced order controller are practically undistinguishable.

11.7 Nonlinear System Simulation The LSDP controller and µ -controller designed are investigated by simulation of the correR sponding non-linear closed-loop system. The simulation is carried out by the Simulink model nls col.mdl using the nonlinear plant model given in Section 11.2. R The Simulink model of the distillation column control system shown in Figure 11.37 allows to carry on a number of simulations for different set points and disturbances. Note that the inputs to the controller are formed as differences between the values of the corresponding variables and their nominal (steady state) values used in the linearization. As opposite, the controller outputs are added to the corresponding nominal inputs in order to obtain the full inputs to the nonlinear model of the column. Before simulation of the system it is necessary to set the model parameters by using the M-file init col.m. Also, the controller is rescaled so that to implement the unscaled input/output variables.

Maximum singular values of the controller transfer matrices

2

10

1

10

0

10

−1

10

−2

10

−3

10

−4

10

−5

10

−6

10

−2

10

−1

10

0

10

1

10

2

3

10 10 Frequency (rad/min)

4

10

5

10

6

10

Fig. 11.34. Frequency responses of the full order and reduced order controllers

The nonlinear system simulation is done for the following reference and disturbance signals. At t = 10 min the feed rate F increases from 1 to 1.2, at t = 100 min the feed composition zF increases from 0.5 to 0.6 and at t = 200 min the set point in yD increases from 0.99 to 0.995.

Robust Control of a Distillation Column

315

The time response of the distillate yD for the case of the reduced-order µ -controller is given in Figure 11.35. It is seen from the Figure that the disturbances are attenuated well and the desired set point is achieved exactly. The time response of the bottom product composition xB for the same controller is given in Figure 11.36. The simulation results show that the robust design method is appropriately chosen and confirm the validity of the uncertain model used. Nonlinear system simulation 0.996

Product composition yD

0.994

0.992

0.99

0.988

0.986

0.984

0

50

100

150 Time (min)

200

250

300

Fig. 11.35. Transient response of the nonlinear system - yD

Nonlinear system simulation 0.015

0.014

Product composition xB

0.013

0.012

0.011

0.01

0.009

0.008

0

50

100

150 Time (min)

200

250

Fig. 11.36. Transient response of the nonlinear system - xB

300

D.-W. Gu

316

SIMULINK Model of the Distillation Column System zF

F

qf1 y1

Graph1

y_D

2.70629 L0

Sum3 y2

colas

Graph2

x_B

Transport Delay1

y3 Distillation column (nonlinear)

Sum4

M_D y4

Transport Delay2

M_B

LV configuration

Comp

3.20629

kf1*[Tf11 0](s)

Comp.

Sum1

V0 −10

Tf12.s+1

0.5

Shaping filter1

Sum9 Gain1

ref_M_D −10 Gain2

0.5 Sum2

Noise1

kf2*[Tf21 0](s) Tf22.s+1

ref_M_B

Sum10

Shaping filter2

Noise2

0.99 Sum5

x’ = Ax+Bu y = Cx+Du

Graph3

Controller

yDs1 0.01 Sum6

xBs1 Sum7 ref_yD

u1 Delta L

Sum8 ref_xB

u2 Delta V

t

Graph4 Clock

Time

R Fig. 11.37. Simulink model of the nonlinear system

11.8 Conclusions The results from the analysis and design of a distillation column control system may be summarized as follows. •

It is possible to use sufficiently low order linearized model of the given nonlinear plant, so that the designed linear controllers allow to achieve satisfactory dynamics of the nonlinear closed loop system. The linearized model is scaled in order to avoid very small or very large signals.

•

The one-degree-of-freedom controller does not allow to meet the time domain and frequency domain specifications which makes it necessary to use two-degree-of-freedom controllers. Two controllers are designed – one by using the H∞ Loop Shaping Design

Robust Control of a Distillation Column

•

317

method and the other by using the µ -synthesis method. Both controllers satisfy the time domain and frequency domain specifications and ensure robust stability of the corresponding closed loop systems. It is impressive how the low order, easily designed loop shaping controller allows to obtain practically the same characteristics of the closed loop systems as the µ -controller, while the latter requires much more experiments for tuning the weighting functions. The nonlinear system simulation results confirm the ability of the loop shaping controller and the reduced order µ -controller to achieve disturbance attenuation and good responses to reference signals. The simulation confirms the validity of the uncertain model used.

Notes and References The distillation column control problem presented in this chapter was introduced by Limebeer [86] as a benchmark problem at the 1991 Conference on Decision and Control. In [86] the uncertainty is defined in terms of parametric gain and delay uncertainty and the control objectives are a mixture of time domain and frequency domain specifications. The problem originates from Skogestad et al. [141] where a simple model of a high purity distillation column was used and uncertainty and performance specifications were given as frequency dependent weighting functions. A tutorial introduction to the dynamics of the distillation column is presented in [140]. A design of two-degree-of-freedom loop shaping controller for the distillation column is presented in [53] where an 8th order model of the column is used. A two-degree-of-freedom controller for the distillation column system is proposed in [95] with a reference model and using µ -synthesis. In that paper, one may find a selection procedure for the weighting functions described in details. Our design differs from the design in [95] in several aspects. First, instead of a 2nd order model with time delay we use a 6th order model which is justified by the results from nonlinear system simulation. Second, we use modified weighting functions in order to obtain better results. In particular, we use performance weighting transfer function matrix with nonzero off-diagonal elements which meets the time domain specifications much better. Also, the control weighting functions are taken as first order, low pass filters. Various design methods have been reported, in addition to the above, to tackle this distillation column problem ( [113, 127, 142, 147, 161]). In [161], the design problem is formulated as a mixed optimization problem. It is well known that control system design problems can be formulated as constrained optimization problems. Design specifications in both the time and frequency domains as well as stability can be naturally formulated as constraints. Numerical optimization approaches can directly be used and a solution obtained, if there is one, will characterize an acceptable design. However, the optimization problems such derived are usually very complicated with many unknowns, many nonlinearities, many constraints, and in most cases, they are multi-objective with several conflicting design aims which need to be simultaneously achieved. furthermore, a direct parameterization of the controller will increase

318

D.-W. Gu

the complexity of the optimization problem. In [161], the H∞ loop shaping design procedure is followed. Instead of direct parameterization of controllers, the pre- and post-weighting functions used to shape the open-loop, augmented system are chosen as design (optimization) parameters. The low order and simple structure of such weighting functions make the numerical optimization much more efficient. The H∞ norm requirement is also included in the cost/constraint set. The stability of the closed-loop system is naturally met by such designed controllers. Satisfactory designs are reported in that paper. [147] extends further the optimization approach in [161] by using Genetic Algorithm to choose the weighting function parameters.

References 1. D. Abramovitch and G. Franklin. A brief history of disk drive control. IEEE Control Systems Magazine, 22:28–42, 2002. 2. D. Abramovitch and G. Franklin. Disk drive control: The early years. In Proceedings of the 15th IFAC Congress. Session T-Th-M12, pages 1–12, Barcelona, Spain, July 2002. CD-ROM. 3. D. Abramovitch, T. Hurst, and D. Henze. An overview of the PES Pareto method for decomposing baseline noise sources in hard disk postion error signals. IEEE Transactions on Magnetics, 34:17–23, 1998. 4. D. Abramovitch, F. Wang, T. Hurst, and G. Franklin. Disk drive pivot nonlinearity modelling Part I: Frequency domain. In Proceedings of the 1994 American Control Conference, pages 2600–2603, Baltimore, MD, June 1994. 5. D.Y. Abramovitch. Magnetic and optical disk control: Parallels and contrasts. In Proceedings of the 2001 American Control Conference, pages 421–428, Arlington, VA, June 2001. 6. V.M. Adamjan, D.Z. Arov, and M.G. Krein. Analytic properties of Schmidt pairs for a Hankel operator and the generalized Schur-Takagi problem. Mat. USSR Sbornik, 15:31– 73, 1971. 7. R.J. Adams, J.M Buffington, A.G. Sparks, and S.S. Banda. Robust Multivariable Flight Control. Springer-Verlag, New York, 1994. 8. U.M. Al-Saggaf and G.F. Franklin. An error bound for a discrete reduced order model of a linear multivariable system. IEEE Transactions on Automatic Control, AC-32:815– 819, 1987. 9. G.J. Balas, J.C. Doyle, K. Glover, A. Packard, and R. Smith. µ -Analysis and Synthesis Toolbox: User’s Guide. MUSYN Inc. and The Mathworks, Inc., 1995. 10. D.S. Bernstein and W.M. Haddad. LQG control with an H ∞ performance bound: A Riccati equation approach. IEEE Transactions on Automatic Control, AC-34:293–305, 1989.

Robust Control of a Distillation Column

319

11. J.H. Blakelock. Automatic Control of Aircraft and Missiles. Wiley, New York, 1991. 12. A.E. Bryson. Control of Spacecraft and Aircraft. Princeton University Press, Princeton, NJ, 1994. 13. H. Buschek. Design and flight test of a robust autopilot for the IRIS-T air-to-air missile. Control Engineering Practice, 11:551–558, 2003. 14. F.M. Callier and C.A. Desoer. Linear System Theory. Springer-Verlag, New York, 1991. 15. B.M. Chen, T.H. Lee, and V. Venkataramanan. Springer-Verlag, Berlin, 2002.

Hard Disk Drive Servo Systems.

16. K.K. Chew and M. Tomizuka. Digital control of repetetive errors in disk drive systems. IEEE Control Systems Magazine, 10:16–20, 1990. 17. B.W. Choi, D.W. Gu, and I. Postlethwaite. Low-order H∞ Sub-Optimal Controllers. IEE Proceedings - Part D, 141:243–248, 1994. 18. R.R.E. de Gaston and M.G. Safonov. Exact calculation of the multiloop stability margin. IEEE Transactions on Automatic Control, AC-33:156–171, 1988. 19. U.B. Desai and D. Pal. A transformation approach to stochastic model reduction. IEEE Transactions on Automatic Control, AC-29:1097–1100, 1984. 20. C.A. Desoer and W.S. Chan. The feedback interconnection of lumped linear timeinvariant systems. Journal of the Franklin Institute, 300:335–351, 1975. 21. C.A. Desoer and M. Vidyasagar. Feedback Systems: Input-Output Properties. Academic Press, New York, 1975. 22. J. Ding, M. Tomizuka, and H. Numasato. Design and robustness analysis of dual stage servo system. In Proceedings of the 2000 American Control Conference, pages 2605– 2609, Chicago, Illinois, June 2000. 23. J.C. Doyle. Analysis of feedback systems with structured uncertainties. IEE Proceedings - Part D, 129:242–250, 1982. 24. J.C. Doyle. Structured uncertainty in control system design. In Proceedings of the 24 IEEE Conference on Decision and Control, pages 260–265, December 1985. 25. J.C. Doyle. A review of µ : For case studies in robust control. In Proceedings of 10th IFAC World Congress, pages 395–402, Munich, Germany, July 1987. 26. J.C. Doyle, B.A. Francis, and A.R. Tannenbaum. Feedback Control Theory. Macmillan Publishing Co., New York, 1992. 27. C. Du, J. Zhang, and G. Guo. Vibration analysis and control design comparison of HDDs using fluid bearing and ball bearing spindles. In Proceedings of the 2002 American Control Conference, pages 1378–1383, Anchorage, AK, May 2002. 28. K.G. Eltohamy. Nonlinear optimal control of a triple inverted pendulum with single control input. International Journal of Control, 69:239–256, 1998. 29. D. Enns. Model Reduction for Control Systems Design. PhD thesis, Department of Aeronautics and Astronautics, Stanford University, Stanford, CA, 1984. 30. D. Enns. Model reduction with balanced realizations: An error bound and a frequency weighted generalization. In Proceedings of the 23rd IEEE Conference on Decision and Control, pages 127–132, Las Vegas, NV, 1984. 31. R.B. Evans, J.S. Griesbach, and W.C. Messner. Piezoelectric microactuator for dual stage control. IEEE Transactions on Magnetics, 35:977–982, 1999.

320

D.-W. Gu

32. M.K.H. Fan and A.L. Tits. Characterization and efficient computation of the structured singular value. IEEE Transactions on Automatic Control, AC-31:734–743, 1986. 33. M.K.H. Fan, A.L. Tits, and J.C. Doyle. Robustness in the presence of mixed parametric uncertainty and unmodeled dynamics. IEEE Transactions on Automatic Control, AC36:25–38, 1991. 34. K.V. Fernando and H. Nicholson. Singular perturbational model reduction of balanced systems. IEEE Transactions on Automatic Control, AC-27:466–468, 1982. 35. K.V. Fernando and H. Nicholson. Singular perturbational approximation for discretetime systems. IEEE Transactions on Automatic Control, AC-28:240–242, 1983. 36. G. Ferreres. A Practical Approach to Robustness Analysis with Aeronautical Applications. Kluwer Academic/Plenum Publishers, New York, 1999. 37. I. Fialho, G.J. Balas, A.K. Packard, J. Renfrow, and C. Mullaney. Gain-scheduled lateral control of the F-14 aircrfat during powered approach landing. Journal of Guidance, Control, and Dynamics, 23:450–458, 2000. 38. B.A. Francis. A Course in H ∞ Control Theory, volume 88 of Lecture Notes in Control and Information Sciences. Springer-Verlag, 1987. 39. G.F. Franklin, J.D. Powell, and M.L. Workman. Digital Control of Dynamic Systems. Addison-Wesley, Menlo Park, CA, third edition, 1998. 40. K. Furuta, K. Kajiwara, and K. Kosuge. Digital control of a double inverted pendulum on an inclined rail. International Journal of Control, 32:907–924, 1980. 41. K. Furuta, T. Ochia, and N. Ono. Attitude control of a triple inverted pendulum. International Journal of Control, 39:1351–1365, 1984. 42. E. Garcia-Benitez, J. Walkins, and S. Yurkovich. Nonlinear control with acceleration feedback for a two-link flexible robot. Control Engineering Practice, 1:989–997, 1993. 43. S. Ge, T. Lee, and G. Zhu. Improving regulation of a single-link flexible manipulator with strain feedback. IEEE Transactions on Robotics and Automation, 14:179–185, 1998. 44. H. Geniele, R. Patel, and K. Khorasani. End-point control of a flexible-link manipulator: Theory and experiment. IEEE Transactions on Control Systems Technology, 5:559–570, 1997. 45. M. Ghanekar, D. Wang, and G. Heppler. Scaling laws for linear controllers of flexible link manipulators characterized by nondimensional groups. IEEE Transactions on Robotics and Automation, 13:117–127, 1997. 46. K. Glover. All optimal Hankel-norm approximations of linear multivariable systems and their l ∞ -error bounds. International Journal of Control, 39:1115–1193, 1984. 47. K. Glover. Multiplicative approximation of linear multivariable systems with L∞ error bounds. In Proceedings of the 1986 American Control Conference, pages 1705–1709, Minneapolis, MN, 1986. 48. K. Glover and J.C. Doyle. State-space formulae for all stabilizing controllers that satisfy an H ∞ norm bound and relations to risk sensitivity. Systems & Control Letters, 11:167– 172, 1988. 49. K. Glover and E.A. Jonckheere. A comparison of two Hankel norm methods for approximating spectra. In C.I. Byrnes and A. Lindquist, editors, Modelling, Identification and Robust Control. North-Holland, Amsterdam, The Netherlands, 1986.

Robust Control of a Distillation Column

321

50. T.B. Goh, Z. Li, B.M. Chen, T.H. Lee, and T. Huang. Design and implementation of a hard disk drive servo system using robust and perfect tracking approach. In Proceedings of the 38th IEEE Conference on Decision and Control, pages 5247–5252, Phoenix, Arizona, December 1999. 51. M. Green. Balanced stochastic realizations. Linear Algebra & its Applications, 98:211– 247, 1988. 52. M. Green. A relative error bound for balanced stochastic truncation. IEEE Transactions on Automatic Control, AC-33:961–965, 1988. 53. M. Green and D.J.N. Limebeer. Linear Robust Control. Prentice-Hall, Englewood Cliffs, NJ, 1995. 54. A.L. Greensite. Analysis and Design of Space Vehicle Flight Control Systems. Spartan, New York, 1970. 55. D.W. Gu, B.W. Choi, and I. Postlethwaite. Low-order stabilizing controllers. IEEE Transactions on Automatic Control, AC-38:1713–1717, 1993. 56. D.W. Gu, S.J. Goh, T. Fitzpatrick, I. Postlethwaite, and N. Measor. Application of an expert system for robust controller design. Control’96, 2:1004–1009, 1996. 57. D.W. Gu, P.Hr. Petkov, and M.M. Konstantinov. Direct Formulae for the H∞ SubOptimal Central Controller. Technical Report 1998-7, Niconet Report, 1998. Available at http://www.win.tue.nl/niconet. 58. D.W. Gu, P.Hr. Petkov, and M.M. Konstantinov. Formulae for discrete H∞ loop shaping design procedure controllers. In Proceedings of the 15th IFAC World Congress, paper 276, Session T-Mo-M15, Barcelona, Spain, July 2002. CD-ROM. 59. D.W. Gu, I. Postlethwaite, and M.C. Tsai. H∞ super-optimal solutions. In C.T. Leondes, editor, Advances in Control and Dynamic Systems, volume 51, pages 183–246. Academic Press, San Diego, CA, 1992. 60. D.W. Gu, M.C. Tsai, and I. Postlethwaite. State-space formulae for discrete-time optimization. International Journal of Control, 49:1683–1723, 1989. 61. L. Guo, H.S. Lee, A. Hudson, and S.-H. Chen. A comprehensive time domain simulation tool for hard disk drive TPI prediction and mechanical/servo enhancement. IEEE Transactions on Magnetics, 35:879–884, 1999. 62. S. Hammarling. Numerical solution of the stable non-negative definite Lyapunov equation. IMA Journal of Numerical Analysis, 2:303–323, 1982. 63. S. Hara, T. Hara, L. Yi, and M. Tomizuka. Two degree-of-freedom controllers for hard disk drives with novel reference signal generation. In Proceedings of the 1999 American Control Conference, pages 4116–4121, San Diego, CA, June 1999. 64. J.W. Helton. Worst case analysis in the frequency domain: The H ∞ approach to control. IEEE Transactions on Automatic Control, AC-30:1154–1170, 1985. 65. J.W. Helton and O. Merino. Classical Control Using H ∞ Methods. Society for Industrial and Applied Mathematics, Philadelphia, PA, 1998. 66. D. Hernandez, S.-S. Park, R. Horowitz, and A.K. Packard. Dual stage track-following servo design for hard disk drives. In Proceedings of the 1999 American Control Conference, pages 4116–4121, San Diego, CA, June 1999. 67. D. Hoyle, R. Hyde, and D.J.N. Limebeer. An H ∞ approach to two-degree-of-freedom design. In Proceedings of the 30th IEEE Conference on Decision and Control, pages 1581–1585, Brighton, UK, December 1991. 68. C.S. Hsu, X. Yu, H.H. Yeh, and S.S. Banda. H∞ compensator design with minimal order observer. In Proceedings of the 1993 American Control Conference, San Francisco, CA, June 1993.

322

D.-W. Gu

69. Y. Huang, P. Mathur, and W.C. Messner. Robustness analysis on a high bandwidth disk drive servo system with an instrumented suspension. In Proceedings of the 1999 American Control Conference, pages 3620–3624, San Diego, CA, June 1999. 70. J. Ishikawa, Y. Yanagita, and T. Hattori ans M. Hashimoto. Head positioning control for low sampling rate systems based on two degree-of-freedom control. IEEE Transactions on Magnetics, 32:1787–1792, 1996. 71. T. Iwasaki and R.E. Skelton. All low order H∞ controllers with covariance upper bound. In Proceedings of the 1993 American Control Conference, San Francisco, CA, June 1993. 72. H. Kajiwara, P. Apkarian, and P. Gahinet. LPV techniques for control of an inverted pendulum. IEEE Control Systems Magazine, 19:44–54, 1999. 73. M. Karkoub, G. Balas, K. Tamma, and M. Donath. Robust control of flexible manipulators via µ -synthesis. Control Engineering Practice, 8:725–734, 2000. 74. C. Kempf, W. Messner, M. Tomizuka, and R. Horovitz. Comparison of four discretetime repetetive control algorithms. IEEE Control Systems Magazine, 13:48–54, 1993. 75. S.W. Kim, B.D.O. Anderson, and A.G. Madievski. Error bound for transfer function order reduction using frequency weighted balanced truncation. Systems & Control Letters, 24:183–192, 1995. 76. M. Kobayashi, T. Yamaguchi, and R. Horowitz. Track-seeking controller design for dual-stage actuator in magnetic disk drives. In Proceedings of the 2000 American Control Conference, pages 2610–2614, Chicago, Illinois, June 2000. 77. C.M. Kozierok. The PC Guide. The Reference Section on Hard Disk Drives. January 2003. Available at http://www.PCGuide.com. 78. S. Kung. A new low-order approximation algorithm via singular value decomposition. In Proceedings of the 18th IEEE Conference on Decision and Control, Ft. Lauderdale, Florida, December 1979. 79. S. Kung and D.W. Lin. Optimal Hankel norm model reduction: Multivariable systems. IEEE Transactions on Automatic Control, AC-26:832–852, 1981. 80. P.J. Larcombe. On the generation and solution of the symbolic, open-loop charactersitic equation for a double inverted pendulum. International Journal of Systems Science, 24:2379–2390, 1993. 81. A.J. Laub. On computing balancing transformations. In Proceedings of the Joint 1980 American Control Conference, pages FA8–E, San Francisco, CA, August 1980. 82. A.J. Laub, M.T. Heath, C.C. Paige, and R.C. Ward. Computation of system balancing transformations and other applications of simultaneous diagonalization algorithms. IEEE Transactions on Automatic Control, AC-32:115–121, 1987. 83. H.S. Lee. Controller optimization for minimum position error signals of hard disk drives. In Proceedings of the 2000 American Control Conference, pages 3081–3085, Chicago, Illinois, June 2000. 84. D.J. Leith and W.E. Leithead. Survey of gain-scheduling analysis and design. International Journal of Control Control, 73:1001–1025, 2000. 85. Y. Li, B. Thang, Z. Shi, and Y. Lu. Experimental study for trajectory tracking of a two-link flexible manipulators. International Journal of Systems Science, 31:3–9, 2000. 86. D.J.N. Limebeer. The specification and purpose of a controller design study. In Proceedings of the 30th IEEE Conference on Decision and Control, pages 1579–1580, Brighton, UK, December 1991. 87. D.J.N. Limebeer, E.M. Kasenally, and J.D. Perkins. On the design of robust two degree of freedom controllers. Automatica, 29:157–168, 1993.

Robust Control of a Distillation Column

323

88. C.-A. Lin and T.-Y. Chiu. Model reduction via frequency weighted balanced realization. CONTROL – Theory and Advanced Technology, 8:341–351, 1992. 89. J.L. Lin. Control System Design for Robust Stability and Robust Performance. PhD thesis, Department of Engineering, University of Leicester, Leicester, UK, May 1992. 90. J.L. Lin, I. Postlethwaite, and D.W. Gu. µ -K iteration: a new algorithm for µ -synthesis. Automatica, 29:219–244, 1993. 91. R. Lind and M. Brenner. Robust Aeroservoelastic Stability Analysis. Flight Test Applications. Springer-Verlag, London, 1999. ISBN 1-85233-096-1. 92. A. Lindquist and G. Picci. On the stochastic realization problem. SIAM Journal of Control and Optimization, 17:365–389, 1979. 93. Yi Liu and B.D.O. Anderson. Singular perturbation of balanced systems. International Journal of Control, 50(4):1379–1405, 1989. 94. A. De Luca and B. Siciliano. Trajectory control of a non-linear one-link flexible arm. International Journal of Control, 50:1699–1715, 1989. 95. P. Lundstr¨om, S. Skogestad, and J.C. Doyle. Two-degree-of-freedom controller design for an ill-conditioned distillation process using µ -synthesis. IEEE Transactions on Control Systems Technology, 7:12–21, 1999. 96. J. Lunz. Robust Multivariable Feedback Control. Prentice Hall, London, 1989. 97. C.C. MacDuffee. The Theory of Matrices. Chelsea Publishing Co., New York, 1950. 98. J.M. Maciejowski. Multivariable Feedback Design. Addison-Wesley, Wokingham, England, 1989. ISBN 0-201-18243-2. 99. D.P. Magee. Optimal filtering to improve performance in hard disk drives: Simulation results. In Proceedings of the 1999 American Control Conference, pages 71–75, San Diego, CA, June 1999. 100. D. McFarlane and K. Glover. A loop shaping design procedure using H∞ synthesis. IEEE Transactions on Automatic Control, AC-37:749–769, 1992. 101. D.C. McFarlane and K. Glover. Robust Controller Design Using Normalized Coprime Factor Plant Descriptions, volume 138 of Lecture Notes in Control and Information Sciences. Springer-Verlag, 1990. 102. G.A. Medrano-Cerda. Robust stabilization of a triple inverted pendulum-cart. International Journal of Control, 68:849–865, 1997. 103. G.A. Medrano-Cerda. Robust computer control of an inverted pendulum. IEEE Control Systems Magazine, 19:58–67, 1999. 104. Z. Meier, H. Farwig, and H. Unbehauen. Discrete computer control of a triple-inverted pendulum. Optimal Control Applications & Methods, 11:157–171, 1990. 105. L. Meirovitch. Elements of Vibration Analysis. McGraw-Hill, New York, 1986. 106. W. Messner and R. Ehrlich. A tutorial on controls for disk drives. In Proceedings of the 2001 American Control Conference, pages 408–420, Arlington, VA, June 2001. 107. D.G. Meyer. Fractional balanced reduction: Model reduction via fractional representation. IEEE Transactions on Automatic Control, AC-35(3):1341–1345, 1990. 108. B. C. Moore. Principle component analysis in linear systems: Controllability, observability, and model reduction. IEEE Transactions on Automatic Control, AC-26:17–31, 1981. 109. M. Morari and E. Zafiriou. Robust Process Control. Prentice-Hall, Englewood Cliffs, NJ, 1989. 110. S. Mori, H. Nishihara, and K. Furuta. Control of an unstable mechanical system. Control of pendulum. International Journal of Control, 23:673–692, 1976.

324

D.-W. Gu

111. G. Murad. Robust Multivariable Control of Industrial Production Processes: A DiscreteTime Multi-Objective Approach. PhD thesis, Department of Engineering, University of Leicester, Leicester, UK, 1995. 112. G. Murad, D-W. Gu, and I. Postlethwaite. A Direct Model Reduction Approach for Discrete-Time Non-Minimal State-Space Systems. Internal Report 94-23, University of Leicester, Leicester, UK, September 1994. 113. G. Murad, D.-W. Gu, and I. Postlethwaite. Robust internal model control of a binary distillation column. In IEEE International Conference on Industrial Technology, pages 194–198, Shanghai, China, December 1996. 114. G. Murad, I. Postlethwaite, D.-W. Gu, and R. Samar. On the structure of an H ∞ twodegree-of-freedom internal model-based controller and its application to a glass tube production process. In Proceedings of the Third European Control Conference, pages 595–600, Rome, September 1995. 115. G. Murad, I. Postlethwaite, D.-W. Gu, and J.F. Whidborne. Robust control of a glass tube shaping process. In Proceedings of the Second European Control Conference, volume 4, pages 2350–2355, Gr¨oningen, The Netherlands, June-July 1993. 116. Z. Nehari. On bounded bilinear forms. Annals of Mathematics, 65:153–162, 1957. 117. W.Y. Ng. Interactive Multi-Objective Programming as a Framework for ComputerAided Control System Design. Lecture Notes in Control and Information Sciences. Springer-Verlag, 1989. 118. R.A. Nichols, R.T. Reichert, and W.J. Rugh. Gain scheduling for H∞ controllers: A flight control example. IEEE Trans. Control Systems Technology, 1:69–53, 2001. 119. International Society Niconet. SLICOT: Numerical Subroutine Library for Control and Systems Theory. Available at http://www.win.tue.nl/niconet, 2001. 120. A. Packard and J.C. Doyle. The complex structured singular value. Automatica, Special Issue on Robust Control, 29:71–109, 1993. 121. A. Packard, M.K.H. Fan, and J.C. Doyle. A power method for the structured singular value. In Proceedings of the 27th IEEE Conference on Decision and Control, pages 2132–2137, Austin, Texas, December 1988. 122. A. Packard and P. Pandey. Continuity properties of the real/complex structured singular value. IEEE Transactions on Automatic Control, AC-38:415–428, 1993. 123. L.Y. Pao and G.F. Franklin. Time-optimal control of flexible structures. In Proceedings of the 29th IEEE Conference on Decision and Control, pages 2580–2581, Honolulu, HI, December 1990. 124. T. Pappas, A.J. Laub, and N.R. Sandell. On the numerical solution of the discrete-time algebraic Riccati equation. IEEE Transactions on Automatic Control, AC-25:631–641, 1980. 125. J. Park and H. Asada. Dynamic analysis of noncollocated flexible arms and design of torque transmission mechanisms. ASME Journal of Dynamic Systems, Measurements and Control, 116:201–207, 1994. 126. L. Pernebo and L.M. Silverman. Model reduction via balanced state space representations. IEEE Transactions on Automatic Control, AC-27:382–387, 1982. 127. I. Postlethwaite, J.L. Lin, and D.W. Gu. Robust control of a high purity distillation column using µ -K iteration. In Proceedings of the 30th IEEE Conference on Decision and Control, Brighton, UK, December 1991. 128. R.M. Redheffer. Remarks on the basis of network theory. Journal of Mathematics and Physics, 28:237–258, 1950. 129. R.M. Redheffer. On a certain linear fractional transformation. Journal of Mathematics and Physics, 39:269–286, 1960.

