Resilient Control of Uncertain Dynamical Systems (Lecture Notes in Control and Information Sciences)

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

303

Springer Berlin Heidelberg NewYork Hong Kong London Milan Paris Tokyo

Magdi S. Mahmoud

Resilient Control of Uncertain Dynamical Systems With 39 Figures and 29 Tables

13

Series Advisory Board A. Bensoussan · P. Fleming · M.J. Grimble · P. Kokotovic · A.B. Kurzhanski · H. Kwakernaak · J.N. Tsitsiklis

Author Prof. Magdi S. Mahmoud UAE University College of Engineering Department of Electrical Engineering P.O. Box 17555-Al Ain Adu-Dhabi United Arab Emirates [email protected]

ISSN 0170-8643 ISBN 3-540-21351-1

Springer-Verlag Berlin Heidelberg New York

Library of Congress Control Number: 2004104464 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 other ways, 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-Verlag. Violations are liable to prosecution under German Copyright Law. Springer-Verlag is a part of Springer Science+Business Media springeronline.com © Springer-Verlag Berlin Heidelberg 2004 Printed in Germany 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: Data conversion by the author. Final processing by PTP-Berlin Protago-TeX-Production GmbH, Berlin Cover-Design: design & production GmbH, Heidelberg Printed on acid-free paper 62/3020Yu - 5 4 3 2 1 0

Dedicated to my Wife Salwa and my Children Medhat, Monda, and Mohamed

Contents

1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 The Fragility Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.2 Robust Control Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.3 Non-fragile (Resilience) Controllers . . . . . . . . . . . . . . . . . . 1.2.4 Parameter Space Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Motivating Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.1 Inverted Pendulum on a Cart . . . . . . . . . . . . . . . . . . . . . . . 1.3.2 Flexible Disk Drive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.3 Simple Heat Exchanger . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.4 Stream Water Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Approaches to Resilient Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.1 Guaranteed-Cost Approach . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.2 The Proposed Framework . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.3 Controller Gain Perturbations . . . . . . . . . . . . . . . . . . . . . . 1.4.4 Multiplicative Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 A Glossary of Terminology and Notations . . . . . . . . . . . . . . . . . . 1.5.1 General Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.2 Functional Differential Equations . . . . . . . . . . . . . . . . . . . 1.6 Outline of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6.2 Chapter Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.7 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1 1 2 2 3 4 4 6 7 9 13 16 17 18 19 19 20 21 21 22 22 23 23 25

2

Resilient Control-Continuous Case . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Model with Norm-Bounded Uncertainties . . . . . . . . . . . . . . . . . . . 2.3 Guaranteed Cost Control I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 Quadratic Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.2 Additive Gain Perturbations . . . . . . . . . . . . . . . . . . . . . . . .

27 27 28 29 30 31

XII

Contents

2.3.3 Multiplicative Gain Perturbations . . . . . . . . . . . . . . . . . . . 2.3.4 Special Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.5 Example 2.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . H∞ Control I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.1 Additive Gain Perturbations . . . . . . . . . . . . . . . . . . . . . . . . 2.4.2 Multiplicative Gain Perturbations . . . . . . . . . . . . . . . . . . . 2.4.3 Special Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.4 Example 2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dynamic Feedback Control I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.1 Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.2 Design Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.3 Example 2.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.4 Example 2.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Model with Convex-Polytopic Uncertainties . . . . . . . . . . . . . . . . . Guaranteed Cost Control II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.7.1 Quadratic Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.7.2 Additive Gain Perturbations . . . . . . . . . . . . . . . . . . . . . . . . 2.7.3 Multiplicative Gain Perturbations . . . . . . . . . . . . . . . . . . . 2.7.4 Special Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.7.5 Example 2.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . H∞ Control II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.8.1 Additive Gain Perturbations . . . . . . . . . . . . . . . . . . . . . . . . 2.8.2 Multiplicative Gain Perturbation . . . . . . . . . . . . . . . . . . . . 2.8.3 Special Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.8.4 Results for Delayless Systems . . . . . . . . . . . . . . . . . . . . . . . 2.8.5 Example 2.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dynamic Feedback Control II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.9.1 Guaranteed Cost Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.9.2 H∞ Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.9.3 Special Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.9.4 Example 2.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Uncertain Model with State-Delays . . . . . . . . . . . . . . . . . . . . . . . . 2.10.1 Resilient Feedback Control . . . . . . . . . . . . . . . . . . . . . . . . . 2.10.2 Error Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.10.3 Robust Stability Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.10.4 Special Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.10.5 Example 2.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.10.6 Example 2.9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

33 34 36 37 37 39 39 41 41 43 44 46 47 47 49 49 51 52 53 55 55 57 58 58 59 60 61 63 65 67 69 71 71 73 74 77 81 81 82

Resilient Control-Discrete Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Model of Discrete-Time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 Gain Perturbations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.2 Quadratic Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

83 83 84 84 85

2.4

2.5

2.6 2.7

2.8

2.9

2.10

2.11 3

Contents

XIII

3.3 Guaranteed Cost Control Synthesis . . . . . . . . . . . . . . . . . . . . . . . . 3.3.1 Additive Gain Perturbations . . . . . . . . . . . . . . . . . . . . . . . . 3.3.2 Multiplicative Gain Perturbations . . . . . . . . . . . . . . . . . . . 3.3.3 Example 3.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.4 Example 3.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 H∞ Control Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Design Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

87 87 89 90 91 91 95 96

4

Resilient Adaptive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 4.2 Continuous-Time Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 4.2.1 Nominal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 4.2.2 Design Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 4.3 Adaptive Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 4.3.1 Known Perturbation Bound . . . . . . . . . . . . . . . . . . . . . . . . 102 4.3.2 Example 4.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 4.3.3 Unknown Gain Perturbation Bound . . . . . . . . . . . . . . . . . 104 4.3.4 Example 4.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 4.3.5 Special Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 4.4 Polytopic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 4.5 Design Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 4.5.1 Known Perturbation Bound . . . . . . . . . . . . . . . . . . . . . . . . 110 4.5.2 Example 4.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 4.5.3 Unknown Gain Perturbation Bound . . . . . . . . . . . . . . . . . 112 4.5.4 Example 4.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 4.5.5 Model Reference State Regulation . . . . . . . . . . . . . . . . . . . 114 4.5.6 Example 4.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 4.5.7 Special Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 4.6 Discrete-Time Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 4.6.1 Stability and Stabilization Results . . . . . . . . . . . . . . . . . . . 119 4.7 Adaptive Stabilization Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 4.7.1 Known Perturbation Bound . . . . . . . . . . . . . . . . . . . . . . . . 122 4.7.2 Unknown Perturbation Bound . . . . . . . . . . . . . . . . . . . . . . 123 4.8 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 4.8.1 Example 4.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 4.8.2 Example 4.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 4.9 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

5

Resilient Linear Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 5.2 A Class of Uncertain Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 5.3 The Resilient Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 5.3.1 A Class of Linear Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 5.3.2 LMI-Based Design Conditions . . . . . . . . . . . . . . . . . . . . . . . 132

XIV

Contents

5.4 5.5

5.6

5.7

5.8 6

5.3.3 Special Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 5.3.4 A Limiting Approach to Kalman Filtering . . . . . . . . . . . . 136 5.3.5 Multiplicative Gain Perturbations . . . . . . . . . . . . . . . . . . . 137 5.3.6 Example 5.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 5.3.7 Example 5.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 Continuous Polytopic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 The Resilient Filtering Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 5.5.1 Stability Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 5.5.2 Delay-Independent Filter Synthesis . . . . . . . . . . . . . . . . . . 143 5.5.3 Multiplicative Gain Perturbations . . . . . . . . . . . . . . . . . . . 146 Resilient Delay-Dependent Filtering . . . . . . . . . . . . . . . . . . . . . . . . 149 5.6.1 A Descriptor Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 5.6.2 Extended Newton-Leibniz Approach . . . . . . . . . . . . . . . . . 154 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 5.7.1 Example 5.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 5.7.2 Example 5.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 5.7.3 Example 5.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 5.7.4 Example 5.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 5.7.5 Example 5.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

Resilient Delay-Dependent Control . . . . . . . . . . . . . . . . . . . . . . . . 171 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 6.2 Continuous-Time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172 6.2.1 Descriptor Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . 172 6.3 Feedback Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 6.3.1 Nominal H2 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 6.3.2 Resilient H2 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 6.3.3 Nominal H∞ Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178 6.3.4 Resilient H∞ Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 6.3.5 Simultaneous Nominal H2 /H∞ Design . . . . . . . . . . . . . . . 180 6.3.6 Simultaneous Resilient H2 /H∞ Design . . . . . . . . . . . . . . . 181 6.3.7 Example 6.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 6.3.8 Example 6.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 6.3.9 Example 6.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 6.4 Continuous Polytopic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 6.4.1 Polytopic H2 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 6.4.2 Polytopic H∞ Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 6.4.3 Simultaneous Polytopic H2 /H∞ Design . . . . . . . . . . . . . . 187 6.4.4 Example 6.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 6.5 Discrete-Time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 6.5.1 System Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 6.5.2 Descriptor Model Transformation . . . . . . . . . . . . . . . . . . . . 189 6.5.3 Robust Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194 6.5.4 Nominal Feedback Stabilization . . . . . . . . . . . . . . . . . . . . . 195

Contents

XV

6.5.5 Resilient Feedback Stabilization . . . . . . . . . . . . . . . . . . . . . 199 6.5.6 Example 6.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 6.5.7 Example 6.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 6.5.8 Example 6.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 6.5.9 Time-Varying Delay-Dependent Stability . . . . . . . . . . . . . 202 6.5.10 Example 6.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204 6.5.11 Example 6.9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 6.6 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 7

Resilient Control-Nonlinear Systems . . . . . . . . . . . . . . . . . . . . . . . 207 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207 7.2 Nonlinear Continuous-Time Systems . . . . . . . . . . . . . . . . . . . . . . . 207 7.3 Robust Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208 7.3.1 Robust Delay-Independent Stability . . . . . . . . . . . . . . . . . 208 7.3.2 Example 7.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211 7.3.3 Robust Delay-Dependent Stability . . . . . . . . . . . . . . . . . . . 211 7.3.4 Example 7.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 7.4 Robust Stabilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 7.4.1 Nominal Feedback Design . . . . . . . . . . . . . . . . . . . . . . . . . . 216 7.4.2 Example 7.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 7.4.3 Example 7.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219 7.4.4 Resilient Feedback Design . . . . . . . . . . . . . . . . . . . . . . . . . . 219 7.4.5 Example 7.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222 7.5 Nonlinear Discrete-Time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 223 7.6 Robust Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224 7.6.1 Example 7.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 7.6.2 Example 7.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 7.7 Robust Stabilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228 7.7.1 Example 7.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229 7.7.2 Example 7.9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229 7.7.3 Example 7.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 7.7.4 Resilient Feedback Design . . . . . . . . . . . . . . . . . . . . . . . . . . 231 7.7.5 Example 7.11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234 7.7.6 Example 7.12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235 7.8 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236

8

Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 8.1 Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 8.1.1 Inequality 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 8.1.2 Inequality 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 8.1.3 Inequality 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238 8.1.4 Inequality 4 (Schur Complements) . . . . . . . . . . . . . . . . . . . 239 8.1.5 Inequality 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240 8.2 Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 8.3 Linear Matrix Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244

XVI

Contents

8.3.1 8.3.2 8.3.3 8.4 Some 8.5 Some 8.5.1 8.5.2

Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 Some Standard Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 245 The S-Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246 Lyapunov-Krasovskii Functionals . . . . . . . . . . . . . . . . . . . . . 247 Formulas on Matrix Inverses . . . . . . . . . . . . . . . . . . . . . . . . . 247 Inverse of Block Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . 248 Matrix Inversion Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . 248

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257

1 Introduction

1.1 Overview Perhaps the problem of designing of optimal and robust controllers has been one of the most active research areas in linear control systems engineering. Over the past two decades or so, several methods have been developed that enable the controller to cope with parametric uncertainties in the plant dynamics. This is true for both types of uncertainties : structured and unstructured. A common denominator of these methods is that they rely on the celebrated YJBK parameterization [1, 7] of all stabilizing controllers for a fixed linear time-invariant plant, which provides a free parameter over which an appropriate function of a closed-loop transfer function may be minimized. Using this framework of parameterization, elegant techniques for minimizing different norms of the closed-loop transfer functions have been formulated. These include H∞ [111, 113], H2 [2] and L1 [13]. Moreover, efficient numerical approaches have been subsequently developed. Strictly speaking, the majority of the foregoing formulations represents a remarkable direction of systems research reflecting robust stability and/or robust performance. The controllers obtained using most robust design approaches are optimal if implemented exactly. It has been observed [43] that modern robust control design methods have been derived with one underlying assumption, that is, the resulting controller from off-line design will be implemented exactly in practice. It can be argued that this is generally a valid point when the main focus is on the plant uncertainty and considering that the controller will be implemented using high-precision hard-ware. Unfortunately, this is not the case in practice. Rather, it is expected that any controller being a part of a closed-loop system must be able to tolerate some uncertainty in its coefficients. Sources of this uncertainties include finite word length, imprecision inherent in analog-digital and digital-analog conversions , finite resolution measuring equipments and round-off errors in numerical computations, to name a few. More importantly, controller redesign becomes a basic operation in fine-tuning of the developed controllers to meet several practical considerations that might

M.S. Mahmoud: Resilient Control of Uncertain Dynamical Systems, LNCIS 303, pp. 1-25, 2004. © Springer-Verlag Berlin Heidelberg 2004

2

1 Introduction

have been overlooked at the beginning. This raises an important issue, that is a robust controller can be very sensitive, or fragile , with respect to errors or perturbations in the controller coefficients. In turn, that brings about a fundamental problem in robust control system design which has been recently termed the fragility problem and hence the design of non-fragile controller opens up as an important research topic that deserve further investigations. Throughout this monograph, we use the terms non-fragile and resilient interchangeably as equivalent terms. We anticipate that resilient controllers to minimize the cost of implementation and allow for on-line tuning of controller parameters .