Robust Control of a Distillation Column

325

130. R.T. Reichert. Dynamic scheduling of modern-robust-control autopilot design for missiles. IEEE Control Systems Magazine, 12:35–42, 1992. 131. W.J. Rugh and J.S. Shamma. Research on gain scheduling. Automatica, 36:1401–1425, 2000. 132. M.G. Safonov. Stability margins of diagonally perturbed multivariable feedback systems. IEE Proceedings - Part D, 129:251–256, 1982. 133. M.G. Safonov and R.Y. Chiang. A Schur method for balanced-truncation model reduction. IEEE Transactions on Automatic Control, AC-34:729–733, 1989. 134. M.G. Safonov, R.Y. Chiang, and D.J.N. Limebeer. Hankel model reduction without balancing – A descriptor approach. In Proceedings of the 26th IEEE Conference on Decision and Control, Los Angeles, CA, December 1987. 135. M.G. Safonov, D.J.N. Limebeer, and R.Y. Chiang. Simplifying the H ∞ theory via loop-shifting, matrix-pencil and descriptor concepts. International Journal of Control, 50:2467–2488, 1989. 136. R. Samar, I. Postlethwaite, and D.-W. Gu. Model reduction with balanced realizations. International Journal of Control, 62:33–64, 1995. 137. R.S. S´anchez-Pe˜na and M. Sznaier. Robust Systems. Theory and Applications. John Wiley & Sons, Inc., New York, 1998. 138. J. Sefton and K. Glover. Pole-zero cancellations in the general H∞ problem with reference to a two block design. Systems & Control Letters, 14:295–306, 1990. 139. R.S. Sezginer and M.L. Overton. The largest singular value of eX Ao e−X is convex on convex sets of commuting matrices. IEEE Transactions on Automatic Control, AC35:229–230, 1990. 140. S. Skogestad. Dynamics and control of distillation columns. A tutorial introduction. Trans. IChemE, 75:1–36, 1997. 141. S. Skogestad, M. Morari, and J.C. Doyle. Robust control of ill-conditioned plants: High purity distillation. IEEE Transactions on Automatic Control, 33:1092–1105, 1988. 142. S. Skogestad and I. Postlethwaite. Multivariable Feedback Control. John Wiley and Sons Ltd, Chichester, UK, 1996. 143. M.S. Spillman. Robust longitudinal flight control design using linear parameter-varying feedback. Journal of Guidance, Control, and Dynamics, 23:101–108, 2000. 144. V. Sreeram, B.D.O. Anderson, and A.G. Madievski. New results on frequency weighted balanced reduction technique. In Proceedings of the 1995 American Control Conference, pages 4004–4009, Seattle, WA, June 1995. 145. A.A. Stoorvogel. The H∞ Control Problem: A State Space Approach. Prentice Hall, Englewood Cliffs, NJ, 1992. 146. R. Sutton, G. Halikias, A. Plummer, and D. Wilson. Modelling and H∞ control of a single-link flexible manipulator. Proceedings of the Institution of Mechanical Engineers, Part I, Journal of Systems and Control Engineering, 213:85–104, 1999. 147. K.S. Tang, K.F. Man, and D.-W. Gu. Structured Genetic Algorithm for Robust H∞ Control Systems Design. IEEE Transactions on Industrial Electronics, 43(5):575– 582, 1996. 148. M.S. Tombs and I. Postlethwaite. Truncated balanced realization of a stable non-minimal state space system. International Journal of Control, 46(4):1319–1330, 1987. 149. V.A. Tsacouridis and G.A. Medrano-Cerda. Discrete-time H∞ control of a triple inverted pendulum with single control input. IEE Proceedings-Control Theory Appl., 146:567– 577, 1999. 150. G.-W. van der Linden and P.F. Lambrechts. H∞ control of an experimental inverted pendulum with dry friction. IEEE Control Systems Magazine, 19:44–50, 1993.

326

D.-W. Gu

151. A. Varga. Balancing free square-root algorithm for computing singular perturbation approximations. In Proceedings of the 30th IEEE Conference on Decision and Control, pages 1062–1065, Brighton, UK, December 1991. 152. A. Varga. On stochastic balancing related model reduction. In Proceedings of the 39th IEEE Conference on Decision and Control, pages 2385–2390, Sydney, Australia, December 2000. 153. M. Vidyasagar and H. Kimura. Robust controllers for uncertain linear multivariable systems. Automatica, 22:85–94, 1986. 154. M. Vidyasagar, H. Schneider, and B. Francis. Algebraic and topological aspects of feedback stabilization. IEEE Transactions on Automatic Control, AC-27:880–894, 1982. 155. D.J. Walker. Relationship between three discrete-time H∞ algebraic Riccati equation solutions. International Journal of Control, 52:801–809, 1990. 156. F. Wang, T. Hurst, D. Abramovitch, and G. Franklin. Disk drive pivot nonlinearity modelling Part II: Time domain. In Proceedings of the 1994 American Control Conference, pages 2604–2607, Baltimore, MD, June 1994. 157. G. Wang, V. Sreeram, and W.Q. Liu. A new frequency-weighted balanced truncation method and an error bound. IEEE Transactions on Automatic Control, 44:1734–1737, 1999. 158. J.F. Whidborne, D.-W. Gu, and I. Postlethwaite. Algorithms for solving the method of inequalities – A comparative study. In Proceedings of the 1995 American Control Conference, June 1995. 159. J.F. Whidborne and G.P. Liu. Critical Control Systems. Research Studies Press Ltd, Taunton, Somerset, UK, 1993. 160. J.F. Whidborne, G. Murad, D.-W. Gu, and I. Postlethwaite. Robust control of an unknown plant – the IFAC 93 benchmark. International Journal of Control, 61:589–640, 1994. 161. J.F. Whidborne, I. Postlethwaite, and D.-W. Gu. Robust controller design using H ∞ loop-shaping and the method of inequalities. IEEE Transactions on Control Systems Technology, 2:455–461, 1994. 162. J.F. Whidborne, I. Postlethwaite, and D.-W. Gu. Multiobjective control system design – a mixed optimization approach. In R.P. Agarwal, editor, Recent Trends in Optimization Theory and Applications, volume 5 of World Scientific Series in Applicable Analysis, pages 467–482. World Scientific, 1995. 163. J.F. Whidborne, I. Postlethwaite, and D.-W. Gu. Robust control of a paper machine. Control Engineering Practice, 3:1475–1478, 1995. 164. M.T. White, M. Tomizuka, and C. Smith. Rejection of disk drive vibration and shock disturbances with a disturbance observer. In Proceedings of the 1999 American Control Conference, pages 4127–4131, San Diego, CA, June 1999. 165. W.N. White and R.C. Fales. Control of a double inverted pendulum with hydraulic actuation: A case study. In Proceedings of the 1999 American Control Conference, pages 495–499, San Diego, CA, June 1999. 166. M. Yamakita, T. Hoshino, and K. Furuta. Control practice using pendulum. In Proceedings of the 1999 American Control Conference, pages 490–494, San Diego, CA, June 1999. 167. D.C. Youla, H.A. Jabr, and J.J. Bongiorno. Modern Weiner-Hopf design of optimal controllers, Part ii: The multivariable case. IEEE Transactions on Automatic Control, AC-21:319–338, 1976.

Robust Control of a Distillation Column

327

168. D.C. Youla, H.A. Jabr, and C.N. Lu. Single-loop feedback stabilization of linear multivariable dynamical plants. Automatica, 10:159–173, 1974. 169. V. Zakian and U. Al-Naib. Design of dynamical and control systems by the method of inequalities. IEE Proceedings - Control & Science, 120:1421–1427, 1973. 170. G. Zames. Feedback and optimal sensitivity: Model reference transformations, multiplicative seminorms and approximate inverses. IEEE Transactions on Automatic Control, AC-26:301–320, 1981. 171. G. Zames and B.A. Francis. Feedback, minimax sensitivity, and optimal robustness. IEEE Transactions on Automatic Control, AC-28:585–600, 1983. 172. P. Zarchan. Tactical and Strategic Missile Guidance, volume 176 of Progress in Astronautics and Aeronautics. AIAA, Washington, DC, 3rd edition, 1998. 173. H.P. Zeiger and A.J. McEwen. Approximate linear realization of given dimension via Ho’s algorithm. IEEE Transactions on Automatic Control, AC-19:153, 1974. 174. K. Zhou and J.C. Doyle. Essentials of Robust Control. Prentice Hall, Upper Saddle River, NJ, 1998.

12 Robust Control of a Hard-D isk Drive Da-Wei Gu

Summary. This chapter considers the design of a robust servo system of hard disk drive(HDD). Three robust design methods are applied, namely, the µ -synthesis, H∞ optimal design and H∞ loop-shaping design procedure (LSDP). After a description of the HDD servo system dynamics in the first section, it is shown in details (in Section 12.2) how to derive the plant model which involves several uncertain parameters. Then we consider the synthesis of continuoustime controllers using the available methods for robust control design. These controllers are compared in aspects of robustness of closed-loop system stability and of performance in the frequency domain and in the time-domain. The design of discrete-time controller is included as well, for two sampling frequencies. Finally, we present the simulation results of the nonlinear continuous-time and discrete-time closed-loop systems, followed by conclusions at the end of the chapter. The prevalent trend in hard disks design is towards smaller disks with increasing larger capacities. This implies that the track width has to be smaller leading to lower error tolerance in the positioning of the read/write heads. The controller for track following has to achieve tighter regulation in the presence of parameter variations, nonlinearities and noises. Hence it is appropriate to use advanced design methods like µ synthesis and H∞ optimization in order to achieve robust stability and robust performance of the closed-loop servo system.

12.1 Hard Disk Drive Servo System The schematic diagram of a typical hard disk drive is shown in Figure 12.1. The disk assembly consists of several flat disks called platters coated on both sides with very thin layer of magnetic material (thin film media). The magnetic material is used to store the data in the form of magnetic patterns. The platters rotate at high speed, driven by a spindle motor. Recently, the spindle speed is 5400 RPM, 7200 RPM or even 10,000 and 15,000 RPM. The data are retrieved from, or recorded onto, the platters by electromagnetic read/write(R/W)

†

This is based in part on Chapter 10 of Robust Control Design with Matlab by D.W. Gu, M. Konstatinov and P. Petkov, Springer, 2005.

M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 329-372, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

330

D. -W. Gu

Spindle Motor

Arm

Voice Coil Motor (VCM) Read/Write Heads

Platter

Fig. 12.1. Schematic diagram of a hard disk drive

heads which are mounted at the bottom of sliders. Today’s hard disks read data using giantmagneto-resistive(GMR) heads and write data with thin film inductive heads. The sliders with read/write heads are mounted onto head arms. The arms are lightweight rigid constructions allowing to be moved rapidly on the platter surface. There is one arm per read/write head and all arms are mounted in a stack assembly that moves over multiple disk surfaces simultaneously. The heads are suspended several micro inches above the disk surface. The appropriate flying height of the heads is achieved thanks to the air flow generated by the spinning disk. The data recorded on the platters are in concentric circles called tracks. Modern hard disks have tens of thousands of tracks resulting in track density as high as 30000 tracks per inch(TPI). Thus the distance between adjacent tracks is of the order of a micro inch. Each track is divided into smaller pieces called sectors which contain 512 bytes of information. There may be several hundreds or even thousands of sectors in a track. The drive density may reach more than 20 GB per platter. The head arms are moved on the surface of the platter by a rotary voice coil actuator frequently called Voice Coil Motor(VCM). The VCM consists of a voice coil, mounted at the end of the head arm assembly, and permanent electromagnets. By controlling the current in the coil, the heads can move in one direction or the other in order to follow precisely the data track. The goal of the hard disk drive servo control system is to achieve a precise positioning of the read/write heads on the desired track (track following mode) while data are being written or read, and a quick transition from one track to another target track (seeking mode). As the

Robust Control of a Hard-Disk Drive

331

drive initiates a seek command, it switches to a seek control algorithm which is a type of timeoptimal (bang-bang) control algorithm. When the head is positioned over the desired track, the drive switches into track following mode. In the present work we consider the servo controller of the track following mode. All modern hard disk drives read the relative position of the head to track directly from the disk media by using a method called embedded servo or sector servo. In this method the servo information is interleaved with user data across the entire surface of all platters. Position information is placed on the disk surfaces in servo frames during the manufacturing of the disk and cannot be rewritten. That is why the disk heads are locked out by the drive controller from writing to the areas where servo information is written. Because of the interleaving of servo information with data, the embedded servo system is a sampled-data system. Increasing the number of servo frames within a track improves the performance of the servo system due to the higher sample rate but limits the maximum data storage. The servo information is read by the same head which reads the data. The position information usually consists of two parts: coarse position giving track number and fine position information relative to each track. The position error, representing the difference between the reference track position and the head position, is measured by making use of position bursts which are part of the servo information. The position bursts are patterns of alternating magnetic polarity written on the disk surface with a particular frequency. The periodic signal, obtained from the read head passing over a burst pattern, has amplitude which is proportional to how much the read head is directly over the burst pattern. The signals corresponding to 2 or 4 bursts of position information are demodulated by a servo demodulator in order to compose the position error signal(PES). This signal is used by the servo controller to change appropriately the voice coil current and hence the read/write heads position. It should be pointed out that the absolute head position is not generally known from the servo information which is read off of the disk. Only a signal proportional to the position error is available, which means that for track following the reference signal is ideally equal to zero. The failure of the head to follow the track on a platter surface faithfully as the disk spins is referred to as runout. The runout could have serious consequences, especially during writing where data in adjacent tracks might be overwritten. There are two types of runout, namely the repeatable runout(RRO) and non-repeatable runout(NRRO). Repeatable runout is caused by both imbalance in the spindle and imperfections in the servowriting process which result in non-circular position information. This information is encoded at the spindle frequency, yielding a repeatable non-circular track for the disk servo to follow. The non-repeatable runout is caused by windage induced actuator arm, slider motion and mechanical vibrations which can arise from various sources such as ball bearing defects, spindle motor vibrations and slider vibrations. The RRO may be reduced through feedforward compensation at the corresponding frequency by using harmonic correctors. The block-diagram of the HDD servo system is shown in Figure 12.2. The R/W heads are moved by a VCM which is driven by the output current ic of a power amplifier(PA). The actual position signal y is compared with the signal yre f which represents the desired head position. For the track following, the reference input yre f is theoretically equal to zero and y(t) appears as error signal. In practice, yre f must be set equal to the signal representing both RRO and NRRO. The digital signal yr is a reference for the desired track and is used during the seeking mode. The error signal is sampled by an analog-to-digital(A/D) converter and serves as part of an input to the digital servo controller Kd which is typically implemented on a DSP chip. The output of the controller is converted to analog form by a digital-to-analog(D/A) converter and amplified by the PA. Since the motor torque is proportional to the voice coil current, the

332

D.-W. Gu

kb −

VC

PA

u D/A

ic

kt

+

Kd +

yr

tm

A/D

td ta

Hd(s)

th

+

1

yref

+

Θ

s

+ y

−

S/H

ω

1

Js

R.tpm.ky

η

Fig. 12.2. Block diagram of the hard disk drive servo system

amplifier is configured as a current source. The exogenous signal td is the torque disturbance due to external shock and vibrations, power amplifier noise, digital-to-analog converter noise, pivot bearing friction and flex cable bias. The increase of the spindle speed increases the air flow inside the disk (windage) which in turn increases the disturbance torque on the actuator. The disturbance is a low frequency signal with spectral content usually below 500 Hz. The position noise signal η includes quantization errors due to servo demodulator noise, finite resolution of analog-to-digital converter, media noise and preamplifier noise. The position noise is a high frequency signal with spectral content usually above 1 kHz. Since the measured PES is contaminated with noise, the true PES, yre f − y, is not available. One of the limitations inherent in the design of servo controllers for high track density HDD is the influence of actuator mechanical resonant modes on the head-positioning servo. If the actuator input contains a periodic component with frequency equal to a resonance frequency, this component may be amplified greatly which results in large off-track deviation of the read/write heads. Usually, the actuator is mechanically designed in such a way that the resonant modes occur at frequencies that will be attenuated by the servo system. However, as servo bandwidth increases to meet higher performance requirements, this attenuation may not be achieved due to mechanical design constraints. It is also important to note that the presence of resonant modes may limit the servo bandwidth via stability margin constraints. With a reduced bandwidth, the servo system may not be able to achieve the desired performance. The rotary actuators of hard disks may have tens of resonances which may lead to a high-order model. However, in practice only 3 to 4 main resonances are taken into account. Usually, these are the first and second torsion modes (in the range of 1500−2500 Hz) as well as the first sway mode (in the range of 8000 − 12000 Hz). A common approach to reduce the effect of resonance modes is to put notch filters in the servo-loop which attenuates or filters out vibrations at selected major resonant frequencies. However, each notch filter introduces phase margin loss at low frequencies thus reducing the system robustness. Also, the presence of uncertainty in the resonant modes may decrease significantly the efficiency of those filters. Our goal in this chapter is to design a robust track following servo control system for a 3.5 inch HDD with track density 25400 TPI. The desired settling time is about 1 ms in the presence of four resonances, several uncertain parameters, position sensing noise and disturbances. The

Robust Control of a Hard-Disk Drive

333

Table 12.1. Rigid body model parameters and tolerances Para- Description meter

Value

arm moment of inertia arm length amplifier gain VCM torque constant back e.m.f. constant tracks per meter position measurement gain coil resistance sense resistance in the power amplifier feedback Lcoil coil inductance emax saturated power amplifier voltage RPM disk rotation rate tw track width J R kPA kt kb t pm ky Rcoil Rs

6.3857 × 10−6 5.08 × 10−2 10.0 9.183 × 10−2 9.183 × 10−2 106 1.2 8.00 0.2

Units

kgm2 m V/V Nm/A Nm/A tracks V/track Ω Ω

0.001 12.0

H V

7200 1

rev/min µm

Tolerance ±10.0% ±0.1% ±0.0% ±10.0% ±10.0% – ±5.0% ±20.0% ±1.0% +0, −15% −0, +5% ±1.0% ±1.0%

parameters of the rigid body model and their tolerances are given in Table 12.1. For dimensions compatibility, the track density is given in tracks per meter, instead of in TPI. We now first consider the derivation of a HDD servo system model. The dynamics of the rotary arm is described by the equation J

d 2Θ = tm + td , dt 2

where J is the arm moment of inertia, Θ is the angle of arm rotation, tm is the VCM torque and td is the disturbance torque. The VCM torque is given by tm = kt ic , where kt is the motor torque constant and ic is the current through the VCM coil. The voice coil has a resistance Rcoil and an inductance Lcoil . An additional current sense resistance Rs is connected in serial with the voice coil to implement a feedback from the power amplifier output. Hence the voice coil admittance is described by the transfer function Gvca (s) =

1/Rc ic (s) = , ec (s) τ s + 1

where ec is the input voltage to the voice coil, τ = Lcoil /Rc and Rc = Rcoil + Rs . The block diagram of the power amplifier with voice coil is shown in Figure 12.3. The input of the voice coil is the difference ec = e p − eb , where e p is the output voltage of the amplifier and eb = kb ω is the back electro motive force (e.m.f.) which is generated during the moving

D.-W. Gu

334

ω kb eb

+ emax

u

kPA −

− ec

ep − e−max

1 1 RC τ s + 1

ic

RS Fig. 12.3. Block diagram of the power amplifier with voice coil

of the coil in the magnetic field. Since the saturation voltage of the amplifier is emax , in the absence of back e.m.f., the amplifier will saturate for an input voltage greater than emax /kPA . In the study limited to linear systems, the amplifier saturation is neglected. The length of the arc, corresponding to the arm rotation angle Θ , is equal to RΘ . For small values of Θ the number of tracks contained in the arc is R · Θ ·t pm. This gives an output signal y = R · t pm · ky · Θ . If we neglect the dynamics of the voice coil, the transfer function of the servo actuator considered as a rigid body is obtained in the form of a double integrator. Such a model oversimplifies the system dynamics and cannot produce reliable results if used in the design. The next step is to take into account the high frequency resonant modes of the head disk assembly represented by the transfer function Hd (s). In the given case Hd (s) consists of four resonant modes and is obtained as 4

Hd (s) =

∑

j=1

b2 j ω j s + b2 j−1 ω 2j s2 + 2ξ j ω j s + ω 2j

(see Figure 12.4). Here ω j , ξ j and b2 j , b2 j−1 are respectively the resonance frequency, the damping coefficient and the coupling coefficients of the j-th mode, for j = 1, ..., 4. The resonance parameters are usually determined experimentally and for the servo system under consideration their values are shown in Table 12.2. It is important to note that all model parameters are known with some tolerances and may vary with the changing of working conditions as well as with the time. Also, the closed-loop system can be very sensitive to the external disturbance td and the position sensing noise η . Both factors would lead to an actual system dynamics far away from the dynamics of the nominal closed-loop system. Thus it is necessary to use control system design methods which ensure the desired closed-loop stability and performance in the presence of uncertain parameters, noises and disturbances.

Robust Control of a Hard-Disk Drive

335

Fig. 12.4. Transfer functions of resonant modes Table 12.2. Resonance parameters and tolerances Parameter ω1 ω2 ω3 ω4 b1 b2 b3 b4 b5 b6 b7 b8 ξ1 ξ2 ξ3 ξ4

Description pivot bearing resonance first torsional resonance second torsional resonance first sway resonance first resonance coupling first resonance coupling second resonance coupling second resonance coupling third resonance coupling third resonance coupling fourth resonance coupling fourth resonance coupling first resonance damping second resonance damping third resonance damping fourth resonance damping

Value 2π 50 2π 2200 2π 6400 2π 8800 0.006 0 0.013 −0.0018 0.723 −0.0015 0.235 −0.0263 0.05 0.024 0.129 0.173

Units rad/s rad/s rad/s rad/s – 1/s – 1/s – 1/s – 1/s – – – –

Tolerance ±5.0% ±12.0% ±8.0% ±15.0% ±7.0% ±7.0% ±10.0% ±7.0% ±5.0% ±10.0% ±5.0% ±10.0% ±5.0% ±8.0% ±10.0% ±10.0%

12.2 Derivation of Uncertainty Model In order to implement robust control design methods we must have a plant model which incorporates uncertain parameters. As it is seen from Tables 12.1 and 12.2, the total number of uncertain parameters is greater than 25, which complicates very much the analysis and design of a HDD servo system. In this study we shall concentrate on those uncertainty parameters which influence the closed-loop system behaviour most. The block diagram of the plant is shown in Figure 12.5. Consider first the derivation of the uncertainty model for the resonant modes. All the four modes have similar transfer functions. A state space model is

336

D.-W. Gu

kb _

u

Gvca(s)

kPA

_

ic

kt

td tm +ta

ω

Hd(s)

th

1

R.tpm

Js

s

ky

y

Rs Fig. 12.5. Block diagram of the plant x˙1 = ω x2 , x˙2 = ω (−x1 − 2ξ x2 + ta ), ya = b1 x1 + b2 x2

ta −

+

ω

. x2

x2

ω

. x1

x1

-2 ξ

Fig. 12.6. Block diagram of a resonant mode

The block diagram corresponding to the state equations of a resonant mode is shown in Figure 12.6. The variations in the frequency ω and the damping coefficient ξ are represented, respectively, by

ω = ω (1 + pω δω ) and

ξ = ξ (1 + pξ δξ ), where ω and ξ are the nominal values, pω and pξ are the maximum relative uncertainties with −1 ≤ δω ≤ 1, −1 ≤ δξ ≤ 1

being the relative variations in these parameters.

The parameter ω may be represented as an upper linear fractional transformation (LFT) in δω ,

ω = FU (Mω , δω ) with Mω =

0 ω pω ω

Robust Control of a Hard-Disk Drive

337

and the parameter ξ may be represented similarly as

ξ = FU (Mξ , δξ ) with Mξ =

uω ta − +

ω

Mω

# 0 ξ . pξ ξ

yω . x2

yξ -2

"

vω x2

ξ

ω

Mω

zω . x1

x1

uξ

Mξ

vξ

Fig. 12.7. Resonant mode with uncertain parameters

The block diagram of a resonant mode with uncertain parameters is given in Figure 12.7. Based on this diagram we derive the following equations uω 0 ω yω = , ta − 2vξ − x1 x˙2 pω ω

vω , x2 yξ uξ 0 ξ = , vξ x2 pω ξ

and

zω x˙1

=

0 ω pω ω

x1 ya = b1 b2 . x2

From these equations we further obtain the perturbed model of a resonant mode in the form of     x1 x˙1  x2   x˙2       −−   −−        yω     = Π  uω   vω   zω       uξ   yξ       −−   −−  ya ta

where

338

D.-W. Gu



0  −ω   −−   −ω   0   0   −− b1

ω −2ωξ −−− −2ωξ ω ξ −−− b2

Π= | 0 pω | pω 0 − −−− −−− | 0 0 | 0 0 | 0 0 − −−− −−− | 0 0

0 −2ω pξ −−− −2ω pξ 0 0 −−− 0

| | − | | | − |

 0 ω   −−   ω  . 0   0   −−  0

yω zω yξ ya

uω vω uξ ta

Fig. 12.8. Block representation of a resonant mode

A block with four inputs and four outputs can be used to represent a resonant mode (Figure 12.8). We now consider other uncertain parameters of the plant. The uncertain parameters of the rigid body model are taken as kt = kt (1 + pkt δkt ), J = J(1 + pJ δJ ) ky = ky (1 + pky δky ) where kt , J, ky are the corresponding nominal parameters and −1 ≤ δkt ≤ 1, −1 ≤ δJ ≤ 1.

−1 ≤ δky ≤ 1,

are variations. These parameters are represented as LFTs in the related uncertainties δkt , δJ , δky: kt = FU (Mkt , δkt ), J = FU (MJi , δJ ), ky = FU (Mky , δky ), where Mkt =

" −pJ 0 kt , MJi = pkt kt −pJ

1 J 1 J

#

, Mky =

0 ky . pky ky

Robust Control of a Hard-Disk Drive

uω1 vω1 u ξ1 ta

yω1 zω1

1

uω2 vω2 u ξ2 ta ukt ykt ic Mkt tm

td ta

339

yξ1 ya1 yω2 zω2

2

uω3 vω3 u ξ3 ta

yξ2 ya2 yω3 zω3

3

th

yξ3 ya3 yω4 zω4

uω4 vω4 u ξ4 ta

4

yξ4 ya4

Fig. 12.9. Representations of uncertain motor torque constant and resonant modes

In Figure 12.9 we show the part of the uncertainty model involving the motor torque constant and the four resonant modes and in Figure 12.10 we show the part of the model involving the arm moment of inertia and position measurement gain.

uJ

th

J

MJi

yJ

uky ω

Θ

R.tpm

ky

Mky

yky y

Fig. 12.10. Block diagram of uncertain arm moment of inertia and position gain Overall, we have twelve uncertain parameters but four of them (ω1 , ω2 , ω3 and ω4 ) are repeated themselves twice. By “puling out” the uncertain parameters from the nominal part of the model, we obtain a perturbed plant model in the form of an upper LFT FU (Gnom , ∆ ) as shown in Figure 12.11 with a 15 × 15 matrix ∆ ,

∆ = diag(δω1 , δω1 , δξ1 , δω2 , δω2 , δξ2 , δω3 , δω3 , δξ3 , δω4 , δω4 , δξ4 , δkt , δJ , δky ), which contains all the uncertain parameters. In the given case the matrix Gnom is obtained by the function sysic using the interconnections shown in Figures 12.5, 12.7, 12.9 and 12.10. The system model is of order 11. The uncertainty model of the HDD servo system is implemented by the M-file mod hdd.m.

340

D.-W. Gu

∆

Gnom

u

y

Fig. 12.11. Plant model in the form of an upper LFT

12.3 Closed-Loop System Design Specifications The design of the HDD servo controller will be first conducted in the continuous-time case. In general it may obtain best possible performance in the continuous-time case which can then be considered as a limit for discrete-time designs. Also, in the continuous-time case it is easier to find appropriate performance weighting functions which again may be implemented in the discrete-time design.

Wu

eu

∆ z

w

r

+

K

u

+

+

d y

Gnom

+

Wp

ey

+ +

Wn

n

M Fig. 12.12. Block diagram of the closed-loop system with performance specifications

The block diagram of the closed-loop system, which includes the feedback structure and the controller as well as the elements representing the model uncertainty and the performance objectives, is shown in Figure 12.12.

Robust Control of a Hard-Disk Drive

341

The system has a reference input (r), an input disturbance (d), a noise (n) and two output costs (ey and eu ). The system M is an ideal model of performance, to which the designed closedloop system tries to match. The rectangle, shown with dash line, represents the plant transfer function matrix G. Inside the rectangle is the nominal model Gnom of the Hard Disk Drive plant and the block ∆ which parameterizes the model uncertainties. The matrix ∆ is unknown, but has a diagonal structure and norm bound k∆ k∞ < 1. The performance objective requires the transfer matrices from r, d and n to ey and eu to be small in the sense of k · k∞ , for all uncertain matrices ∆ under consideration. The position noise signal is obtained by passing the unit-bounded signal n through the weighing transfer matrix Wn . The transfer matrices Wp and Wu are used to reflect the relative significance of the different frequency ranges for which the performance is required. Hence, the performance objective can be recast, with possible slight conservativeness, as that the k · k∞ of the transfer function matrix from r, d and n to ey and eu is less than 1. It is straightforward to show that   r ey Wp (So GK − M) Wp So G −Wp So GKWn   d = (12.1) eu Wu Si K −Wu KSo G −Wu KSoWn n where Si = (I + KG)−1 , So = (I + GK)−1 are the input and output sensitivities, respectively. Note that So G is the transfer function between d and y. This objective is similar to the usual mixed S/KS sensitivity optimization and it would meet both robust stability and performance criteria by incorporating performance specifications in the matching model M. The six functions to be minimized are described in Table 12.3.

Table 12.3. H∞ functions to be minimized Function Description Wp (So GK − M) Weighted difference between the ideal and actual closed-loop systems Wp So G Weighted disturbance sensitivity Wp So GKWn Weighted noise sensitivity Wu Si K Weighted control effort due to reference Wu KSo G Weighted control effort due to disturbance Wu KSoWn Weighted control effort due to noise

The controller synthesis problem of the Hard Disk Drive Servo System is to find a linear, output feedback controller K(s) which has to ensure the following properties of the closedloop system.

Nominal performance The closed-loop system achieves nominal performance if the performance objective is satisfied for the nominal plant model. The nominal performance objective is to satisfy the inequality

342

D.-W. Gu

Wp (So Gnom K − M) Wp So Gnom −Wp So Gnom KWn

<1

Wu Si K −Wu KSo Gnom −Wu KSoWn ∞

(12.2)

Robust stability

The closed-loop system achieves robust stability if the closed-loop system is internally stable for each possible plant dynamics G = FU (Gnom , ∆ ).

Robust performance The closed-loop system must remain internally stable for each G = FU (Gnom , ∆ ) and in addition the performance criterion

Wp (So GK − M) Wp So G −Wp So GKWn

<1 (12.3)

Wu Si K −Wu KSo G −Wu KSoWn ∞

should be satisfied for each G = FU (Gnom , ∆ ). Thismeans  that the structured singular value,   w z r    corresponding to the transfer function matrix from   d  to ey (in (Figure 12.12) should eu n ∆ 0 be less than 1, with regard to where ∆F is a fictitious 3 × 2 complex uncertainty 0 ∆F block. In addition to those requirements it is desirable that the controller designed would have acceptable complexity, i.e., it is of reasonably low order. According to the above considerations, the aim of the design is to determine a controller K, such that for all stable perturbations ∆ with k∆ k∞ < 1, the perturbed closed-loop system remains stable and the performance objective is satisfied for all such perturbations.

12.4 System Interconnections The internal structure of the eighteen-input, seventeen-output system, which is saved in the variable sys ic, is shown in Figure 12.13. The inputs and outputs of the uncertainties are saved in the variables pertin and pertout, the reference, the disturbance and the noise in the variables ref, dist, noise, respectively, and the control signal in the variable control. Both variables pertin and pertout have fifteen elements and ref, dist, noise, y, y c, e y and e u are scalar variables. The open-loop connection is obtained by the M-file olp hdd. The schematic diagram showing the specific input/output ordering for the variable sys ic is shown in Figure 12.14. The block-diagram used in the simulation of the closed-loop system is shown in Figure 12.15. The corresponding closed-loop interconnection, which is saved in the variable sim ic, is obtained by the M-file sim hdd. The schematic diagram showing the specific input/output ordering for the variable sim ic is shown in Figure 12.16.

Robust Control of a Hard-Disk Drive pertout {1-15}

pertin {1-15}

dist

+

Gnom

−

y

e_y

Wp +

+

control

e_u

Wu +

y_c

Wn

noise

M

−

ref

+

Fig. 12.13. Block diagram of the open-loop system with performance specifications

pertin{1} pertin{2} pertin{3} pertin{4}

pertin{5} pertin{6} pertin{7} pertin{8} pertin{9} pertin{10} pertin{11} pertin{12} pertin{13} pertin{14} pertin{15} ref dist noise

control

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

sys_ic

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

pertout{1} pertout{2}

pertout{3} pertout{4} pertout{5} pertout{6} pertout{7} pertout{8} pertout{9} pertout{10} pertout{11} pertout{12} pertout{13} pertout{14} pertout{15} ey eu yc

Fig. 12.14. Schematic diagram of the open-loop connection

12.5 Controller Design in Continuous Time There are a few further “hard” constraints for the controller design, as listed below.