1.2 The Fragility Problem 1.2.1 Introduction It is generally known that feedback control systems that are designed based on dynamic optimization of single performance criteria or to guarantee robustness with respect to parameter variations of the plant, by and large, may require precise or highly accurate controllers. In connection with the design of statefeedback gain K [1], it is reported that one needs a safety margin for K away from the boundaries of Kγ for the case of implementation inaccuracies. Additionally, in connection with the choice of structure for two-degree-offreedom controllers [124], it is stated that it may be necessary to single one (a structure) out on the basis of system sensitivity to controller uncertainty. To clarify the foregoing points further, consider the classical example [16] of designing an amplifier with open loop gain Ao to be insensitive to variation in Ao . For this purpose, the amplifier is embedded in a feedback configuration with a controller of fractional gain α in the feedback path and the goal is to ensure that the loop gain A β much greater than one. With reference to Figure 1.1, the closed-loop transfer function is

T =

Ao 1 + α Ao

(1.1)

and the sensitivity of T with respect to Ao is given by T = SA o

1 1 + α Ao

(1.2)

T It is readily seen that SA → 0 as Ao becomes very large. On the other hand, o the sensitivity of T with respect to the controller fractional gain α is given by

SαT =

α Ao 1 + α Ao

(1.3)

1.2 The Fragility Problem

3

Fig. 1.1. Amplifier with Feedback

In this case, we see that SαT → 1 as Ao becomes very large. The immediate conclusion [16] is that while a robust design guarantees zero sensitivity against changes in the open-loop gain Ao , it is 100% sensitive to changes in controller fractional gain α. As pointed out in [43], when one attempts to optimize a single performance index it is expected that small changes in controller parameters may lead to performance deterioration of other indeces. Typically a controller designed based on gain margin optimization is in effect very fragile with respect to phase margin. 1.2.2 Robust Control Problem In an engineering setting, the fragility problem discloses the issue of accuracy of controller implementation to the extent that it brings about a trade-off between implementation accuracy and performance deterioration. It is therefore crucial to address and understand all the effects of controller uncertainties in the implementation of robust controllers which optimize a prescribed performance measure in linear dynamical systems. Consider the classical control system of Figure 1.2. The robust design problem is: Given a linear plant P with additive uncertainties ∆P , it is required to find a controller K which internally stabilizes the family of plants P + ∆P and satisfies a given performance measure. There are different algorithms that provide an answer to this problem [1, 7]. Most of the time the focus is on structured uncertainties in the plant to represent the effect of general time-varying parameters whose exact values are unknown but which are known to belong to a given set. By and large, the available algorithms do not incorporate the problems associated with the implementation of uncertain controllers.

4

1 Introduction

Fig. 1.2. Plant with uncertain elements

1.2.3 Non-fragile (Resilience) Controllers At this stage, it is quite reasonable for various practical purposes to restrict attention to structured uncertainties in the controllers. Allowing some tuning procedures for some of the controller parameters will eventually lend the designer additional degrees of freedom to achieve a safety margin with respect to sampling methods, roundoff errors, to name a few. Therefore a more realistic robustness problem would be the one incorporating both plant uncertainties and controller uncertainties as depicted in Figure 1.3. In case of unstructured uncertainties in the plant and using well-known synthesis techniques (like linear quadratic optimization), the resulting controllers exhibit a poor stability margin if not implemented exactly. The term fragile controllers has been introduced in [43] to designate the class of highly accurate controllers. We remark that the fragility is recorded regardless of whether the designed controllers are optimal when implemented using their nominal parameters. It has been further suggested in [43] to overcome the fragility problem to: 1) Develop synthesis methods which incorporates some structured uncertainties in the controller and then search for the ”best” solution guaranteeing a compromise between optimality and fragility. 2) Employ a useful parameterization of the controller. 1.2.4 Parameter Space Design To illustrate the use of controller parameterization, consider Figure 1.4. Let ∆

p = [p1 , p2 , ...., ps ]

1.2 The Fragility Problem

5

Fig. 1.3. Uncertain plant with gain perturbations

denote the vector of real parameters representing the plant with p = po being the nominal value of the plant parameters and let x be the vector of real parameters representing the controller. We limit discussions to a class of parametric uncertainties represented by p ∈ P = {p : p− ≤ pj p+ j j , j = 1, 2, ..., s} ∆

Using a controller with fixed dynamic order, then ∆

x = [x1 , x2 , ...., xr ] denote the vector of real parameters representing the controller Additionally, the performance of the control system is characterized by a vector ∆

a = [a1 , a2 , ...., aq ] of real numbers which are functions of the plant and controller parameters: aj = aj (x, p),

j = 1, 2, ....q

The vector a(x, p) is referred to as the closed-loop performance vector. Suppose that the desired value of the closed-loop performance vector is ∆

d = [d1 , d2 , ...., dq ] ∈ IRn the components of which could represent the coefficient of a desired closed-loop transfer function. In this case, the goal of control system would be choosing the controller parameter vector x such that the set of inequalities aj (x, po ) = dj ,

j = 1, 2, ....q

6

1 Introduction

Fig. 1.4. General plant with uncertain elements

This represents a set of equations the solution of which yields the solution to model matching under nominal conditions. For practical convenience, we relax the desired closed-loop performance to a box-like set + D = {d : d− j ≤ dj ≤ dj , j = 1, 2, ....q ∆

which means that the robust design problem is to choose x if possible so that the following set of inequalities + d− j ≤ aj (x, p) ≤ dj , j = 1, 2, ....q

are satisfied for all p ∈ P. Indeed, the size of the desired box D would have to be adjusted to ensure satisfactory closed-loop performance. It should be observed that the foregoing formulation is carried out in the frequency-domain and its solution using linear programming techniques has been demonstrated [44].

1.3 Motivating Examples In this section, we provide some real-life examples to illustrate the fragility issue and related concepts. The first example deals with an open-loop unstable system (an inverted pendulum). The second example treats a continuoustime system with time-varying parameters (flexible disk drive with read/write head). In the remaining two examples, we shift attention to dynamical models with time-delays. Thus in the third example, we examine a simplified model of heat exchanger with output time-lag arising from finite measurement of

1.3 Motivating Examples

7

terminal temperature. Finally, in the fourth example we discuss a model of water quality in which time-delay phenomena is an intrinsic property due to mass conservation. Throughout the examples, we draw heavily on the use of MATLAB software in simulation. 1.3.1 Inverted Pendulum on a Cart In the course of parameter space designs of robust controllers , it has been shown [1] in case of car steering, cranes and flight control that small gain requirement (due to actuator constraints) frequently leads into a cusp of the stable region. This brings about a version of the fragility trap. In particular, it is not trivial to escape from there because almost all controller gain perturbations eventually lead to instability. To shed more light on this issue and related ideas, we consider the state-feedback design problem of an inverted pendulum shown schematically in Figure 1.5.

Fig. 1.5. An Inverted Pendulum

A linearized state-space model is given by     01 0 0 0  0 0 a23 0   b2     x˙ =  0 0 0 1 x +  0  u b4 0 0 a43 0 where −m g (m + mc ) g , a43 = mc mc  −1 1 = , a23 = mc mc 

a23 = b2

It can easily verified that the open-loop characteristic polynomial

(1.4)

8

1 Introduction

∆o (s) =  mc s4 − (m + mc ) g s2 has double poles s1,2 at origin and two other real poles at  s3,4 = ± (m + mc ) g/(mc ) Using the state-feedback u = −[k1 ,

k2 , k3 ,

k4 ] x

the closed-loop characteristic polynomial can be expressed as ∆c (s) = a4 s4 + a3 s3 + a2 s2 + a2 s + a0 where the coefficients are given by a4 =  mc , a3 = k4 − k2  , a2 = k3 − k1  − g (m + mc ) a1 = g k2 , a0 = g k1 Since  > 0, g > 0, mc > 0, m > 0mc > 0, it follows from Hurwitz theory that the necessary and sufficient conditions for stability are the following algebraic inequalities: kj > 0 j = 0, ...., 4 , k4 > k2  , k3 > k1  + (m + mc ) g g k2 [k3 − k1  − g(m + mc )][k4 − k2 ] > g 2 k22  mc + 2 m2c [k4 − k2 ]2 It is readily seen that the foregoing inequalities are difficult to cope with. Instead, we will explore some relevant special cases. Initially, we proceed to stabilize the system by shifting the positive real root to the stability boundary Re s = 0 using the feedback vector with smallest norm ||k||22 = k12 + k22 + k23 + k42 . The resulting gain vector turns out to be ko t = [0, 0, (m + mc ) g, 0] but the closed-loop characteristic polynomial becomes ∆co (s) = s4 implying a pole at origin with order 4. In attempting to use a marginal stabilization with a feedback vector kt = [ε1 ,

ε2 , (m + mc ) g, ε4 ]

it follows that the problem is not trivial. An alternative method would be to attempt a step-wise shifting by first placing the poles at to double imaginary poles (through manipulation of ε1 , ε2 ) then move the double roots into the left half plane (through manipulation of ε3 , ε4 ). This shows that the problem of minimum norm state-feedback stabilization for an unstable plant leads to a very fragile controller . Moreover, this controller is located in a cusp of the stability region of the four-dimensional

1.3 Motivating Examples

9

Table 1.1. Computational Results of Parameter Variations Case k % Mp yss ||k|| y1 [2 2 36 4] 12.37 0.3334 36.3318 y2 [3 3 54 6] 8.82 0.2498 54.4977 y3 [4 4 72 8] 6.90 0.1998 72.6636

gain space. A possible design route is to construct a singular path out of this cusp in several steps of pole shifting. In the sequel, we perform computer simulation with the following data g = 10,  = 1/3, mc = 4/5, m = 1/2 The results of some technical properties are summarized in Table 1.1 The initial unity feedback response of the system with nominal state feedback vector k = [1, 1, 18, 2] is plotted in Figure 1.6. The step response of with different gain vectors are displayed in Figure 1.7

Fig. 1.6. Initial closed-loop response

1.3.2 Flexible Disk Drive Many dynamical systems including disk-drive read/write head have some flexibility between the sensor and actuator locations, which causes problems in the design of a control system [23]. Therefore, it is important to reflect this in the dynamical model. To elaborate on this point, consider the schematic of Figure 1.9 for a typical disk-drive read/write head system . Let Ih , Im stand for the head and motor inertias, respectively. Assume that the flexible

10

1 Introduction

Fig. 1.7. Different closed-loop responses

shaft has stiffness and damping coefficients k and b, respectively. Ignoring the disturbance torque and letting the angular displacements of the motor and the read head be θm , θr , then a state-space model of the system takes the form:

Fig. 1.8. Disk drive with flexible read head

z = [θm θr  0  0 z˙ =   −k/I1 k/I2

θ˙m θ˙r ]t , u = Tc    0 0 1 0   0 0 1   z +  0  u  1/I1  −b/I1 k/I1 b/I1  b/I2 −k/I2 −b/I2 0

1.3 Motivating Examples

y = [0 1 0 0]z

11

(1.5)

It is readily evident that system (1.5) represents an affine uncertain system [26] with four uncertain parameters in the system matrix. In the sequel, we consider the relative coefficients k(t)/I1 , k(t)/I2 , b(t)/I1 and b(t)/I2 are time-varying parameters over the intervals [0.5 → 2], [0.5 → 2], [0.2 → 1], [0.2 → 1], respectively. The nominal values are taken as k(t)/I1 |nominal = 1.25 , k(t)/I2 |nominal = 1.25 b(t)/I1 |nominal = 0.6 , b(t)/I2 |nominal = 0.6 Our objective hereafter is twofold: First is to study the sensitivities of the time-varying parameters on the dynamical behavior of the controlled system. Second is to examine the impact of the drift in the nominal controller gain factor. For the first objective, we proceed to design a nominal linear-quadratic static state-feedback controller Ko . We do this by using the MATLAB software system with state and controller weighting matrices Q = 10 × I4 , R = 1. The resulting controller gain is given by Ko = [26.6016, 16.8363, 6.3469, 17.8069] , ||k|| = 36.7216 By considering the nominal, maximum and minimum values of the parameters, the ensuing output responses are displayed for the following cases: 1) When all the parameters k(t)/I1 , k(t)/I2 , b(t)/I1 and b(t)/I2 vary simultaneously. The results are depicted in Figure 1.9 2) When the parameter k(t)/I1 varies and others are kept at their nominal values, he results are depicted in Figure 1.10 3) When the parameter k(t)/I2 varies and others are kept at their nominal values, he results are depicted in Figure 1.11 4) When the parameter b(t)/I1 varies and others are kept at their nominal values, he results are depicted in Figure 1.12 and finally 5) When the parameter b(t)/I2 varies and others are kept at their nominal values, he results are depicted in Figure 1.13 A summary of the effects on the controller gain (measured by ||k|| ) and the associated cost (measured by the trace of the Riccati matrix) is presented in Table 1.2 from which it is observed that the factor b/I2 has the most significant effect on the controller gain. To examine the fragility of the controller, we take the parameters at their maximum values and consider an affine family of perturbed controllers described by K = Ko (1 + δk ) where δk is the gain perturbation in the nominal coefficients. Observe that the form of K implies that each gain component has the same relative uncertainty range. By employing MATLAB facilities, we studied the impact of varying δk

12

1 Introduction

Fig. 1.9. Output Response with Variations . max value, .... nominal, − − −min value

in

All

Parameters:

−. −

Fig. 1.10. Output Response with Variations in k/I1 : −. − . max value, − − −min value

on the closed-loop quadratic stability. It is found that if δk is greater than 0.18, the quadratic stability is lost. Now by changing the gain vector over the domain −0.18 ≤ δk ≤ 0.8 and evaluating the associated cost, it is found that the ’critical’ gain is given by Kc = [29.1234, 13.4092, 11.0102, 25.3117] and the associated cost is 1.2478 × 103 implying that there is 13.07% degradation in the design cost as a price paid to guarantee non fragility of the state-feedback controller. Therefore, this example illuminates an important

1.3 Motivating Examples

13

Fig. 1.11. Output Response with Variations in k/I2 : −. − . max value, − − −min value

Fig. 1.12. Output Response with Variations in b/I1 : −. − . max value, − − −min value

issue in controller synthesis, that is, a trade-off exists between system performance and fragility of the controller. 1.3.3 Simple Heat Exchanger In a simple heat exchanger , steam enters the chamber at the top and cooler steam leaves at the bottom. There is a constant flow of water into a series of pipes to pick up the heat from the steam. The relevant variables here are the outflow steam temperature Ts and the out flowing water temperature Tw .