343

344

D.-W. Gu

pertout {1-15}

pertin {1-15}

dist

Gnom

+

y

+

control

+

y_c

K

Wn

noise

−

+

ref

Fig. 12.15. Block diagram of the closed-loop system Peak closed-loop gain Open-loop gain Steady state error Settling time Closed-loop bandwidth Gain Margin Phase margin

< > < < > > >

4 dB 20 dB at 100 Hz 0.1 µ m 1.5 ms 1000 Hz 5 dB 40 deg

The designed control system must achieve good disturbance rejection and noise attenuation. In addition, it is necessary to have control action smaller than 1.2 V in order to avoid the power amplifier saturation. To design the controller we shall use µ -synthesis, H∞ optimization and H∞ loop-shaping design procedure (LSDP) in this exercise. In the case of µ -synthesis and H∞ optimization design, we have to specify the model transfer function M and the weighting transfer functions Wn , Wp and Wu . The model transfer function is chosen so that the time response to the reference signal would have an overshoot less than 20 % and a settling time less than 1 ms. A possible model satisfying the requirements is 1 . M= 3.75 × 10−9 s2 + 1.2 × 10−4 s + 1 The noise shaping function Wn is determined on the basis of the spectral density of the position noise signal. In the given case it is taken as the high pass filter Wn = 6 × 10−4

0.1s + 1 0.001s + 1

Robust Control of a Hard-Disk Drive

pertin{1} pertin{2} pertin{3} pertin{4}

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

pertin{5} pertin{6} pertin{7} pertin{8} pertin{9} pertin{10} pertin{11} pertin{12} pertin{13} pertin{14} pertin{15} ref dist noise

control

sym_ic

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

345

pertout{1} pertout{2}

pertout{3} pertout{4} pertout{5} pertout{6} pertout{7} pertout{8} pertout{9} pertout{10} pertout{11} pertout{12} pertout{13} pertout{14} pertout{15} y control yc

Fig. 12.16. Schematic diagram of the closed-loop connection whose output has a significant spectral content above 500 Hz. For this shaping filter, the position noise signal is only 0.6 mV in the low frequency range but it is 60 mV in the high frequency range which corresponds to a position error of 5 % of the track width. The weighting functions Wp and Wu have to be chosen so as to ensure an acceptable tradeoff between the nominal performance and the robust performance of the closed-loop system. They are selected in the course of the µ -synthesis, since this particular design method allows to achieve maximum performance of the perturbed, closed-loop system.

12.5.1 µ -Design In the µ -synthesis it is necessary to specify individually the inputs and outputs of the uncertainty blocks. In this exercise only the inputs and outputs of the uncertainty in the rigid body model (i.e., the parameters kt , J and ky ) will be considered. The inclusion of the uncertainties of resonant modes would make the D − K iterations difficult to converge. This confirms that the resonant modes may create difficulties in the controller design. However, those resonant modes (with nominal values) are included in the plant dynamics and will be considered, with parametric variations, in the assessment of designed controllers in Section 12.6. The block diagram of the closed-loop system used in the µ synthesis is shown in Figure 12.17. Denote by P(s) the transfer matrix of the seven-input, six-output, open-loop system

346

D.-W. Gu

u kt uJ uky r d n

ykt yJ yky ey eu

nominal_dk

K

u

yc

Fig. 12.17. Block diagram of the closed-loop system with µ controller

nominal dk and let the block structure of the uncertainty ∆P be defined by ∆r 0 : ∆r ∈ R3×3 , ∆F ∈ C3×2 ∆P := 0 ∆F The first block of this uncertainty matrix corresponds to the block ∆r containing the uncertainties in the rigid body model. The second block ∆F is a fictitious uncertainty block which is used for the performance requirements in the framework of the µ approach. In order to satisfy the robust performance requirements it is necessary to find a stabilizing controller K(s), such that for each frequency ω ∈ [0, ∞] the structured singular value satisfies the condition

µ∆P [FL (P, K)( jω )] < 1. The fulfillment of this condition guarantees robust performance of the closed-loop system, i.e.

Wp (So GK − M) Wp So G −Wp So GKWn

< 1.

(12.4)

Wu Si K −Wu KSo G −Wu KSoWn ∞

The µ -synthesis is conducted by using the M-file ms hdd along with the file dk hdd describing the interconnection of the design. It should note that the robust performance achieved during the D − K iteration is only with respect to the uncertainties in the rigid body model, since only these uncertainties are taken into account in the design. Hence it is necessary to make additional robust stability and robust performance analysis which takes into account the other uncertainties. The closed-loop system performance specifications are reflected by the weighting performance function Wp (s). Three performance weighting functions are considered in the design. They are s2 + 8 × 104 s + 108 , Wp1 (s) = 10−4 2 s + 7 × 104 s + 2.5 × 104 Wp2 (s) = 10−4 and Wp3 (s) = 10−4

s2 + 4 × 105 s + 2.5 × 109

s2 + 3.9 × 105 s + 6.25 × 105

,

s2 + 1.15 × 106 s + 3.6 × 1010 . s2 + 1.05 × 106 s + 9 × 106

Robust Control of a Hard-Disk Drive

347

Inverse Performance Weighting Functions 80

70

60

Magnitude (dB)

50

40

30

20

W−1 p1 W−1 p2 W−1

10

p3

0 −4 10

−2

10

0

2

10

10

4

10

6

10

Frequency (Hz)

Fig. 12.18. Frequency response of the inverse of Wp

In Figure 12.18 we show the frequency responses of the inverses of those three weighting −1 −1 −1 functions, i.e., Wp1 , Wp2 and Wp3 . It is seen that in all three selections, the aim is to achieve small difference between the system and model outputs, and small effect of the disturbance on the system outputs. This will ensure good tracking of the reference input and small error due to low-frequency disturbances. Changing the performance weighting function from Wp1 to Wp3 moves the inverse weighting frequency response to the right (to higher frequencies) which forces the system to match the model in a larger frequency range. The control weighting function is usually chosen as high pass filters in order to ensure that the control action will not exceed 1.2 V . Again, three such weighting functions are considered in the design and listed below: Wu1 (s) = 10−6 Wu2 (s) = 10−6 and

0.385s2 + s + 1 , 10−4 s2 + 2 × 10−3 s + 1 0.55s2 + s + 1

10−4 s2 + 2.1 × 10−3 s + 1

4.05s2 + s + 1 . 10−4 s2 + 2 × 10−3 s + 1 Those three control weighting functions are paired with the three performance weighting functions, in the given order, in the µ -synthesis. The final choice of the appropriate performance and control weighting functions is obtained by comparing the results from the corresponding µ -designs. (Six DK iterations are used in each case.) The robust stability and robust performance analysis conducted by the file mu hdd gives the results shown in Table 12.4. Wu3 (s) = 3 × 10−6

348

D.-W. Gu

Table 12.4. Robust stability and robust performance for three controllers Controller Order Robust stability Robust performance µmax µmax 1 38 0.3730 0.436 2 38 0.412 0.549 3 30 0.577 0.904

In the first and third design cases, corresponding to the weighting functions (Wp1 ,Wu1 ) and (Wp3 ,Wu3 ), we take the controllers from the 6th D − K iteration and in the second case we take the controller obtained at the 5th D − K iteration. The closed-loop system achieves robust stability and robust performance for all three controllers; however, the best results, in terms of gain and phase margins, come from the first controller. This can also be seen from the corresponding µ -values. The gain and phase margins for all three design cases are listed in Table 12.5. Table 12.5. Gain and phase margins for three controllers Controller Gain margin Phase margin dB deg 1 10.9 59.0 2 8.9 50.5 3 8.9 36.9

The closed-loop transient responses are shown in Figure 12.19, for a simulated runout of 1 track width (1µ m) and a torque disturbance td = 0.0005 Nm. It is seen from Figure 12.19 that the first controller gives a response with the smallest undershoot (about 20 %), but this response is the slowest one. The third controller gives the fastest response but the undershoot in this case is the largest one (about 50 %). Figure 12.20 shows that as a result of the appropriate tuning of the control weighting functions all three controllers produce control action whose amplitude is slightly less than 1.2 V. The comparison of the transient responses to disturbance, shown in Figure 12.21, reveals that the worst disturbance rejection is found in the first controller case and the best in the third controller case. This is a result of the tightest closed-loop bandwidth of 2 kHz (measured at −3 dB) in the first controller case compared with the bandwidth of 3.2 kHz in the third controller case (see Figure 12.22). Note that the largest peak of the magnitude is found in the third controller case, which results in the largest undershoot of the transient response. The results obtained from different weighting functions show that moving the frequency response of the inverse performance weighting function to the right would lead to larger closedloop system bandwidth, and consequently, faster time-responses of the closed-loop system, though may introduce larger over(under)shoot. However, at the same time, this may reduce the robustness of the closed-loop system. Hence, one has to compromise between the different objectives in the design. In the present design case, it seems that the second controller leads

Robust Control of a Hard-Disk Drive

349

Closed−loop transient response 1.2

First controller Second controller Third controller

1

Position Error Signal (tracks)

0.8

0.6

0.4

0.2

0

−0.2

−0.4

−0.6

0

0.5

1

1.5

2 Time (secs)

2.5

3

3.5

4 −3

x 10

Fig. 12.19. Transient responses of three µ -controllers Control action due to reference First controller Second controller Third controller

1

0.8

0.6

u (V)

0.4

0.2

0

−0.2

−0.4

0

0.5

1

1.5

2

2.5 Time (secs)

3

3.5

4

4.5

5 −4

x 10

Fig. 12.20. Control actions of three µ -controllers

to a good trade-off between the requirements in terms of transient response, disturbance rejection and robustness. Hence we will use the weighting functions Wp2 and Wu2 in both the µ -synthesis and H∞ design.

D.-W. Gu

Transient response to disturbance 0.05

0

Position Error Signal (tracks)

−0.05

−0.1

−0.15

−0.2

−0.25

First controller Second controller Third controller −0.3

0

0.5

1

1.5

2 Time (secs)

2.5

3

3.5

4 −3

x 10

Fig. 12.21. Transient responses to disturbance of three µ -controllers

Closed−loop magnitude plot 10

8

6

4

Magnitude (dB)

350

2

0

−2

−4

−6

−8

−10 2 10

First controller Second controller Third controller 3

10 Frequency (Hz)

4

10

Fig. 12.22. Closed-loop magnitude plots of three µ -controllers

Robust Control of a Hard-Disk Drive

351

The aim of the design in this case is to find an H∞ (sub)optimal, output controller for the interconnection shown in Figure 12.23 in which we exclude the inputs and outputs of the uncertainty block. The variable hin ic which corresponds to the transfer function P of the

r d n

ey eu

P

K

u

yc

Fig. 12.23. Closed-loop system with H∞ controller open-loop system is obtained by the command line

hin_ic = sel(sys_ic,[16:18],[16:19])

The H∞ optimal control minimizes the ∞-norm of FL (P, K) in respect to the transfer function K of the controller. In the given case FL (P, K) (as the transfer function matrix in (12.2)) is the nominal closed-loop transfer function matrix from the reference, disturbance and noise signals (the variables ref, dist and noise) to the weighted outputs e y and e u. The design is conducted using the M-file hinf hdd.m, which computes a (sub)optimal H∞ control law for a given open-loop system. The value of γ is chosen 10% higher than the minimum possible value. The controller such obtained is of 18th order.

12.5.3 H∞ Loop-shaping Design The design of a robust control for the Disk Drive System can be successfully accomplished using the H∞ loop-shaping design procedure (LSDP) as well, as described in this sub-section. Note that in the case of H∞ LSDP, we do not use the performance weighting function implemented in the cases of µ and H∞ designs. Instead, we use a pre-filter W1 and a post-filter W2 in order to shape appropriately the frequency response of the augmented, open-loop transfer function W2 GW1 . In the present case we choose a pre-filter with transfer function W1 = 4

0.05s + 1 . s

The gain of 4 is chosen to ensure a steady-state error, due to the disturbance, less than 10% of the track width. Larger gain leads to smaller steady-state errors but possibly worse transient

352

D.-W. Gu

response. The post-filter is taken simply as W2 = 1. The magnitude plots of the original and shaped systems are shown in Figure 12.24. The design of the H∞ LSDP controller uses the M-file lsh hdd which implements the function ncfsyn. The controller obtained is of order 13. Frequency responses of the plant and shaped plant 200

Initial plant Shaped plant

Magnitude (dB)

150

100

50

0

−50 −2 10

−1

10

0

10

1

10 Frequency (Hz)

2

10

3

10

4

10

Fig. 12.24. Magnitudes of the original and shaped systems

12.6 Comparison of Designed Controllers The comparison of the closed-loop system with µ , H∞ and H∞ LSDP controllers begins with the robust stability and performance analysis. The robust stability is tested on the upper 15 × 15 block of the closed-loop transfer function matrix. To achieve robust stability it is necessary that the µ -values are less than 1 over the frequency range. In Figure 12.25 we compare the structured singular values, for the robust stability analysis, of the closed-loop systems with the three controllers (µ , H∞ , and H∞ LSDP). It shows that all the three closed-loop systems achieve the robust stability. The best robustness is obtained by the µ controller. The nominal performance is tested on the bottom 3 × 2 block of the closed-loop transfer matrix. The comparison of the nominal performance for the three controllers, in Figure 12.26, shows that the performance in the H∞ LSDP controller case over the low frequency range is much worse than that in the other two cases. This is a consequence of the fact that the performance specifications used in the µ design and H∞ design are not explicitly adopted in the design of the H∞ LSDP controller. The larger magnitude over the low frequencies leads to an expectation of worse steady state errors.

Robust Control of a Hard-Disk Drive

353

Robust stability for three controllers 1

µ−controller H controller ∞

0.9

Loop shaping controller

0.8

Upper bound of µ

0.7

0.6

0.5

0.4

0.3

0.2

0.1 1 10

2

3

10

10 Frequency (Hz)

4

10

5

10

Fig. 12.25. Robust stability of the closed-loop systems Nominal performance for three controllers

1

10

0

10

−1

Magnitude

10

−2

10

−3

10

µ−controller H controller ∞ Loop Shaping controller

−4

10

1

10

2

10

3

10 Frequency (Hz)

4

10

5

10

Fig. 12.26. Nominal performance of the closed-loop systems

The robust performance of the closed-loop system is studied also by the aid of the µ -analysis. The closed-loop transfer function matrix has 18 inputs and 17 outputs. The first 15 inputs/outputs correspond to the 15 channels connecting the perturbation matrix ∆ , while the last 3 inputs and 2 outputs correspond to the weighted sensitivity of the closed-loop system. Hence, for a µ -analysis of the robust performance the block-structure of the uncertainty should consist of a 15 × 15 diagonal real uncertainty block and a 3 × 2 complex uncertainty block.

354

D.-W. Gu

∆P :=

∆ 0 0 ∆F

: ∆ ∈ C15×15 , ∆F ∈ C3×2

Robust performance for three controllers

0

Upper bound of µ

10

µ−controller H controller ∞

Loop shaping controller 1

10

2

10

3

10 Frequency (Hz)

4

10

5

10

Fig. 12.27. Robust performance of the closed-loop systems

The robust performance (in respect to the uncertainty and performance weighting functions) is achieved if and only if for each frequency µ∆P (.), computed for the closed-loop frequency response, is less than 1. The robust performance test for all controllers is shown in Figure 12.27. Again, the H∞ LSDP controller shows large µ -values over the low frequency range. The robust stability and robust performance analysis also shows that the worse results may possibly occur over frequencies around the resonant frequencies. The Bode plots of the closed-loop systems with three controllers are shown in Figures 12.28 and 12.29. It is seen that the system with the H∞ LSDP controller has the largest bandwidth which may lead to a fast transient response. A larger bandwidth, however, may also lead to larger effect of noises and resonances. The plot of the output sensitivity to disturbances (Figure 12.30) shows that in the low frequency range the influence of the disturbance on the system output in the case of the H∞ LSDP controller is the largest. Better disturbance attenuation for this controller may be achieved by choosing higher gain in the pre-filter. This will lead, however, to greater overshoot in the transient response. The sensitivity to disturbance in the cases of µ and H∞ controllers reaches maximum value in the frequency range from 300 Hz to 700 Hz which is inside of the closedloop bandwidth. This means that the closed-loop system will be susceptible to disturbances over that frequency range. It is interesting to notice that the sensitivity of the H∞ LSDP controller over the same range is 8 dB lower. The output sensitivity to noise is shown in Figure 12.31. (Note that the sensitivity is in respect to the unit-bounded noise, the input of the noise shaping filter.) The lowest sensitivity to noise

Robust Control of a Hard-Disk Drive

355

Closed−loop magnitude plot 10

8

6

Magnitude (dB)

4

2

0

−2

−4

−6

−8

µ−controller H controller ∞

Loop Shaping controller −10 1 10

2

3

10

10

4

10

Frequency (Hz)

Fig. 12.28. Magnitudes of the closed-loop systems Closed−loop phase plot 50

0

−50

−100

Phase (deg)

−150

−200

−250

−300

−350

−400

µ−controller H∞ controller Loop Shaping controller

−450 1 10

2

3

10

10

4

10

Frequency (Hz)

Fig. 12.29. Phase Plots of the closed-loop systems

has achieved by the system with µ controller and the largest sensitivity by the system with the H∞ LSDP controller. The Bode plots of the three controllers are compared in Figures 12.32 and 12.33. It is seen that the H∞ LSDP controller has a very low gain in the range from 1 Hz to 500 Hz which is the reason for the weak attenuation of the disturbances over that range. The transient responses of the closed-loop systems are obtained by using the file clp hdd. In Figure 12.34 we show the transient responses of the closed-loop systems to a reference

356

D.-W. Gu Output sensitivity to disturbance 60

50

Magnitude (dB)

40

30

20

10

0

µ−controller H controller ∞

Loop Shaping controller −10 −2 10

−1

0

10

1

10

2

10 Frequency (Hz)

3

10

10

4

10

Fig. 12.30. Output sensitivity to disturbance Output sensitivity to noise −15

−20

−25

Magnitude (dB)

−30

−35

−40

−45

−50

−55

µ−controller H controller

−60

∞

Loop Shaping controller −65 −1 10

0

10

1

2

10

10

3

10

4

10

Frequency (Hz)

Fig. 12.31. Output sensitivity to noise

signal, equivalent to 1 track. While for the H∞ and H∞ LSDP controllers the undershoot is about 60%, it is only 28% for the µ controller. The settling time for the H∞ LSDP, µ and H∞ controllers is 0.8 ms, 1 ms and 1.5 ms, respectively. The control actions of the three controllers are shown in Figure 12.35. For all controllers the control signal amplitude does not exceed 1.2 V, as required. In Figure 12.36 we show the system response to a step torque disturbance td = 0.0005 Nm (equivalent to a force of 9.8 × 10−3 N applied to the Disk Head Assembly). The transient error

Robust Control of a Hard-Disk Drive

357

Controller magnitude plot 30

µ−controller H∞ controller Loop Shaping controller

20

10

Magnitude (dB)

0

−10

−20

−30

−40

−50 −2 10

−1

10

0

10

1

10

2

10 Frequency (Hz)

3

10

4

10

5

10

6

10

Fig. 12.32. Magnitude plots of three controllers Controller phase plot 150

µ−controller H controller ∞

Loop Shaping controller 100

Phase (deg)

50

0

−50

−100

−150 −2 10

−1

10

0

10

1

10

2

10 Frequency (Hz)

3

10

4

10

5

10

6

10

Fig. 12.33. Phase plots of three controllers

for the H∞ LSDP controller has the smallest undershoot (11.5 %) but has a nonzero steady state error. This reveals that in the given case the LSDP controller does not have integrating action in respect to the disturbance. The transient error for µ and H∞ controllers is less than 17% of the track width and the steady state error is practically equal to zero. The output response to the position sensing noise is simulated for a noise signal with amplitude which does not exceed 60 mV (5 % of the track width). This signal is obtained at the output of the noise shaping filter whose input is chosen as a sequence of uniformly distributed random

358

D.-W. Gu Closed−loop transient response 1.2

1

0.8

Position Error Signal (tracks)

0.6

0.4

0.2

0

−0.2

−0.4

µ−controller H∞ controller

−0.6

Loop shaping controller −0.8

0

0.5

1

1.5

2

2.5

Time (secs)

−3

x 10

Fig. 12.34. Transient responses with three controllers Control action due to reference 1.5

µ−controller H controller ∞

Loop shaping controller 1

u (V)

0.5

0

−0.5

−1

0

0.5

1

1.5

2

2.5 Time (secs)

3

3.5

4

4.5

5 −4

x 10

Fig. 12.35. Control actions of three controllers

numbers in the interval [−1, 1]. In the case of the µ controller the output due to noise is less than 1.9 % of the track width. The largest output to the noise (2.3 % of the track width) is seen in the H∞ LSDP controller case due to the largest closed-loop bandwidth of this controller. The comparison of the robust stability and robust performance for the three controllers, as well as the comparison of the corresponding frequency and transient responses shows that it is reasonable to conclude that the µ controller is preferable.

Robust Control of a Hard-Disk Drive

359

Transient response to disturbance 0.02

0

−0.02

Position Error Signal (tracks)

−0.04

−0.06

−0.08

−0.1

−0.12

−0.14

µ−controller H controller

−0.16

∞

Loop shaping controller −0.18

0

0.5

1

1.5

2 Time (secs)

2.5

3

3.5

4 −3

x 10

Fig. 12.36. Transient disturbance responses of three controllers

12.7 Controller Order Reduction The controller obtained by the µ -synthesis is initially of 38th order. It is useful to reduce as much as possible the controller order which will simplify the implementation and increase the reliability. To do this we use the system balancing followed by optimal Hankel approximation. There is a clear gap between the 12th and 13th Hankel singular values. Hence, the order of the µ controller is reduced to 12. Further reduction of the controller order leads to deterioration of the closed-loop performance. In Figures 12.37 and 12.38, we compare the Bode plots of the full-order and reduced-order controllers. The corresponding plots practically coincide with each other, which implies similar performances in the closed-loop systems. In particular, the transient responses of the closed-loop system with full-order and that with the reduced-order controller are practically undistinguishable.

D.-W. Gu

Controller magnitude plots 30

Full order (n = 32) controller Reduced−order (n = 12) controller 20

Magnitude (dB)

10

0

−10

−20

−30

−40 −2 10

−1

10

0

1

10

10 Frequency (Hz)

2

10

3

10

4

10

Fig. 12.37. Magnitude plots of full and reduced-order µ controllers Controller phase plots 150

Full order (n = 32) controller Reduced−order (n = 12) controller

100

50

Phase (deg)

360

0

−50

−100

−150 −2 10

−1

10

0

10

1

10 Frequency (Hz)

2

10

3

10

4

10

Fig. 12.38. Phase plots of full and reduced-order µ controllers

Robust Control of a Hard-Disk Drive

361

12.8 Design of Discrete-time Controller In general there are two approaches to design a discrete-time servo controller. The first approach is to sample the already designed continuous-time controller at a given sampling frequency fs = 1/Ts . This may be accomplished by the M-file dcl hdd.m which utilizes the Robust Control Toolbox function samhld. It is assumed that the control action calculation requires one sampling period Ts . This introduces a pure delay equal to Ts which is approximated by a rational transfer function using the command pade. The resultant sampleddata, closed-loop system is simulated by using the function sdtrsp. This approach gives satisfactory results for sufficiently high sampling frequency (say, 100 kHz in the given case). The second method is to sample the continuous-time, open-loop system (including the weighting filters) and to design directly a discrete-time controller by using H∞ optimization (implementing the function dhinfsyn) or µ -synthesis (implementing the function dkit). The choice of the sampling frequency in the discrete-time case has a strong influence on the closed-loop system performance. A low sampling frequency limits the system bandwidth and would deteriorate the transient performance such as the disturbance rejection. On the other hand, the increase of the sampling frequency would complicate the controller implementation and raise the cost of HDD. Further on we consider the µ -synthesis of the discrete-time controller for two sampling rates - 24 kHz and 36 kHz. In both cases we use the same performance weighting function Wp2 (s) = 10−4

s2 + 4 × 105 s + 2.5 × 109 , s2 + 3.9 × 105 s + 6.25 × 105

utilized already in the continuous-time design. Depending on the sample rate we use two different control weighting functions Wu1 (s) = 10−6

4s2 + 2s + 1 2 × 10−3 s2 + 2 × 10−3 s + 1

(for fs = 24 kHz) and Wu2 (s) = 10−6

1.04s2 + 2s + 1 . 7.5−5 s2 + 2 × 10−3 s + 1

(for fs = 36 kHz). This allows in both cases to obtain control signals which do not exceed 1.2 V. The noise shaping filter is the same as in the continuous-time case. The sampling of the extended open-loop system for the given sampling rate is conducted by the M-file dlp hdd.m. The discrete-time µ -synthesis is accomplished by the file dms hdd.m which is used in conjunction with the auxiliary file ddk hdd.m. The file ddk hdd.m sets the structure of the uncertainty and the parameter values of the D − K iteration. As in the continuous-time µ -synthesis, only the rigid body uncertain parameters are taken into account. The frequency is set on the unit circle in the interval [0, π ]. In the discrete-time case it is also necessary to add the operator

DISCRETE_DK = 1;

362

D.-W. Gu

in the file ddk hdd.m. The results from the µ synthesis for fs = 24 kHz shows that for the chosen weighting functions the closed-loop system almost achieves robust performance at the fifth D−K iteration (µmax = 1.06) but the closed-loop response is relatively slow and the undershoot is large (44 %). The maximum control amplitude is 1.19 V. To obtain better results it is necessary to increase the sampling frequency.

Robust stability

0

10

−1

10

−2

µ

10

−3

10

−4

10

µ−upper bound µ−lower bound

−5

10

1

10

2

10

3

10 Frequency (Hz)

4

10

5

10

Fig. 12.39. Robust stability of fs = 36 kHz design

We now present in more details the results from the µ synthesis at fs = 36 kHz. In this case an appropriate controller is obtained after three D − K iterations and the maximum robust performance achieved is µmax = 1.071. In Figures 12.39 and 12.40, we show the µ -plots, obtained by the M-file dmu hdd.m, for robust stability analysis and robust performance analysis, respectively. In both plots the worst results are seen around the second resonant frequency of 2.2 kHz. The closed-loop transient response is obtained by the file dsl hdd.m which uses the function sdtrsp. The function sdtrsp computes also the control signal obtained at the output of the discrete-to-analog converter. The closed-loop transient response is shown in Figure 12.41 and the corresponding control action in Figure 12.42. The undershoot of the transient response is less than 36 % and the peak amplitude of the control is less than 1.05 V. Transient response to disturbance is shown in Figure 12.43. Overall, the results obtained are almost as good as the results obtained with the continuous-time, µ controller.

Robust Control of a Hard-Disk Drive Robust performance

1

10

0

µ

10

−1

10

µ−upper bound µ−lower bound

−2

10

1

2

10

3

10

4

10 Frequency (Hz)

5

10

10

Fig. 12.40. Robust performance of fs = 36 kHz design

Closed−loop sample−data transient response 1.2

1

Position Error Signal (tracks)

0.8

0.6

0.4

0.2

0

−0.2

−0.4

0

0.5

1

1.5

2

2.5 Time (secs)

3

3.5

4

4.5

Fig. 12.41. Transient response of fs = 36 kHz design

5 −3

x 10

363

D.-W. Gu

Controller output due to reference 1.5

1

u (V)

0.5

0

−0.5

−1

−1.5

0

0.1

0.2

0.3

0.4

0.5 Time (secs)

0.6

0.7

0.8

0.9

1 −3

x 10

Fig. 12.42. Control action of fs = 36 kHz design

Transient response to disturbance 0.1

0.05

0

Position Error Signal (tracks)

364

−0.05

−0.1

−0.15

−0.2

−0.25

0

0.5

1

1.5

2

2.5 Time (secs)

3

3.5

4

4.5

5 −3

x 10

Fig. 12.43. Transient response to disturbance of fs = 36 kHz design

Robust Control of a Hard-Disk Drive

365

12.9 Nonlinear System Simulation In order to obtain a realistic idea about the behavior of the designed system, the nonlinR For this aim, two models ear, closed-loop servo system is simulated by using Simulink . are developed, namely c hdd.mdl for the continuous-time system and d hdd.mdl for the sampled-data system. In the simulation we take into account the amplifier saturation which was neglected in the design so far. Both models allow to simulate the closed-loop system for different reference, disturbance and noise signals. Before simulating the continuous-time system it is necessary to assign the model parameters by using the M-file init c hdd.m. The sampled-data model involves the discrete-time controller, 16-bit analog-to-digital converter with maximum input voltage 2.5 V and 16-bit digital-to-analog converter with maximum output voltage 10 V. It is assumed that the discrete-time controller is implemented on a digital signal processor(DSP) with word length of 64 bits. These parameters are set prior to the simulation by using the M-file init d hdd.m. It is assumed that the control action calculation requires one sampling period Ts . R In Figure 12.44 we show the Simulink model d hdd.mdl of the nonlinear, sampled data, closed-loop system. As in the linear case, the transient responses of the nonlinear closed-loop system are obtained for a simulated runout of 1 track width (1µ m) and torque disturbance td = 0.0005 Nm. In Figure 12.45 and in Figure 12.46 we compare the results from the simulation of the continuous-time and discrete-time nonlinear systems. The continuous-time controller is the reduced-order µ controller in Section 12.7 and the discrete-time controller is the controller designed at the sampling frequency of 36 kHz. The transient responses of the nonlinear system are close to the corresponding responses of the linear system due to the small input signals (amplitude less than 1.2 V). It should be mentioned that the controllers designed are appropriate for small reference signals (equivalent to one track). For larger references the amplifier saturates and it is necessary to implement an appropriate seeking algorithm.

366

D.-W. Gu

SIMULINK model of the nonlinear sample−data Hard Disk Drive Servo System

num(s) den(s)

Control

Mode 1

y

num(s)

To Workspace

Control1

u

1/Rc

Kpa PA

i

tau.s+1 Saturation

Admittance

Kt

Tm

den(s) Mode 2

1/J

1 s

ω

Θ

1 s

y −K−

Gain1 Arm

num(s)

Integrator1

Integrator2

Position gain

den(s)

Rs

Mode 3 Resistance

Td Disturbance

num(s) den(s) Mode 4 Kb Back e.m.f.