14

1 Introduction

Fig. 1.13. Output Response with Variations in b/I2 : −. − . max value, − − −min value Table 1.2. Computational Results of Parameter Variations Case 1) N ominal 1) M ax 1) M in 2) M ax 2) M in 3) M ax 3) M in 4) M ax 4) M in 5) M ax 5) M in

||k|| 36.7216 53.7593 27.3143 36.9879 37.1318 45.3366 42.6883 36.7623 36.9250 56.3917 29.6593

T r[P ] 1.1036 × 103 2.0343 × 103 0.7811 × 103 1.1545 × 103 1.1107 × 103 1.4413 × 103 1.6835 × 103 1.1039 × 103 1.1166 × 103 2.4018 × 103 0.7542 × 103

The heat process can be controlled by the steam inlet valve openingAs and monitored by the downstream water temperature Tm . Following [23], a simple transfer function of the heat exchanger has the form Ts (s) Ks e−d s = As (s) (1 + τ1 s) (1 + τ2 s)

(1.6)

where Ks is a flow coefficient of the inlet valve, d is the time delay in output measurement and the time constants τ1 , τ2 are functions of the heat exchanger parameters. Alternatively, a state-space model would be     T˙s −α1 α1 Ts α4 = + (1.7) As α2 α3 Tw 0 T˙w Tm (t) = Tw (t − d)

(1.8)

1.3 Motivating Examples

15

where α1 , ..., α4 are scalars related to the time constants of the exchanger. To study the behavior, we select d = 5 , τ1 = 10 , τ2 = 60 The control objective is to design controller for the heat exchanger using the Smith predictor and pole placement [23]. Thus we choose to place the control poles at −0.05 ± j 0.087 and the estimator poles at −0.15 ± j 0.26. We have used Simulink to simulate the response of the closed-loop system (see Figure 1.14).

Fig. 1.14. Simulink diagram of closed-loop system

Standard computations showed that the state-feedback and estimator gains are given by Kc = [K11 K12 ] = [5.2

− 0.17 , Lte = [L11 L12 ] = [0.18 4.2]

We have considered small variations in Kc and the ensuing results are presented in Tables 1.3 and 1.4. The corresponding output responses are displayed in Figures 1.15 and 1.16. It is quite clear that small change in the controller coefficients has some effects on the output response. Similar simulations were performed with the Le coefficients but the results was almost negligible.

16

1 Introduction Table 1.3. Computational Results of Gain Variations-k11 Case y1 y2 y3 y4

Gain 6.5 -0.17 7.8 -0.17 9.1 -0.17 10.4 -0.17

Mp 0.2311 0.2835 0.3329 0.3777

Outputss 1.0224 1.0458 1.0631 1.0764

tp 35 32 30 28

Table 1.4. Computational Results of Gain Variations-k12 Case y1 y2 y3 y4

Gain 5.2 -0.2125 5.2 -0.2250 5.2 -0.2975 5.2 -0.3401

Mp 0.1920 0.1951 0.2286 0.2491

Outputss 0.9892 0.9894 0.9899 0.9905

tp 40 40 40 40

Fig. 1.15. Output response due to different gain coefficients-Case I

1.3.4 Stream Water Quality For all practical purposes, it is crucial to preserve the standards of water quality in streams. This can be measured by the concentrations of some water biochemical constituents. Let z(t), q(t) be the concentrations per unit volume of biological oxygen demand (BOD) and dissolved oxygen (DO), respectively, at time t. Under the simplifying assumptions that the stream has a constant flow rate and the water is well mixed. Further, we assume that there exists a τ > 0 such that the (BOD,DO) concentrations entering at time t are equal to the corresponding concentrations τ time units ago. Using mass balance concentration and employing a linearization about an equilibrium operating point, the growth of (BOD,DO) can be expressed as:

1.4 Approaches to Resilient Control

17

Fig. 1.16. Output response due to different gain coefficients-Case II



z(t) ˙ q(t) ˙



   z(t) kcd kod z(t − τ ) kce 0 + = q(t) q(t − τ ) −kde −kre −kdd −krd   mse 0 uz (t) + 0 mre uq (t) 

(1.9)

where kce , ... , mre are composite rates some of which are unknown and uz (t), uq (t) are appropriate effluent controls. Using representative data on a single reach of the River Nile, system (1.9) has the form      z(t) ˙ 1 0 z(t) −0.45 −0.50 z(t − τ ) = + q(t) ˙ −3 −2 q(t) −0.15 −0.10 q(t − τ )   1.2 0 uz (t) + (1.10) 0 1.4 uq (t) With τ + = 0.65, the nominal state feedback control    uz (t) 1.9237 −0.0307 z(t) = uq (t) −0.1145 0.8415 q(t) renders the water quality system asymptotically stable. Introducing 15% variations in the gain matrix coefficients can be easily shown to destabilize the system for 0 < τ ≤ 0.57.

1.4 Approaches to Resilient Control In this section, we outline some of the available approaches to resilient control design. Initially, from scanning the literature it is easy to observe that the

18

1 Introduction

majority of practical methods for control system design include the following ingredients: 1) A fixed structure controller with controller design parameters. It might be argued that smaller number of parameters would be an advantage in adjustment and tuning procedure. However with the growing software capabilities, versatile simulation packages are readily available so that large number of control parameters could be easily accommodated. 2) Performance specification in terms of multiple indices. In this way, several design objectives could be included for proper assessment of the desired behavior. 3) Selection of control system components including sensors and actuators followed by controller implementation. Let the controller parameters and the plant parameters be represented by q and p, respectively. Suppose that the performance specifications are formulated in terms of a set of N algebraic inequalities gj (r, p, q), j = 1, ..., N, where r is a generic variable stands for t in the time domain or ω in the frequency domain. Then, broadly speaking, the general objective is find the set Q of admissible parameter vector values that meet all of the performance specifications, that is gj (r, p, q) > 0,

r ∈ R , p ∈ P , j = 1, ..., N

(1.11)

Indeed, the foregoing family of inequalities encompass the procedure outlined in section (2.4). There are some techniques for the computations of the set Q in the frequency domain including A) Symbolic Quantifier-Elimination Techniques [15] These techniques are based on the development of quantifier-free logic formula involving polynomial inequalities and logic connectives. B) Interval Arithmetic Techniques [50] Interval arithmetic methods aim at exploring the range of functionals gj (ω, p, q) in the frequency domain and hence test for positivity. It further use brach-and-bound algorithms for the determination of the design parameter set Q. C) Statistical Learning Techniques [114] These probabilistic methods can be used to determine the design parameter set Q in an approximate manner with probabilistic guarantees. 1.4.1 Guaranteed-Cost Approach The main rationale behind the guaranteed-cost (GC) approach is to design feedback controllers that guarantee prescribed levels of performance for all plant parameter vectors p ∈ P [10]. In this case, the linear model of the plant has the form x(t) ˙ = A(p) x(t) + B(p) u(t)

(1.12)

1.4 Approaches to Resilient Control

19

where x(t) ∈ IRn is the state vector; u(t) ∈ IRm is the control input. A fixed state-feedback controller u(t) = K x(t) such that a prescribed level of integral-quadratic performance measure

∞  t x (t) Q x(t) + ut (t) R u(t) , Q > 0 , R > 0 (1.13) J = 0

is guaranteed for all possible plant vectors p ∈ P, where K is the unknown gain matrix. To employ this approach in the design of non-fragile controllers, one has to assume that the unknown feedback gain matrix is a function of the controller-uncertainty parameter vector, that is K = K(q), q ∈ Q. In this case, GC methods can be used to guarantee a level of performance with uncertainties in both p and q [18, 38, 41]. 1.4.2 The Proposed Framework In this monograph, we take a more general perspective by focusing on a widerclass of linear dynamical systems described by x(t) ˙ = A(∆p ) x(t) + Ad (∆p )x(t − τ ) + B(∆p )u(t) + Γ w(t) y(t) = C(∆p )x(t) + Cd (∆p )x(t − τ ) + Ψ w(t) z(t) = D(∆p )x(t) + F (∆p )u(t) + Φw(t)

(1.14)

where x(t) ∈ IRn is the state vector; u(t) ∈ IRm is the control input; w(t) ∈ q is the disturbance input which belongs to L2 [0, ∞); y(t) ∈ p is the measured output; z(t) ∈ r is the controlled output and τ ∈ [0, τ ∗ ] is an unknown time-varying delay factor satisfying 0 ≤ τ ≤ τ ∗ , 0 ≤ τ˙ ≤ τ + where the bounds τ ∗ , τ + < 1 are known constants and ∆p stands for the plant uncertainties. In terms of controller uncertainties ∆c , the feedback gain is represented by K(∆c ). A block diagram representation is shown in Figure 1.15. System (1.14) contains state and output time-varying delays. Throughout the book we will closely examine the development of control methods that generate various forms of the feedback gain K(∆c ) including state-feedback (instantaneous and delayed), dynamic output feedback , observers , adaptive controllers and filters . Interestingly enough, the proposed framework will encompass all previously published work [18, 38, 41, 118, 119, 120, 121, 122] and extend them in various technical ways 1.4.3 Controller Gain Perturbations In the subsequent chapters of the book we have to characterize the perturbed gain K(∆c ). Below, we describe some of the models that we are going to deal with. These are essentially two types of models:

20

1 Introduction

Fig. 1.17. Uncertain plant with time-delay and gain perturbations

Additive Type: K(∆c ) = Ko + ∆K(t) , ∆K(t) = Ha ∆c (t) Ea

(1.15)

where Ko is the nominal gain matrix, Ha and Ea are known constant matrices of appropriate dimensions and ∆c (t) is the uncertain parameter matrix which ∆ belongs to a compact set such that ∆c ∈ ∆ = {∆c : ∆tc ∆c ≤ I}. In the course of design, the matrices Ha and Ea could be manipulated to examine the robustness of the proposed method. A special case of (1.15) is that ∆K is a constant perturbation matrix to be determined in the part of controller synthesis [18] using appropriate numerical optimization. 1.4.4 Multiplicative Type K(∆c ) = Ko + ∆K(t) , ∆K(t) = Hm ∆c Em Ko

(1.16)

where Hm and Em are known constant matrices of appropriate dimensions. With Ko = {kij }, i = 1, ..., m, j = 1, ..., n, an alternative form of (1.16) is [18]:   k11 (1 + δ11 ) · · · k1n (1 + δ1n )   .. .. .. Ko + ∆K =   . . . km1 (1 + δm1 ) · · · kmn (1 + δmn )    k11 δ11 · · · k1n δ1n k11 · · · k1n     .. .. .. =  ... . . . ...  +   . . . km1 · · · kmn km1 δm1 · · · kmn δmn 

1.5 A Glossary of Terminology and Notations

−1 < δ¯ij ≤ δij ≤ δˆij < 1

21

(1.17)

where δij represents the relative percentage drift from the nominal entries of the gain matrix Ko .

1.5 A Glossary of Terminology and Notations In this section, we assemble the terminologies and notations to be adopted throughout the book with the objective of paving the way to the technical development of subsequent chapters. These terminologies and notations are quite standard in the scientific media and only vary in form or character. 1.5.1 General Terms As a start, matrices as n × m dimensional arrays of elements with n-rows and m-columns are represented by capital letters while vectors as n-tuples or columns (unless otherwise specified) and scalars (single elements) are represented by lower case letters. We use IR, IR+ , IRn and IRn×m to denote the set of reals, positive real numbers, real n-tuples (vectors) and real n×m matrices, respectively. Alternatively, IRn is called the Euclidean space and is equipped ∆  with the vector-norm as ||x|| = [x21 + · · · x2n ]. The terms f (t), g(s) denote, respectively, scalar-valued functions of the real variables t and s. The quantities x, ˙ x ¨ are the first and the second derivative of x with respect to time, respectively. The symbols [., .], (., .], (., .) denote, respectively, closed, semiclosed, and open intervals; that is t ∈ (a, b] ⇒ a < t ≤ b. The open left-half (≡ {s : Re(s) ≤ 0}, the open proper left-half (≡ {s : Re(s) < 0} and the open proper right-half (≡ {s : Re(s) > 0} of the complex plane are represented by C I †, C I − and C I + , respectively. We use Up ∈ n×n and Uk ∈ m× to denote, respectively, the set of uncertain plant perturbations ∆A of the nominal dynamical system A and the set of uncertain controller perturbations ∆K of the nominal controller gain K. The Lebsegue space L2 [0, ∞) consists of square integrable functions on the interval [0, ∞) and equipped with the norm ∆

||p||2 =



∞

1/2 pt (τ ) p(τ ) dτ

(1.18)

0

For any square matrix W of arbitrary dimension n×n, let W t , W −1 , λ(W ), r(W ), tr(W ), det(W ), ρ(W ) and ||W || denote, respectively, the transpose, the inverse, the spectrum, the rank, the trace, the determinant, the spectral radius and the induced norm defined by 1/2

∆ ||W || = λ W W t (1.19) We use W > 0 (≥, <, ≤ 0) to denote a symmetric positive definite (positive semidefinite, negative, negative semidefinite matrix W with λm (W ) and

22

1 Introduction

λM (W ) being the minimum and maximum eigenvalues of W . Frequently, I stands for the identity matrix with appropriate dimension, W † denotes the pseudo-inverse of W and diag(W1 , · · · , Wp ) stands for the block-diagonal matrix   W1 0 · · · 0  0 W2 · · · 0    (1.20)  0 ··· ··· 0  0 · · · · · · Wp 1.5.2 Functional Differential Equations I 0], IRn ) denotes the Banach space of continuous vector Let C I n,τ = C([−τ, functions mapping the interval [−τ, 0] into IRn with the topology of uniform convergence and designate the norm of an element φ ∈ C I n,τ by ∆

||φ||∗ =

sup ||φ(θ)||2

(1.21)

θ∈[−τ,0]

If α ∈ IR, d ≥ 0 and x ∈ C([α I − τ, α + d], IRn ) then for any t ∈ [α, α + d], I be defined by xt (θ) := x(t + θ), −τ ≤ θ ≤ 0. If D ⊂  × C, I we let xt ∈ C f : ID → IRn is a given function, the relation x(t) ˙ = f (t, xt ) is a retarded functional differential equation (RFDE) on ID where xt (t), t ≥ to denotes the restriction of x(·) to the interval [t − τ, t] translated to [−τ, 0]. Here, τ > 0 is termed the delay factor. In the sequel, if α ∈ IR, d ≥ 0 and x ∈ C([α I − τ, α + d], n ) then for any t ∈ [α, α + d], we let xt ∈ C I be defined ∆ by xt (θ) = x(t + θ), −τ ≤ θ ≤ 0 . In addition, if ID ⊂ IR × C, I f : ID → IRn is given function, then the relation x(t) ˙ = f (t, xt )

(1.22)

is a retarded functional differential equation (RFDE) on ID where xt , t ≥ t0 denotes the restriction of x(.) on the interval [t − τ, t] translated to [−τ, 0]. Here τ > 0 is termed the state-delay factor. A function x is said to be a solution of (1.22) on [α − τ, α + d] if there α ∈ IR and d > 0 such that x ∈ C([α I − τ, α + d], IRn ), (t, xt ) ∈ ID, t ∈ [α, α + d]

(1.23)

and x(t) satisfies (1.22) for t ∈ [α, α+d]. For a given α ∈ IR, φ ∈ C, I x(α, φ, f ) is said to be a solution of (1.22) with initial value φ at α.

1.6 Outline of the Book The chief objective of this monograph is to provide a complete description of resilient control and filtering theory. In our viewpoint, this theory virtually

1.6 Outline of the Book

23

stands as robust redesign approach and based thereon, it should be distinguished from robust control theory since the latter focuses mainly on parametric uncertainties. Thus this monograph brings together advanced methods for resilient control and filtering of dynamical systems. For generality in exposition, the main focus is on time-delay models as they capture additional features over conventional models described by ordinary differential equations. In writing up the different topics, emphasis is primarily placed on major developments attained thus far and then reference is made to other related work. Overall, the monograph is a suitable reference for graduate-students and researchers from wide-spectrum of engineering disciplines, science and mathematics. 1.6.1 Methodology Throughout the monograph, our methodology is composed of five-steps: • • • • •

Mathematical Modeling in which we discuss the main ingredients of the state-space model under consideration. Definitions and/or Assumptions here we state the definitions and/or constraints on the model variables to pave the way for subsequent analysis. Analysis and Examples this signifies the core of the respective sections and subsections which contains some solved examples for illustration. Results which are provided most of the time in the form of theorems, lemmas and corollaries. Remarks which are given to shed some light of the relevance of the developed results vis-a-vis published work.

In the sequel, theorems(lemmas, corollaries) are keyed to chapters and stated in italic font with bold titles, for example, Theorem 3.4 means Theorem 4 in Chapter 3 and so on. For convenience, we have grouped the reference in one major bibliography cited towards the end of the book. Relevant notes and research issues are offered at the end of each Chapter for the purpose of stimulating the reader. We hope that this way of articulating the information will attract the attention of a wide-spectrum of readership. 1.6.2 Chapter Organization The material covered is divided into 7 chapters whereby continuous-time results go in parallel with discrete-time results. All the developed results are conveniently expressed in LMI feasibility problem. It is fair to mention that

24

1 Introduction

almost all the analytical developments hereafter are new and original contributions to the control of time-delay systems. In Chapter 2 we start our guided tour by considering the design of resilient feedback controllers for a class of linear continuous-time systems with parametric uncertainties and state-delay. The main vehicle in the discussions to follow is the use of state-space modeling and design tools. Our effort treats both cases of norm-bounded and convex-bounded. The controller gain perturbations are either of additive and multiplicative forms. The control objective is to design resilient feedback stabilization schemes using guaranteed cost control and H∞ control approaches. Necessary and sufficient conditions for closed-loop robust quadratic stability are derived as opposed to the available sufficient conditions in the literature . The solution of all design problems under consideration is cast as convex optimization over linear matrix inequalities. The case of state-variable measurements is initially considered leading to resilient static feedback control. Then we move to examine the case of outputvariable measurements where the gain matrices of a dynamic output feedback control are constructed using an appropriate design procedure. The objective of Chapter 3 is to develop resilient controllers for a class of discrete-time systems with norm-bounded uncertainties and controller gain variations. The focus is on the development of resilient controllers for uncertain discrete-time systems with both types of gain variations. Necessary and sufficient conditions are established such that the resulting closed-loop feedback control system is quadratically stable for all admissible perturbations and uncertainties. These conditions are conveniently expressed in the form of linear matrix inequalities (LMIs). The feedback stabilization schemes are based on guaranteed cost control and H∞ control approaches. Chapter 4 addresses the problem of resilient adaptive control problem for classes of uncertain continuous-time and discrete-time systems with statedelays against controller gain variations. In the continuous case, design results on both norm-bounded and convex-bounded parametric uncertainties are derived. In the discrete-case, design results on norm-bounded parametric uncertainties are developed. Adaptive control schemes are constructed when gain perturbation bounds are known and then extended to accommodate unknown norm-bounded perturbations. To complete the design profile, special cases of delayless continuous and discrete systems are provided. The problem of robust nonfragile filtering for a class of linear uncertain continuous-time systems is investigated in Chapter 5. Initially we focus on norm-bounded uncertainties. We consider additive filter gain variations to reflect the imprecision in filter implementation. As a limiting procedure, the case of nonfragile Kalman filter is derived. All the developed results are conveniently extended to the case of multiplicative filter gain variations. Later on, we investigate the problem of designing a resilient L2 − L∞ filter for a class of linear uncertain state delay systems with uncertainties which belong to a convex bounded polytopic domain. The objective is to derive tractable synthesis conditions for the resilient design of full-order and reduced-order

1.7 Notes and References

25

filters such that a prescribed energy-to-peak disturbance-attenuation level is attained for all admissible uncertainties and gain perturbations. Both additive and multiplicative gain perturbations are considered. It is established that the filter design problems can be obtained from the solution of convex optimization problem over linear matrix inequalities. The design results are developed for both delay-independent and delay-dependent cases. In the latter case, two approaches are used: descriptor and extended Newton-Leibniz. In Chapter 6, we direct attention to delay-dependent methodologies for continuous-time and discrete-time systems by exhibiting the delay-dependence dynamics in the design procedure. For the continuous-time case, we develop a descriptor approach to simultaneous H2 /H∞ control design for a class of time-delay systems with additive and multiplicative controller gain perturbations. The delay factor is considered unknown-but-bounded constant or timevarying. The cases of nominal and convex polytopic uncertainties are treated. In the discrete-time case, delay-dependent stability and feedback stabilization methods are developed for the nominal, robust and resilience cases. We end the monograph by Chapter 7 in which we move further to examine classes of nonlinear continuous-time and discrete-time systems with statedelay. We develop an LMI-based analysis and design procedures to check primarily into the robust stability of both continuous-time and discrete-time systems. Then we address the robust stabilization using nominal and resilient feedback designs. In both cases the trade-off between the size of the controller gains and the bounding factors is illuminated and incorporated into the design formalism. Seeking computational convenience, all the developed results are cast in the format of family of LMIs. An appendix containing some relevant mathematical tools and basic results is provided.

1.7 Notes and References In addition to the numerous papers and articles on time-delay systems, there are a number of reference books which might have some connection to the topics to be discussed in this book. This includes [8, 81]. The discussions presented thus far follow the work of [7, 12, 15, 16]. It is hoped that this introductory chapter has succeeded in motivating the readers to the upcoming topics. We have made every effort to produce the book as a self-contained examination of the background and progress of resilient control and filtering of uncertain dynamical systems.

2 Resilient Control-Continuous Case

2.1 Introduction An integral part of robust control system design methods has been based on using a fixed quadratic Lyapunov function in order to guarantee robust stability [17, 73]. An implicit assumption inherent in these design methods is that the controller will be implemented exactly. In the presence of uncertain parameters, it is often desirable to perform the control system design not only to ensure stability but also to guarantee an adequate level of system performance. This brings about the notion of guaranteed cost control (GCC) [103] and H∞ -control [17]. When attending to the controller implementation based on different control design methods, it turns out that the controllers are very sensitive ”fragile” with respect to errors in the controller coefficients. This is due to imprecision in analog-digital conversion, fixed word length, finite resolution instrumentation and numerical roundoff errors. It is considered beneficial that the designed (nominal) controllers should be capable of tolerating some level of controller gain variations. This illuminates the controller fragility problem as discussed in Chapter 1 and in the discussions to follows we address methods to overcome this problem. This chapter considers the design of resilient feedback controllers for a class of linear continuous-time systems with parametric uncertainties and statedelay . The main vehicle in the discussions to follow is the use of state-space modeling and design tools. Our effort will treat both cases of norm-bounded and convex-polytopic. The controller gain perturbations are either of additive and multiplicative forms. The control objective is to design resilient (nonfragile) feedback stabilization schemes using guaranteed cost control and H∞ control approaches. Throughout this chapter, necessary and sufficient conditions for closed-loop robust quadratic stability are derived as opposed to the available sufficient conditions in the literature . The solution of all design problems under consideration is cast as convex optimization over linear matrix inequalities (LMIs) . This ensures the full use of convenient computational environment based on available commercial software. In the sequel, the case of

M.S. Mahmoud: Resilient Control of Uncertain Dynamical Systems, LNCIS 303, pp. 27-82, 2004. © Springer-Verlag Berlin Heidelberg 2004

28

2 Resilient Control-Continuous Case

state-variable measurements is initially considered leading to state-feedback control. Then we move to examine the case of output-variable measurements where the gain matrices of a dynamic output feedback control are constructed using an appropriate design procedure. Several system examples are provided to illustrate the theoretical developments.

2.2 Model with Norm-Bounded Uncertainties A schematic of the problem setup is displayed in Figure 2.1 which shows a plant P subjected to uncertainties ∆p and a controller K having gain perturbations ∆c . Let the plant be represented by the class of linear continuous-time systems with state-delay : (Σ∆ ) :

x(t) ˙ = [Ao + ∆Ao (t)]x(t) + [Ad + ∆Ad (t)]x(t − τ ) + Bo u(t) + Γ w(t) = A∆o (t)x(t) + A∆d (t)x(t − τ ) + Bo u(t) + Γ w(t) y(t) = Co x(t) + Cd x(t − τ ) + Ψ w(t) z(t) = Do x(t) + Fo u(t) + Φw(t)

(2.1)

where x(t) ∈ IRn is the state vector; u(t) ∈ IRm is the control input; w(t) ∈ IRq is the disturbance input which belongs to L2 [0, ∞); y(t) ∈ IRp is the measured output; z(t) ∈ IRr is the controlled output and τ ∈ [0, τ ∗ ] is an unknown time-varying delay factor satisfying 0 ≤ τ ≤ τ ∗ , 0 ≤ τ˙ ≤ τ + where the bounds τ ∗ , τ + < 1 are known constants. The matrices Ao , Bo , Ad , Co , Do , Cd , Fo , Γ, Φ, Ψ are known real constant matrices of appropriate dimensions which describe the nominal system of (Σ∆ ). In this part, the matrices ∆Ao (t), ∆Ad (t) are real, time-varying matrix functions representing the norm-bounded parameter uncertainties and represented by: [∆Ao (t) ∆Ad (t)] = Ma ∆p (t)[Na Nd ] ∆

∆p (t) ∈ ∆p (t) = {∆p (t) : ∆tp (t)∆p (t) ≤ I ∀ t}

(2.2)

where Ma ∈ IRn×α , Na ∈ IRβ×n and Nd ∈ IRβ×n are known real constant matrices and the elements of ∆p (t) are being Lebesgue measurable. The initial condition is specified as x(0), x(s) = xo , φ(s) , where φ(.) ∈ L2 [−τ, 0]. We remark here that the norm-bounded uncertainty structure considered here has been widely adopted in robust control [73, 80, 82]. It covers the usual matching condition as a special case by setting N ≡ I. Observe that the unit upperbound condition on ∆(t) does not cause any loss of generality. Indeed ∆p (t) can always be normalized by appropriately selecting the matrices M and N .

2.3 Guaranteed Cost Control I

29

Fig. 2.1. Uncertain Plant with Controller Perturbations

2.3 Guaranteed Cost Control I Associated with system (Σ∆ ) is the cost function :

∞ [xt (t)Qx(t) + ut (t)Ru(t)] dt J=

(2.3)

0

where 0 < Q = Qt ∈ IRn×n and 0 < R = Rt IRm×m are given weighting matrices. Consider for the time being that we have a state-feedback controller u(t) = Ko x(t) with Ko being the static nominal (designed) gain matrix but the actual implemented controller is assumed to be u(t) = [Ko + ∆K(t)] x(t) = K∆ x(t)

(2.4)

where ∆K(t) represents the gain perturbation . Definition 2.1. Consider system (Σ∆ ) with cost function (2.3) and w(t) ≡ 0. The state-feedback control law (2.4) is said to be a guaranteed cost control (GCC) with quadratic cost matrix P > 0 given matrices 0 < Q = Qt , 0 < W = W t and 0 < R = Rt , if the following LMI  t P A∆ + At∆ P + K∆ RK∆ + Q + W P A∆d < 0 ˆ • −W ˆ = (1 − τ + )W (2.5) W has a feasible solution with respect to P for all admissible uncertainties ∆A(t) satisfying (2.2) and gain perturbations ∆K(t) We note that the foregoing definition holds for arbitrary gain variations. In the absence of controller perturbations ∆K(t) ≡ 0, then Definition 2.1 reduces to the well known definition of guaranteed cost control for uncertain continuous-time systems [81].