Out1

In1

Controller

ref Out1

t Clock

Reference

Noise1

To Workspace2

R Fig. 12.44. Simulink model of the nonlinear sampled data system

Output

Robust Control of a Hard-Disk Drive

Nonlinear system response to reference 0.2

0

Position Error Signal (tracks)

−0.2

−0.4

−0.6

−0.8

−1

−1.2

Continuous−time controller Discrete−time controller −1.4

0

0.5

1

1.5

2

2.5 Time (secs)

3

3.5

4

4.5

5 −3

x 10

Fig. 12.45. Transient responses of the nonlinear systems

Nonlinear system response to disturbance 0.1

0.05

Position Error Signal (tracks)

0

−0.05

−0.1

−0.15

−0.2

Continuous−time controller Discrete−time controller −0.25

0

0.5

1

1.5

2

2.5 Time (secs)

3

3.5

4

4.5

5 −3

x 10

Fig. 12.46. Disturbance responses of the nonlinear systems

367

368

D.-W. Gu

12.10 Conclusions The experience gained in the design of the HDD servo controllers makes it possible to derive the following conclusions: •

•

•

•

• •

The implementation of designed µ , H∞ and H∞ LSDP controllers in the HDD servo system gives satisfactory results in respect to robustness and performance. All three controllers ensure robust stability of the closed-loop system. The best robust performance is achieved by using the µ and H∞ controllers. The implementation of the H∞ LSDP controller gives the fastest transient response and the corresponding design is less complicated. This controller, however, leads to the worst performance in the low frequency range which results in a large steady state error. In the given case the best trade-off between the robustness and transient response requirements is achieved by using the µ controller which is, to some extent, due to the specially chosen weighting functions. The number of the original uncertain parameters is very large (more than 25 in the given case). This complicates the derivation of the uncertainty model and requires heavy computation demands. That is why it is necessary to investigate the parameter importance in respect to the robustness and performance in order to reduce their number to an acceptable value. However, in the evaluation of the design, it is better to take into account all the possible uncertainties to ensure a satisfactory design in a real case. In the µ -synthesis, the order of the resultant controller depends on the order of the plant, of the weighting functions and of the scaling diagonal elements approximations. The designed controllers are usually of high orders, which complicates the implementation of the controller. Hence, an order reduction should usually be considered right after the controller design. Most controllers used in the HDD designs are of order between 8 and 15. Good disturbance attenuation requires sufficiently large closed-loop bandwidth. This may, however, lead to difficulties in achieving robust stability and robust performance in the presence of resonant modes. Some resonances whose frequencies are much higher than the closed-loop bandwidth and thus seem innocent may even actually destroy the robust stability of the system. In such cases it is necessary to increase the damping of these modes by using techniques of passive/active damping. The presence of resonant modes may require sufficiently high sampling rates in the case of using discrete-time controller. It is important to stress that better results in respect to the transient response (overshoot and settling time) are difficult to obtain for the current plant parameters. If higher performance demands are required, it is necessary to change the HDD parameters, for instance to increase the VCM torque constant.

Notes and References The history of the Hard Disk Drive control is presented in the fascinating papers of Abramovitch and Franklin [1, 2]. An excellent survey on similarities and contrasts in the magnetic and optical disk controls is given in [5]. In [20] one may find an attractive description of the HDD construction and functioning.

Robust Control of a Hard-Disk Drive

369

The book of Chen, Lee and Venkataramanan [6] is the only book up to the moment which is entirely devoted to HDD servo systems and contains rich information related to the design of such systems. A tutorial on HDD control can be found in [106]. A detailed model of the HDD servo system, which has very much influenced the model used in this chapter, is presented in the book of Franklin, Powell and Workman [11, Chap. 14]. Below 100 Hz the rotary actuator dynamics is affected by pivot bearing nonlinearity which is known under the name “stick-slip” . It has a strong effect particularly in the case of small disk drives with lower actuator inertia. Analysis and simulation of this phenomenon are presented in [4, 25]. An important step in the design of HDD servo systems is the reliable estimation of the various disturbances and noises acting on the system. Methods for such estimations are described in [3, 9, 13]. The harmonic compensation used to reduce RRO is considered in [7, 18]. The track seeking and track following modes require different control algorithms. Track seeking algorithms are described in [11, 24]. The switching of control mode from track seeking to track following should be smooth so that the residual vibration of the read/write head suspension is minimal. There are several control algorithms which work for both track seeking and following, see for instance [14, 17]. Such algorithms utilize two-degree-of-freedom (2DOF) controllers in which case the track seeking is accomplished by using a feedforward controller along with a reference trajectory. Apart of the track seeking and track following the HDD contains also a spindle velocity control loop. The purpose of this loop is to control the air flow over the disk in order to guarantee the appropriate flying height of the read/write head. This is a low frequency control loop and its design does not represent a serious difficulty. Further expansion of the closed-loop bandwith of the HDD control system may be achieved by using the so called dual stage servos which consist of a low bandwidth coarse actuator (the usual VCM) and a high bandwidth fine actuator. The fine actuator is with a small stroke and may be implemented as a piezoelectric transducer(PZT) [10]. The design of dual stage servos is considered in [8, 15, 19]. Other important aspects of the analysis and design of HDD servo systems are presented in [12, 16, 21, 22, 26], to name a few.

References 1. D. Abramovitch and G. Franklin. A brief history of disk drive control. IEEE Control Systems Magazine, 22:28–42, 2002. 2. D. Abramovitch and G. Franklin. Disk drive control: The early years. In Proceedings of the 15th IFAC Congress. Session T-Th-M12, pages 1–12, Barcelona, Spain, July 2002. CD-ROM. 3. D. Abramovitch, T. Hurst, and D. Henze. An overview of the PES Pareto method for decomposing baseline noise sources in hard disk postion error signals. IEEE Transactions on Magnetics, 34:17–23, 1998. 4. D. Abramovitch, F. Wang, T. Hurst, and G. Franklin. Disk drive pivot nonlinearity modelling Part I: Frequency domain. In Proceedings of the 1994 American Control Conference, pages 2600–2603, Baltimore, MD, June 1994.

370

D.-W. Gu

5. D.Y. Abramovitch. Magnetic and optical disk control: Parallels and contrasts. In Proceedings of the 2001 American Control Conference, pages 421–428, Arlington, VA, June 2001. 6. B.M. Chen, T.H. Lee, and V. Venkataramanan. Hard Disk Drive Servo Systems. Springer-Verlag, Berlin, 2002. 7. K.K. Chew and M. Tomizuka. Digital control of repetetive errors in disk drive systems. IEEE Control Systems Magazine, 10:16–20, 1990. 8. J. Ding, M. Tomizuka, and H. Numasato. Design and robustness analysis of dual stage servo system. In Proceedings of the 2000 American Control Conference, pages 2605– 2609, Chicago, Illinois, June 2000. 9. C. Du, J. Zhang, and G. Guo. Vibration analysis and control design comparison of HDDs using fluid bearing and ball bearing spindles. In Proceedings of the 2002 American Control Conference, pages 1378–1383, Anchorage, AK, May 2002. 10. R.B. Evans, J.S. Griesbach, and W.C. Messner. Piezoelectric microactuator for dual stage control. IEEE Transactions on Magnetics, 35:977–982, 1999. 11. G.F. Franklin, J.D. Powell, and M.L. Workman. Digital Control of Dynamic Systems. Addison-Wesley, Menlo Park, CA, third edition, 1998. 12. T.B. Goh, Z. Li, B.M. Chen, T.H. Lee, and T. Huang. Design and implementation of a hard disk drive servo system using robust and perfect tracking approach. In Proceedings of the 38th IEEE Conference on Decision and Control, pages 5247–5252, Phoenix, Arizona, December 1999. 13. L. Guo, H.S. Lee, A. Hudson, and S.-H. Chen. A comprehensive time domain simulation tool for hard disk drive TPI prediction and mechanical/servo enhancement. IEEE Transactions on Magnetics, 35:879–884, 1999. 14. S. Hara, T. Hara, L. Yi, and M. Tomizuka. Two degree-of-freedom controllers for hard disk drives with novel reference signal generation. In Proceedings of the 1999 American Control Conference, pages 4116–4121, San Diego, CA, June 1999. 15. D. Hernandez, S.-S. Park, R. Horowitz, and A.K. Packard. Dual stage track-following servo design for hard disk drives. In Proceedings of the 1999 American Control Conference, pages 4116–4121, San Diego, CA, June 1999. 16. Y. Huang, P. Mathur, and W.C. Messner. Robustness analysis on a high bandwidth disk drive servo system with an instrumented suspension. In Proceedings of the 1999 American Control Conference, pages 3620–3624, San Diego, CA, June 1999. 17. J. Ishikawa, Y. Yanagita, and T. Hattori ans M. Hashimoto. Head positioning control for low sampling rate systems based on two degree-of-freedom control. IEEE Transactions on Magnetics, 32:1787–1792, 1996. 18. C. Kempf, W. Messner, M. Tomizuka, and R. Horovitz. Comparison of four discretetime repetetive control algorithms. IEEE Control Systems Magazine, 13:48–54, 1993. 19. M. Kobayashi, T. Yamaguchi, and R. Horowitz. Track-seeking controller design for dual-stage actuator in magnetic disk drives. In Proceedings of the 2000 American Control Conference, pages 2610–2614, Chicago, Illinois, June 2000. 20. C.M. Kozierok. The PC Guide. The Reference Section on Hard Disk Drives. January 2003. Available at http://www.PCGuide.com. 21. H.S. Lee. Controller optimization for minimum position error signals of hard disk drives. In Proceedings of the 2000 American Control Conference, pages 3081–3085, Chicago, Illinois, June 2000. 22. D.P. Magee. Optimal filtering to improve performance in hard disk drives: Simulation results. In Proceedings of the 1999 American Control Conference, pages 71–75, San Diego, CA, June 1999.

Robust Control of a Hard-Disk Drive

371

23. W. Messner and R. Ehrlich. A tutorial on controls for disk drives. In Proceedings of the 2001 American Control Conference, pages 408–420, Arlington, VA, June 2001. 24. L.Y. Pao and G.F. Franklin. Time-optimal control of flexible structures. In Proceedings of the 29th IEEE Conference on Decision and Control, pages 2580–2581, Honolulu, HI, December 1990. 25. F. Wang, T. Hurst, D. Abramovitch, and G. Franklin. Disk drive pivot nonlinearity modelling Part II: Time domain. In Proceedings of the 1994 American Control Conference, pages 2604–2607, Baltimore, MD, June 1994. 26. M.T. White, M. Tomizuka, and C. Smith. Rejection of disk drive vibration and shock disturbances with a disturbance observer. In Proceedings of the 1999 American Control Conference, pages 4127–4131, San Diego, CA, June 1999.

13 Modelling and Control of Railway Vehicle Suspensions Argyrios C. Zolotas and Roger M. Goodall

Summary. This chapter uses a railway vehicle as an example of a mechanical dynamic system to which control can be applied in a manner that yields significant benefits from an engi neering and operational viewpoint. The first part describes the fundamentals of railway vehi cles and their dynamics: the normal configuration, the suspension requirements, how they are modelled and an overview of the types of control concept that are currently applied or under consideration. The second part provides a case study of controller design issues.

13.1 Overview of Railway Vehicle Dynamics and Control 13.1.1 Railway Vehicles: Conventional Configuration Railway vehicles employ steel wheels running on tracks with steel rails, which provide the support and guidance functions. The interface between the two is established at contact point(s) between the wheels and rail surface, and both the vehicle configuration and the track greatly influence how vehicles behave [1]. Most modern passenger-carrying railway vehicles have the configuration shown schematically in Figure 13.1, which gives simplified side-view and end-view diagrams. The vehicle body is supported by two bogies using relatively soft secondary suspensions to provide isolation from the track-induced vibrations i.e., to provide a good ride quality. Each bogie has four wheels arranged in two pairs, where each pair is rigidly connected via a common axle (known as the solid-axle wheelset) such that the two wheels have to rotate at the same speed. The wheelsets are connected to the bogie via primary suspension elements: these are much stiffer than in the secondary suspension and are designed to satisfy the vehicle’s stability and guidance requirements. Both primary and secondary suspensions are provided in the vertical, lateral and longitudinal directions, and extra stiffness is often added in the secondary roll suspension. These suspensions mainly comprise passive springs and dampers connected in parallel and/or series, but airbags providing better performance are commonly used for the secondary suspension on modern passenger vehicles. M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 373- 412, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

374

A.C. Zolotas and R.M. Goodall

Fig. 13.1. Simplified side-view and end-view diagrams of a bogie vehicle

13.1.2 Suspension Design Requirements Fundamentally there are three things that a suspension needs to do: • • •

support the (changing) weight of the vehicle provide guidance so that the vehicle follows the intended path provide isolation to give a satisfactory ride quality

The first two requirements are satisfied by a suspension that is relatively stiff, although in both cases this is a characteristic that is only really required at low frequencies, significantly less than 1 Hz. The third requirement calls for a suspension that is soft, but the track irregularities that cause vibrations of the vehicle body don’t become significant until around 1 Hz and higher. Hence there is clearly a design trade-off to be achieved in the suspension design. This trade-off is clarified by further consideration of the inputs. The main dynamic excitations, in addition to the weight which changes rather slowly and is therefore essentially a quasi-static problem, are the track inputs, which can be divided into two types: •

Deterministic inputs – Isolated features – steps, dips, short ramps, etc. – Intended inputs – gradients, curves, etc. having well-defined characteristics Design requirement: constraint on suspension deflection

•

Random inputs – Irregularities and imperfections (roughness of track) – Characterised by a power spectrum (track velocity spectrum approximates to white noise) Design requirement: satisfactory ride quality

In fact, designing any suspension is a non-trivial, multi-objective problem, particularly when combined with the dynamic complexity mentioned in the next subsection, and an active suspension brings additional issues. The overall process is summarised by Figure 13.2. The inputs mentioned above are shown on the left, all of which occur in the vertical, lateral and roll direction.

Modelling and Control of Railway Vehicle Suspensions

375

Actuators

Vehicle system

Sensors

Controller

Track features (deterministic)

Body acceleration (minimise)

Track irregularities (stochastic)

Suspension deflection (constrain)

Load changes

Stability (constrain) Curving performance (optimise)

Fig. 13.2. Design requirements

The outputs are shown on the right, and four items are shown. The acceleration levels on the vehicle body, which represents the quality of ride, are to be minimised, and as mentioned the suspension movements must not become too large i.e., they must be constrained. It is also necessary to ensure that there is a minimum margin of stability, a constraint which applies mainly to the wheelset dynamics determined by the primary suspension. The other primary suspension characteristic is its curving performance: essentially this is concerned with the guidance function, and is mainly about minimising any wear of the wheels and the rails, so an optimisation is needed. There are therefore three input types to consider, two output measures to minimise and two design constraints to meet, and as indicated for an active system it is necessary to choose a set of actuators , identify what measurements can practically be made to choose the set of sensors , then design a controller to satisfy the multiple objectives.

13.1.3 Modelling of Suspensions (for Applying Control) Railway vehicles are dynamically-complex multi-body systems. Each mass within the system has six dynamic degrees of freedom corresponding to three displacements (longitudinal, lateral and vertical) and three rotations (roll, pitch and yaw). Each degree of freedom results in a second-order differential equation and hence 6 × N differential equations will be necessary to represent the system mathematically, where N is the number of masses. For a conventional bogie vehicle, there are seven main masses (one vehicle body, two bogies and four wheelsets) and therefore a total of 42 second-order differential equations will be required if no constraints are considered. In addition, wheel-rail contact presents a highly non-linear dynamic/kinematic problem adding extra complexity to the already complex system. In the modelling, Newton’s law is applied to every degree of freedom of vehicle body and bogies, and external forces/torques are applied through suspension components. For design purposes, the suspensions can be largely considered as linear components, and can readily be R More thorough modelling for generated using control design software such as MATLAB . simulation will require the inclusion of non-linearities due to factors such as dead-band, hardening/softening and Coulomb friction elements, and nowadays will normally be undertaken using one of a number of 3D modelling packages which will incorporate the full complexity and non-linearity, including effects such as body flexibility that are essential for properly assessing ride quality, for example. It is of course essential that such modelling software can support the integration of the controller into the mechanical system.

376

A.C. Zolotas and R.M. Goodall

The forces on the wheelset arise from so-called “creepages” between the wheel and rail, small relative velocities which arise because of elastic deformation of the steel at the point of contact and which apply in both the longitudinal and the lateral directions. Modelling of these effects is rather complex and non-linear [2], but it is a useful simplification to consider the effects separately for the two translational directions with linearized creep coefficients. Using this assumption, dynamic responses can be adequately described by deriving the creepages and using these coefficients to determine the corresponding creep forces generated at the contact interface. For applying control to complex systems such as these, it is important to distinguish between the design model and the simulation model. The design model is a simplified version used for synthesis of the control strategy and algorithm, whereas the simulation model is a more complex version used to test fully the system performance. For example, there is a relatively weak coupling between the vertical and lateral motions of a vehicle and, depending on the objectives, only selected degrees of freedom need to be included in the design model. To study the vertical response, it would be adequate to include the bounce, pitch and sometimes roll degrees of freedom of the components. For the lateral response, the lateral, yaw and sometimes roll degrees of freedom are sufficient. In studies of the longitudinal dynamics, the longitudinal, pitch and roll degrees of freedom may be included in the model. As a common practice, the vehicle model is partitioned into side-view, plan-view and end-view models. The side-view model is concerned with the bounce, pitch and longitudinal degrees of freedom; the plan-view model deals with the lateral and yaw motions and the end-view model covers the bounce and roll motions. The complete vehicle model can be assembled by determining which masses and directions of motion are to be included, and then creating a corresponding set of simultaneous differential equations that can be represented in matrix form.

Vehicle body Secondary suspension

z

m

c

f a Actuator

k

z'

m Primary suspension

k'

c'

Track/road input

zt

Fig. 13.3. Two-mass suspension example

As an example, the two-mass vertical model shown in Figure 13.3 can either be represented in block diagram form as shown by Figure 13.4, or by the following equations:

Modelling and Control of Railway Vehicle Suspensions

fa

Primary suspension displacement

zt

-

+

k +sc

+

1/m

377

Secondary suspension force

1/s2

z

z'

-

+

k+sc

+

1/m

1/s2

Secondary suspension displacement

Fig. 13.4. Block diagram representation of two-mass suspension m z¨ + c(˙z − z˙′ ) + k(z − z′ ) = fa

(13.1)

m′ z¨′ + c(˙z′ − z˙) + c′ (˙z′ − z˙t ) + k(z′ − z) + k′ (z′ − zt ) = − fa

(13.2)

which can in turn be written using a matrix equation

M z¨ +Cz˙ + Kz = Ct z˙t + Kt zt + Ba fa z = z z′ T m 0 c −c k −k M= ; C = ; K = 0 m′ −c (c + c′ ) −k (k + k′ ) This can subsequently be converted into state-space form as follows:      −1   −1    z˙ z¨ −1C −M −1 K −M −1 K −M M Ct M Ba t  ′  z¨′  z ˙  0     0     1 0 0 0 0     z +  z˙  =     0 1   ′   0  fa +  0  w 0 0 0 z   z˙′  1 0 0 1 0 0 −10−6 zt z˙t

(13.3) (13.4) (13.5)

(13.6)

As observed in the previous subsection, there is a design trade-off between the suspension deflection and the ride quality, the latter being quantified by a frequency-weighted r.m.s. acceleration, and so it is primarily these two quantities that are required as outputs. Accordingly the corresponding output equation to give the secondary suspension deflection and the acceleration of the secondary mass is   z˙  z˙′    A1 j z¨  z  + B1 fa + G1 w = (13.7) ′   0 0 1 −1 0  ′  z−z 0 0 z zt

Here the track input is a scalar, i.e., w = [˙zt ], and the actuator (control) input is also a scalar u = [ fa ], but in general it is obvious that this kind of representation can be extended to accommodate multiple inputs. The example doesn’t include the more complicated modelling of the wheelset dynamics [2], i.e., it doesn’t include creep forces etc., but the general principle can readily include these aspects.

378

A.C. Zolotas and R.M. Goodall

13.1.4 Control Concepts Active control can be applied to both secondary and primary suspensions. The control can be directed towards the response to the deterministic (intended) inputs such as curves or gradients, or towards the response to irregularities, or to a combination of the two. Although in principle this offers a very broad range of possibilities, in practice it’s possible to provide more focussed categorisation of the major opportunities as follows: •

Secondary suspension control (a) Tilting trains – enables higher speeds through curves via improved response to deterministic track features. (b) Active secondary suspensions – provides better ride quality i.e., an improved response to track irregularities.

•

Primary suspension (control of wheels and wheelsets) (a) Active steering – gives improved curving leading to reduced wheel and rail wear (improved response to deterministic track features). (b) Active stability/guidance – provides improved stability and/or higher speed (improved response to track irregularities).

In practice tilting and active secondary suspensions are sufficiently distinct in an operational sense that separate descriptions are appropriate, whereas the technology required to achieve the two primary suspension options is very similar and so they are described together, even though the control strategies will be very different.

Tilting Trains The basic idea is to lean the vehicles inwards on curves to reduce the acceleration felt by the passengers [3]. However this acceleration as a vehicle passes onto a railway curve does not rise suddenly: there is a transition from the straight to the curve, usually lasting around 2 seconds, which is a deliberate design feature so that passengers are not made uncomfortable by too sudden an application of sideways acceleration. Normally the track is leaned inward or “canted” to reduce the lateral acceleration experienced by the passengers, and this also increases steadily through the transition - Figure 13.5. At higher speeds the curving acceleration rises, and the transition will also be more severe because the duration of the transition will reduce, that is unless the track is changed. This is where tilt comes in, to bring the acceleration back to the level it was before. However, it is not only what happens in the steady curve that is important, but also the dynamic response during the transition. Ideally the tilt angle of the body should rise progressively, perfectly aligned both with the onset of curving acceleration and the rising cant angle. Some means of providing tilt of the vehicle body relative to the bogie is achieved using a combination of a tilting mechanism and a set of actuators. Essentially this achieves active control of the secondary roll suspension, but for tilting the aim is only to respond to curve inputs so as to reduce the lateral acceleration perceived by the passengers; the response to lateral track irregularities is largely unaffected and characterised by the components in the secondary suspension. To avoid motion sickness it is now normal not to compensate fully for the curving

Modelling and Control of Railway Vehicle Suspensions

379

Acceleration Perceived acceleration

Tilt Cant

Time

Straight Transition

Curve

Fig. 13.5. Tilting concept

acceleration, with typically around 60-70% compensation being used. It is particularly important to respond quickly in curve transitions, i.e., as well as providing appropriate steady-state curving performance, and achieving this without degrading the straight track ride quality is a non-trivial issue. Control can either use a feed-forward system where the track features are determined and used to provide an appropriate command input for the tilt actuation system, or a feedback approach in which lateral accelerations on the vehicle body or bogie are measured, or a combination of the two.

Active Secondary Suspensions The essential theoretical concepts involved were identified back in the 1970s, and are largely based upon the use of so-called “skyhook” damping in which the actuators are made to appear like “virtual dampers” connected to an absolute datum (see Figure 13.6), which enables significantly higher levels of damping to be introduced into the suspension dynamic modes without compromising the performance at higher frequencies [4]. Active control also means that characteristics of the different modes (vertical, pitch, lateral, yaw, roll) can be influenced in a much more flexible manner, for example the rotational modes can be made significantly softer than the translational modes, which further improves the suspension performance. The improvements can be used either to provide a better ride quality, or to achieve higher speed of operation, or to enable the use of a lower track quality.

"Skyhook" damper

Sensor Actuator

Controller

Fig. 13.6. Active secondary suspension concept

380

A.C. Zolotas and R.M. Goodall

Active Primary Suspensions Before discussing the use of active elements, it’s necessary to understand how the conventional railway wheelset works. It consists of two coned or otherwise profiled wheels rigidly connected by an axle. On straight track the wheelset runs in a centralised position, but when a curve is encountered the wheelset naturally moves outwards; this causes the outer wheel to run on a larger radius and the inner on a smaller radius. Being connected by the axle the wheels must still rotate at the same rotational speed, so the outer wheel moves faster along the track, and the effect is to make the wheelset go around the curve, i.e., a natural mechanical steering action. However there is a problem, which arises when the dynamics of the wheelset are assessed. The combination of the profiled wheels and the creep forces mentioned previously is to create an oscillatory system, a combined lateral and yaw motion known as “hunting”. Adding mechanical dampers doesn’t provide stability, and the normal solution is to have two wheelsets connected to a bogie frame by means of longitudinal springs. These springs create nonconservative creep forces at the wheel-rail contact point and stabilise the hunting motion. However, on a curve these stabilising longitudinal springs produce forces which interfere with the natural curving action of the wheelset. The result is that on the tighter curves the wheel flange will be in contact with the side of the rail, causing wear of the wheels, wear of the rails and often significant amounts of noise. There is therefore a difficult design trade-off: stiff springs give stable high-speed running, but poor curving; soft springs mean that the curving performance is better, but stable running is only possible at low speeds.

+

-

-

+

Fig. 13.7. Active primary suspension concept

The variety of possibilities for controlling the wheels and wheelsets is very large, and based upon an analysis of possible configurations (mechanical scheme, axle type, control objective and actuation technology) [5]. However the most basic scheme is shown in Figure 13.7, in which longitudinal actuators between the bogie frame and the ends of the wheelsets are controlled in a differential sense to apply a yaw torque to each wheelset. This torque can either be used to provide a steering action on curves, or it can be used to provide stability to the wheelset without affecting the natural curving, or a combination of the two.

Modelling and Control of Railway Vehicle Suspensions

381

13.2 Case Study: Control of Secondary Suspensions - Tilting Trains The concept of tilting trains and associated general control concepts have already been discussed in Section 13.1.4. In this case study we will encounter specific issues on the modelling, control objectives, assessment and control design of a tilting train.

13.2.1 Historical Facts on Tilt Control

disturbances: curvature, cant, lateral track irregularities

(measured lateral acceleration)

0 (zero) -

Controller

Vehicle Dynamics (plant model)

yacc

Bogie accel. 1 preview effect

actuator

scaling

k/g LPF

+ -

K(s) demand Tilt angle 1 actuator

k/g

LPF

+

LPF

+

-

K(s) demand Tilt angle 2

(suspension roll) θs -

(effective cant deficiency angle)

+

1 g

Vehicle 1

Vehicle 2

actuator k/g

(equivalent cant deficiency angle)

-

K(s) demand Tilt angle 3

Vehicle 3

... to vehicle 4, etc Digitally transmitted

(a) Intuitive partial nulling control (early-type)

(b) Command driven with precedence control (commercial)

Fig. 13.8. Tilt control schemes

Early tilt control systems were based solely upon local-per vehicle measurements (Figure13.8a), however it proved impossible at the time to get an appropriate combination of straight track and curve transition performance. Interactions between suspension and controller dynamics (with the sensors being within the control loop) led to control limitations and stability problems. Since then, tilt controllers have evolved in an incremental sense, the end result of which is a control structure which is not optimised from a system point of view. The industrial norm nowadays is to utilise precedence control devised in the early 1980s as part of the UK’s Advanced Passenger Train development [4]. In this scheme a bogie-mounted accelerometer from the vehicle in front is used to provide “precedence” (a priori track information) (via appropriate inter-vehicle cable/signalling connections), carefully designed so that the delay introduced by the filter compensates for the preview time corresponding to approximately a vehicle length (Figure 13.8b). Normally nowadays a single command signal would be generated from the first vehicle and transmitted digitally with appropriate time delays down the train. Consequently the velocity and the direction of travel are important factors for the correct operation of the tilt system. The command-driven with precedence strategy proved to be successful and it is nowadays used by most tilting train manufacturers. However it is a complex scheme, direction-sensitive,

A.C. Zolotas and R.M. Goodall

382

signal connections between trains are required, while the tilt system performance can be optimised for a specific route operation. Moreover, leading vehicles have inferior performance due to lack of precedence. Nevertheless achieving a satisfactory local tilt control strategy remains an important issue because of the system simplifications and more straightforward failure detection.

13.2.2 Tilting Vehicle Modelling The modelling of the baseline railway vehicle is based upon a linearised end-view model version (Figure 13.9). It includes the lateral and roll dynamics of both the body and the bogie plus a state from the airspring dynamics, resulting to a 9th order model.