30

2 Resilient Control-Continuous Case

2.3.1 Quadratic Stability Considering the GCC approach, the following theorem provides a link between the notion of quadratic cost matrix and an upper bound on the cost function J. Theorem 2.2. Consider system (Σ∆ ) with cost function (2.3). Suppose that the state-feedback control law (2.4) with controller gain perturbations ∆K(t) is a quadratic GCC with matrix P > 0. Then the resulting closed-loop system with w(t) ≡ 0 x(t) ˙ = A∆ (t)x(t) + A∆d (t)x(t − τ )

(2.6)

is quadratically stable for all admissible uncertainties satisfying (2.3) with the cost function bounded by

0 J ≤ xto P xo + xt (s) W x(s) ds (2.7) −τ

Proof (⇒) : If P > 0 is a quadratic cost matrix for system (Σ∆ ) and cost function (2.3), it follows from (2.5) and the Schur complements operations that   t t t −1 t ˆ x P A∆ + A∆ P + K∆ RK∆ + Q + W + P A∆d W A∆d P x < 0 ∀ x = 0 , ∀ ∆(t) ∈ ∆p (t)

(2.8)

Hence, system (Σ∆ ) is quadratically stable . Next, introduce the Lyapunov-Krasovskii functional

0 xt (s)W x(s) ds V (x(t)) = xt (t)P x(t) + −τ

Standard matrix manipulations along the solutions of (2.6) yield:   V˙ (x) = xt (t) At∆ P + P A∆ + W x(t) + 2xt P A∆d x(t − τ )  − xt (t − τ )(1 − τ˙ )W x(t − τ ) ≤ xt (t) At∆ P + P A∆ + W    −1 t t t ˆ + P A∆d W A∆d P x(t) < − x (t) Q + K∆ RK∆ x(t) From which we conclude that   t t x (t) Q + K∆ RK∆ x(t) < V˙ (x)

(2.9)

(2.10)

Integrating over the period t ∈ [0, ∞) and using 2.3, we get J ≤ V (xo ) − V (x(∞))

(2.11)

2.3 Guaranteed Cost Control I

31

By (2.8), system (Σ∆ ) is quadratically stable thereby leading to V (x(t)) → 0 as t → ∞ and therefore (2.11) reduces to

J ≤ V (xo ) = xto P xo +

0

xt (s)W x(s) ds

(2.12)

−τ

(⇐) : Let system (Σ∆ ) be quadratically stable. It then follows on using the Schur complement operations that there exist 0 < P = P t such that t ˆ −1 At∆d P + Q + W < 0 P A∆ + At∆ P + K∆ RK∆ + P A∆d W ∀ ∆(t) ∈ ∆(t) (2.13)

Hence, one can find some µ > 0 such that the following inequality t ˆ −1 At∆d P + Q + W ] = µ−1 [P A∆ + At∆ P + K∆ RK∆ + P A∆d W ¯ˆ −1 t ¯ t ¯ ¯+W ¯ < 0 A∆d P + Q RK∆ + P¯ A∆d W P¯ A∆ + At∆ P¯ + K∆

∀ ∆(t) ∈ ∆p (t)

(2.14)

This inequality implies that, with the help of Schur complements, there exists ¯ = µ−1 W, Q ¯ = µ−1 Q, R ¯ = µ−1 R, such that the matrix matrices W −1 P¯ = µ P is a quadratic cost matrix for system (Σ∆ ). This completes the proof. ∇∇∇ 2.3.2 Additive Gain Perturbations It is assumed that the gain perturbation matrix ∆K(t) has the form: ∆K(t) = Ha ∆c Ea

(2.15) ∆

where Ha and Ea are known constant matrices and ∆c (t) ∈ ∆c (t) = {∆c (t) : ∆tc (t)∆c (t) ≤ I ∀ t} is the uncertainty parameter matrix. In the sequel, we establish LMI-based results for GCC synthesis such that the closed-loop controlled system achieve quadratic stability for all admissible uncertainties ∆p (t) ∈ ∆p (t). For simplicity in exposition, we define the matrices: Af = Ao + Bo Ko , ∆Af (t) = ∆A(t) + Bo ∆K(t) A∆ = Af + ∆Af (t)

(2.16)

The following theorem provides a necessary and sufficient condition for GCC under additive gain perturbations. Theorem 2.3. Consider system (Σ∆ ) with cost function (2.3). There exist state feedback gain Ko such that the control law (2.4) with additive gain perturbations (2.15) is a GCC with a quadratic cost matrix P > 0 given matrices 0 < Q = Qt , 0 < W = W t and 0 < R = Rt if and only if there exist matrices

32

2 Resilient Control-Continuous Case

0 < X = X t , Y, Z, L and some parameters ε > 0,  > 0, µ > 0, such that the following LMIs   Y t R+ Ad + t t Ξ M B H ZN LE XQ XW a a  a a o a ZNat Nd µBo Ha Hat Bot      • −εI 0 0 0 0 0 0 0     • • −I 0 0 0 0 0 0     • • • −εI 0 0 0 0 0     • • • • −I 0 0 0 0 <0    • • • • • −Q 0 0 0     • • • • • • −W 0 0     ˆ −W +   • • • • • • • 0 t   εNd Nd     −R+ • • • • • • • • t µRHa Ha R  − µ ≥0 (2.17) • −µ are feasible with respect to X and Y where Ξa = Ao X + Bo Y + XAto + Y t Bot . The feedback gain is Ko = Y X −1 . Proof: By Definition 2.1 and [73], it follows that inequality (2.5) holds if and only if the matrix inequality   P (Ao + Bo Ko )+   t  (Ao + Bo Ko )t P Kot R P Ad  P Ma P Ma     + ε−1  0   0  +W + Q    • −R 0  0 0 ˆ • • −W   t   t t  t t  t t  t Na P Bo Ha Ea Na Ea P Bo Ha + ε  0   0  + −1  RHa   RHa  +   0   0  Ndt Ndt 0 0 0 0   P (Ao + Bo Ko )+ t   (A o + Bo Ko ) P + t  P Ad +   W + Q + ε−1 P Ma Mat P −1 Ko R+ t  t    +−1 P Bo Ha Hat Bo P +  P Bo Ha Ha R εNd Na     < 0 εNat Na + Eat Ea (2.18) =     −R+   0 •   −1 RHat Ha R    ˆ+  −W • • εNd Ndt holds some parameters ε > 0,  > 0. By the congruent transformation diag[P −1 , I, I], P −1 = X, Ko P −1 = Y, −1 = µ, εP −1 = Z, P −1 = L and followed by Schur complement operations we obtain the equivalent LMI conditions of (2.17). ∇∇∇

2.3 Guaranteed Cost Control I

33

2.3.3 Multiplicative Gain Perturbations Here, it is assumed that the gain perturbation matrix ∆K(t) has the form: ∆K(t) = Hm ∆c Em Ko

(2.19) ∆

where Hm and Em are known constant matrices and ∆c (t) ∈ ∆c (t) = {∆c (t) : ∆tc )∆c (t) ≤ I ∀ t} is the uncertainty parameter matrix. The following theorem builds on Theorem 2.3 and provides necessary and sufficient conditions for GCC under multiplicative gain perturbations. Theorem 2.4. Consider system (Σ∆ ) with cost function (2.3). There exist state feedback gain Ko such that the control law (2.4) with multiplicative perturbations (2.19) is a GCC with a quadratic cost matrix P > 0 given matrices 0 < Q = Qt , 0 < W = W t and 0 < R = Rt if and only if there exist matrices 0 < X = X t , Y, Z, L and some parameters ε > 0,  > 0, µ > 0, such that the following LMIs   Y t R+ Ad + t t Ξ M B H ZN GE XQ XW t a m  a a o m Bot  ZNat Nd µBo Hm Hm    • −εI  0 0 0 0 0 0 0    • • −I  0 0 0 0 0 0    • •  • −εI 0 0 0 0 0    • •  • • −I 0 0 0 0  <0  • •  • • • −Q 0 0 0    • •  • • • • −W 0 0     ˆ −W +  • •  • • • • • 0 t   εNd Nd     −R+ • • • • • • • • t µRHm Hm R  − µ ≥0 (2.20) • −µ are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 . Proof: Similar to Theorem 2.3, it follows that inequality (2.5) holds if and only if the matrix inequality   P (Ao + Bo Ko )+   t  (Ao + Bo Ko )t P Kot R P Ad  P Ma P Ma     + ε−1  0   0  +W + Q    • −R 0  0 0 ˆ • • −W   t  t   t t  t t  t t Na P Bo Hm Ea P Bo Ha Na Ea + ε  0   0  + −1  RHm   RHa  +   0   0  0 0 Ndt Ndt 0 0

34

2 Resilient Control-Continuous Case



 P (Ao + Bo Ko )+   (Ao + Bo Ko )t P + t  P Ad +   W + Q + ε−1 P Ma Mat P −1 Ko R+ t    P Bo Ha Ha R εNdt Na  t  +−1 P Bo Hm Hm  B P + o   t t t   εN N + K E E K = a a o m m o    −R+   0 • −1 t   RH H R  a a    ˆ −W +  • • εNd Ndt

< 0

(2.21)

holds some parameters ε > 0,  > 0. By the congruent transformation diag[P −1 , I, I], P −1 = X , Ko P −1 = Y, −1 = µ εP −1 = Z, P −1 Kot = G and followed by Schur complement operations we obtain the LMI conditions of (2.20). ∇∇∇ 2.3.4 Special Cases We provide hereafter, resilient control results on some systems of particular interests. In the absence of system uncertainties (M ≡ 0, Na ≡ 0, Nd ≡ 0) we obtain the nominal system (Σo ) : x(t) ˙ = Ao x(t) + Ad x(t − τ ) + Bo u(t)

(2.22)

for which implementation of Theorem 2.3 reduces to: Theorem 2.5. Consider system (Σ∆ ) with cost function (2.3). There exist state feedback gain Ko such that the control law (2.4) with additive gain perturbations (2.15) is a GCC with a quadratic cost matrix P > 0 given matrices 0 < Q = Qt , 0 < W = W t and 0 < R = Rt if and only if there exist matrices 0 < X = X t , Y, Z, L and parameters ε > 0,  > 0, µ > 0 such that the following LMIs   Y t R+ t  Ξa Bo Ha LEa XQ XW Ad µBo Ha Hat Bot     • −I 0  0 0 0 0    •  • −εI 0 0 0 0    • <0 • • −Q 0 0 0    •  • • • −W 0 0    •  ˆ • • • • −W 0     −R+ • • • • • • µRHa Hat R  − µ ≥0 (2.23) • −µ are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 .

2.3 Guaranteed Cost Control I

35

On the other hand, implementation of Theorem 2.4 becomes: Theorem 2.6. Consider system (Σ∆ ) with cost function (2.3). There exist state feedback gain Ko such that the control law (2.4) with multiplicative gain perturbations (2.19) is a GCC with a quadratic cost matrix P > 0 given matrices 0 < Q = Qt , 0 < W = W t and 0 < R = Rt if and only if there exist matrices 0 < X = X t , Y, Z, L and parameters ε > 0,  > 0, µ > 0 such that the following LMIs   Y t R+ t Ξ B H GE XQ XW A d t m  a o m µBo Hm Hm Bot      • −I 0 0 0 0 0     • • −εI 0 0 0 0   <0  • • • −I 0 0 0     • • • • −Q 0 0     • ˆ • • • • −W 0     −R+ • • • • • • t µRHm Hm R  − µ ≥0 (2.24) • −µ are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 . Alternatively, by suppressing the delay term, we deal with the system (ΣM ) :

x(t) ˙ = A∆o (t)x(t) + Bo u(t) + Γ w(t) y(t) = Co x(t) + Cd x(t − τ ) + Ψ w(t) z(t) = Do x(t) + Fo u(t) + Φw(t)

(2.25)

for which the following results hold. Theorem 2.7. Consider system (ΣM ) with cost function (2.3). There exist state-feedback gain Ko such that the control law (2.4) with additive gain perturbations (2.15) is a GCC with a quadratic cost matrix P > 0 given matrices 0 < Q = Qt and 0 < R = Rt if and only if there exist matrices 0 < X = X t , Y, Z, L and some parameters ε > 0,  > 0, µ > 0, such that the following LMIs   Y t R+ t t  Ξa Ma Bo Ha ZNa LEa XQ µBo Ha Hat Bot      • −εI 0 0 0 0 0     • • −I 0 0 0 0   <0  • • • −εI 0 0 0     • • • • −I 0 0     • • • • • −Q 0     −R+ • • • • • • µRHa Hat R  − µ ≥0 (2.26) • −µ

36

2 Resilient Control-Continuous Case

are feasible with respect to X and Y where Ξa = Ao X + Bo Y + XAto + Y t Bot . The feedback gain is Ko = Y X −1 . Theorem 2.8. Consider system (ΣM ) with cost function (2.3). There exist state feedback gain Ko such that the control law (2.4) with multiplicative gain perturbations (2.19) is a GCC with a quadratic cost matrix P > 0 given matrices 0 < Q = Qt and 0 < R = Rt if and only if there exist matrices 0 < X = X t , Y, Z, L and some parameters ε > 0,  > 0, µ > 0, such that the following LMIs   Y t R+ t t Ξ XQ M B H ZN GE t a m  a a o m Bot  µBo Hm Hm    • −εI  0 0 0 0 0    • • −I  0 0 0 0    • • <0 • −εI 0 0 0    • •  • • −I 0 0    • •  • • • −Q 0     −R+ • • • • • • t µRHm Hm R  − µ ≥0 (2.27) • −µ are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 . Remark 2.9. It is significant to observe that in the absence of controller gain perturbations (Ha ≡ 0, Hm ≡ 0, Ea ≡ 0, Em ≡ 0), Theorems 2.5 and 2.6 recover the LMI-based necessary and sufficient conditions for GCC with quadratic cost matric P [92]. Finally, it should be observed that the nice structures of LMIs (2.17), (2.20), (2.23), (2.24), (2.26) and (2.27) have been enabled through the clever use of linearization (slack) variables [25]. Additionally, Theorems 2.7 and 2.8 improves drastically upon the results of [118, 121] from different technical aspects. 2.3.5 Example 2.1 Consider a continuous-time system of the type (2.1) with     0 3 1 −0.45 0.3 Ao = , Bo = , Γ = , Ma = , Hm = 0.2 −4 −5 0 0.90 0   0 0.2 Nat = , Co = [0.5 0.5] , Do = 1 , Ha = 1 0.1 Ea = [0.3 0.6] , Em = 0.4   0 0.1 0 Ad = , Ndt = , τ + = 0.65 −0.2 −0.3 0.7 A summary of the numerical computations is given in Table 2.1, from which it is clear that the case of multiplicative gain perturbations yields less cost bound than additive gain perturbations.