Vehicle body d1

d1

d1 FR

y+v

c.o.g

F"

δa

krz

h g1

crz

P

Fg

Fy1

h1

Fz1

h1

h g1

Fy2

ksy csy cpz

kpy

h2

y+ b

θ+b

h3

Wheelset

h g2

Gv

F'

θv+

Vehicle bogie c.o.g kpz

T

T"

kvr

kaz ksz

d'

d2

cpy

θ+o

(a) End-view model

y+ o

FR

Fy1

Fz1

d1

F"

Fy3 Fz3

T" d2

Fz2 d1

Gb

Fy2 h3

h g2

θο

Fz2

T h2

Fg

Fz4 Fy4 d2

(b) Free body diagram

Fig. 13.9. Tilting vehicle end-view diagram with actuation system

A pair of linear airsprings represents the vertical suspensions, which only contribute to the roll motion of the vehicle (the vertical degrees of freedom are ignored). The model also contains the stiffness of an anti-roll bar connected between the body and the bogie frame. Detailed wheelset dynamics were not included for simplicity. To provide active tilt a rotational displacement ideal actuator, is included in series with the roll stiffness (‘active anti-roll bar (ARB)’ [6]). The ARB-system is assumed to provide up to a maximum tilt angle of 10 degrees. The advantages of active ARBs results from their relative simplicity, i.e., small weight increase, low cost, easily fitted as an optional extra to build or as a retro-fit. The mathematical models of increasing complexity, via Newton’s laws, were developed to encapsulate the lateral and roll dynamics of the tilting vehicle system. The equations of motion

Modelling and Control of Railway Vehicle Suspensions

383

are given below with all variables and parameter values provided in Appendix A. For the vehicle body: mv y¨v = −2ksy (yv − h1 θv − yb − h2 θb ) − 2csy (y˙v − h1 θ˙v − y˙b − h2 θ˙b ) −

mv v2 + mv gθo − hg1 mv θ¨o R

(13.8)

ivr θ¨v = −kvr (θv − θb − δa ) + 2h1 {ksy (yv − h1 θv − yb − h2 θb ) +csy (y˙v − h1 θ˙v − y˙b − h2 θ˙b )} +mv g(yv − yb ) + 2d1 {−kaz (d1 θv − d1 θb ) −ksz (d1 θv − d1 θr )} − ivr θ¨o

(13.9)

For the vehicle bogie: mb y¨b = 2ksy (yv − h1 θv − yb − h2 θb ) + 2csy (y˙v − h1 θ˙v − y˙b − h2 θ˙b ) −2kpy (yb − h3 θb − yo ) − 2cpy (y˙b − h3 θ˙b − y˙o ) −

mb v2 + mb gθo − hg2 mb θ¨o R

(13.10)

ibr θ¨b = kvr (θv − θb − δa ) + 2h2 {ksy (yv − h1 θv − yb − h2 θb ) +csy (y˙v − h1 θ˙v − y˙b − h2 θ˙b )} −2d1 {−kaz (d1 θv − d1 θb ) − ksz (d1 θv − d1 θr )} +2d2 (−d2 kpz θb − d2 cpz θ˙b )

+2h3 {kpy (yb − h3 θb − yo ) + cpy (y˙b − h3 θ˙b − y˙o )} − ibr θ¨o

(13.11)

for the (additional) airspring state: (ksz + krz ) ksz krz θ˙r = − θr + θv + θb + θ˙b crz crz crz

(13.12)

Local track references were used, and both the translation and rotation of these reference axes associated with curves were allowed for in the equations (F ′′ , F¯ ′′ and T ′′ , T¯ ′′ . Moreover, (13.9) includes an end moment effect, F ′ = mv g(yv −yb ), modelling the roll effect of the body weight due to the lateral displacement of its centre of gravity on the curve. However, this effect was neglected in the case of the bogie mass owing to the high stiffness of the primary suspensions. The complexity of the system is clearly shown by the set of equations of motion. Note also that substantial coupling exists between the lateral and roll motions which result in two sway modes combining both lateral and roll movement, and their centres located at points other than the vehicle centre of gravity. An ‘upper sway’ mode, its node appears above the body c.o.g., with predominantly roll movement; and a ‘lower sway’ mode, its node located below the body c.o.g., characterised predominantly by a lateral motion. The modal analysis of the vehicle is shown in Table 13.1, with the modes being close to the industrial-norm. For system analysis and control design, the system is written in the state space form x˙ = Ax + Bu + Bw w

(13.13)

y = Cx + Hw

(13.14)

384

A.C. Zolotas and R.M. Goodall Table 13.1. ARB vehicle model dynamic modes Mode

Body lower sway Body upper sway Bogie lateral Bogie roll Airspring mode

Damping (%) Frequency (Hz)

16.5 27.2 12.4 20.8 100.0

0.67 1.50 26.80 11.10 3.70

where, T x = yv θv yb θb y˙v θ˙v y˙b θ˙b θr , u = [δa ], . . . T w = R−1 θo θ˙o θ¨o yo y˙o

(13.15) (13.16)

For simulation purposes only, disturbance signals θo , θ˙o , yo should be incorporated into the A matrix (in this case the stochastic track includes the filtering effects of the wheelset). It should be noted, that the necessary C and D output matrices, for control design, can be formed from the relevant rows (depending on the required outputs) of the given A and B matrices. A number of outputs is available such as displacements, velocities, accelerations of the vehicle body and bogies and also displacements, velocities of the active elements. A series of transient tests ensure that the vehicle behaves in a similar manner to its full scale equivalent (real) vehicle (for the passive model, with actuator inactive). The track profiles, both deterministic and stochastic, used for this purpose can be seen in Section 13.2.3. In this passive (non-tilting) case, a nominal vehicle speed of vo = 45 ms (162 km h ) is assumed, v2

and the designed cant deficiency at this speed is Ro − gθo = 5.83o or 1.0( sm2 ). Figure 10(a) shows the lateral acceleration and corresponding roll angle for the body mass. The lateral acceleration level is what the passengers would experience on the curve transition, and it is provided by a lateral accelerometer placed on the vehicle body c.o.g. The peak value is 13.0%g, while the steady-state value is around 11.93%g at a forward vehicle speed of 45(m/s). The increase in lateral acceleration is because the lateral suspension acts significantly lower than the body centre of gravity, and as a consequence the body roll-outwards on curves (steadystate 1.0o ). In the case of the bogie mass, Figure 10(b), the roll-out is less (steady state value of 0.43o ) due to the stiffer primary suspensions. The steady-state level of the bogie lateral acceleration of 10.7%g is closer to the cant deficiency for which the track was designed by the civil engineers. Note that more high frequency components are now present due to the harsh environment of the bogie system. Figure 10(c) shows a comparison between the lateral displacements of the vehicle body and bogie. The effect of the two different sets of suspensions is clearly evident. The soft secondary suspensions cause a displacement of 36.5(mm) in steady-state, while the primary suspensions owing to the high stiffness have hardly been displaced by 0.3(mm) in steady-state. It is also important to test the behaviour of the vehicle model on the straight track irregularities, which are the primary cause of ride quality degradation. Figure 10(d) presents the time histories for the lateral acceleration of both the body and the bogie travelling on straight track. The nominal vehicle speed is assumed 45(m/s). The bogie due to its harsh environment has

Modelling and Control of Railway Vehicle Suspensions

Lateral acceleration

Lateral acceleration 1.5

1.5

accel. (ms−2)

accel. (ms−2)

1

0.5

0

−0.5

0

200

400

600

800

1000

0

0

200

400

track (m)

Roll angle

800

1000

1200

800

1000

1200

1000

1200

1000

1200

0.2

0

−0.5

−1

200

0

400

600

800

1000

0

−0.2

−0.4

−0.6

1200

0

200

400

track (m)

600

track (m)

(a) Vehicle body

(b) Vehicle bogie

Secondary suspension

Body Lateral Acceleration

0.01

1

0

0.5 −2

accel. (ms )

deflection (m)

600

track (m)

Roll angle

angle (degrees)

angle (degrees)

1

0.5

−0.5

1200

0.5

−1.5

385

−0.01

−0.02

−0.03

−0.04

−1

200

0

400

−4

x 10

600

800

1000

0

200

400

600

800

track (m)

track (m)

Primary suspension

Bogie Lateral Acceleration 5

2.5

−2

0

−2

−4

−1.5

1200

accel. (ms )

deflection (m)

2

0

−0.5

0

200

400

600

800

1000

1200

track (m)

(c) Suspension deflections on det. track

0

−2.5

−5

0

200

400

600

800

track (m)

(d) Lateral acceleration on stochastic track @ 45(m/s)

Fig. 13.10. Vehicle body/bogie time histories @ 45(m/s)

an R.M.S. value of lateral acceleration equal to 16.4%g, while the soft secondary suspensions filter out a large amount of high frequencies and leave the body with an R.M.S. lateral acceleration of 2.932%g.

13.2.3 Tilt Control Requirements and Assessment Approach Requirements The performance of tilt control systems on the curve transitions is critical; most importantly the passenger ride comfort provided by the tilting vehicle should not be (significantly) degraded compared to the non-tilting vehicle speeds. The main objectives of any tilt control system are:

386

A.C. Zolotas and R.M. Goodall

1. to provide an acceptably fast response to changes in track cant and curvature (deterministic track features) 2. not to react significantly to track irregularities (stochastic track features) However, any tilt control system directly controls the secondary suspension roll angle (i.e., the inclination of the body) and not the vehicle lateral acceleration. Hence, there is a fundamental trade-off between the vehicle curve transition response and straight track performance. Moreover, for reasons of human perception, designers utilise partial tilt compensation. In such a case the passenger will still experience a small amount of acceleration on steady curve, in order to minimise motion sickness phenomena. From a control design point of view the objectives of the tilt system can be translated (in terms of shaping the loop) as: increasing the response of the system at low frequencies (deterministic track features), while reducing the high frequency system response (stochastic track features) and maintaining stability. Note that in this study the tilt controllers should provide 60% tilt compensation on steady curve at high speed (for reasons of passenger motion sickness as discussed earlier in the chapter), i.e., the curving acceleration at the increased speed is reduced to 40% of the non-tilting case (at the same speed) rather than to zero. Although this strictly is known as partial-nulling tilt, for simplicity we will refer to it as nulling tilt.

Tilt Control Assessment As mentioned previously, the performance of the tilt controller on the curve transition is critical, and their assessment is based upon a combination of two approaches, i.e., the ‘PCT factors’ and the ‘ideal tilting’ assessment. The former is based upon a comprehensive experimental/empirical study which provides the percentage of (both standing and seated) passengers who feel uncomfortable during the curve transition. The latter method emphasises the assessment of the control system performance by determining the deviations from the concept of “ideal tilting”, i.e., the tilt action follows the specified tilt compensation in an ideal manner, defined on the basis of the maximum tilt angle and cant deficiency compensation factor. This combination of parameters is optimised via the PCT factors approach to choose a basic operating condition. The procedure follows a minimisation approach of dynamic effects relative to tilt angles, roll velocities and lateral accelerations (Figure 13.11); more details can be found in [7]. •|y¨m − y¨mi |, the deviation of the actual lateral acceleration y¨m from the ideal lateral acceleration y¨mi , in the time interval between 1s before the start of the curve transition and 3.6s after the end of the transition. • θ˙m − θ˙mi , the deviation of the actual absolute roll velocity θ¨m from the ideal absolute roll velocity θ¨mi , in the time interval between 1s before the start of the curve transition and 3.6s after the end of the transition. Regarding the straight track case the ‘rule-of-thumb’ currently followed by designers is to allow a lateral ride quality degradation of the tilting train by no more than a specified margin of between 7.5% − 10% compared with the non-tilting vehicle.

Modelling and Control of Railway Vehicle Suspensions

387

Fig. 13.11. “Ideal Tilting”- Calculation of deviation of actual from ideal responses for acceleration and roll velocity [7]

Track Inputs The deterministic track input employed for simulation purposes consists of a curved section with a radius of 1000m superimposed by a maximum track cant angle of 155mm(i.e., 6o ), with a tilting speed of 209km/h over a track length of 1200m. The stochastic track inputs represent the irregularities in the track alignment on both straight track and curves, and these 2 m2 were characterised by an approximate spatial spectrum equal to Ωfl 3v ( cycle/m ) with a lateral track roughness {Ωl } of 0.33·10−8 m [12].

s

13.2.4 Conventional Tilt Control This section deals with the early-type conventional nulling control scheme and the current practice followed by industry i.e., command-driven with precedence approach.

Classical Nulling Control Strategy In this intuitively formulated control strategy (as seen from Figure 13.8a), which is a classical application of SISO feedback control, the signal from a body-mounted lateral accelerometer provides a measurement of the curving acceleration experienced by the passengers. The controller would drive the feedback signal to zero and therefore give 100% compensation, so a

388

A.C. Zolotas and R.M. Goodall

portion of the suspension roll angle (actual tilt angle) is included in the feedback path, chosen to provide the required 60% compensation. The control input comprises an angular displacement (δa ) provided by a rotary actuator in series with the anti-roll bar, which in turn provides a torque to the vehicle body. Note that the reference signal is ‘zero’ i.e., the system is subject to track disturbances only. The feedback ′ which is given by signal for partial tilt compensation is the effective cant deficiency angle θdm y¨vm ′ (13.17) θdm − λ2 θ2sr = −λ1 g where y¨vm is the lateral acceleration felt by the passengers as measured from an accelerometer on the body c.o.g (13.18), and θ2sr is the secondary suspension roll angle (13.19). v2 − g (θo + θv ) + y¨v R θ2sr = θv − θb y¨vm =

(13.18) (13.19)

Note the effect of the deterministic track included in (13.18). Factors λ1 , λ2 ensure partial tilt and for 60% compensation need to be set to 0.615 and 0.385 respectively, taking in account bogie roll-out in (13.19). Remark: The sign of the feedback signal is inverted for correct application of negative feedback based upon the chosen axes system. All lateral motions (y) are positive inwards to the curve center and all roll motions (θ ) are positive clockwise. However, the lateral accelerometer (a mass on a spring) measures positive acceleration outwards (13.18). Hence, the acceleration is translated into a positive cant deficiency angle⋆ , combined with a portion of the suspension roll, and is then fed back (note that λ1 y¨vm g ≥ λ2 θ2sr ). The combined signal will thus be a positive (w.r.t. the choice of directions) angle, which if fed back using negative feedback will cause the controller to provide a negative tilt angle (i.e., anti-clockwise rotation), consequently destabilising the system. Thus inversion of the sign of the feedback signal is necessary to guarantee proper body rotation for correct compensation. Conclusions about the stability of the closed-loop may be drawn by investigating the openloop frequency response of the system. The nominal open-loop frequency responses from ′ , and y := θ u := δa to y1 := θdm 2sr can be seen in Figure 13.12. Note that, while gain reduction 2 is required to stabilise the closed-loop system, the opposite applies in the case of fast tilt response. Hence, there must be a compromise between the tilt response and the ride quality. A closer investigation of the open-loop poles and zeros of the transfer function Gy1 u (s) = ′ θdm (s) δa (s)

reveals that, while the system is open-loop stable, is also non-minimum phase due to the existence of two RHP zeros at (s − 29.4) and (s − 6.02), Table 13.2. The pole-zero map of the uncompensated system Gy1 u (s) can be seen in Figure 13.13. The presence of RHP-zeros imposes a fundamental limitation on control, and high controller gains induce closed-loop instability. Integral action is necessary to guarantee zero effectivecant deficiency angle on steady curve.

We design a classical PI controller K pi (s) = kg 1 + s1τi , with the proportional term limiting the extra phase-lag at increasing frequency. The settings for kg and τi were adjusted to suffice both the deterministic and stochastic criteria, i.e., kg = 0.225, τi = 0.4 s (these can be slightly further tuned via optimization). Table 13.3 presents the overall controller assessment, while ⋆

Cant deficiency defines the difference between the existing degree of cant and the degree required to fully eliminate the effect of centrifugal force at maximum allowable speed.

Modelling and Control of Railway Vehicle Suspensions Actuator input to actual tilt (secondary suspension roll)

Actuator input to inferred tilt (effective cant deficiency) 40

40

0 dB

0 dB 0.25 dB

30

0.25 dB

30

0.5 dB

0.5 dB 1 dB

20

−1 dB

3 dB

10

−3 dB

6 dB 0

−6 dB

−10

−12 dB

−20 dB

−20

Open−Loop Gain (dB)

Open−Loop Gain (dB)

20

1 dB

−1 dB

3 dB

10

−3 dB

6 dB 0

−6 dB

−10

−12 dB

−20 dB

−20

−30

−30

−40 dB

−40

−40 dB

−40

−50

−50

−60 −540

389

−60 dB −495

−450

−405

−360

−315

−270

−225

−180

−135

−90

−45

0

−60 −360

−60 dB −315

−270

(a) Tilt model open-loop responses (Nichols)

−225

−180

−135

−90

−45

0

Open−Loop Phase (deg)

Open−Loop Phase (deg)

(b) Effective cant deficiency openloop

Fig. 13.12. Tilt angle open-loop Table 13.2. Open-loop poles and zeros of Gy1 u ( jω ) OL poles

OL zeros

-20.87 ± 167.34j -2.41 ± 125.94j -14.57 ± 68.38j 29.36 ± 0.00j -2.56 ± 9.03j -40.73± 0.00j -0.69 ± 4.12j -26.18± 0.00j -23.22± 0.00j 6.02± 0.00j -3.83 ± 3.13j Figure 13.14 and Fig. 13.15 present the simulation results on curved track and the correspond-

Table 13.3. Early-type nulling PI control assessment @ 58(m/s) D ETERMINISTIC - steady-state - R.M.S. deviation error - peak value Roll gyroscope - R.M.S. deviation - peak value (PCT-factor) - peak jerk level - standing - seated S TOCHASTIC Passenger comfort - R.M.S. passive (equiv.) - R.M.S. active - degradation Lateral accel. (actual vs ideal)

n/a 5.555 19.510 0.032 0.086 10.286 71.411 22.640

(%g) (%g) (%g) (rad/s) (rad/s) (%g/s) (% of passng) (% of passng)

3.778 (%g) 3.998 (%g) 5.802 (%)

390

A.C. Zolotas and R.M. Goodall Pole-Zero map of Gy1 u ( jω ) 200

150

100

non−minimum phase

Imag Axis

50

0

−50

−100

−150

−200 −50

−40

−30

−20

−10

0

10

20

30

Real Axis Fig. 13.13. Pole-zero map of uncompensated OL ing frequency responses respectively. Clearly the response is very slow with the steady-state values for acceleration, roll angle and roll rate profiles not met. Note that, due to the vehicle body inertia at the start and the end of curve, the body roll angle initially has an inverse response and then rises slowly up to the required steady-state value. The difference between the control input δa and the body roll θv is due to the torque imposed by the secondary suspension subject to the curving forces on the vehicle body. Thus, additional control effort is demanded to overcome this extra torque and regulate the body roll (via the suspension roll portion) to the desired value. There is also great difficulty in controlling the roll rates appropriately, which has a detrimental effect on the overall system performance. Note that the secondary suspension deflection is given by x2d f l = {yv − h1 θv − (yb + h2 θb )}.

Moreover, the complementary sensitivityi T , in Figure 15(b), shows that the control action h rads in the frequency range 0.4( s ), 4( rads s ) does not affect the system, which causes a slow response (i.e., insufficient bandwidth). Moreover, higher frequency components enter owing to h i rads the sway mode resonances in the interval 4( rads ), 10( ) , with the control action incapable s s

of improving the performance due to |S| > 1. The uncompensated and compensated open-loop responses are presented in Figure 15(a), with the latter emphasising the OL integral action at low frequencies. The difficulties in the early-type classical nulling strategy are the following: (i) the sensor exists within the control loop resulting in interactions between controller and suspension dynamics, and (ii) the system is also non-minimum phase and there is a stability threshold point set by the non-minimum phase zero, which result in major limitations in the controller design. Later on in the section we will see how we can improve the performance of nulling-type schemes with the use of modern control methods.

Command-driven with Precedence Control The problems with the early-type nulling schemes and also with heavy low pass filtering in attempting local command-driven strategies (to reduce the effect of bogie high frequency com-

Modelling and Control of Railway Vehicle Suspensions

Body roll gyroscope (abs roll rates) @ 58 m s

Body lateral acceleration @ 58 m s 2

6

passive active actual active "ideal"

1.5

4

1

2

deg rate ( s )

accel. ( m2 ) s

actual "ideal"

0.5

0

0

−2

−0.5

−4

−1

200

0

400

600

800

1000

−6

1200

0

200

400

track (m)

600

800

1000

1200

track (m)

(a) Passenger acceleration

(b) Roll rate

Tilt angles @ 58 m s

Secondary lateral suspension deflection

12

0.02

active command (u)

10

391

active ideal profile

0

active @ 58m/s passive @ 45m/s

8

deflection (m)

angle (degrees)

−0.02

6

active actual (θ

)

2sr

4

−0.04

−0.06

2

passive (θ2sr)

−0.08

0

−2

0

200

400

600

800

track (m)

1000

1200

−0.1

0

200

400

600

800

1000

1200

track (m)

(c) Tilt angles

(d) Suspension Deflection

Fig. 13.14. Early-type nulling PI scheme for deterministic track

ponents in the feedback signal used), led to the currently used by industry command-driven with precedence scheme (the concept can be seen in Figure 13.8(b)). For illustrative purposes we consider the command-driven with precedence scheme in Figure 13.16. It employs an accelerometer on the leading bogie of the leading vehicle (a preview of 29m was assumed) provided the curving acceleration signal, passed via a 0.45Hz low pass filter. The signal is then processed to provide 60% compensation for the lateral acceleration (K is set to 0.6). The actuator is controlled using an additional feedback of tilt angle (i.e., the suspension roll). For appropriate comparison to the local tilt strategies, the filter delay was chosen to match the precedence time, however this can be changed to emphasise precedence information if necessary. The transfer function of the LP filter is given by HLP2 (s) =

w2c2 , wc2 = 2π × 0.45, ζ2 = 0.707 s2 + 2ζ2 wc2 + w2c2

(13.20)

392

A.C. Zolotas and R.M. Goodall Bode Diagrams T , S, KS

Nichols Chart 20

40

0 dB

uncompensated compensated

0.5 dB 1 dB

−1 dB

3 dB

10

−3 dB

6 dB

0

−10

−20

K⋅S

S

T

−30

−40

0

−6 dB

−10

−12 dB

−50 −60 90

compensated GM = 3.15 at 7.84 (rad/s) PM = 40.58o at 4.95 (rad/s)

−20

−30

−40 dB

−40

−50

−60 −540

S

−20 dB

0

Phase (deg)

Open−Loop Gain (dB)

0.25 dB

uncompensated GM = −9.6 at 8.64 (rad/s) PM = −86o at 16.9 (rads/s)

20

Mag. (dB)

30

10

K⋅S

−90

T −180

−270

−360

−60 dB −495

−450

−405

−360

−315

−270

−225

−180

−135

−90

−45

0

Open−Loop Phase (deg)

−450 −2

10

−1

10

0

10

1

10

2

10

3

10

Frequency ( rads sec )

(a) Uncompensated and compensated OL

(b) Complementary Sensitivity, Sensitivity and Control Sensitivity

Fig. 13.15. Early-type nulling PI system frequency responses

and the time delay introduced is tdLP =

2ζ2 wc2 w2c2

, which for the current case is 0.5s. Thus for

the precedence to match the filter delay, it takes l = 58( ms ) × 0.5s = 29m precedence, i.e., approximately 1.5 vehicle length. Note that the tilting response for the leading vehicle will unavoidably be too late. 1 The leading vehicle controller was chosen as K pi (s) = 1 + s0.5 . For the trailing vehicle, the controller designed to actively tilt the body is a PI compensator in series with a low-pass filter (LPF)⋆⋆ , in order to remove high frequencies from the secondary suspension roll (those are introduced due to the bogie roll contribution). The overall controller transfer function is 1.5 + s0.75 400 Ktotal (s) = × 2 s0.5 s + 28.28s + 400 The compensated and uncompensated open loop together with the overall compensator frequency response can be viewed in Figure 17(a). The corresponding sensitivity and complementary sensitivity of the closed loop system are presented in Figure 17(b), where it is evident that the control action influences the system over a wider range of frequencies compared to the classical nulling case. A set of time-domain results for the deterministic track case is shown in Figure 13.18, where it is obvious that the precedence scheme is superior to the classical nulling approach. The tilt controller performance is presented in Table 13.4, and it is closer to the ideal performance as expected in all cases. In the stochastic case there is an improvement in ride quality by the active system, because the precedence time matches the filter delay meaning that the reference and track input are uncorrelated, thus the tilt command will compensate for long wavelengths.

⋆⋆

The LP filter will be redundant in the case of actuator dynamics, i.e., actuator will have limited bandwidth by default.

Modelling and Control of Railway Vehicle Suspensions

393

Precedent Vehicle LPF

K/g

0.45Hz

Track

r+

e

Controller K2(s)

-

δa u

tilt command

Vehicle Dynamics Local Vehicle

Track

..

..

yvm

e

r+ -

θ2sr

Controller K1(s)

δa u

suspension roll

y bm

Vehicle Dynamics Precedent Vehicle

Local Vehicle (current)

.. yvm

θ2sr suspension roll

precedence effect

tilt command

K/g

LPF

29.0m

0.45Hz

(a) Block Diagram

direction of travel 1.5 Vehicle length preview (~29m) Leading

Trailing Tilt comm.

vehicle body

xxxxxxxxxxxx xxxxxxxxxxxx xxxxxxxxxxxx

tilt mechanism

xxxxxxxxxxxx xxxxxxxxxxxx

xxxxxxxxxxxx xxxxxxxxxxxx

xxxxxxxxxxxx xxxxxxxxxxxx

xx

bogie

bg accel

speed=58(m/s)

K/g

-s0.5 e

Low-Pass 0.45Hz Track Data

for simulation

(b) Interpretation for simulation purposes Fig. 13.16. Command-driven with precedence approach

Emphasising more precedence information improves the deterministic performance, subject of course to the amount of precedence used, i.e., too much precedence (over-precedence) can be disastrous for the normal operation of the train (tilt action will apply on straight track segments much sooner than the intended start of the curve!). In addition, the amount of precedence used will influence the stochastic ride quality either in a positive or negative way depending upon the correlation of the signals (in the case where the precedence time differs from the filter delay, the reference and track input signals are no longer uncorrelated). It should be noted that, even in the precedence schemes, sensors located on each vehicle (i.e., local sensors) are used to ensure the correct operation of the overall tilting system (the sensors are always present for safety purposes).

A.C. Zolotas and R.M. Goodall

394

Bode Diagrams OL (δa ) to θ2sr

Bode Diagrams T , S

Mag. (dB)

0

Mag. (dB)

50

uncompens compensator compensated

50

−50

0

T S

−50

−100 −100

−150 −2 10

−1

10

0

1

10

2

10

10

3

10

−150 0

100

Phase (deg)

Phase (deg)

S −90

−180

−270

0

−100

T

−200

−300 −360 −1

10

0

10

1

10

Frequency ( rads sec )

2

10

(a) Compensated and uncompensated OL

−400 −2 10

−1

10

0

1

10

10

Frequency ( rads sec )

2

10

3

10

(b) Complementary Sensitivity, Sensitivity

Fig. 13.17. Command-driven with precedence system frequency responses Table 13.4. Command-driven with precedence - assessment @ 58(m/s) D ETERMINISTIC - steady-state - R.M.S. deviation error - peak value Roll gyroscope - R.M.S. deviation - peak value (P-factor) - peak jerk level - standing - seated S TOCHASTIC Passenger comfort - R.M.S. passive (equiv.) - R.M.S. active - degradation Lateral accel. (actual vs ideal)

9.53 1.54 12.18 0.018 0.104 6.80 47.62 13.455

(%g) (%g) (%g) (rad/s) (rad/s) (%g/s) (% of passng) (% of passng)

3.78 (%g) 3.31 (%g) -12.12 (%)

13.2.5 Nulling-Type Tilt Via Robust Control Techniques In this section we will see how to improve the performance nulling-type schemes by employing robust control methods, in order to achieve comparable results to the command-driven with precedence scheme. In particular the case study involves an LQG/LTR approach [11] and a multiple objective H∞ /H2 scheme [12].

LQG/LTR Nulling-Type Tilt Control Linear Quadratic Gaussian control is well documented in [8], [10], defining the following state-space plant model

Modelling and Control of Railway Vehicle Suspensions

395

Body roll gyroscope (abs roll rates) @ 58 m s

Body lateral acceleration @ 58 m s 1.2

6

actual "ideal"

active actual active "ideal" prof passive actual

1 4

0.8

deg rate ( s )

accel. ( m2 ) s

2

0.6

0.4

0

0.2 −2

0 −4

−0.2

−0.4

200

0

400

600

800

1000

−6

1200

0

200

400

track (m)

600

800

1000

1200

track (m)

(a) Passenger acceleration

(b) Roll rate

Tilt angles @ 58 m s

Secondary lateral suspension deflection

14

0.01

θ passive actual v θv active actual

12

0

u command "ideal" profile

−0.01

10

−0.02

deflection (m)

angle (degrees)

8

6

4

−0.03

−0.04

−0.05

2 −0.06

0

−0.07

−2

−0.08

active @ 58m/s passive @ 45m/s −4

0

200

400

600

800

1000

1200

−0.09

0

track (m)

200

400

600

800

1000

1200

track (m)

(c) Tilt angles

(d) Suspension Deflection

Fig. 13.18. Command-driven with precedence on deterministic track

x˙ = Ax + Bu + Γ w

(13.21)

y = Cx + v,

(13.22)

where w, v are (ideally) white uncorrelated process and measurement noises which excite the system, and are characterised by covariance matrices W,V respectively. The separation principle can be applied to first find the optimal control u = −Kr x which minimises (13.23) Z T 1 J = lim E [xT Qx + uT Ru]d τ , (13.23) T →∞ T 0 where Kr = R−1 BT X and X is the positive semi-definite solution of the following Algebraic Riccati Equation (ARE) I A −BR−1 BT X −I =0 (13.24) T X −Q −A

396

A.C. Zolotas and R.M. Goodall

Next find the optimal state estimate xˆ of x where xˆ = Axˆ + Bu + K f (y −Cx) ˆ (13.25) n o to minimise E [x − x] ˆ T [x − x] ˆ . The optimal Kalman gain is given by K f = YCT V −1 and Y is the positive semi-definite solution of the following ARE AT −CT V −1C I Y −I =0 (13.26) Y −Γ W Γ T −A

Weighting matrices Q (pos. semidefn.), R (pos. defn.) for control, and W (pos. semidefn.), V (pos. defn.) for estimation, can be tuned to provide the desired result. Note that it is also possible to follow the dual procedure, i.e., solve for the state estimate sub-problem and next for the optimal gain sub-problem (although this is not considered in our study). For synthesizing the tilt controller we consider a simple extension to the classical nulling ′ , with approach in an optimal control framework by deriving the SISO model, from δa to θdm all disturbance signals set to zero. Thus, on measurement is effectively use for the Kalman filter. We will synthesize the controller using the weighting matrices Q, R, W, V purely as tuning parameters until an appropriate design is obtained. In particular, the structure of the LQG tilt compensator is found by shaping the principal gains of the system, i.e., return ratios. First the LQR is synthesized via Q, R to obtain a satisfactory a satisfactory return ratio −Kr (sI − A)−1 B, with the Kalman Filter designed via W,V such that the return ratio at the input of the compensated plant converges sufficiently close to −Kr (sI − A)−1 B over the frequency range of interest (Loop Transfer Recovery to recover as much as possible of the robust properties of LQR). For disturbance rejection and/or reference tracking (which is zero in this case), the system ′ . should be augmented using an extra state, the integral of the effective cant deficiency θdm This approach will produce an optimal controller with integral action [8]. Hence, the system is becomes x˙ A 0 x B = + u (13.27) x˙′ C′ 0 x′ 0 R

′ and C′ is the selector matrix for integral action and is found from θ ′ = C′ x. where x′ = θdm dm The control signal has the form x u = − K p Ki (13.28) x′

We start with the simplest possible choices for Q, R 0 0 Q = 9×9 9×1 , R = 1 01×9 qi

(13.29)

thus adjusting only the weight of the integral state, and imposing no constraints on the remaining states. Figure 13.19 illustrates the return ratio −Kr (sI − A)−1 B for various qi . We choose the return ratio for qi = 100 with a crossover of approx 20rad/s, to recover for in the next steps. However a simple calculation of the transmission zeros for the design plant reveals a non-minimum phase zeros at approximately 6.0rad/s (this

Modelling and Control of Railway Vehicle Suspensions

397

is characteristic for such a setup in tilting trains [12]), thus making full recovery cumbersome. For illustration, the usual LTR procedure is followed up to the limit of recovery allowed from the non-minimum phase zero (usually the achievable bandwidth of the system is less than half of the RHP zero frequency [10]).

−1

K (sI−A) B; q =[.1 1 5 10 100 500 1e3] r

i

60

Singular Values (dB)

40

20

0

−20

−40

increasing q

i

−60

−80 −1 10

0

10

1

10 Frequency (rad/sec)

2

10

3

10

Fig. 13.19. Return ration −Kr (sI − A)−1 B for various qi Moreover, for the design of the Kalman Filter, the zero eigenvalue of the augmented A matrix needs to be placed just to the left of the origin for the solution to exist. For controller implementation this should move back to the origin for proper integration. equal to B, 1 respectively (still for the SISO model). Setting Γ = B refers to any (virtual) disturbances on the plant acting via the input, rather than the actual track disturbances from the track. The sensor noise covariance matrices is set to V = 1 (can be reduced to characterise better quality measurements), while the process noise covariance matrix is set to W = Wo + wI ˜ (Wo = 0). Figure 13.20 illustrates the amount of recovery at plant output for increasing values of w. ˜ It is seen that there is no point in recovering after w˜ = 500 as appropriate integral action is already recovered and the actual crossover limit is placed by the non-minimum phase zero. Figure 13.21 presents the principal gains of the designed closed loop system w.r.t. sensitivity S and complementary sensitivity T . The bandwidth of the system is rather low (approx. 1rad/s due to the NMP zero) however there is some compensation at higher frequencies which cater for few stochastic components (on straight track ride). The synthesized LQG controller realization is given by s A − BKr − K f C K f Klqg = −Kr 0

(13.30)

398

A.C. Zolotas and R.M. Goodall

LQG/LTR @ input for w=[.1 1 5 10 100 500 1e3] 200

150

Singular Values (dB)

100

50

−1

K (sI−A) B r

0

−50

increasing

K (s)G(s) lqg

−100

increasing

−150

−200 −5 10

0

5

10 Frequency (rad/sec)

10

Fig. 13.20. LTR at plant output for increasing w˜ and is 10th order and its frequency response can be seen in Figure 13.22. This can be further reduced either in an open loop or closed sense [10], however at the expense of performance quality.

Table 13.5. LQG nulling-type control assessment @ 58(m/s) D ETERMINISTIC - steady-state - R.M.S. deviation error - peak value Roll gyroscope - R.M.S. deviation - peak value (PCT-factor) - peak jerk level - standing - seated S TOCHASTIC Passenger comfort - R.M.S. passive (equiv.) - R.M.S. active - degradation

Lateral accel. (actual vs ideal)

9.53 4.09 17.3 0.033 0.101 8.98 65.76 19.95

(%g) (%g) (%g) (rad/s) (rad/s) (%g/s) (% of passng) (% of passng)

3.78 (%g) 3.97 (%g) 4.83 (%)

The time domain results for the lateral acceleration felt by the passengers and the related body tilt angle can be seen in Figure 13.23, while the performance assessment of the controller is presented in Table 13.5. It is seen that although the LQG-based is a simple straightforward

Modelling and Control of Railway Vehicle Suspensions

399

Principal Gains 10

Singular Values (dB)

0

−10

−20

sensitivity

compl. sensitivity

−30

−40

−50

−60 −3 10

−2

10

−1

10

0

1

10 10 Frequency (rad/sec)

2

10

3

10

Fig. 13.21. Designed system Sensitivity and Compl. Sensitivity with LQG controller optimal extension of the classical nulling scheme, the performance is much improved (emphasizing robustness with the additional damping injected). It is worth mentioning that the performance of the controller can be further improved by using extra sensor information, i.e., passenger acceleration, body roll rate, vehicle yaw rate [13]. The design system in this case will be non-square with more outputs than inputs, as a result the LQG controller will be non-square (more inputs than outputs). However, we can still synthesize via the separation principle, but make the system square for LTR [8].

13.2.6 Multi-objective H∞ /H2 Nulling-Type Control Via LMIs It is well known that the H∞ norm of a system represents the worst-case energy transfer between (bounded energy) disturbances to (bounded-energy) regulated outputs, and as a result can be conservative when disturbances are naturally modelled as persistent or white noise signals. In cases where the interests falls upon minimising the RMS value of a regulated output, the H2 norm [10] of the corresponding closed-loop transfer function is a more appropriate measure of stochastic performance. This section considers a multiple-objective H∞ /H2 via LMIs [14] design method for the local nulling-type tilt control [15]. Details on some preliminary concepts related to the following design procedure can be found in Appendix B. The design objectives are formulated as an optimisation problem, defined in the generalisedregulator setting shown in Figure 13.24, where P(S) and K(s) are the generalised plant (inclusive of all weighting factors) and the controller to be designed. The vector of external disturbances was set to w = [w1 w2 ]T , where w1 denotes ( R1 ) the (deterministic) track-curvature (low-frequency) disturbance signal and w2 is the (stochastic) lateral track position (higherfrequency) signal yo . Scaling factors Wi1 ,Wi2 emphasise the relative weighting between the

400

A.C. Zolotas and R.M. Goodall Singular Values 60

Singular Values (dB)

40

20

0

−20

−40

−60 −4 10

−2

10

0

2

10 10 Frequency (rad/sec)

4

10

Fig. 13.22. LQG controller principal gain

two disturbances for the design. Note that in this case we choose to employ two outputs y1 and y2 which are the measured body lateral acceleration and the secondary suspension roll angle, respectively (chosen for the output vector to replicate the sensors used in the basic classical nulling control). It is worth noting that the system transfer function with the two aforementioned measurements is not NMP. It is very important to meet both deterministic (curve track) and stochastic (straight track) requirements, thus the following multi-objective optimisation problem was formulated min α kW1 Tz1 w k2∞ + β kW2 Tz2 w k22

K∈S

(13.31)

in which S denotes the set of all internally stabilising controllers. The first regulated output ′ . For z1 for infinity-norm minimisation, was chosen as the effective cant deficiency z1 = θdm the minimisation of the 2-norm, z2 was chosen as the control input u denoting the actuator roll angle δa . Regulating z1 to zero corresponds to 60% tilt compensation and thus attains the desired (steady-state) level of acceleration on steady state curve. Tzi w (i = 1, 2) denotes the (closed-loop) transfer functions between signals w and z1 , z2 respectively. Multi-objective optimisation typically refers to the joint optimisation of a vector consisting of two or more functions, typically representing conflicting objectives. Common types of multi-objective optimisation problems include “Pareto-optimal” (non-inferior) optimality criteria, minimax optimality criteria, etc. In the context of this exercise, the term “multi-objective” refers simply to the fact that the cost function of the optimisation problem involves two different types of norms, capturing the deterministic and stochastic objectives of the design. The two different norms

Modelling and Control of Railway Vehicle Suspensions Body acceleration 2

actual ideal

acceleration (m/s2)

1.5

1

0.5

0

−0.5

−1

0

200

400

600

800

1000

1200

track (m)

(a) Lateral acceleration

Body roll (tilt) 9

actual ideal

8

7

tilt (deg)

6

5

4

3

2

1

0

−1

0

200

400

600

800

1000

1200

track (m)

(b) Tilt angle Fig. 13.23. Lateral acceleration and tilt profile on curved track with LQG that are used here are the 2-norm and the infinity-norm. Thus, typical examples of multi-objective problems in our context include:

401

402

A.C. Zolotas and R.M. Goodall

z1 z2

W1

Wi1

W2

Wi

P

w1 w2

2

u y=[y1 y2]T

K Fig. 13.24. Generalised Regulator configuration for multi-objective control

1. Constrained w minimisation: w Minimise wW1 Txy w2 subject to kW2 Tzw k∞ < γ , 2. Unconstrained w minimisation: w Minimise β wW1 Txy w2 + α kW2 Tzw k∞ , and 3. Feasibility problem: Find a stabilising K(s) (if one exists) such that w w wW1 Txy w ≤ γ1 and kW2 Tzw k ≤ γ2 ∞ 2 Txy and Tzw represent two general closed-loop transfer functions, weighted via W1 and W2 . Scalars α and β , in (13.31), are positive definite design parameters which may be used to shift the emphasis of the optimisation problem between the minimisation of the kTz1 w k∞ term (deterministic objective) and the kTz2 w k2 term (stochastic objective). The frequency-domain weights W1 and W2 have been chosen as: W1 (s) = 104 W2 (s) = 0.5

s 200 + 1 s 0.0001 + 1 s3 + 1.59s2 + 0.58s + 0.06

s3 + 13.81s2 + 38.4s + 2.98

(13.32) (13.33)

W1 is essentially a low-pass filter with a very low pole cut-off frequency (10−4 rads s ) and high gain at low frequencies, Figure 25(a). Thus W1 emphasises minimisation of the kTz1 w k∞ term in the low frequency range and effectively enforces integral control on the regulated output rads (z1 ). W2 is a high-pass filter with pole (10 rads s ) and zero (0.2 s ) cut-off frequencies. A rads rads lead/lag network is also included in W2 , in the range of [.1 s , 6 s ], which found to have a positive effect on controller design (by enhancing the cross-over frequency of W1 ,W2 ), Figure 25(a). By limiting the high-frequency components of the control input (z2 ), effectively places a limit on the closed-loop bandwidth of the system, which in turn limits the RMS acceleration on straight track (stochastic case). Additional benefits include a smoother control signal and improved robustness properties of the controller when the effects of uncertainty in P(s) and

Modelling and Control of Railway Vehicle Suspensions

403

in the actuator dynamics are taken into account. Moreover, the relative weighting between w1 and w2 were simply set to unity, i.e. W 0 10 Wi = i1 = (13.34) 0 Wi2 01 Thus the energy of either of the signals is equally incorporated in the cost function. Increasing either Wi1 or Wi2 with respect to the other will put more emphasis on the deterministic or the stochastic track respectively. However, the current choice of Wi provides the best results.

Bode diagram

Singular values

60

10

W1 (s)

0

20

−10

mag (dB)

mag (dB)

40

W2 (s) 0

−20

−20

−30

−40

−40

−60 −3 10

−2

10

−1

10

0

10

1

2

10

Frequency (rads/s)

10

3

10

−50 −3 10

(a) Weighting filters

−2

10

−1

10

0

10

1

10

Frequency (rads/s)

2

10

3

10

(b) Controller principal gains

Fig. 13.25. Multi-objective H∞ /H2 LMI approach scheme R The minimisation problem in (13.31) was solved in MATLAB using the LMI toolbox [14], i.e., representing the problem in a set of Linear Matrix Inequalities and follow a convex optimisation approach. This technique has very attractive computational properties and is widely used in systems and control theory. For controller design, the generalised plant was formulated as follows

x˙ = Ax + B1 w + B2 u z∞ = C∞ x + D∞1 w + D∞2 u

(13.35)

z2 = C2 x + D21 w + D22 u

(13.36)

y = Cy x + Dy1 w

(13.37)

where all matrices can be formed based upon the state space model of the system and the specifications in the generalised plant discusses earlier in this section. The controller can be R then found by using MATLAB function hinfmix(). The optimisation problem was solved for a few combinations of the α and β parameters and the results can be seen in Table 13.6. The results shown in the table clearly illustrate the fundamental trade-off between the deterministic and the stochastic objectives of the design. As expected, increasing the value of β relative to α places more emphasis on the stochastic aspects of the design, and as a result the RMS acceleration on straight track is further

404

A.C. Zolotas and R.M. Goodall Table 13.6. α -β combinations for the H∞ /H2 problem

α β Ride Quality-Degrad.(%) Deviations-Determ.(%g) 1 1 21.7 1.95 1 2.5 10 2.15 1 5 4.95 2.37 1 10 3.4 2.62 1 20 2.1 2.9 Ride Quality-Degrad.: ride-quality degradation @58m/s of active system compared to passive @58m/s (straight track) Deviations-Determ.: RMS acceleration deviation from the ideal response of an ideal tilting controller @58m/s (curved track)

reduced. This is at the expense of deterministic performance and, therefore, the curved track response becomes slower (larger deviations from the ideal tilt response). Since it is required that stochastic performance deteriorates by no more than 7.5% compared to the passive system, R the “best” design was obtained for α = 1 and β = 5. The result returned in MATLAB for the “best” configuration is shown below (summary)

Optimization of 1.000 * G^2 + 5.000 * H^2 : ---------------------------------------------Solver for linear objective minimization under LMI constraints Iterations : Best objective value so far 1 . . . 16 * switching to QR 17 . . . . 35 36 1.018621e+005 37 1.018621e+005 38 6.830550e+004 39 6.830550e+004 40 5.029532e+004 41 5.029532e+004 42 4.475980e+004 43 4.226245e+004 *** new lower bound: 3159.072336 44 4.067042e+004 45 3.932891e+004 .

Modelling and Control of Railway Vehicle Suspensions

405

. . 62

3.493807e+004 new lower bound: 3.437031e+004 63 3.488358e+004 *** new lower bound: 3.441555e+004 64 3.483742e+004 *** new lower bound: 3.450646e+004 Result: feasible solution of required accuracy best objective value: 3.483742e+004 guaranteed relative accuracy: 9.50e-003 f-radius saturation: 88.219\% of R = 1.00e+008 Guaranteed Hinf performance: 1.54e+002 Guaranteed H2 performance: 4.68e+001 ***

Note that, in the first few iterations, the algorithm does not find any solutions, however the solution converges soon after. The resulting controller is of 2-input/1-output dimension due to the two measurements used in the formulation. The singular value plot is shown in Figure 25(b), by definition depicting the largest singular of the two output/one input system transfer function. Note that the controller order is equal to 13, i.e., 9 states from the train model + 4 states from the weights. However, it can be easily reduced down to a 7th order equivalent, e.g., via balanced truncation [10], with minimal degradation in performance. The performance of the designed system is assessed in Table 13.7, where it can be seen that it is significantly improved compared to the classical nulling and LQG nulling-type control schemes. This exercises illustrate the usability of employing two measurements (compared to only one in the LQG scheme) and the effectiveness of distinguishing the design objectives in the cost function. For completeness the associated time history analysis for the design track is presented in Figure 13.26.

Table 13.7. H∞ /H2 multi-objective LMI approach @ 58(m/s) D ETERMINISTIC - steady-state - R.M.S. deviation error - peak value Roll gyroscope - R.M.S. deviation - peak value (PCT-factor) - peak jerk level - standing - seated S TOCHASTIC Passenger comfort - R.M.S. passive (equiv.) - R.M.S. active - degradation

Lateral accel. (actual vs ideal)

9.53 2.37 13.66 0.023 0.101 7.07 51.7 14.93

(%g) (%g) (%g) (rad/s) (rad/s) (%g/s) (% of passng) (% of passng)

3.78 (%g) 3.96 (%g) 4.95 (%)

406

A.C. Zolotas and R.M. Goodall Body roll gyroscope (abs roll rates) @ 58 m s

Body lateral acceleration @ 58 m s 0.15

1.4

actual

ideal

1.2

0.1 1

0.05

rate ( rads s )

accel. ( m2 ) s

0.8

0.6

ideal 0.4

actual

0

−0.05

0.2

0

−0.1

−0.2 −0.15

−0.4

−0.6

0

200

400

600

800

1000

−0.2

1200

0

200

400

600

track (m)

800

1000

1200

track (m)

(a) Passenger acceleration

(b) Body Roll rate Tilt angles @ 58 m s

0.16

ideal 0.14

0.12

angle (rads)

0.1

0.08

actual 0.06

0.04

0.02

0

−0.02

0

200

400

600

800

1000

1200

track (m)

(c) Tilt angles Fig. 13.26. Design track time history results for H∞ /H2 design

13.2.7 Case Study Remarks This exercise has considered the design of local tilt controllers (a form of secondary railway suspension control) based upon advanced control concepts. The problems with earlytype classical nulling approaches has been presented, and briefly discussed the currently-used precedence strategy. It has been shown that by using modern control methods, i.e., LQG and H∞ based schemes, the performance of nulling-type controllers can be significantly improved. Problems with the two model-based schemes include high-order controller size and the choice of weighting functions. Controller reduction can be employed for the former, while the latter requires a realistic setup of the design problem (to reduce the complexity of choosing the structure of the weights) and usually designer experience (this being the case for the majority of engineering applications).

Modelling and Control of Railway Vehicle Suspensions

407

13.3 Appendix A- Tilting Train Parameter Values and Notation

yv , yb , yo θv , θb , θr δa θo , R v mv ivr mb ibr g kaz ksz krz crz ksy csy kvr kpz cpz kpy cpy d1 d2 h1 h2 h3 hg2 hg1

Lateral displacem. of body, bogie, track Roll displacement of body, bogie, airspring reservoir ARB applied tilt Track cant, curve radius Vehicle forward speed Half body mass, 19,000(kg) Half body roll inertia, 25,000(kgm2 ) Bogie mass, 2,500(kg) Bogie roll inertia, 1,500(kgm2 ) gravitational acceleration, 9.81(ms−2 ) — Values per bogie side — Airspring area stiffness, 210,000( N m) Airspring series stiffness, 620,000( N m) Airspring reservoir stiffness, 244,000( N m) Airspring reservoir damping, 33,000( Ns m) Secondary lateral stiffness, 260,000( N m) Secondary lateral damping, 33,000( Ns m) Nm Anti-roll bar stiffness/bogie, 2,000,000( rad ) N Primary vertical stiffness, 2,000,000( m ) Primary vertical damping, 20,000( Ns m) Primary lateral stiffness, 35,000,000( N m) Primary lateral damping, 16,000( Ns m) Airspring semi-spacing, 0.90(m) Primary vertical suspension semi-spacing, 1.00(m) 2ndary lateral susp. height(body cog), 0.9(m) 2ndary lateral susp. height(bogie cog), 0.25(m) Primary lateral susp. height(bogie cog), -0.09(m) Bogie cog height(rail level), 0.37(m) Body cog height(rail level), 1.52(m)

13.4 Appendix B- H∞ Based Controllers: Preliminaries 13.4.1 Basic Notation A continuous time, linear time invariant, state space system is given by x(t) ˙ = Ax(t) + Bu(t)

(13.38)

y(t) = Cx(t) + Du(t)

(13.39)

p×n

p×m

n×m , C ∈ R and D ∈ Re . The above state space system is charwhere A ∈ Rn×n e e , B ∈ Re acterised by the following transfer function with dimension p × n

408

A.C. Zolotas and R.M. Goodall

G(s) = C(sI − A)−1 B + D

(13.40)

This thesis adopts the following conventional state-space representation to represent G(s) A B s G(s) = (13.41) CD Note also that the complex conjugate of G(s) is given by T T s −A −B G∗ (s) = GT (−s) = CT DT

(13.42)

and if G(s)G∗ (s) = I = G∗ (s)G(s) for all s ∈ jRe , then G(s) is said to be all-pass. Finally if matrix D is invertible, then G−1 (s) is given by⋆⋆⋆ −1 −1 s A − BD C BD −1 G (s) = −D−1C D−1

(13.43)

• Frequency Domain Spaces and Norms This part introduces the meaning of frequency domain spaces and norms of real rational, matrix valued, transfer functions. For a more comprehensive study the reader is referred to Zhou and Doyle [9]. Let R denote the space of all real rational transfer function matrices. The L2 /H2 norm of G(s) is given by r Z 1 ∞ kGk2 , (13.44) tr (G∗ ( jω )G( jω )) d ω 2π −∞ which is used to define the following spaces (i). RL 2 refers to the space of all real rational transfer function matrices with no poles on the imaginary axis and is characterised by a finite L2 norm. (ii). RH 2 defines the space of all transfer function matrices in RL 2 with no poles in Re(s) > 0. The L∞ /H∞ norm of G(s) is given by kGk∞ , sup σ [G( jω )] ω ∈Re

(13.45)

and (i). RL ∞ refers to the space of all real rational transfer function matrices with no poles on the imaginary axis (with finite L∞ norm). (ii). RH ∞ defines the space of all transfer function matrices in RL ∞ with no poles in Re(s) > 0. ⋆⋆⋆

Using the matrix inversion lemma: −1 −1 −1 −1 −1 (A1 + A2 A3 A4 )−1 = A−1 1 − A1 A2 (A4 A1 A2 + A3 ) A4 A1 .

Modelling and Control of Railway Vehicle Suspensions

409

Also, the H∞ norm of a stable transfer function G(s) is its largest input/output RMS gain kGk∞ , sup

u∈L2 u6=0

kykL2 kukL2

(13.46)

where L2 is the space of signals having finite energy and y is the output of the system G for a given input u. Thus, for any input u of unit energy, the output energy in y is bounded by the H∞ norm of G(s).

• Linear Fractional Transformations The basic concept of Linear Fractional Transformations is outlined in this section. Linear Fractional Transformations (LFT) are frequently used in the area of H∞ optimisation as well as in other areas of control theory. They do represent a means of standardising a wide variety of feedback arrangements [10]. Let P(s) define a transfer function matrix with the following state-space representation   A B1 B2 s  P(s) = C1 D11 D12  (13.47) C2 D21 D22

which can be also partitioned as

s

P(s) = where

P11 P12 P21 P22

(13.48)

Pi j (s) = Ci (sI − A)−1 B j + Di j

z(s) ¾

(13.49)

¾ w(s)

P (S) ¾

¾

y(s)

u(s) - K(S)

-

Fig. 13.27. The Generalised Regulator Configuration Referring to Figure 13.27, which presents the generalised regulator configuration, the (lower) linear fractional transformation† of P and K is given by †

There is also the concept of the upper LFT which is employed in representing uncertainties in a system [10].

410

A.C. Zolotas and R.M. Goodall FL (P, K) , P11 + P12 K(I − P22 K)−1 P21

(13.50)

for det(I −P22 K) 6= 0. P(s) represents the “generalised plant”, consisting of the nominal model G(s) combined with all frequency weightings appropriately chosen to shift the emphasis with frequency between different design objectives. In addition, the signals are: u the control variables, w the exogenous inputs such as disturbances wd and commands r, y the measured variables and z the regulated outputs, i.e., the signals need to minimise to meet the design objectives. In fact FL (P, K) represents the transfer function between w and z in Figure 13.27, i.e. h i z(s) = P11 + P12 K(I − P22 K)−1 P21 w(s) (13.51) H∞ and H2 control implies the minimisation of the H∞ -norm and the H2 -norm of FL (P, K) respectively.

Modelling and Control of Railway Vehicle Suspensions

411

References 1. S. Iwnicki (Ed.) (2006) Handbook of Railway Vehicle Dynamics. Taylor and Francis 2. A. Wickens (2003) Fundamentals of Rail Vehicle Dynamics. Swets and Zeitlinger 3. R.M. Goodall RM and S. Brown (2001) Tilt technology still evolving as the cost falls. Railway Gazette International. 521–525 4. R.M. Goodall (1999) Tilting Trains and Beyond – the Future for Active Railway Suspensions: Part 1 Improving Passenger Comfort. Computing and Control Engineering Journal. August 153–159 5. R.M. Goodall, S. Bruni S and T.X. Mei (2005) Concepts and prospects for activelycontrolled railway running gear. Procs 19th IAVSD Symposium. Milano 6. J.T. Pearson, R.M. Goodall and I. Pratt (1998) Control System Studies of an Active AntiRoll Bar Tilt System for Railway Vehicles. Proceedings of the Institution of Mechanical Engineers Part F. 212(F1):43–60 7. R.M. Goodall, A.C. Zolotas and J. Evans (2000) Assessment of the Performance of Tilt System Controllers. Proceedings of the Railway Technology Conference IMechE. C580/028/2000. Birmingham UK. November 231–239 8. J.M. Maciejowski (1989) Multivariable Feedback Design. Addison-Wesley 9. K. Zhou and J.C. Doyle (1998) Essentials of Robust Control. Prentice-Hall 10. S. Skogestad and I. Postlethwaite (2000) Multivariable Feedback Control: Analysis and Desgn. Wiley and Sons 11. A.C. Zolotas, G.D. Halikias, R.M. Goodall and J. Wang (2006) Model Reduction Studies in LQG Optimal Control Design for High-Speed Tilting Railway Carriages. Proceedings of the 2006 American Control Conference. Minneapolis MN USA. June 1796–1801 12. A.C. Zolotas (2002) Advanced Control Strategies for Tilting Trains. PhD Thesis, Loughborough University, UK 13. A.C. Zolotas and R.M. Goodall (2005) Improving the tilt control performance of high-speed railway vehicles: an LQG approach. Proceedings of the 16th IFAC World Congress. Prague 14. P. Gahinet, A. Nemirovski, A.J. Laub and M. Chilali (1994) LMI Control Toolbox. Natick, MA, The MathWorks. 15. A.C. Zolotas, G.D. Halikias and R.M. Goodall (2000) A Comparison of Tilt Control Approaches for High Speed Railway Vehicles. Proceedings of the 14th International Conference on Systems Engineering ICSE 2000, Coventry UK. 2:632–636

14 Case Study on Anti-w indup Compensation Micro-actuator Control in a Hard-D isk Drive Guido Herrmann, Matthew C. Turner, and Ian Postlethwaite

Summary. This chapter demonstrates the use of anti-windup compensation in the control loop of a micro-actuator which is nominally controlled by a linear, discrete robust controller. The micro-actuator is part of a hard disk drive dual-stage servo-control system for positioning of the read/write head. The actuator inputs are constrained to retain the micro-actuator’s displacement range of less than 0.4 µ m for mechanical protection. In the first part of the chapter, the anti-windup compensation scheme exemplifies an approach suggested by Weston & Postlethwaite [29]. Here, the scheme is posed as a discrete full-order compensator and the closed loop analysis uses a generalized circle citerion approach. The design of the compensator is posed in LMI-form. In the second part of the chapter, it is shown how the linear micro-actuator control loop with anti-windup compensation is incorporated into the non-linear servo-control scheme for positioning of the read/write head in a hard disk drive.

14.1 Introduction Micro-actuators have been gaining in importance in practical systems during recent years. For instance, it is nowadays a target to integrate the actuator, sensor and associated electronics for powerful computations into one device of micro or nano-scale. These technologies can be useful in optical communication systems, in electromechanical signal processing systems or in healthcare systems, such as BIO-MEMS for microchip-lab diagnostics or micro-devices for therapeutic targeting and delivery. Mechanical micro-actuators for nano-positioning are of interest to the University of Leicester because of their importance in the servo-controller used in the positioning of the read-write head in hard disk data-storage systems [5, 8, 11]. The micro-actuator is being used in a dual-stage control system to achieve high-banwidth positioning control. Significant research effort on hard-disk drive (HDD)-servo techniques has been invested in the area of dual-stage servo control [14, 17, 18]. The reason for this is the continuous increase in the track density and in storage capacity of HDDs. Recently, a track density of more than 420 kTPI (TPI- track per inch) has been demonstrated (see [2, 23]) in a laboratory environment M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 413- 430, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

414

G. Herrmann, M.C. Turner, and I. Postlethwaite

Fig. 14.1. Schematic of Hard Disk Drive with PZT-actuator

and 149 kTPI density HDDs are nowadays available in the consumer market. In addition to the significant increase in data density, the increased demand for higher data rates requires the improved performance of the head-positioning servo system. A promising way to meet these demands is to augment the conventional voice coil motor (VCM) actuator with a second-stage, high bandwidth micro-actuator. Dual-stage servo systems in HDDs are now a feasible alternative to the single stage VCM-servo system. PZT-based micro-actuators using PZT-elements embedded in the head suspension are popular, e.g., the ‘FUMA’-actuator in [19] (Figure 14.1). However, the displacement range of secondary actuators is very limited, typically less than 1-2 µ m, and the input signal for the actuator is limited to prevent damage. In dual-stage servosystems, the two actuators have to deal with the following servo-tasks: seek/settling and track following. Seek/settling control has to ensure a fast movement of the read/write head from one track to another. For track following, high bandwidth controllers are necessary to ensure good error rejection capabilities to counteract disturbances. Therefore, the primary VCM-actuator is required for large displacements and the secondary actuator provides large bandwidth. For servo-control of such a dual-stage actuator system, the method of [11, 12] is employed. It is based on the well known decoupled dual-stage controller structure of [13], where design and stability for the primary VCM-control loop and the secondary PZT-control loop are guaranteed independently. It will be shown [11, 12] that the primary and secondary loop remain stable independently regardless of the seek/settling/track-following control method used for the VCM-loop, providing the secondary control loop is stable in the presence of saturation limits. Hence, it seems logical to use in the secondary loop an AW compensator to guarantee overall large and small signal stability despite actuator saturation limits. AW compensators are most suitable, as they have been developed to retain the nominal performance of the original control system [7,9,20], e.g., high-bandwidth tracking in the servo-control of hard disk drives. This case study on AW-compensation will consider the following issues: 1. Controller design for the micro-actuator loop and wind-up problems due to actuator limits.

Case Study on Anti-windup Compensation

415

2. Discrete anti-windup compensator design and anti-windup compensation for the microactuator control loop. 3. The micro-actuator control loop as part of the overall seek-settling scheme for the hard disk drive. We will consider, for reasons of simplicity, the micro-actuator loop first.

14.2 The Micro-actuator Control Loop and Windup Problems

Log. Magn. [µ m/V]

PZT−model

10

0

10

−1

10

−2

10

1

10

2

3

10 Frequency [Hz]

10

4

Phase [deg]

200

0

−200

−400 1

10

2

10

3

10 Frequency [Hz]

4

10

Fig. 14.2. Continuous model of a PZT-actuator The model of a micro-actuator is usually very simple. Ignoring high frequency resonances, the model of the micro-actuator could be considered to be just a constant gain (Figure 14.2). However, in practice these resonances (at 6.5 kHz and 9.6 kHz in Figure 14.2) are important. They are usually subject to phase and gain uncertainty and the centre frequency of a resonance might even shift due to the influence of temperature. This usually forces the control engineer to design the open and closed loop bandwidth lower than 50 % of the smallest significant resonance. Hence, for our micro-actuator, an open loop crossover frequency of no more than 3 kHz is deemed possible. It must be also noted that hard disk drives are usually controlled using digital technology. The sampling frequency is directly correlated to the rotational speed of the hard disk and the servo-information which is placed along the concentric data tracks. The amount of servoinformation will limit the available space for user data. Hence, it is necessary to limit the amount of servo-information which in turn limits the sampling frequency (Figure 14.3). In

416

G. Herrmann, M.C. Turner, and I. Postlethwaite

Fig. 14.3. Hard-disk with servo-information and data tracks

our case, we have used a sampling frequency of 27 kHz which is reasonable for modern hard disk drives. However, this sampling frequency is rather low, since micro-actuators often also have resonance frequencies at 20 kHz and above (not depicted here) i.e., above the Nyquist frequency of 13.5 kHz. Thus, a hard disk drive servo system is a sampled-data control system for which it is necessary to use discrete or sampled-data control methods to design the servocontroller. Notch filters 0

Log. Magn. [µ m/V]

10

−1

10

−2

10

1

10

2

10

10

3

10

4

Frequency [Hz]

Phase [deg]

0

−50

−100

−150

−200 1

10

10

2

3

10

Frequency [Hz]

Fig. 14.4. Frequency response of discrete notch filter

4

10

Case Study on Anti-windup Compensation

sys_notch_d1*sys_notch_d2

417

PZT Saturation1

NotchFilter

Zero−Order Hold1

PZT actuator

R R Fig. 14.5. MATLAB Simulink

model of PZT-actuator with notch filter and protective saturation element

In the hard disk drive industry, it is common practice to counteract resonances by simple notch filters. In our case, this may be done by using two discrete notch filters which are digitally implemented (see Figure 14.4). Hence, these notch filters are easily augmented at the input of the PZT-actuator via a digital-to-analogue unit (DAU). In general, a saturation is placed between the filter and the DAU to protect the actuator from any damage due to high voltage R peaks (see Figure 14.5 for a MATLAB respresentation). The saturation limits are also tuned in such a way as to protect the micro-actuator from too large displacements. The micro-actuator augmented with the notch filters is easily modelled by a second order system PPZT (z) (see Figure 14.6): −0.07997z − 0.05363 z2 − 0.7805z + 0.3097

PPZT =

(14.1)

Log. Magn. [µ m/V]

Discrete PZT−model with notch

−1

10

−2

10

1

10

2

10

10

3

10

4

Frequency [Hz]

Phase [deg]

200

0

−200

−400

1

10

10

2

3

10

4

10

Frequency [Hz]

Fig. 14.6. Zero-order hold discretization of PZT-model with notch (line); Second order model of PZT-actuator (dash) The first step in the closed loop controller design for this micro-actuator is to design a suitable linear controller K(z) based on robust µ -design methods. The design problem is standard.

418

G. Herrmann, M.C. Turner, and I. Postlethwaite

We will minimize the µ -value for the following control problem with two blocks [ WT T ]:

h

Ws S Wu SK

i

and

 where S =

1 1+KPPZT

and T =

KPPZT 1+KPPZT

 Ws S  Wu SK  WT T

(see Figure 14.7).

Fig. 14.7. µ -control problem The weights (as shown in Figure 14.8) are given by WT =

1.3637(z + 0.04294) −0.25664(z − 2.578) , Ws = , Wu = 0.000001. z + 0.6843 z − 0.9605

R HowThe design of the controller is easily achieved using the µ -toolbox of MATLAB . R ever, a small trick has to be used. MATLAB does not provide µ -tools for discrete control problems. Hence, the discrete control problem has to be converted into a continuous-time problem using a bilinear transformation [3, 5]. Note that the bilinear transformation creates an equivalent continuous-time µ -problem. The discrete controller can then be obtained from the solution to the continuous-time control problem by again using the bilinear transformation. The discrete controller will always be stabilizing and µ -(sub)-optimal with the same µ -value as that obtained from the continuous-time design. √ Thus, for our design problem, a controller is obtained using DK-iteration for a µ -value of 2. Balanced truncation methods then leads to a third order controller which perfectly matches the originally designed controller. The discrete controller of third order is:

K(z) =

−6.584z3 − 1.403z2 + 3.15z − 2.079 z3 + 0.4509z2 − 0.9169z − 0.4211

The control loop is easily simulated with a sine sweep in the frequency range from 10 Hz to 500 Hz for the position demand. For a small demand amplitude of 0.2 µ m, the controller is able

Case Study on Anti-windup Compensation

419

Log. Magn. [µ m/V]

µ−design weights

1

10

0

10

1

10

2

10

10

3

10

4

Frequency [Hz]

Phase [deg]

50

0

−50

−100

−150

−200 1

10

10

2

3

10

4

10

Frequency [Hz]

Fig. 14.8. µ -control problem weights (WT (line), Ws (dashed)) to follow this demand without significant phase lag but with a small amplitude loss (Figure 14.9). However, for a large demand amplitude of 1 µ m, the actuator saturation limit of 1.54 V is activated, limiting the actuator displacement range. Thus, in this case, the controller is not able to track the controller demand as the actuator displacement is limited to a magnitude of less than 0.5 µ m. Hence, the actuator starts to oscillate and the controller exhibits a lag (Figure 14.10). These two issues of resonance and lag can be easily resolved by introducing an antiwindup compensator to recover performance quickly following saturation. The anti-windup compensator of interest to us will be discussed next.