2.4 H∞ Control I

37

Table 2.1. Computational Results of Example 2.1 P erturbation Theorem Q R Jo W Ko Additive 2.3 3 0 1.5 1.823 5 0 -1.5947 -2.1769 03 05 Multiplicative 2.4 4 0 1.5 0.976 6 0 -2.6092 -1.9205 04 06

2.4 H∞ Control I In this section, we extend the results attained in the forgoing section to the case of H∞ control. For a given a prespecified H∞ -norm level γ, it is required to develop conditions for the state-feedback controller (2.4) with controller gain perturbations (2.15) or (2.19) to render the resulting closed-looped system quadratically stable and ||z||22 ≤ γ ||w||22 for all admissible uncertainties ∆tp (t)∆p (t) ≤ I. Following [81, 92], we have the following definition: Definition 2.10. Consider system (Σ∆ ). The state-feedback control law (2.4) with controller gain perturbations (2.15) or (2.19) is said to be a H∞ with disturbance attenuation γ > 0 if and only if there exists a matrix 0 < P t = P such that the following LMI   P A∆ + At∆ P + W P A∆d P Γ + Dkt Φ   < 0 ˆ (2.28) • −W 0 • • −γ 2 I + Φt Φ has a feasible solution for all admissible uncertainties ∆A(t) and ∆K(t) satisfying (2.2) and (2.15) or (2.19), where Dk = Do + Fo Ko . By a parallel development, the corresponding results for controller gain variations of both types are established by the following theorems. 2.4.1 Additive Gain Perturbations For the additive type of gain perturbations, we have Theorem 2.11. Consider system (Σ∆ ) with cost function (2.3). There exist a state feedback gain Ko such that the control law (2.4) with additive gain perturbations (2.15) is a H∞ control with disturbance attenuation γ > 0 if and only if there exist matrices 0 < X = X t , Y, Z, L and some parameters ε > 0,  > 0, such that the following LMIs

38

2 Resilient Control-Continuous Case

  Ξa   •   •   •   •   •    •    •

 Ad + Γ + XDot Φ ZNat Nd +Y t Fot Φ    −εI 0 0 0 0 0 0   • −I 0 0 0 0 0   • • −εI 0 0 0 0  <0 • • • −I 0 0 0   • • • • −W 0 0   ˆ −W +  • • • • • 0 t  εNd Nd  2 −γ I+  • • • • • • Φt Φ Ma Bo Ha ZNat LEat XW

(2.29)

are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 . Proof: By Definition 2.10 and [73], it follows that inequality (2.28) holds if and only if the matrix inequality   P (Ao + Bo Ko )+  (Ao + Bo Ko )t P P Ad P Γ + Dkt Φ    t   P Ma P Ma   +W   + ε−1  0   0    ˆ • −W 0   0 0 2  γ I+  • • Φt Φ   t  t   t t  t t  t t P Bo Ha Ea P Bo Ha Na Ea Na   +  0   0  0 0 + ε  Ndt   Ndt  + −1  0 0 0 0 0 0   P (Ao + Bo Ko )+  (Ao + Bo Ko )t P +   P Ad + P Γ + Dot Φ   W + ε−1 P Ma Mat P  t t t    +−1 P Bo Ha Hat Bo P + εNa Nd +Ko Fo Φ     < 0 εNat Na + Eat Ea (2.30) =     ˆ − W +   • 0   εNd Ndt   2  −γ I+  • • Φt Φ holds some parameters ε > 0,  > 0. By the congruent transformation diag[P −1 , I, I], P −1 = X and using

Ko P −1 = Y, −1 = µ, εP −1 = Z, P −1 = L

followed by Schur complement operations we obtain the equivalent LMI conditions of (2.29). ∇∇∇

2.4 H∞ Control I

39

2.4.2 Multiplicative Gain Perturbations For the multiplicative perturbation type, we have the following result Theorem 2.12. Consider system (Σ∆ ) with cost function (2.3). There exist state-feedback gain Ko such that the control law (2.4) with multiplicative gain perturbations (2.19) is a H∞ control with disturbance attenuation γ > 0 if and only if there exist matrices 0 < X = X t , Y, Z, L and some parameters ε > 0,  > 0, such that the following LMIs   Ad + Γ + XDot Φ t t  Ξa Ma Bo Hm ZNa GEm XW ZNat Nd +Y t Fot Φ     • −εI  0 0 0 0 0 0    • • −I  0 0 0 0 0    • •  • −εI 0 0 0 0    • • <0 • • −I 0 0 0 (2.31)    • •  • • • −W 0 0     ˆ+ −W  • •  • • • • 0 t   εN N d d    −γ 2 I+  • • • • • • • Φt Φ are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 . 2.4.3 Special Cases Theorem 2.13. Consider system (Σo ) with cost function (2.3). There exist state-feedback gain Ko such that the control law (2.4) with additive gain perturbations (2.15) is a H∞ control with disturbance attenuation γ > 0 if and only if there exist matrices 0 < X = X t , Y, Z, L and some parameters ε > 0,  > 0, such that the following LMIs   Γ + XDot Φ t  Ξa Bo Ha LEa XW Ad +Y t Fot Φ     • −I 0  0 0 0    •  • −εI 0 0 0  <0 (2.32)  •  • • −W 0 0    •  ˆ • • • −W 0    −γ 2 I+  • • • • • Φt Φ are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 . Theorem 2.14. Consider system (Σo ) with cost function (2.3). There exist state-feedback gain Ko such that the control law (2.4) with multiplicative gain perturbations (2.19) is a H∞ control with disturbance attenuation γ > 0 if

40

2 Resilient Control-Continuous Case

and only if there exist matrices 0 < X = X t , Y, Z, L and some parameters ε > 0,  > 0, such that the following LMIs   Ad + Γ + XDot Φ t Ξ B H GE XW m  a o m ZNat Nd +Y t Fot Φ     • −I  0 0 0 0    •  • −I 0 0 0  <0 (2.33)  •  • • −W 0 0    •  ˆ • • • −W 0    −γ 2 I+  • • • • • Φt Φ are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 . Theorem 2.15. Consider system (ΣM ) with cost function (2.3). There exist a state-feedback gain Ko such that the control law (2.4) with additive gain perturbations (2.15) is a H∞ control with disturbance attenuation γ > 0 if and only if there exist matrices 0 < X = X t , Y, Z, L and some parameters ε > 0,  > 0, such that the following LMIs   t t t Γ + XDo Φ Ξ M B H ZN LE a a o a a a +Y t F t Φ   o    • −εI 0  0 0 0    • • −I  0 0 0  <0 (2.34)  • •  • −εI 0 0    • •  • • −I 0   2  −γ I+  • • • • • Φt Φ are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 . Theorem 2.16. Consider system (ΣM ) with cost function (2.3). There exist state feedback gain Ko such that the control law (2.4) with multiplicative gain perturbations (2.19) is a H∞ control with disturbance attenuation γ > 0 if and only if there exist matrices 0 < X = X t , Y, Z, L and some parameters ε > 0,  > 0, such that the following LMIs   t t t Γ + XDo Φ  Ξa Ma Bo Hm ZNa GEm +Y t Fot Φ     • −εI  0 0 0 0    • • −I  0 0 0  <0 (2.35)  • •  • −εI 0 0    • •  • • −I 0   2  −γ I+  • • • • • Φt Φ are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 .

2.5 Dynamic Feedback Control I

41

2.4.4 Example 2.2 Consider a continuous-time system of the type (2.1) with     −2 1 1 1 0.2 Ao = , Bo = , Γ = , Ma = , Hm = 0.3 1 1 0 1 0.2   0.3 0.5 Nat = , Co = [0.5 0.5] , Do = 1 , Ha = 0.2 0.5 Ea = [0.5 0.4] , Em = 0.3   0.2 0.1 0.1 t Ad = , Nd = , τ + = 0.75 0.3 −0.1 0.5 A summary of the numerical computations is given in Table 2.2, from which it is clear that the case of multiplicative gain perturbations yields less noise attenuation level. Table 2.2. Computational Results of Example 2.2 P erturbation Theorem Q R γ W Ko Additive 2.11 6 1 2.2 3.454 5 0 0.5210 1.7842 1 3 05 Multiplicative 2.12 10 2 4.1 2.016 6 0 1.017 -3.114 2 5 06

2.5 Dynamic Feedback Control I Unlike the foregoing two sections dealing with the design of resilient static feedback controller, we examine hereafter the problem of designing full-order resilient dynamic output feedback controller such that the resulting closedloop time-delay system achieves quadratic stability for all admissible uncertainties. For this purpose, the controller dynamics have the form: (Σ∆C ) :

x˙ c (t) = [Ac + ∆Ac (t)]xc (t) + [Bc + ∆Bc (t)]y(t) = A∆c (t)x(t) + B∆c (t)y(t) u(t) = [Kc + ∆Kc (t)]xc (t) = K∆c xc (t)

(2.36)

where xc (t) ∈ IRnc is the controller state vector. One possible representation of the controller perturbations is given by [∆Ac (t) ∆Bc (t)] = Mc ∆c (t)[Nc Nb ] , ∆Kc (t) = Hc ∆c (t)Ec ∆

∆c (t) ∈ ∆c (t) = {∆c (t) : ∆tc (t) ∆c (t) ≤ I}

(2.37)

42

2 Resilient Control-Continuous Case

Now the problem of interest is phrased as follows: Given the uncertain system (Σ∆ ) and the dynamic outputfeedback controller (Σ∆C ) with controller perturbations (2.37), it is required to determine the controller matrices Ac , Bc , Kc such that the closed-loop controlled system achieve quadratic stability for all admissible uncertainties satisfying ∆p (t) ∈ ∆p (t) , ∀t and for all controller perturbations ∆c (t) ∈ ∆c (t) , ∀t. By augmenting systems (Σ∆ ) and (Σ∆C ), the closed-loop time-delay system can be written as (Σ∆DC ) :

˙ = A∆ξ (t)ξ(t) + A∆dξ (t)ξ(t − τ ) + Γξ w(t) ξ(t)

(2.38)

where A∆ξ (t) = Aξo + ∆Aξ (t) , A∆dξ (t) = Adξo + ∆Adξ (t)   ∆Ao Bo ∆Kc Ao Bo Kc Aξo = , ∆Aξ (t) = Bc Co Ac ∆Bc Co ∆Ac   Γ x(t) ∈ IRn+nc , Γξ = ξ(t) = B∆c Ψ xc (t)   0 Ad 0 ∆Ad Adξo = , ∆Aξd = 0 Bc Cd 0 ∆Bc Cd

(2.39)

By Definition (2.1), it follows that system (Σ∆DC ) is quadratically stable for all admissible uncertainties satisfying ∆p (t) ∈ ∆p (t) , ∀t if given matrices Q, W, R, the matrix inequality   A∆ξ X + X At∆ξ + A t  X K∆ξ RK∆ξ X + X (Q + W)X ∆dξ  < 0 (2.40) ˆ • −W holds. In turn, this is equivalent to the case that the matrix inequality  A∆ξ X + X At∆ξ X Λ∆ <0 (2.41) • −Υ is feasible with respect to X , Ac , Bc and Kc , where K∆ξ = Kξo + ∆Kξ = [0 Kc ] + [0 t Λ∆ = [K∆ξ Q W X −1 A∆dξ ]

Hc ∆ c Ec ]

= Λo + ∆Λ t = [Kξo Q W X −1 Adξo ] + ˆ] [∆Kξt 0 0 X −1 ∆Adξ ] , Υ = diag[R Q W W

(2.42)

In the sequel, we adopt a parameterization procedure to cast the nonlinear matrix inequality (2.42) into an alternative LMI characterization which amenable to reliable software computations.

2.5 Dynamic Feedback Control I

43

2.5.1 Parameterization The following intermediate result is first established: Lemma 2.17. The matrix inequality (2.41) is feasible with respect to X , Ac , Bc and Kc if and only if the following matrix inequality is feasible in with respect to X , Ac , Bc , Kc and V   −(V + V t ) V t A∆ξ + X V t Λ∆ V t  • −X 0 0  <0  (2.43)  • • −Υ 0  • • • −X Proof: By rewriting (2.43) in the form   0 X 0 0  • −X 0 0  t t t    • • −I 0  + Ω VΠ + Π V Ω • • • −X Ω = [−I A∆ξ Λ∆ I] , Π = [I 0 0 0]

(2.44)

(2.45)

and observing that the explicit bases of the null spaces of Ω and Π are given by     A∆ξ Λ I 000  I 0 0 I 0 0    (2.46) NΩ =   0 I 0  , NΠ =  0 I 0  0 0I 00I We apply Lemma (8.8) of the Appendix with respect to matrix V in (2.44) and notice that the existence of V, X satisfying (2.44) is equivalent to the existence of X satisfying the inequality   A∆ξ X + X At∆ξ − X X Λ∆ X  (2.47) • −Υ 0  < 0 • • −X Inequality (2.47) is equivalent to (2.42) by a Schur complement argument. ∇∇∇ Now, given some scalars ε1 > 0, .... , ε7 > 0, we introduce the matrices:   0 Ao + f1 Bo Kc Ad + f3 , A¯dξo = A¯ξo = Bc Co Ac + f2 0 Bc Cd + f4  0 ¯ ct Q W X −1 A¯dξo ] ¯ ct = , Λ¯o = [K K Kct + f5 −1 t t t t t t f1 = ε−1 1 Ma Ma + ε1 Na Na + ε3 Co Nb Nb Co + ε4 Bo Hc Hc Bo (2.48) −1 −1 t t t t f2 = ε2 Mc Mc + ε2 Nc Nc + ε3 Mc Mc + ε4 Ec Ec

−1 t t t t t f3 = ε−1 6 Ma Ma + ε6 Nd Nd , f4 = ε7 Mc Mc + ε7 Co Nb Nb Co t t f5 = ε−1 (2.49) 5 Ec Ec + ε 5 Hc Hc

44

2 Resilient Control-Continuous Case

In the manner of Theorem 2.3, it is a now straightforward task to show by applying Inequality 1 of the Appendix and performing Schur complement operations with the linearizing relation Z t = V t X −1 that inequality (2.43) holds if and only if the following matrix inequality  ¯ t V t Q V t W Z t A¯dξo V t  −(V + V t ) V t A¯ξo + X V t K c  • −X 0 0 0 0 0     • • −R 0 0 0 0     • • • −Q 0 0 0  (2.50)  <0   • • • • −W 0 0    ˆ • • • • • −W 0  • • • • • • −X is feasible with respect to the matrices X , Z, Ac , Bc , Kc and V and the scalar parameters ε1 > 0, .... , ε7 > 0. 2.5.2 Design Procedure Our purpose hereafter is to derive tractable synthesis conditions for the resilient dynamic output feedback controller. To simplify the discussions to follow, we start by defining the following matrices   Qx 0 Xx Xd ∈ IR2n×2n , X = ∈ IR2n×2n Q= • Qc • Xc   Vx Vd Wx 0 ∈ IR2n×2n , W = ∈ IR2n×2n V= • Wc Vg Vc  Zx Zd ∈ IR2n×2n Z= (2.51) Zg Zc It should be observed that (2.50) implies that V > 0 and thus we can assume that Vg > 0 and Vc > 0. Next we introduce the matrices Ω1 = Vgt Vc−t Vg , Ω2 = Vgt Vc−t Vdt

(2.52)

With (2.49) and (2.51), the general matrix inequality (2.50) becomes:   −Ξa Ξb Ξc Ξd Ξe Ξf V t  • −X 0 0 0 0 0     •  • −R 0 0 0 0    • <0 • • −Q 0 0 0 (2.53)    •  • • • −W 0 0    • ˆ 0  • • • • −W • • • • • • −X where

2.5 Dynamic Feedback Control I



t





45



Vxt Qx Vgt Qc Vdt Qx Vct Qc  t Vx (Ao + f1 ) + Vgt Bc Co + Xx Vxt Bo Kc + Vgt (Ac + f2 ) + Xd Ξb = Vdt (Ao + f1 ) + Vgt Bc Co + Vdt Vdt Bo Kc + Vct (Ac + f2 ) + Xc  t  t Zx (Ad + f3 ) Zgt (Bc Cd + f4 ) Vx Wx Vgt Wc (2.54) , Ξ Ξe = = f Zdt (Ad + f3 ) Zct (Bc Cd + f4 ) Vdt Wx Vct Wc

Ξa =

Vx + Vxt Vd + Vg Vg + Vdt Vc + Vct



, Ξc =

Vgt (Kct + f5 ) Vct (Kct + f5 )

, Ξd =

To cast inequality (2.53) into an LMI format, we perform a congruent transformation  diag I Vc−1 Vg , I Vc−1 Vg , I, I I, I I, I Vc−1 Vg , I Vc−1 Vg followed by a linearization procedure defined by ˆc = Vgt Bc , K ˆ c = Kc Vc−1 Vg Aˆc = Vgt Ac Vc−1 Vg , B Xˆd = Xd Vc−1 Vg , Xˆc = Vgt Vc−t Xc Vc−1 Vg , Xˆx = Xx

(2.55)

Ψa = Vxt Bo Kc Vc−1 Vg , Ψb = Vgt Vc−t Bo Kc Vc−1 Vg Ψe = Vgt f2 Vc−1 Vg , Ψd = Vct Kct , Ψf = Zgt Bc Cd Ψg = Zct Bc Cd , Ψc = Vgt Kct

(2.56)

The result of the two consecutive transformations is finally expressed by the following LMI:   −Πa Πb Πc Ξd Ξe Πf V t  • −Xˆ 0 0 0 0 0      • • −R 0 0 0 0   <0  • • • −Q 0 0 0 (2.57)     • • • • −W 0 0    • ˆ 0  • • • • −W • • • • • • −X where

 Ψc + Vgt f5 Vx + Vxt Ω1 + Ω2t , Π = c Ω1 + Ω2 Ω1 + Ω2t Ψd + Vct f5  t ˆc Co + Xˆx Ψa + Aˆc + Ψe + Xˆd V (A + f1 ) + B Πb = ˆxt o ˆc Co + Xˆ t Ψb + Aˆc + Ψe + Xˆc Vd (Ao + f1 ) + B d   t Zx (Ad + f3 ) Ψf + Zgt f4 Xˆx Xˆd ˆ , X = Πf = Zdt (Ad + f3 ) Ψg + Zct f4 • Xˆc 

Πa =

(2.58)

We observe that all the foregoing linearizations are reversible transformations ensuring that the ensuing resulting conditions are equivalent. In brief, the following lemma summarizes the main result.

46

2 Resilient Control-Continuous Case

Lemma 2.18. The nonlinear matrix inequality (2.41) is feasible with respect to X , Ac , Bc and Kc if and only if the LMI (2.57) is feasible in with respect ˆc , K ˆ c , Vx , Ω1 , Ω2 , Ψa , Ψb to Xˆ , Xˆ , Z, Aˆc , B Ψ c , Ψd , Ψf , Ψg . The computation of the controller gains proceeds as follows: Step 0: Rescale the variables εj , j = 1, .., 5 by defining σj = εj /(1 + εj ) so that εj > 0 is equivalent to σj ∈ (0, 1). Assign a uniform grid on each σj . Step 1: For each grid point of {σ1 , ..., σ5 }, solve the LMI (2.57) to obtain a feasible solution. Step 2: Solve the factorization problem Ω1 = Vgt Vc−t Vg to determine Vc and Vg . Step 3: Solve   −t   −1 ˆc Vg 0 Vg Vc 0 Aˆc B Ac Bc = ˆc 0 Kc 0 0 I 0 I K to determine the triplet Ac , Bc , Kc . Remark 2.19. Unlike other direct dynamic feedback control schemes [79, 82, 84, 91, 94, 90] using partitioning and Riccati inequalities, the matrix inequality (2.43) offers new potential since it does not involve product of the Lyapunov matrix X and system matrices A¯ξo and A¯dξo . Remarkably enough, the employment of linearization transformation paved the way to casting the feedback design problem to a convex optimization over LMIs as the design procedure. 2.5.3 Example 2.3 Consider a third-order system of the type (2.2) with       0 1 −2 −3 −1 Ao =  1 0 0  , Bo =  1  , Γ =  0  , Hm = 0.2 1 1 0 1 0   0.6 Nat =  0.85  , Co = [1 0 1] , Do = 0.8 0.45     0.4 0.3 Ha =  0.4  , Ea = [0.4 0.6 0.5] , Ma =  0.4  0.4 0.2     0 1 −1 0.1 Ad =  1 0 1  , Ndt =  0.5  , τ + = 0.5 , Em = 0.5 −1 −1 0 0.7 By invoking the MATLAB toolbox, the ensuing numerical results are summarized in Tables 2.3 and 2.4.

2.6 Model with Convex-Polytopic Uncertainties

47

Table 2.3. Results of Guaranteed Cost Control Ko Jo W P erturbation Theorem Q R 1 0 0 1.0 1.9165 3 1 0 -11.4557 -9.2338 0.2573 Additive 2.3 031 030 005 023 Multiplicative 2.4 2 1 0 3.5 1.7556 4 0 1 -8.0337 -14.1035 0.3359 031 051 026 014 Table 2.4. Results of H∞ Control P erturbation Theorem Q R γ W Ko Additive 2.3 8 0 0 2.5 5.1653 6 1 0 -6.0339 9.7356 -1.1233 051 050 024 008 Multiplicative 2.4 9 1 0 4.0 7.2677 7 0 1 -7.1145 11.1035 -2.0914 071 071 016 027

2.5.4 Example 2.4 Given the data for a second-order system     0 1 0 0.5 0.1 Ao = , Bo = , Γ = , Ma = , Nb = 0.5 1 −2 1 0.5 0  0.7 Mc = , Co = [1 0] , Do = 2 , Ec = 0.4 , Nc = [0.4 0.4] , Hc = [1 1] 0.6    0 0 0.1 0.2 t t Ad = , Nd = , Na = , τ + = 0.6 0.1 0.1 0 0.2 For selected values of Q, R, and W, a summary of the numerical computations is given in Table 2.5. It should be recorded that these are obtained after few iterations of the prescribed design procedure.

2.6 Model with Convex-Polytopic Uncertainties With reference to Figure 2.1, let the plant be represented by the class of linear continuous-time systems with state-delay : (ΣJ ) :

x(t) ˙ = Ao x(t) + Ad x(t − τ ) + Bo u(t) + Γ w(t) y(t) = Co x(t) + Cd x(t − τ ) z(t) = Do x(t) + Fo u(t) + Φw(t)

(2.59) (2.60) (2.61)

where x ∈ IRn is the system state, u(t) ∈ IRp is the control input, y(t) ∈ IRp is the measured output, z(t) ∈ IRr is the controlled output and τ is a time-delay

48

2 Resilient Control-Continuous Case Table 2.5. Computational Results of Dynamic Controller Q R W Ac I2 3 3 I2 -9.4069 -0.0031 9.3975 -2.1165 4I2 7 9 I2 -10.1091 -0.1356 7.3599 -3.6295 2I2 6 3 I2 -10.088 -0.3561 8.8358 -2.6115 10I2 5 I2 -11.1034 -0.0031 8.2469 -2.1165

Bc -12.5987 17.6889 -8.8566 16.9116 -10.2477 13.3492 -6.4557 11.8236

Kc -11.0419 -0.0081 -10.1934 -0.7723 -12.3547 -0.8010 -17.9002 -0.7452

factor such that 0 < τ ≤ β , 0 < τ˙ ≤ α < 1. The corresponding state-space matrices contain uncertainties represented by a real convex bounded polytopic model of the type     Ao Ad Bo A(oλ) A(dλ) B(oλ) ∆  Co Cd 0  ∈ Sλ =  C(oλ) C(dλ) 0  = D(oλ) 0 F(oλ) Do 0 Fo    N Aoj Adj Boj    ,λ ∈ Λ (2.62) λj Coj Cdj 0 j=1 Doj 0 Foj where Λ is the unit simplex  ∆

Λ=

(λ1 , · · · , λN ) :

N 

 λ j = 1 , λj ≥ 0

(2.63)

j=1

Define the vertex set IN = {1, ..., N }. Throughout this section, we use {Ao , · · · , Fo } to imply generic system matrices and {Aoj , · · · , Foj , j ∈ IN} to represent the respective values at the vertices. The initial condition is specified as x(0), x(s) = xo , φ(s) , where φ(.) ∈ L2 [−τ, 0]. The cost function associated with system (ΣJ ) is: Jj =

N

 j=1

0

∞

λj [xt (t)Qj x(t) + ut (t)Rj u(t)] dt ,

j ∈ IN

(2.64)

where 0 < Qj = Qtj and 0 < Rj = Rjt are given weighting matrices. For a given state-feedback controller u(t) = Ko x(t) with Ko being the nominal controller gain, the actual implemented controller due to various practical sources is assumed to be u(t) = [Ko + ∆K(t)] x(t) = K∆ x(t)

(2.65)

where ∆K(t) represents the gain perturbation, which is assumed to be either of two forms (2.15) or (2.19).

2.7 Guaranteed Cost Control II

49

Definition 2.20. Consider system (ΣJ ) with cost function (2.64) and w(t) ≡ 0. The state-feedback control law (2.65) with controller gain perturbations (2.15) or (2.19) is said to be a guaranteed cost control (GCC) with quadratic cost matrix Pj > 0, j ∈ IN if given matrices 0 < Qj = Qtj , 0 < Rj = Rjt , , 0 < Wj = Wjt , the following matrix inequalities  t Pj A∆j + At∆j Pj + K∆ Rj K∆ + Qj + Wj Pj Adj ˆj < 0 • −W ˆ j = (1 − τ + )Wj , j ∈ IN W (2.66) has a feasible solution with respect to Pj , j ∈ IN, perturbations satisfying (2.15) or (2.19)

for all admissible gain

In the discussions to follow, the problem of interest in this section is as follows: Given the uncertain system (ΣJ ) with cost function (2.64) along with the state-feedback control law (2.65) with controller gain perturbations (2.15) or (2.19), it is required to determine the feedback gain matrix Ko such that the closed-loop controlled system achieve quadratic stability over the vertex set IN. Proceeding in parallel to the previous sections, we will derive expressions for gain matrix Ko , using guaranteed cost and H∞ approaches.

2.7 Guaranteed Cost Control II In this section, we establish results for GCC synthesis with both types of controller perturbations. Subsequently, we derive some important cases. We start by establishing the stability result 2.7.1 Quadratic Stability Considering the GCC approach, the following theorem provides a link between the notion of quadratic cost matrix and an upper bound on the cost function Jj , j ∈ IN. Theorem 2.21. Consider system (ΣJ ) with cost function (2.64). Suppose that the state-feedback control law (2.65) with controller gain perturbations (2.15) or (2.19) is a quadratic GCC with matrix P > 0. Then the resulting closed-loop system with w(t) ≡ 0 x(t) ˙ = A∆j (t)x(t) + Adj (t)x(t − τ )

(2.67)

is quadratically stable over the vertex set IN with the cost function bounded by Jj ≤

N  j=1



λj xto Pj xo +

0

−τ

xt (s) Wj x(s) ds

(2.68)

50

2 Resilient Control-Continuous Case

Proof (⇒) : If Pj > 0, j ∈ IN is a quadratic cost matrix for system (ΣJ ) and cost function (2.64), it follows from (2.65) and the Schur complements that   t ˆ −1 Atdj Pj x < 0 xt Pj A∆j + At∆j Pj + K∆ Rj K∆ + Qj + Wj + Pj Adj W j ∀ x = 0 , j ∈ IN

(2.69)

Hence, system (ΣJ ) is quadratically stable. Next, introduce the Lyapunov-Krasovskii functional N 

V (x) =

 λj xt (t)Pj x(t) +



0

−τ

j=1

xt (s)Wj x(s) ds .