14.3 Anti-windup Compensation for Discrete Linear Control Systems We consider the system in Figure 14.11 as introduced in [21, 29] for continuous-time antiwindup (AW) compensation and in [7] for discrete-time control systems. P(z) is the discrete plant and K(z) is the nominal linear discrete controller. The AW-compensator is given by the blocks, M(z) − I and PM(z). The anti-windup compensation is designed using the free parameter M(z). The control signal is constrained by the limits of a saturation function. The saturation function is defined as sat(u) := [sat1 (u1 ), . . . , satm (um ))]T (14.2) where sati (ui ) := sign(ui ) × min {|ui |, u¯i } and u¯i > 0 is the i’th saturation limit. The following identity holds

420

G. Herrmann, M.C. Turner, and I. Postlethwaite

0.2

0.15

0.1

time [s]

0.05

0

−0.05

−0.1

−0.15

−0.2 0

0.01

0.02

0.03

0.04 0.05 0.06 postion [µm]

0.07

0.08

0.09

0.1

Fig. 14.9. Control result for 0.2 µ m demand amplitude (demand (dashed), actuator-position (line)) 1

0.8

0.6

0.4

time [s]

0.2

0

−0.2

−0.4

−0.6

−0.8

−1 0

0.01

0.02

0.03

0.04 0.05 0.06 postion [µm]

0.07

0.08

0.09

0.1

Fig. 14.10. Control result for 1 µ m demand amplitude (demand (dashed), actuator-position (line)) Dz(u) = u − sat(u)

(14.3)

Case Study on Anti-windup Compensation

421

Fig. 14.11. Conditioning with M(z)

Fig. 14.12. Equivalent representation of conditioning with M(z)

where Dz(u) is the deadzone function. For the deadzone nonlinearity Dz(u), it follows that there exists a diagonal matrix W such that Dz(u)′W (u − Dz(u)) ≥ 0, u ∈ Rm

(14.4)

This inequality will be used later for stability analysis employing the S-procedure [2]. This is the discrete-time version of the continuous time configuration introduced in [21, 29]. As in [21, 29], it is straightforward to show that, using the identity (14.3), Figure 14.11 can be re-drawn as Figure 14.12. These diagrams essentially cast the anti-windup problem as one of choosing an appropriate M(z). However, notice that Figure 14.12 reveals an attractive decoupling into nominal linear system, nonlinear loop and disturbance filter which motivates a useful approach to synthesizing M(z). As noted in [29], most linear conditioning schemes can be interpreted in the framework of Figure 14.11. In [7], the mapping T : ulin 7→ yd was picked as a measure of the anti-windup

422

G. Herrmann, M.C. Turner, and I. Postlethwaite

compensator’s performance (Figure 14.12). We shall choose to minimise the l2 gain, kT ki,l2 , in our anti-windup synthesis. The detectable plant, P(z), has the following state-space description

P(z) ∼

x p (k + 1) = A p x p (k) + B p um (k) y(k) = C p x p (k) + D p um (k)

(14.5)

where x p ∈ Rn p is the plant state, um ∈ Rm is the actual control input to the plant and y ∈ Rq is the output which is fed back to the controller. We are making the assumption that the plant P(z) is asymptotically stable i.e., |λmax (A p )| < 1, for global stability of the non-linear control system with anti-windup compensation. As in [29] for continuous-time, M(z) can be chosen as a coprime factor of P(z). So if P(z) = N(z)M −1 (z), we can search for a coprime factor of P(z) such that the anti-windup closedloop has the best performance in terms of the gain of T . This approach is also related to that of [20] and, to a lesser extent, that of [16]. To achieve full-order stabilisation we would like to choose coprime factors, which share the same state space and are of order equal to that of P(z). Employing Figure 14.12, such coprime factorisations can be characterised by

  Ap + BpF Bp M(z) − I ∼ F 0  N(z) Cp + D p F D p

(14.6)

where u(k) ˜ = Dz(ulin (k) − ud (k)). Note that these equations are parameterised by the free parameter F and therefore we attempt to choose F such that kT ki,l2 is minimised. Theorem 14.1. There exists a dynamic compensator Θ (z) = [M(z)′ − I (P(z)M(z))′ ]′ which solves strongly the anti-windup problem if there exist matrices Q > 0,U = diag(µ1 , . . . µm ) > 0, L ∈ R(m+q)×m and a scalar γ > 0 such that the following linear matrix inequality is satisfied  −Q −L′ 0 QC′p + L′ D′p QA′p + L′ B′p ′ ′   ⋆ −2U I UD p UB p   <0  ⋆ ⋆ − γ I 0 0     ⋆ 0 ⋆ ⋆ −γ I ⋆ ⋆ ⋆ ⋆ −Q 

(14.7)

Furthermore, if this inequality is satisfied, a suitable F for the coprime factorisation (14.6) achieving kT ki,l2 < γ , is given by F = LQ−1 . Proof. Let us choose a Lyapunov function candidate as V (k) = x(k)′ Px(k) > 0. We define the Lyapunov difference as ∆ V (k) := V (k + 1) −V (k). Next we consider the function ∆ V˜ (k) which is defined as 1 ∆ V˜ (k) := ∆ V (k) + 2u(k) ˜ ′W [ulin (k) − ud (k) − u(k)] ˜ + kyd (k)k2 − γ kulin (k)k2 γ

(14.8)

This function is a combination of the Lyapunov difference (first term), the sector bounds from (14.4) associated with the deadzone nonlinearity (second term), and two final terms which ensure we have a certain level of l2 performance. If we can ensure that equation (14.8) is negative definite, we have

Case Study on Anti-windup Compensation

423

1. Asymptotic stability When ulin (k) = 0 and

1 + kyd (k)k2 < 0 ∆ V˜ (k) = ∆ V (k) + 2u(k) ˜ ′W [ulin (k) − ud (k) − u(k)] ˜ } γ {z | | {z } ≥0

(14.9)

≥0

for [x(k) u(k) ˜ ulin (k)] 6= 0 asymptotic stability is implied from ∆ V (k) < 0. 2. l2 gain < γ Summing ∆ V˜ (k) from 0 to ∞ gives: ∞

∞

k=0

k=0

1

˜ ′W [ulin (k) − ud (k) − u(k)] ˜ + kyd k2l − γ kulin k2l ∑ ∆ V (k) + 2 ∑ u(k) γ 2

2

<0

(14.10)

As ∑∞ k=0 ∆ V (k) = V (∞) −V (0) we get ∞ 1 V (∞) −V (0) + 2 ∑ u(k) ˜ ′W [ulin (k) − ud (k) − u(k)] ˜ + kyd k2l2 − γ kulin k2l2 < 0 (14.11) γ | {z } k=0 | {z } >0 ≥0

for [x(k) u(k) ˜ ulin (k)] 6= 0, which implies kyd k2l2 < γ 2 kulin k2l2 + V (0) and therefore kT ki,l2 < γ . This procedure of extending the Lyapunov difference by some extra non-negative terms is the well established S-procedure [2]. Strictly speaking m + 1 terms are included. For the sector non-linearity (14.4), m terms are included, while the m + 1-th term is contributed by the l2 gain constraint. It is well known that the inclusion of one term only, creates an equivalent inequality for both conditions to be satisfied. However, once more than one term is considered only sufficiency is achieved and conservatism is introduced. Furthermore, the sector bound as given in (14.4) may also introduce conservatism as discussed earlier. These facts are wellknown and accepted trade-offs [2, 4] as they allow practically feasible solutions (e.g., [10]). Substituting for x(k), ud (k) and yd (k) in (14.8), we have that ∆ V˜ (k) < 0 if and only if  ′   VF11 VF12 0 x(k) x(k)  u(k)   ⋆ VF22 W   u(k)  < 0, ˜ ˜ ⋆ ⋆ −γ I ulin (k) ulin (k) | {z } 

[x(k) u(k) ˜ ulin (k)] 6= 0

(14.12)

VF

where

1 VF11 = (A p + B p F)′ P(A p + B p F) − P + (C p + D p F)′ (C p + D p F) γ 1 VF12 = (A p + B p F)′ PB p − F ′W + (C p + D p F)′ D p γ 1 ′ VF22 = −2W + D p D p + B′p PB p . γ

(14.13) (14.14) (14.15)

G. Herrmann, M.C. Turner, and I. Postlethwaite

424

The remainder of the proof has to show that VF < 0 is equivalent to (14.7). This follows by standard Schur complement and congruence transformation arguments. The Schur complement implies that (14.12) holds if and only if 

 −P −F ′W 0 C′p + F ′ D′p A′p + F ′ B′p ′ ′  ⋆ −2W W  Dp Bp    ⋆ <0 γ I 0 0 ⋆ −    ⋆  ⋆ ⋆ −γ I 0 ⋆ ⋆ ⋆ ⋆ −P−1

(14.16)

The next step is to multiply the left and right of the matrix inequality with 

 0 0  0 <0 0 I

(14.17)

 0 P−1C′p + P−1 F ′ D′p P−1 A′p + P−1 F ′ B′p  I W −1 D′p W −1 B′p  <0 0 0 −γ I   ⋆ −γ I 0 ⋆ ⋆ −P−1

(14.18)

−P−1 0  ⋆ W −1   ⋆ ⋆   ⋆ ⋆ ⋆ ⋆

0 0 I ⋆ ⋆

0 0 0 I ⋆

so that 

−P−1 −P−1 F ′  ⋆ −2W −1   ⋆ ⋆   ⋆ ⋆ ⋆ ⋆

If we define L = FP−1 , U = W −1 and Q = P−1 , then (14.7) follows. Hence, considering equation (14.9), internal stability follows. The non-linear operator kT ki, l2 has a finite l2 gain with an upper bound of γ > 0 which is easily implied from equation (14.11). Furthermore, note that as there is no ‘direct feedthrough’ term in the nonlinear loop, wellposedness is ensured. ⊓ ⊔⊓ ⊔

14.4 Anti-windup Compensation for the Micro-Actuator It is now straightforward to design an anti-windup compensator for the linear control scheme of Section 14.2. The model used for controller design (14.1) delivers an anti-windup compensator of (14.6) where Ap =

0.7805 −0.6195 0.5000 , Bp = , C p = −0.1599 −0.2145 , D p = 0 0.5000 0 0

and a suitable state feedback gain F is given by

Case Study on Anti-windup Compensation

425

F = −1.2403 1.7618 .

The l2 gain has been minimized to a value of γ = 0.2616. The application of this computed AW-compensator to the control system of Section 14.2 with saturation limit improves the closed loop as illustrated in Figure 14.13. The oscillations observed without AW-compensation are minimized and the lag seems to be eradicated (see Figures 14.13 and 14.10) for the sine sweep with 1 µ m amplitude.

1

0.8

0.6

0.4

time [s]

0.2

0

−0.2

−0.4

−0.6

−0.8

−1 0

0.01

0.02

0.03

0.04 0.05 0.06 postion [µm]

0.07

0.08

0.09

0.1

Fig. 14.13. Control result for 1 µ m demand amplitude (demand (dashed), actuator-position (line)) Thus, the AW-compensation has achieved a significant improvement in tracking performance when one considers the fact that the micro-actuator has a displacement limit of about ±0 5µ m

14.5 The Micro-actuator Control Loop as Part of a Hard-Disk-Drive Servo-System The micro-actuator control is to be incorporated into the control scheme for the servo-control of a hard disk drive actuator. The micro-actuator is directly embedded in the suspension of the voice-coil motor (VCM) actuator (Figure 14.1). A linear model of the actuator [PPZT PVCM ] is given by the superposition of the two models of the PZT-actuator, PPZT , and the VCMactuator, PVCM . The VCM actuator can be modelled as a double integrator which is affected by high frequency resonances and friction dynamics at low frequencies. It is usually controlled by a (proximate)time-optimal control scheme [3, 6, 11], which achieves fast seeking for fast positioning of

426

G. Herrmann, M.C. Turner, and I. Postlethwaite

Fig. 14.14. Simplified seek/settle/trackfollowing scheme

Fig. 14.15. Equivalent representation of the servo scheme

the read/write head, while a smooth transfer to a linear track-following scheme needs to be achieved for track-following (see [3, 6, 11] for a more detailed explanation). To achieve a higher track-following controller bandwidth, the secondary PZT-actuator is used. The constraints on both actuators (in particular for the PZT-actuator) need to be incorporated into the servo-control scheme. The approach is to use a decoupled control scheme as from [13], which incorporates the traditional time-optimal seek-settling scheme for the VCM-actuator [6, 11]

Case Study on Anti-windup Compensation

427

Fig. 14.16. Simplified seek/settle/trackfollowing scheme with observer 2.5

postion [µm]

2

1.5

1

0.5

0 0

0.005

0.01

0.015

0.02 time [s]

0.025

0.03

0.035

0.04

Fig. 14.17. Control result for 2 µ m seek step without AW-compensation (demand (dashed), actuator-position (line)) and allows an independent controller design for the PZT-micro-actuator so that the controller design of Sections 14.2–14.3 can be reused. The control scheme works on the assumption that the absolute position of the PZT-micro-actuator is measurable (see Figure 14.14). Employing the superposition principle, it is easily seen that the control scheme of Figure 14.14 is equivalent to the controller configuration of Figure 14.15, where the control loops for the PZT and the VCM-actuator are decoupled. Hence, both controllers can be designed

428

G. Herrmann, M.C. Turner, and I. Postlethwaite 2.5

postion [µm]

2

1.5

1

0.5

0 0

0.005

0.01

0.015

0.02 time [s]

0.025

0.03

0.035

0.04

Fig. 14.18. Control result for 2 µ m seek step with AW-compensation(demand (dashed), actuator-position (line)) independently in terms of stability. For the implementation of the scheme, it is necessary to use an observer which obtains the absolute position of the PZT-actuator (see Figure 14.16). It is easily verified that the observer and controller separate due to the linearity of the plant. Using now a non-linear seek-settle control scheme for the VCM-actuator as presented in [6,11] and the linear control loop from Sections 14.2–14.3 with and without anti-windup compensation for the controller of the PZT-actuator, it is easily verified that the AW-compensation scheme significantly improves on seeking performance (see Figures 14.17 and 14.18). The controller with AW-compensation for a 2 µ m-seek step prevents a very large overshoot and allows fast settling to assume track following of the read/write head as quickly as possible.

14.6 Summary The control system of a PZT-micro-actuator has been investigated over its feasible range i.e., the available displacement range of the micro-actuator. It has been shown that anti-windup compensation can improve the tracking of demands which are not feasible. It limits controller lags and suppresses high frequency resonances for a practically valid hard disk drive microactuator. In connection with the recently established dual-stage track-seek/following scheme [6, 11], it has been shown that the AW-compensator suppresses large overshoots and enhances settling speed.

References 1. S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear Matrix Inequalities in System and Control Theory. Society for Industrial and Applied Mathematics, 1994. 2. C. Du, L. Xie, G. Guo, J. Zhang, Q. Li, and B. Hredzak. A generalized kyp lemma based control design and application for 425 ktpi servo track writing. In Proceedings of the 2006 American Control Conference, Minneapolis, Minnesota, USA, 2006.

Case Study on Anti-windup Compensation

429

3. G. F. Franklin, J. D. Powell, and M . L. Workman. Digital Control of Dynamic Systems. Addison-Wesley Publishing Company, Reading, Massachusetts, 2nd edition, 1990. 4. G. Grimm, J. Hatfield, I. Postlethwaite, A. R. Teel, M. C. Turner, and L. Zaccarian. Antiwindup for stable linear systems with input saturation: An LMI based synthesis. IEEE Transactions on Automatic Control, 48(9):1509–1525, 2003. 5. G. Herrmann and G. Guo. HDD dual-stage servo-controller design using a µ -analysis tool. Control Engineering Practice, 12(3):241–251, 2003. 6. G. Herrmann, B. Hredzak, M. C. Turner, I. Postlethwaite, and G. Guo. Improvement of a novel dual-stage large-span track-seeking and track-following method using anti-windup compensation. In Proceedings of the 2006 American Control Conference, Minneapolis, Minnesota, USA, 2006. 7. G. Herrmann, M. C. Turner, and I. Postlethwaite. Discrete time and sampled data antiwindup synthesis: stability and performance. International Journal of Systems Science, 37(2):91–113, 2006. 8. G. Herrmann, M. C. Turner, I. Postlethwaite, and G. Guo. Practical implementation of a novel anti-windup scheme in a hdd-dual-stage servo-system. IEEE/ASME Transactions on Mechatronics, 9(3):580–592, 2004. 9. G. Herrmann, M.C. Turner, and I. Postlethwaite. Discrete time anti-windup - part 2: extension to sampled data case. In Proc. of the European Control Conf., 2003. 10. G. Herrmann, M.C. Turner, I. Postlethwaite, and G. Guo. Application of a novel antiwindup scheme to a hdd-dual-stage actuator. Proceedings of the American Control Conference, 2003. 11. B. Hredzak, G. Herrmann, and G. Guo. A proximate-time-optimal-control design and its application to a hard disk drive dual-stage actuator system. IEEE Transactions on Magnetics, 42(6):1708 – 1715, 2005. 12. B. Hredzak, G. Herrmann, and G. Guo. Short and long-span track seek control for hard disk drive dual-stage servo actuators. In Proceedings of IEEE Industrial Electronics Society Conference, Raleigh, North Carolina, 2005. 13. M. Kobayashi and R. Horowitz. Track seeking control for hard disc dual-stage servo systems. IEEE Transactions on Magnetics, 37(2):949–954, 2001. 14. M. Kobayashi, S. Nakagawa, T. Atsumi, and T. Yamaguchi. High-bandwidth servo control designs for magnetic disc drives. In Proc. IEEE/ASME Intern. Conf. Adv. Intel. Mechatr., Como, Italy, pages 1124–1129, 2001. 15. M.V. Kothare, P.J. Campo, M. Morari, and C.N. Nett. A unified framework for the study of anti-windup designs. Automatica, 30(12):1869–1883, 1994. 16. S. Miyamoto and G. Vinnicombe. Robust control of plants with saturation nonlinearity based on comprime factor representations. Proc. IEEE Conference on Decision and Control, pages 2838–2840, 1996. 17. K. Mori, T. Munemoto, H. Otsuki, Y. Yamaguchi, and K. Akagi. A dual-stage magnetic disk drive actuator using a piezoelectric device for high track density. IEEE Trans. on Magnetics, 27(6):5298–5300, 1991. 18. S. J. Schroeck, W. C. Messner, and R. J. McNab. On compensator design for linear timeinvariant dual-input single-output systems. IEEE Trans. Mechatr., 6(1):50–57, 2001. 19. M. Tokuyama, T. Shimizu, H. Masuda, S. Nakamura, M. Hanya, O. Iriuchijima, and J. Soga. Development of a ϕ -shaped actuated suspension for 100-ktpi hard disk drives. IEEE Transactions on Magnetics, 37(4):1884–1886, 2001.

430

G. Herrmann, M.C. Turner, and I. Postlethwaite

20. M. C. Turner, G. Herrmann, and I. Postlethwaite. Discrete-time anti-windup: Part 1 - stability and performance. In Proceedings of the European Control Conference, Cambridge, UK, 2003. 21. M.C. Turner and I. Postlethwaite. A new perspective on static and low order anti-windup synthesis. International Journal of Control, 77(1):27–44, 2004. 22. P.F. Weston and I Postlethwaite. Analysis and design of linear conditioning schemes for systems containing saturating actuators. IFAC Nonlinear Control System Design Symposium, 1998. 23. J. Zheng, C. Du, G. Guo, Y. Wang, J. Zhang, Q. Li, and B. Hredzak. Phase lead peak filter method to high tpi servo track writer with microactuators. In Proceedings of the 2006 American Control Conference, Minneapolis, Minnesota, USA, 2006.

15 Enhancing Immune System Response Through Optimal Control Robert F. Harrison

15.1 Introduction The aim of this chapter is to introduce you to some ideas about modelling non-linear systems from first principles, about their behaviour and how it can be analyzed, and finally, about how they might be made to behave in a desirable way. As a vehicle I have chosen what I hope you will find to be an interesting problem which is outside the mainstream of technology-based control but is, I believe, part of an area of increasing importance for the near future. The topic is the application of control theory in a biological setting – the optimal enhancement of the immune system. In what follows I will first outline how models of systems of this type can be constructed largely through deduction. They belong to a class of models called Lotka-Volterra Models which are widely used to describe the dynamics of interacting populations such as predatorprey systems, epidemics and, in our case, the concentrations of various biological quantities in the immune system. Next we will look briefly at the behaviour of this class of dynamics. Finally we look at a difficult problem – how control theory can be used to enhance the response of the immune system by allowing us to compute an optimal policy that will prevent organ failure in conditions where, left to itself, organ damage or death would naturally occur. The case study builds on a series of recent papers by Stengel and colleagues [1–3] with the introduction of a newly developed approach to non-linear optimal control [4] which has not yet become well known but appears to have a number of advantages over conventional solutions. Throughout I will try to maintain a “systems viewpoint” and avoid too much biology. I will also avoid too much mathematics other than stating the key equations and providing references to more detail. Those of you versed in linear quadratic optimal control will recognize the form of the non-linear generalization.

15.2 Lotka-Volterra Equations The simplest model of two competing populations in a predator/prey relationship was proposed by the Italian mathematician, Vito Volterra (1860–1940) in 1926, while trying to explain the cyclic nature of fish stocks in the Adriatic Sea [5]. Independently, the Austrian born ecologist and mathematician, Alfred Lotka (1880-1949), proposed a similar model at about the same time [6]. M.C. Turner et al. (Eds.): Mathe. Methods for Robust & Nonlin. Ctrl., LNCIS 367, pp. 431- 444, 2007. springerlink.com © Springer-Verlag Berlin Heidelberg 2007

432

R.F. Harrison

Imagine an island that is home to only two species with no invaders. The prey population, goats, say, has an unlimited food supply which is uniformly available throughout the island, while the predator population, wolves, eats only goats. Goats and wolves are spread uniformly across the island. The uniform distributions mean that we do not have to complicate matters by taking location into account and can describe the populations by differential equations. The situation can be described by a coupled pair of differential equations . Here we assume there is no competition between members of the same group. These are given by

dg = ag − bwg; g(0) = g0 dt dw = −cw + dgw; w(0) = w0 dt

(15.1)

where the variables, g, w, denote the size of the prey & predator populations, respectively. The parameter, a, is the relative birth rate of the goats and c, the relative death rate of the wolves. The product term, −bwg represents the rate of decrease in the population of goats as a result of encountering wolves and +dgw, the rate of increase in the wolf population. The assumption of uniform distribution and random encounters suggests that these rates are proportional to the product of the size of the populations. All quantities, parameters and variables are nonnegative. Let’s examine the equations in more detail to see if they do, indeed, describ e the qualitative situation. First, in the absence of any wolves w = 0 the number of goats increases exponentially⋆ , i.e., solve g˙ = ag; g0 with a, g0 > 0. We have seen this recently on the Hebridean islands of North and South Uist in the context of hedgehogs. Only the motor car is a serious threat to the hedgehog there, and the Uists have very little traffic. From the initial introduction of four hedgehogs their numbers have expanded to around 5,000 [7]. The bird-life of the islands has been devastated by the explosion in numbers to the extent that culling and “repatriation” has been necessary. In the absence of goats, the wolf population will dwindle to nothing because we assume that wolves can only eat goats, i.e., solve w˙ = −cw; w0 with c, w0 > 0. Figure 15.1 shows the typical ⋆⋆ time evolution of the two populations for initial values g0 = 1000, w0 = 10 demonstrating its cyclic nature. Figure 15.2 depicts the phase plane for the same initial conditions. Notice how the phase plane divides naturally into four quadrants corresponding to periods: (A) when goats are increasing while wolves decrease; (B) both populations are growing; (C) predation takes over and the goats are severely reduced resulting in a large rise in wolves, and finally (D) the wolf population reduces owing to a lack of food, slowing the rate of goat reduction. The point marked with a square at dc , ba = (4000, 80) is an equilibrium point of the system.

⋆ ⋆⋆

Don’t forget this is a model – we have not taken the finiteness of resources such as land area and food into account . Parameter values of a = 0.04, b = 0.0005, c = 0.2,d = 0.00005 are chosen in this abstract example.

Optimal Control of the Immune System

433

12000

10000

Population Size

8000

6000

4000

2000

0

0

10

20

30

40

50

60

70

80

90

100

Time

Fig. 15.1. Evolution of goat (solid) and wolf (dashed) populations from an initial state of g0 = 1000, w0 = 10. This depicts one cycle.

15.2.1

Analysis of Equilibria

One of the easiest tools for analyzing non-linear dynamics is known as Lyapunov’s First or Indirect Method. This tells us about dynamic behaviour in the locality of the (possibly many) equilibrium points defined by x˙ = f (x) = 0 – the overbar is used to denote a fixed-point. From this type of analysis it is sometimes possible to infer wider behaviour although global methods are more reliable. Lyapunov’s Second or Direct Method provides a means for doing this although, in general, it can be difficult to apply. A good source of information on this topic is [8]. To apply Lyapunov’s Indirect Method we must find the (static) equilibrium points⋆⋆⋆ of the system, linearize about them and examine the eigenvalues of the Jacobian matrices evaluated at the equilibrium points. Real Eigenvalues: If all eigenvalues are real and negative (positive) then all states asymptotically approach (leave) equilibrium, monotonically. These are known as stable (unstable) nodes. If the eigenvalues have a mixture of signs then some states will approach equilibrium and some will diverge, again monotonically, resulting in so-called saddle-points. Complex Eigenvalues: Here, the eigenvalues must occur in conjugate pairs and those with negative (positive) real parts will converge (diverge) with a cyclical form. These are known as stable (unstable) foci. Again, a mixture can occur. ⋆⋆⋆

The points where the state derivative is zero, i.e., the system can be said to be “at rest” .

434

R.F. Harrison 600

500

Wolves

400

300 D

C

A

B

200

100

0

0

2000

4000

8000

6000

10000

12000

Goats

Fig. 15.2. Phase-plane depiction of one cycle of the evolution of goat and wolf populations from an initial state of g0 = 1000, w0 = 10 Imaginary Eigenvalues: Often associated with purely oscillatory behaviour known as a centre but in fact, indeterminate. What this tells us is that the higher-order terms in the Taylor expansion of f may play an important role. So, for our system, defining x1 as the number of goats and x2 as the number of wolves, we can write its two-dimensional state-space form, x˙ = f (x), thus

x˙1 = ax1 − bx2 x1 ; x1 (0) = x10

(15.2)

x˙2 = −cx2 + dx1 x2 ; x2 (0) = x20 giving two fixed points, (x1 , x2 ) =

c a d, b

and the origin. The Jacobian matrix is given by

a − bx2 −bx1 dx2 −c + dx1

(15.3)

√ At the first equilibrium the two eigenvalues of the Jacobian are given by λ = ± j ac. For a linear system this corresponds to a centre but in fact we can conclude nothing about the stability of the dynamics although, in our case, they do happen to cycle. At the origin we clearly have λ = a and λ = −c defining a saddle-point. Obviously the origin corresponds to the extinction of both species but the saddle suggests that extinction is not easily arrived at naturally.

Optimal Control of the Immune System

435

The situation described in equation (15.2) is not very realistic but it does in troduce the ideas of how two species can interact. Obviously the ideas can be extended. For instance, if food for the goats becomes short there would be intra-species competition. This is easily accounted for by introducing a term proportional to x12 in the first equation in (15.2). Because this will have the effect of reducing goat numbers, it must be subtracted. Likewise, culling of wolves could be introduced as a control variable on the RHS of the second equation – a negative constant, for instance, to indicate the rate of removal. Introduction of any number of additional species is straightforward and for added realism, time delays can be introduced to account for gestation times – obviously the consumption of a goat does not instantaneously result in a new wolf. However, while simulation of time-delay systems is relatively straightforward, analysis and control is vastly more advanced. We won’t consider it here. Figure 15.3 shows what happens when there is competition between goats for food (a term −ex12 is added to the first equation in (15.2). We see that a new equilibrium is reached instead of the continuous cycling (a dynamic equilibrium). This occurs at (4000, 72) after about 2500 time units (e=0.000001) so, for the farmer’s investment of the original 1000 goats a steadystate population of 4000 arises even though some wolves have had to be fed.

500 450 400 350

Wolves

300 250 200 150 100 50 0

0

2000

4000

6000

8000

10000

12000

Goats

Fig. 15.3. Phase-plane depiction of the goats & wolves populations when goats compete for food

15.3 Optimal Control The topic of “optimal control” is enormous so I will restrict the problem to that of choosing a control policy to minimize a quadratic function of the system states and control effort subject

436

R.F. Harrison

to non-linear dynamics. While this is a restricted statement of the optimal control problem, it is widely applicable in practical circumstances. We will solve it by generalizing the ideas of linear quadratic control to this class of system to create a sequence of approximations to the control trajectory that can be shown to converge to the global optimum. These ideas are quite recent but provide, for the first time, an easily computed and implemented solution to this class of feedback control problem [4].

15.3.1

Linear, Time-Varying Quadratic Optimal Control

To keep things simple and because, for the purpose of these notes, it is all we shall need, I shall only look at the regulator problem under the assumption that all states are available for feedback. I’ll say a bit more about this rather strong assumption at the end. The problem of deriving the optimal linear feedback for a linear, time-varying (i.e., the A and B matrices of the state-space form are functions of time) system under quadratic cost is well understood (see, e.g., [9]). Consider the time-varying linear dynamics x˙ = A(t)x + B(t)u; x (t0 ) = x0

(15.4)

where x(t) is the n-dimensional state vector and u, the m-dimensional control vector. A(t) and B(t) are matrix-valued functions of time of appropriate dimension. The objective is to choose the control law that minimizes the quadratic cost functional over the interval t0 –t f 1 1 J (u) = xT t f Fx t f + 2 2

Z tf t0

xT (t) Q(t)x (t) + uT (t) R(t)u (t) dt

(15.5)

It is well-known that the solution to this problem is reached via a two-point boundary value problem which, owing to linearity, happens to lead to a full-state, linear feedback formulation given by uopt (t) = −R−1 (t)BT P(t)x(t)

(15.6)

where P(t) is the positive-definite solution of the matrix Riccati equation

˙ P = −Q(t) − P(t)A(t) − AT (t)P(t) + P(t)B(t)R−1 (t)BT (t)P(t)

(15.7)

with P t f = F.

Since only the terminal condition for equation (15.7) is given, this must must first be solved backwards in time and stored so that it can be used in the forward dynamics. This can be done by substituting τ = t f − t and solving an initial value problem. Clearly, optimal control of a time varying system and/or over a finite time horizon is not causal and questions arise as to its application in a real-time setting. It turns out that the state evolution of a wide class of non-linear dynamical system can be computed as the limit of a sequence of linear, time-varying approximations and that this can be exploited to find the optimal control policy using the above idea [4, 10].

Optimal Control of the Immune System

437

The class of dynamical system of interest is of the form x˙ = f (x, u) ; x (t0 ) = x0

(15.8)

with f (0, 0) = 0 and smooth in some sense† . It is necessary that the system (15.8) can be re-written in a form resembling the familiar, linear, state-space equations, thus x˙ = A (x) x + B (x, u) u; x(t0 ) = x0

(15.9)

where A is an appropriately dimensioned, non-linear, matrix-valued function of the state only and likewise B, except that it is allowed to be a function of the control as well. The fact that A may only be a function of the state and not the control imposes a further, small, restriction on the dynamics. The objective is to find the control law that minimizes the cost functional 1 1 J (u) = xT t f Fx t f + 2 2

Z tf t0

xT (t) Qx (t) + uT (t) Ru (t) dt

(15.10)

with F, Q positive semi-definite and R positive-definite matrices of appropriate order. Note that to simplify notation I have chosen these to be constant but the theory allows them to be functions of the state vector‡ . From here on I have also suppressed the dependence of quantities on t Analogously to the linear time-varying optimal feedback control problem, it can be shown that the sequence of approximating solutions, given by the following, converges globally to the optimal solution [4]. u[i] = R−1 BT x[i−1] , u[i−1] P[i] x[i]

(15.11)

for i ≥ 0, where P[i] is the solution of

P˙ [i] = −Q − P[i] A x[i−1] − AT x[i−1] P[i] +P[i] B x[i−1] , u[i−1] R−1 BT x[i−1] , u[i−1] P[i]

(15.12)

with P[i] t f = F. It follows that the sequence of state trajectories is given by

x˙ [i] = A x[i−1] − B x[i−1] , u[i−1] R−1 BT x[i−1] , u[i−1] P[i] x

(15.13)

with x[i] (t0 ) = x0 . The crucial point here is that the non-linear problem has been reduced to a sequence of linear, time-varying problems that is guaranteed to converge to the correct solution. All that remains is to start the sequence off, which can be done by taking x[i−1] = x0 and u[i−1] = 0 for i = 0. The algorithmic steps for computing the control law are as follows. † ‡

Locally Lipschitz continuous. This means that, unlike many approaches, ASRE is not restricted to quadratic functions of the states and controls in the cost function – these can be very general indeed.

438 • • •

R.F. Harrison

Compute P[0] (t) from (15.12) with i = 0 and x[i−1] = x0 and u[i−1] = 0. Substitute P[0] (t) and compute x[0] (t) from (15.13) with A (x0 ), B (x0 , 0). Apply (15.12) followed by (15.13) repeatedly, until convergence.

15.4 Immune System Dynamics What, then, has the Lotka-Volterra approach to do with the immune system and how can we apply optimal control to assist it in responding to attack by microbes? The answer is that we can use the same intuitive arguments to come up with a plausible model for the dynamics of the immune system which then fits nicely into the formulation required for the ASRE or other optimal control strategies. Now we have different populations – typically concentrations of organisms and chemicals that interact by coming into contact with one another. We shall see how these ideas relate to modelling the immune system. What happens when the body is attacked by an external agent or pathogen ? Roughly speaking, plasma cells specialized to the particular antigens associated with these pathogens are produced. The plasma cells produce antibodies that bind to the antigens resulting in the destruction of the pathogen in some way. Much more detail is given in [11] and in the series of papers by Stengel and colleagues [1–3]. In the model we shall study we define x1 x2 x3 x4

– pathogen concentration – plasma cell concentration – antibody concentration – measure of organ health (0 – healthy, 1 – dead)

and for the controls u1 u2 u3 u4

– pathogen killer – plasma cell enhancer – antibody enhancer – organ health enhancer

The model proposed in [11] and modified through the addition of controls [1] is given by dx1 dt dx2 dt dx3 dt dx4 dt

= (a11 − a12 x3 ) x1 + b1 u1

(15.14)

= a21 (x4 ) a22 x1 x3 − a23 (x2 − x2 ) + b2 u2 = a31 x2 − (a32 + a33 x1 ) x3 + b3 u3 = a41 x1 − a42 x4 + b4 u4

which is clearly of the Lotka-Volterra type. All disease parameters, ai j > 0. For instance, taking the first equation governing the concentration of pathogens, in the absence of antibodies

Optimal Control of the Immune System

439

(x3 ) or pathogen killer (u1 ) the pathogen concentration is free to grow exponentially, however, the interaction with the antibodies has a moderating effect as in the goats and wolves example. Similar reasoning underpins, qualitatively, the processes behind the rest of the equations. One significant departure, though, is the introduction of the function a21 (x4 ). This has the effect of reducing the ability of the vital organs to produce plasma cells as they becomes more diseased and is chosen to be a21 (x4 ) =

cos (π x4 ) , 0 ≤ x4 < 1 0, 2 ≤ x4

1 2

which means that for a healthy organ there is little degradation while for a dead organ the excess plasma cell production decays exponentially to its normal level. Evidently the parameters have to be determined and this can be done from biochemical analysis or, in principle, through system identification. For this example we will use a plausible set as provided by [1]: a11 = a12 = a23 = a31 = a42 = b2 = b3 = 1, b1 = b4 = −1, a22 = 3, a32 = 1.5 and a33 = 0.5. The initial conditions for equation (15.8) determine the uncontrolled respo nse of the states. In all cases we assume an initially healthy organ (x4 (0) = 0), a normal level ofplasma cells

31 (x2 (0) = x2 = 2), and an equilibrium level of antibody x3 (0) = x3 = aa32 x2 . The initial value of x1 determines the strength of the pathogen attack. In [1] four cases of interest are identified

Case 1 Sub-clinical – no medical intervention required, (x1 (0) = 1.50) Case 2 Clinical – needs medical assessment, (x1 (0) = 2.00) Case 3 Chronic – pathogen level goes into steady state, (x1 (0) = 2.57) Case 4 Lethal – organ health is damaged beyond repair and pathogen levels diverge, (x1 (0) = 3.00) These situations can be simulated by judicious choice of initial conditions for x1 (see above). Figure 15.4 demonstrates the natural response of the immune system when attacked by the different levels of pathogen. h i It is easy to verify that uT = 0 0 0 0 , xT = 0 x2 aa3132x2 0 is an equilibrium point satisfying x˙ = 0. We can therefore define a new state vector, z = x − x and a new centred RHS, f (z, u), so that the equilibrium of the centred system lies at the origin as required by the ASRE theory. This gives dz1 dt dz2 dt dz3 dt dz4 dt

= ((a11 − a12 x3 ) − a12 z3 ) z1 + b1 u1

(15.15)

= a21 (z4 ) a22 z1 z3 + a21 (z4 ) a22 x3 z1 − a23 z2 + b2 u2 = a31 z2 − (a32 + a33 z1 ) z3 + a31 x2 − (a32 + a33 z1 ) x3 + b3 u3 = a41 z1 − a42 z4 + b4 u4

which has a non-unique decomposition into the required form of equation (15.9) where A (z) is given by

440

R.F. Harrison 6 plasma conc.

pathogen conc.

6

4

2

0

5

0

0

5

10

0

5

10

1 organ health

antibody conc.

2

0

10

3

2

1

0

4

5

0

10

0.5

0

Fig. 15.4. Uncontrolled response for the four cases, Sub-clinical (solid), Clinical (dashed), Chronic (chained) & Lethal (dotted)



 0 0 ((a11 − a12 x3 ) − a12 z3 ) 0  a21 (z4 ) a22 (z3 + x3 ) −a23 0 0   A (z) =   a33 x3 a31 − (a32 + a33 z1 ) 0  a41 0 0 −a42

(15.16)

and B (z) = B = diag (bi ). Clearly, A (z) and B satisfy the required Lipschitz condition.

15.4.1

Optimal Enhancement of the Immune Response

It is worth asking why anyone would want to apply control to a problem such as this. The over-riding reason is to make sure that as much external agent as needed is given, in the right dose and at the right time, but no more. Many drugs are no more than poisons that we take because the danger they pose is less than that of the illness. Antibiotics fall into this class as do many others. So the driver for this type of research is taking the guesswork out of drug treatment. To demonstrate the effectiveness of the proposed method, we must make some design choices. It is also important to realize that both the states and the controls must be constrained to remain non-negative. Being a new development, ASRE is not yet able explicitly to cope with such constraints§ , however, the dynamics under consideration here are sufficiently benign that the constraints are typically satisfied by the closed-loop system.

§

A version of the theory under control constraints is currently under development.

Optimal Control of the Immune System

441

To illustrate how individual therapies (each control applied separately) can be used, we examine the application of pathogen killer, u1 , alone, as this was found to be most effective in an open-loop strategy [1]. Of course, a variety of combinations can be used as well. Note that the method described here provides a feedback strategy directly, whereas in [2, 3] a separate method based on the idea of neighbouring optimal solutions is used. This involves linearizing about the optimal open-loop trajectory and obtaining the time-varying (approximate) optimal policy. Feedback is important because the open-loop optimal control is unable to account for different initial conditions or for any disturbances that might enter the system. We repeat the situation dealt with by open-loop control in [1] (Figure 2) with q11 = q44 = 1 f11 = f44 = 1 and b2 = b3 = b4 = 0 so that only the pathogen killer is used with R = r = 1. Figure 15.5 shows how the system recovers from all four conditions. The dotted curves correspond to the solid curves in Figure 2 of [1].

6 plasma conc.

pathogen conc.

6

4

2

0

0

5

0

5

10

0

5

10

1 organ health

antibody conc.

2

0

10

3

2

1

0

4

0

5

10

0.5

0

Fig. 15.5. Optimal administration of pathogen killer for the four cases, Sub-clinical (solid), Clinical (dashed), Chronic (chained) & Lethal (dotted). The lethal case using an open-loop optimal strategy is shown in Figure 2 of [1] (solid).

To emphasize the importance of feedback, we examine the lethal situation and compare the open-loop solution computed for an initial value of pathogen concentration of 3.0 but applied to a case when the initial concentration is, in fact, 4.5. Figure 4 of [3] demonstrates that the open-loop strategy would fail to accommodate the mismatch and would lead to organ death. In that paper, the approximate neighbouring optimal solution is applied and it can be seen that this rectifies the situation. I have used the ASRE method on the same problem and Figure 15.6 shows how it is able to cope with the perturbation. By eye it is clear that, in this case, the neighbouring optimal trajectory [3] and the ASRE solution are virtually indistinguishable.

442

R.F. Harrison 6 plasma conc.

pathogen conc.

6

4

2

0

0

5

0

5

10

0

5

10

1 organ health

antibody conc.

2

0

10

3

2

1

0

4

0

5

10

0.5

0

Fig. 15.6. Comparison of open-loop response (solid) with the optimal feedback response (dashed) for a perturbed initial pathogen concentration. c.f. [3] (Figure 4) where the no longer optimal open-loop solution is seen to lead rapidly to organ failure.

What I have done here is to establish the principle that the ASRE method can deliver a viable feedback control policy in spite of it being unable explicitly to handle the positivity constraints on states and controls. To do this I have used the abstract example of [1] with arbitrary choices of F, Q and R. Clearly, even in this case, there is much experimentation to be done to discover an acceptable set of weighting matrices and manipulated variables.

15.4.2

Some Practical Considerations

There are some pretty strong assumptions in the development of the control strategy and I want to deal with two of them. The first, is relatively easy and is to do with the question of causality. The finite-time optimal control problem does not have a causal solution, i.e., information from the future is required at the current time to compute the optimal value of control input. This is compounded in the ASRE approach because the computations are iterative and cannot happen instantaneously. In general, therefore, a real-time optimal policy is not attainable. However, when the system dynamics are “slow ” relative to the speed of the computer used to perform the calculations which will usually be the case for immune systems, then all computations can be performed within the time required. This may be more problematic for “fast” dynamics. Handling unexpected disturbances in the interval requires a re-computation over the remaining time from the revised state – effectively a re-start. Another way of dealing with uncertainty, whether in the model or in external disturbances, is to use an explicit robust control strategy. The ASRE technique has been used in this way

Optimal Control of the Immune System

443

to provide robust rejection of unexpected disturbances and model error in a time-domain H ∞ setting [12]. Application of this idea to the enhancement of immune response is underway. The major difficulty associated with any control signal that makes use of th e system state is that exact measurements of the states are rarely, if ever, available. In the case of linear dynamics, the celebrated Kalman-Bucy filter provides a solution, whereby, given a good model of the dynamics and some knowledge of any measurement and process noise entering the system of interest¶ , an optimal estimate of the states can be computed causally, in real-time [13]. In the non-linear situation, it is usual to adopt the extended Kalman filter (EKF) which accommodates the non-linearity by a local linearization along the estimated state trajectory. While this no longer possesses optimality and may not even converge, the EKF has been found to work well in practice and its use is widespread [13]. More recently attention has been given to other types of state estimator, the most practical being the so-called unscented Kalman filter (UKF) [14]. This provides a more accurate estimate of the state error covariance matrix than the EKF and has been shown to perform better in a number of cases. More recently still, the ASRE technique has been applied to the problem of state estimation by appealing to the duality of the control and filtering problems [15]. While this goes some way toward solving the problem, it must be recognized that since the state estimates are the limit of a sequence of linear, time-varying approximations, such a method cannot properly represent the effects of non-linearity on the noise processes. In [2] a time-varying Kalman filter is established through linearization along the closed-loop trajectory and this is shown to work well in the current setting.

15.5 Conclusion I hope that I have given you a flavour of how relatively simple dynamic models can be derived that explain, with some degree of plausibility, quite complex behaviour and how they can be made to behave in a desirable way through the application of optimal control theory. As I stated earlier, it is my belief that the application of such methods has much to offer in the biomedical and life sciences and that we should expect to see applications leaving the computer laboratory and going live in the next decade or so.

References 1. R.F. Stengel, R. Ghigliazza, N. Kulkarni, and O. Laplace. Optimal control of innate immune response. Optimal Control Applications and Methods, 23:91–104, 2002. 2. R.F. Stengel and R. Ghigliazza. Stochastic optimal therapy for enhanced immune response. Mathematical Biosciences, 191:123–142, 2004. 3. R.F. Stengel, R. Ghigliazza, and N. Kulkarni. Optimal enhancement of immune response. Bioinformatics, 18:1227–1235, 2002. 4. T. C¸imen and S.P. Banks. Global optimal feedback control for general non-linear systems with non-quadratic parformance criteria. Systems & Control Letters, 53:327–346, 2004.

¶

Linear time-varying or invariant dynamics are admissible with uncorrelated, Gaussian distributed process and measurement noise of known covariance.

444

R.F. Harrison

5. V. Volterra. Variations and fluctuations of the number of individuals in animal species living together. In R.N. Chapman, editor, Animal Ecology. McGraw-Hill, New York, 1931. 6. A.J. Lotka. Elements of Physical Biology. Williams & Wilkins Co., Baltimore, 1925. 7. Anonymous. Uist Hedgehogs: Factsheet No. 3. World Wide Web, http://www.snh. org.uk/pdfs/news/nw-uwp03.pdf, 2002. 8. J.-J. E. Slotine and W. Li. Applied Nonlinear Control. Prentice Hall, New Jersey, 1991. 9. B. Friedland. Control System Design: An Introduction to State-Space Methods. McGrawHill, Singapore, 1987. 10. T. C ¸ imen and S.P. Banks. Nonlinear optimal trackingcontrol with application to supertankers for autopilot design. Automatica, 40:1845–1863, 2004. 11. A. Asachenkov, G. Marchuk, R. Mohler, and S. Zuev. Disease Dynamics. Birkh¨auser. 12. S.F. Fahmy and S.P. Banks. Robust H ∞ control of uncertain non-linear dynamical systems via linear time-varying approximations. Non-linear Analysis, 63:2315–2327, 2005. 13. A.H. Jazwinski. Stochastic Processes and Filtering Theory. Academic Press, New York, 1970. 14. S.J. Julier and J.K. Uhlmann. A new extension of the kalman filter to non-linear systems. In The Proceedings of Aerosense: The 11th International Symposium on Aerospace/Defense Sensing, Simulation and Controls, Orlando, 1997. 15. T. C¸imen and S.P. Banks. Stochastic optimal control of partially observable non-linear systems. In CDROM Preprints of the 16th IFAC World Congress.

Lecture Notes in Control and Information Sciences Edited by M. Thoma, M. Morari Further volumes of this series can be found on our homepage: springer.com Vol. 367: Turner M.C.; Bates D.G. (Eds.) Mathematical Methods for Robust and Nonlinear Control 444 p. 2007 [978-1-84800-024-7]

Vol. 355: Zhang H.; Xie L.: Control and Estimation of Systems with Input/Output Delays 213 p. 2007 [978-3-540-71118-6]

Vol. 366: Bullo F.; Fujimoto K. (Eds.) Lagrangian and Hamiltonian Methods For Nonlinear Control 398 p. 2007 [978-3-540-73889-3]

Vol. 354: Witczak M.: Modelling and Estimation Strategies for Fault Diagnosis of Non-Linear Systems 215 p. 2007 [978-3-540-71114-8]

Vol. 365: Bates D.; Hagström M. (Eds.) Nonlinear Analysis and Synthesis Techniques for Aircraft Control 360 p. 2007 [978-3-540-73718-6]

Vol. 353: Bonivento C.; Isidori A.; Marconi L.; Rossi C. (Eds.) Advances in Control Theory and Applications 305 p. 2007 [978-3-540-70700-4]

Vol. 364: Chiuso A.; Ferrante A.; Pinzoni S. (Eds.) Modeling, Estimation and Control 356 p. 2007 [978-3-540-73569-4] Vol. 363: Besançon G. (Ed.) Nonlinear Observers and Applications 224 p. 2007 [978-3-540-73502-1] Vol. 362: Tarn T.-J.; Chen S.-B.; Zhou C. (Eds.) Robotic Welding, Intelligence and Automation 562 p. 2007 [978-3-540-73373-7] Vol. 361: Méndez-Acosta H.O.; Femat R.; González-Álvarez V. (Eds.): Selected Topics in Dynamics and Control of Chemical and Biological Processes 320 p. 2007 [978-3-540-73187-0] Vol. 360: Kozlowski K. (Ed.) Robot Motion and Control 2007 452 p. 2007 [978-1-84628-973-6]

Vol. 352: Chiasson, J.; Loiseau, J.J. (Eds.) Applications of Time Delay Systems 358 p. 2007 [978-3-540-49555-0] Vol. 351: Lin, C.; Wang, Q.-G.; Lee, T.H., He, Y. LMI Approach to Analysis and Control of Takagi-Sugeno Fuzzy Systems with Time Delay 204 p. 2007 [978-3-540-49552-9] Vol. 350: Bandyopadhyay, B.; Manjunath, T.C.; Umapathy, M. Modeling, Control and Implementation of Smart Structures 250 p. 2007 [978-3-540-48393-9] Vol. 349: Rogers, E.T.A.; Galkowski, K.; Owens, D.H. Control Systems Theory and Applications for Linear Repetitive Processes 482 p. 2007 [978-3-540-42663-9]

Vol. 359: Christophersen F.J. Optimal Control of Constrained Piecewise Affine Systems 190 p. 2007 [978-3-540-72700-2]

Vol. 347: Assawinchaichote, W.; Nguang, K.S.; Shi P. Fuzzy Control and Filter Design for Uncertain Fuzzy Systems 188 p. 2006 [978-3-540-37011-6]

Vol. 358: Findeisen R.; Allgöwer F.; Biegler L.T. (Eds.): Assessment and Future Directions of Nonlinear Model Predictive Control 642 p. 2007 [978-3-540-72698-2]

Vol. 346: Tarbouriech, S.; Garcia, G.; Glattfelder, A.H. (Eds.) Advanced Strategies in Control Systems with Input and Output Constraints 480 p. 2006 [978-3-540-37009-3]

Vol. 357: Queinnec I.; Tarbouriech S.; Garcia G.; Niculescu S.-I. (Eds.): Biology and Control Theory: Current Challenges 589 p. 2007 [978-3-540-71987-8] Vol. 356: Karatkevich A.: Dynamic Analysis of Petri Net-Based Discrete Systems 166 p. 2007 [978-3-540-71464-4]

Vol. 345: Huang, D.-S.; Li, K.; Irwin, G.W. (Eds.) Intelligent Computing in Signal Processing and Pattern Recognition 1179 p. 2006 [978-3-540-37257-8] Vol. 344: Huang, D.-S.; Li, K.; Irwin, G.W. (Eds.) Intelligent Control and Automation 1121 p. 2006 [978-3-540-37255-4]

Vol. 341: Commault, C.; Marchand, N. (Eds.) Positive Systems 448 p. 2006 [978-3-540-34771-2] Vol. 340: Diehl, M.; Mombaur, K. (Eds.) Fast Motions in Biomechanics and Robotics 500 p. 2006 [978-3-540-36118-3] Vol. 339: Alamir, M. Stabilization of Nonlinear Systems Using Receding-horizon Control Schemes 325 p. 2006 [978-1-84628-470-0] Vol. 338: Tokarzewski, J. Finite Zeros in Discrete Time Control Systems 325 p. 2006 [978-3-540-33464-4] Vol. 337: Blom, H.; Lygeros, J. (Eds.) Stochastic Hybrid Systems 395 p. 2006 [978-3-540-33466-8] Vol. 336: Pettersen, K.Y.; Gravdahl, J.T.; Nijmeijer, H. (Eds.) Group Coordination and Cooperative Control 310 p. 2006 [978-3-540-33468-2]

Vol. 326: Wang, H.-S.; Yung, C.-F.; Chang, F.-R. H∞ Control for Nonlinear Descriptor Systems 164 p. 2006 [978-1-84628-289-8] Vol. 325: Amato, F. Robust Control of Linear Systems Subject to Uncertain Time-Varying Parameters 180 p. 2006 [978-3-540-23950-5] Vol. 324: Christofides, P.; El-Farra, N. Control of Nonlinear and Hybrid Process Systems 446 p. 2005 [978-3-540-28456-7] Vol. 323: Bandyopadhyay, B.; Janardhanan, S. Discrete-time Sliding Mode Control 147 p. 2005 [978-3-540-28140-5] Vol. 322: Meurer, T.; Graichen, K.; Gilles, E.D. (Eds.) Control and Observer Design for Nonlinear Finite and Infinite Dimensional Systems 422 p. 2005 [978-3-540-27938-9]

Vol. 335: Kozłowski, K. (Ed.) Robot Motion and Control 424 p. 2006 [978-1-84628-404-5]

Vol. 321: Dayawansa, W.P.; Lindquist, A.; Zhou, Y. (Eds.) New Directions and Applications in Control Theory 400 p. 2005 [978-3-540-23953-6]

Vol. 334: Edwards, C.; Fossas Colet, E.; Fridman, L. (Eds.) Advances in Variable Structure and Sliding Mode Control 504 p. 2006 [978-3-540-32800-1]

Vol. 320: Steffen, T. Control Reconfiguration of Dynamical Systems 290 p. 2005 [978-3-540-25730-1]

Vol. 333: Banavar, R.N.; Sankaranarayanan, V. Switched Finite Time Control of a Class of Underactuated Systems 99 p. 2006 [978-3-540-32799-8] Vol. 332: Xu, S.; Lam, J. Robust Control and Filtering of Singular Systems 234 p. 2006 [978-3-540-32797-4]

Vol. 319: Hofbaur, M.W. Hybrid Estimation of Complex Systems 148 p. 2005 [978-3-540-25727-1] Vol. 318: Gershon, E.; Shaked, U.; Yaesh, I. H∞ Control and Estimation of State-multiplicative Linear Systems 256 p. 2005 [978-1-85233-997-5]

Vol. 331: Antsaklis, P.J.; Tabuada, P. (Eds.) Networked Embedded Sensing and Control 367 p. 2006 [978-3-540-32794-3]

Vol. 317: Ma, C.; Wonham, M. Nonblocking Supervisory Control of State Tree Structures 208 p. 2005 [978-3-540-25069-2]

Vol. 330: Koumoutsakos, P.; Mezic, I. (Eds.) Control of Fluid Flow 200 p. 2006 [978-3-540-25140-8]

Vol. 316: Patel, R.V.; Shadpey, F. Control of Redundant Robot Manipulators 224 p. 2005 [978-3-540-25071-5]

Vol. 329: Francis, B.A.; Smith, M.C.; Willems, J.C. (Eds.) Control of Uncertain Systems: Modelling, Approximation, and Design 429 p. 2006 [978-3-540-31754-8] Vol. 328: Loría, A.; Lamnabhi-Lagarrigue, F.; Panteley, E. (Eds.) Advanced Topics in Control Systems Theory 305 p. 2006 [978-1-84628-313-0]

Vol. 315: Herbordt, W. Sound Capture for Human/Machine Interfaces: Practical Aspects of Microphone Array Signal Processing 286 p. 2005 [978-3-540-23954-3]

Vol. 327: Fournier, J.-D.; Grimm, J.; Leblond, J.; Partington, J.R. (Eds.) Harmonic Analysis and Rational Approximation 301 p. 2006 [978-3-540-30922-2]

Vol. 313: Li, Z.; Soh, Y.; Wen, C. Switched and Impulsive Systems 277 p. 2005 [978-3-540-23952-9]

Vol. 314: Gil’, M.I. Explicit Stability Conditions for Continuous Systems 193 p. 2005 [978-3-540-23984-0]

Mathematical Methods for Robust and Nonlinear Control: EPSRC Summer School (Lecture Notes in Control and Information Sciences)

Singular Control Systems (Lecture Notes in Control and Information Sciences)

Networked Control Systems (Lecture Notes in Control and Information Sciences)

Ensuring Control Accuracy (Lecture Notes in Control and Information Sciences)

Nonlinear Analysis and Synthesis Techniques for Aircraft Control (Lecture Notes in Control and Information Sciences)

Robust Control via Variable Structure and Lyapunov Techniques (Lecture Notes in Control and Information Sciences)

L2-Gain and Passivity Techniques in Nonlinear Control (Lecture Notes in Control and Information Sciences)

Perspectives in Robust Control (Lecture Notes in Control and Information Sciences)

Robust Flight Control: A Design Challenge (Lecture Notes in Control and Information Sciences)

Feedback Control, Nonlinear Systems, and Complexity (Lecture Notes in Control and Information Sciences)

Assessment and Future Directions of Nonlinear Model Predictive Control (Lecture Notes in Control and Information Sciences)

Nonlinear and Adaptive Control: NCN4 2001 (Lecture Notes in Control and Information Sciences)

Linear and Nonlinear Iterative Learning Control (Lecture Notes in Control and Information Sciences) (v. 291)

Nonlinear Control in the Year 2000: Volume 1 (Lecture Notes in Control and Information Sciences)

Advances in the Control of Nonlinear Systems (Lecture Notes in Control and Information Sciences)

Nonlinear Control in the Year 2000, Volume 2 (Lecture Notes in Control and Information Sciences)

Nonlinear Control Systems: An Algebraic Setting (Lecture Notes in Control and Information Sciences)

Control of Indefinite Nonlinear Dynamic Systems: Induced Internal Feedback (Lecture Notes in Control and Information Sciences)

Robust Control for Uncertain Networked Control Systems with Random Delays (Lecture Notes in Control and Information Sciences)

Nonlinear Control Systems: An Algebraic Setting (Lecture Notes in Control and Information Sciences, Volume 242)

Nonlinear Model Predictive Control: Towards New Challenging Applications (Lecture Notes in Control and Information Sciences)

Nonlinear Control of Wheeled Mobile Robots (Lecture Notes in Control and Information Sciences)

Fuzzy Control and Filter Design for Uncertain Fuzzy Systems (Lecture Notes in Control and Information Sciences)

Control and Observer Design for Nonlinear Finite and Infinite Dimensional Systems (Lecture Notes in Control and Information Sciences)

Recent Advances in Control and Optimization of Manufacturing Systems (Lecture Notes in Control and Information Sciences)

Topics in Time Delay Systems: Analysis, Algorithms and Control (Lecture Notes in Control and Information Sciences)

Robust Stabilization in the Gap-topology (Lecture Notes in Control and Information Sciences) issue 150

Recent Advances in Learning and Control (Lecture Notes in Control and Information Sciences)

Fault Tolerant Control Design for Hybrid Systems (Lecture Notes in Control and Information Sciences)

Robustness in Identification and Control (Lecture Notes in Control and Information Sciences)

Signal Processing for Control (Lecture Notes in Control and Information Sciences)

Mathematical Methods for Robust and Nonlinear Control: EPSRC Summer School (Lecture Notes in Control and Information Sciences)

Singular Control Systems (Lecture Notes in Control and Information Sciences)

Networked Control Systems (Lecture Notes in Control and Information Sciences)

Ensuring Control Accuracy (Lecture Notes in Control and Information Sciences)

Nonlinear Analysis and Synthesis Techniques for Aircraft Control (Lecture Notes in Control and Information Sciences)

Robust Control via Variable Structure and Lyapunov Techniques (Lecture Notes in Control and Information Sciences)

L2-Gain and Passivity Techniques in Nonlinear Control (Lecture Notes in Control and Information Sciences)

Perspectives in Robust Control (Lecture Notes in Control and Information Sciences)

Robust Flight Control: A Design Challenge (Lecture Notes in Control and Information Sciences)

Feedback Control, Nonlinear Systems, and Complexity (Lecture Notes in Control and Information Sciences)

Assessment and Future Directions of Nonlinear Model Predictive Control (Lecture Notes in Control and Information Sciences)

Nonlinear and Adaptive Control: NCN4 2001 (Lecture Notes in Control and Information Sciences)

Linear and Nonlinear Iterative Learning Control (Lecture Notes in Control and Information Sciences) (v. 291)

Nonlinear Control in the Year 2000: Volume 1 (Lecture Notes in Control and Information Sciences)

Advances in the Control of Nonlinear Systems (Lecture Notes in Control and Information Sciences)

Nonlinear Control in the Year 2000, Volume 2 (Lecture Notes in Control and Information Sciences)

Nonlinear Control Systems: An Algebraic Setting (Lecture Notes in Control and Information Sciences)

Control of Indefinite Nonlinear Dynamic Systems: Induced Internal Feedback (Lecture Notes in Control and Information Sciences)

Robust Control for Uncertain Networked Control Systems with Random Delays (Lecture Notes in Control and Information Sciences)

Nonlinear Control Systems: An Algebraic Setting (Lecture Notes in Control and Information Sciences, Volume 242)

Nonlinear Model Predictive Control: Towards New Challenging Applications (Lecture Notes in Control and Information Sciences)

Nonlinear Control of Wheeled Mobile Robots (Lecture Notes in Control and Information Sciences)

Fuzzy Control and Filter Design for Uncertain Fuzzy Systems (Lecture Notes in Control and Information Sciences)

Control and Observer Design for Nonlinear Finite and Infinite Dimensional Systems (Lecture Notes in Control and Information Sciences)

Recent Advances in Control and Optimization of Manufacturing Systems (Lecture Notes in Control and Information Sciences)

Topics in Time Delay Systems: Analysis, Algorithms and Control (Lecture Notes in Control and Information Sciences)

Robust Stabilization in the Gap-topology (Lecture Notes in Control and Information Sciences) issue 150

Recent Advances in Learning and Control (Lecture Notes in Control and Information Sciences)

Fault Tolerant Control Design for Hybrid Systems (Lecture Notes in Control and Information Sciences)

Robustness in Identification and Control (Lecture Notes in Control and Information Sciences)

Signal Processing for Control (Lecture Notes in Control and Information Sciences)

Recommend Documents