Evaluation of V˙ (x) along the solutions of (2.67) yields: V˙ (x) =

N 

  t λj x (t) A∆j Pj + Pj A∆j + Wj x(t) + 2xt Pj Adj x(t − τ ) t

j=1

 − x (t − τ )(1 − τ˙ )Wj x(t − τ ) ≤ x (t) At∆j Pj + Pj A∆j + Wj  −1 t ˆ + Pj Adj Wj Adj Pj x(t) < t

−

N 

t



 t

λj x (t)

Qj +

t K∆ Rj K∆

x(t)

(2.70)

j=1

From which we conclude that   N  t λj xt (t) Qj + K∆ Rj K∆ x(t) < V˙ (x)

(2.71)

j=1

Integrating over the period t ∈ [0, ∞) and using 2.64, we get Jj ≤ V (xo ) − V (x(∞))

(2.72)

By (2.69), system (Σ∆ ) is quadratically stable thereby leading to V (x(t)) → 0 as t → ∞ and therefore (2.72) reduces to ∆

Jj ≤ V (xo ) =

N  j=1



t λj xo Pj xo +

0

−τ

t

x (s)Wj x(s) ds

, j ∈ IN (2.73)

(⇐) : Let system (ΣJ ) be quadratically stable. It then follows on using the Schur complements operations that there exist 0 < Pj = Pjt such that t ˆ −1 Atdj Pj < 0 Pj A∆j + At∆j Pj + K∆ Rj K∆ + Qj + Wj + Pj Adj W j ∀ j ∈ IN (2.74)

2.7 Guaranteed Cost Control II

51

Hence, one can find some µj > 0, j ∈ IN such that the following inequality t t ˆ −1 t µ−1 j [Pj A∆j + A∆j Pj + K∆ Rj K∆ + Qj + Wj + Pj Adj Wj Adj Pj ] = t ¯ ¯j + W ¯ j + P¯j Adj W ˆ −1 Atdj P¯j < 0 P¯j A∆j + At∆j P¯j + K∆ Rj K∆ + Q j ∀ j ∈ IN (2.75)

This inequality implies that, with the help of Schur complements, there exists ¯ j = µ−1 Wj , Q ¯ j = µ−1 Rj , such that the matrix ¯ j = µ−1 Qj , R matrices W j j j −1 P¯j = µj Pj is a quadratic cost matrix for system (ΣJ ). This completes the proof. ∇∇∇ Next, we establish results for GCC synthesis with both types of controller perturbations. 2.7.2 Additive Gain Perturbations The following theorem provides a necessary and sufficient condition for GCC under additive gain perturbations. Theorem 2.22. Consider system (ΣJ ) with cost function (2.64). There exists a state feedback gain Ko such that the control law (2.65) with additive gain perturbations (2.15) is a GCC with a quadratic cost matrix Pj > 0, j ∈ IN given matrices 0 < Qj = Qtj , 0 < Wj = Wjt , 0 < Rj = Rjt , j ∈ IN if and only if there exist matrices 0 < Xj = Xjt , Lj , j ∈ IN, V, W and some parameters j > 0, µj > 0, j ∈ IN such that the following LMIs   Xj − V+ W t+ t Ξ X Q X W L E B H A aj j j j j j oj a dj a   Aoj W t µj Boj Ha Hat    • −W − W t  0 0 0 0 0 0   t  •  • −R + µ H H 0 0 0 0 0 j j a a    • <0 • • −W 0 0 0 0 j    •  0 0 0 • • • −Q j    •  • • • • − I 0 0 j    • 0  • • • • • −j I ˆj • • • • • • • −W  −j µj ≥0 (2.76) • −µj are feasible with respect to Xj , W, V where t Ξaj = Aoj V + Boj W + V t Atoj + W t Boj

(2.77)

The feedback gain is Ko = W V −1 . Proof: By Definition 2.20 and using Inequality 1 of the Appendix in the manner of [73], it follows that inequality (2.66) holds if and only if the matrix inequality

52

2 Resilient Control-Continuous Case



 Pj (Aoj + Boj Ko )+   t  (Aoj + Boj Ko )t Pj Kot Pj Adj  Pj Boj Ha Pj Boj Ha     + −1  Ha   Ha  +Wj + Qj j    • −Rj 0  0 0 ˆj • • −W  t   t t Ea Ea + j  0   0  = 0 0   Pj (Aoj + Boj Ko )+  (Aoj + Boj Ko )t Pj +  Kot +   −1 t t Pj Adj   Wj + Qj + j Ea Ea P B H H  j oj a a j   t  +−1  < 0 (2.78) j Pj Boj Ha Ha Boj Pj     −R + j  0  • t   −1 j Ha Ha ˆ • • −W j holds for some parameter j > 0. By the congruent transformation diag[Pj−1 , I, I], Pj−1 = Xj with

−1 = Lj −1 j = µj , j Pj

and applying Lemma (8.5) of the Appendix along with Schur complement operations we obtain the equivalent LMI conditions of (2.76). ∇∇∇ 2.7.3 Multiplicative Gain Perturbations The following theorem builds on Theorem 2.22 and provides necessary and sufficient conditions for GCC under multiplicative gain perturbations. Theorem 2.23. Consider system (Σ∆ ) with cost function (2.64). There exists a state feedback gain Ko such that the control law (2.65) with multiplicative gain perturbations (2.19) is a GCC with a quadratic cost matrix Pj > 0, j ∈ IN given matrices 0 < Qj = Qtj , 0 < Wj = Wjt , 0 < Rj = Rjt , j ∈ IN if and only if there exist matrices 0 < Xj = Xjt , Gj , j ∈ IN, V, W and some parameters j > 0, µj > 0, j ∈ IN such that the following LMIs   Xj − V W t+ t Xj Qj Xj Wj Gj Ea Boj Hm Adj  t  Ξaj +Aoj W t µj Boj Hm Hm   t  • −W − W 0 0 0 0 0 0    t  • • −Rj + µj Hm Hm 0 0 0 0 0     • 0 0 0 0  • • −Qj <0    • • • • −W 0 0 0 j    • • • • • −j I 0 0     • • • • • • −j I 0  ˆj • • • • • • • −W

2.7 Guaranteed Cost Control II



−j µj • −µj

53

≥0

(2.79)

are feasible with respect to Xj , W, V. The feedback gain is Ko = W V −1 . Proof: Follows from Theorem 2.22 by substituting Ha → Hm and Ea → Em Ko and introducing the linearization j Pj−1 Ko = Gj . ∇∇∇ 2.7.4 Special Cases We collect hereafter some important cases of special interest: Theorem 2.24. Consider system (ΣJ ) with cost function (2.60). There exist a GCC state feedback gain Ko with a quadratic cost matrix Pj > 0, j ∈ IN given matrices 0 < Qj = Qtj , 0 < Wj = Wjt , 0 < Rj = Rjt , j ∈ IN if and only if there exist matrices 0 < Xj = Xjt , j ∈ IN, V, W such that the following LMIs   Ξaj Xj − V + Aoj W t W t Xj Qj Xj Wj Adj −W − W t 0 0 0 0   •   0 0  • −Rj 0  • <0 (2.80)  • • • −Qj 0 0      • • • • −Wj 0 ˆj • • • • • −W are feasible with respect to Xj , W, V. The feedback gain is Ko = W V −1 . Proof: Follows from Theorem 2.22 by setting Ea ≡ 0, Ha ≡ 0. Consider the continuous-time delay-less system (ΣN ) :

x(t) ˙ = Ao x(t) + Bo u(t)

∇∇∇ (2.81)

where Ao and Bo have the convex-polytopic representation in (2.62). We have the following theorems Theorem 2.25. Consider system (ΣN ) with cost function (2.64). There exists a state feedback gain Ko such that the control law (2.65) with additive gain perturbations (2.15) is a GCC with a quadratic cost matrix Pj > 0, j ∈ IN given matrices 0 < Qj = Qtj , 0 < Rj = Rjt , j ∈ IN if and only if there exist matrices 0 < Xj = Xjt , Lj , j ∈ IN, V, W and some parameters j > 0, µj > 0, j ∈ IN such that the following LMIs   Ξaj Xj − V + Aoj W t W t + µj Boj Ha Hat Xj Qj Lj Eat Boj Ha  • −W − W t 0 0 0 0    t  • • −Rj + µj Ha Ha 0 0 0   <0  • • • −Qj 0 0     • • • • −j I 0  • • • • • −j I  −j µj ≥0 (2.82) • −µj are feasible with respect to Xj , W, V. The feedback gain is Ko = W V −1 .

54

2 Resilient Control-Continuous Case

Theorem 2.26. Consider system (ΣN ) with cost function (2.64). There exists a state feedback gain Ko such that the control law (2.65) with multiplicative gain perturbations (2.19) is a GCC with a quadratic cost matrix Pj > 0, j ∈ IN given matrices 0 < Qj = Qtj , 0 < Rj = Rjt , j ∈ IN if and only if there exist matrices 0 < Xj = Xjt , Gj , j ∈ IN, V, W and some parameters j > 0, µj > 0, j ∈ IN such that the following LMIs   t t Ξaj Xj − V + Aoj W t W t + µj Boj Hm Hm Xj Qj Gj Em Boj Hm  • −W − W t 0 0 0 0    t  • + µ H H 0 0 0  • −R j j m m  <0  • • • −Qj 0 0     • 0  • • • −j I • • • • • −j I  −j µj ≥0 (2.83) • −µj are feasible with respect to Xj , W, V. The feedback gain is Ko = W V −1 . Theorem 2.27. Consider system (ΣN ) with cost function (2.64). There exist a GCC state feedback gain Ko with a quadratic cost matrix Pj > 0, j ∈ IN given matrices 0 < Qj = Qtj , 0 < Rj = Rjt , j ∈ IN if and only if there exist matrices 0 < Xj = Xjt , j ∈ IN, V, W such that the following LMIs 

 Ξaj Xj − V + Aoj W t W t Xj Qj  • −W − W t 0 0   <0  • • −Rj 0  • • • −Qj

(2.84)

are feasible with respect to Xj , W, V. The feedback gain is Ko = W V −1 . Remark 2.28. It is significant to observe that Theorems 2.22 and 2.23 are new contributions to time-delay systems since they provide LMI-based necessary and sufficient conditions for GCC with quadratic cost matrix P . Additionally, Theorems 2.25 through 2.27 improve drastically upon the results of [118, 121] from various technical aspects including convex-bounded parametric uncertainties, stability characterization, LMI formulations, to name a few. Interestingly enough, by considering LMIs (2.79) and (2.80) it is quite evident that our approach has yield similar structure for both additive and multiplicative perturbations. This helps in the computational processing phase. Finally, it should be observed that the nice structures of LMIs (2.76), (2.79), (2.80) and (2.82)-(2.84) have been facilitated through the clever use of linearization (slack) variables [25].

2.8 H∞ Control II

55

2.7.5 Example 2.5 Consider a continuous-time system of the type (2.59) with    00 0.4 0 −1 −1 Ao1 = , Bo1 = , Ad1 = 01 0 0.6 0 −0.9    0.1 0 −0.7 −0.6 1 0.3 , Ad2 = , Bo2 = Ao2 = 0 0.8 0.2 −1.4 0 1.5    −0.2 0 0.2 0 0.1 −1.2 , Bo3 = , Ad3 = Ao3 = 0 −0.5 0 1.3 −0.2 −1.1   0.5 1 , Hm = , Ea = [1 0] , Em = [0 1] , τ + = 0.65 Ha = 0.7 0.4 A summary of the numerical computations is given in Table 2.6, from which it is clear that the cases without gain perturbations yields less cost bound than those with gain perturbations. Figures 2.2-2.3 provide the response of the controlled system states for cases of additive and multiplicative gain perturbations. Table 2.6. Computational Results of Example 2.5 M odel P erturbation Theorem Q (2.1) Additive 2.22 3 0 0 3 (2.1) Multiplicative 2.23 4 0 0 4 (2.1) No 2.24 8 0 0 8 (3.5) Additive 2.25 5 0 0 5 (3.5) Multiplicative 2.26 6 0 0 6 (3.5) No 2.27 10 0 0 10

R Jo W Ko 1.5 1.823 5 0 2.9556 -1.9873 05 1.5 1.795 6 0 2.9724 -1.8857 06 3 0.685 6 0 2.8390 -1.5940 06 1.5 1.561 6 0 3.2017 -1.6102 06 1.5 1.545 6 0 3.1335 -1.5935 06 4 0.584 6 0 3.0245 -1.6857 06

2.8 H∞ Control II In the sequel, we extend the results attained in the forgoing section to the case of H∞ control. For a given a prespecified H∞ -norm level γ, it is required to develop conditions for the state-feedback controller (2.61) with controller gain perturbations (2.15) or (2.19) to render the resulting closed-loop system quadratically stable and ||z||22 ≤ γ ||w||22 for all admissible uncertainties over the vertex IN. Following [81, 92], we have the following definition:

56

2 Resilient Control-Continuous Case

Fig. 2.2. Response of First System State x1 : Example 2.5

Fig. 2.3. Response of Second System State x2 : Example 2.5

Definition 2.29. Consider system (ΣJ ). The state-feedback control law (2.65) with controller gain perturbations (2.15) or (2.19) is said to be a H∞ with disturbance attenuation γ > 0 if given a matrix 0 < Wjt = Wj , j ∈ IN there exists a matrix 0 < Pjt = Pj , j ∈ IN such that the following LMI 

 ˆ tΦ Pj A∆j + At∆j Pj + Wj Pj Adj Pj Γ + D j   ˆj 0 • −W 2 t • • −γ I + Φ Φ

< 0

(2.85)

has a feasible solution for all admissible uncertainties over the vertex IN and ˆ j = Doj + Foj Ko . gain perturbations ∆K(t) (2.15) or (2.19), where D

2.8 H∞ Control II

57

The following theorems summarize the corresponding results for controller gain variations of both types. We provides them in two consecutive subsections beginning with the additive type. 2.8.1 Additive Gain Perturbations Theorem 2.30. Consider system (ΣJ ) with additive gain perturbations (2.15). Given matrices 0 < Wj = Wjt , there exists an H∞ controller of the type (2.65) with disturbance attenuation γ > 0 if and only if there exist matrices 0 < Xj = Xjt , Lj , j ∈ IN, V, W and some parameters j > 0, µj > 0, j ∈ IN such that the following LMIs   t Γ + VDoj Φ Xj − V+ W t+ t Ξ X W L E B H A t  aj Aoj W t µj Boj Ha Hat j j j a oj a dj +W t Foj Φ     • −W − W t  0 0 0 0 0 0     −Rj +  •  • 0 0 0 0 0 t   µj Ha Ha    • <0 • • −Wj 0 0 0 0    •  0 0 0 • • • −j I    •  • • • • −j I 0 0    •  ˆj • • • • • −W 0   2  −γ I+  • • • • • • • Φt Φ  −j µj ≥0 (2.86) • −µj are feasible with respect to Xj , W, V. The feedback gain is Ko = W V −1 . Proof: By Definition 2.29 and [73], it follows that inequality (2.85) holds if and only if the matrix inequality   Pj (Aoj + Boj Ko )+ ˆ tΦ   (Aoj + Boj Ko )t Pj Pj Adj Pj Γ + D j     +Wj  +   ˆ • −Wj 0   2  γ I+  • • Φt Φ   t   t t  t P Bo Ha Ea P Bo Ha Ea        0 0 0 0  +  −1 j j 0 0 0 0

58

2 Resilient Control-Continuous Case

 Pj (Aoj + Boj Ko )+  (Aoj + Boj Ko )t Pj P Γ + Dt Φ   Pj Adj j t t oj  t  E E +W +  +Ko Foj Φ  j j a a   −1 t    P B H H B P =  j j oj a a oj j    ˆ • −Wj 0   2  −γ I+  • • Φt Φ 

< 0

(2.87)

holds some parameters j > 0. Proceeding like Theorem 2.3, we use the = congruent transformation diag[Pj−1 , I, I] with Pj−1 = Xj and −1 j −1 µj , j Pj = Lj and applying Lemma (8.2) followed by Schur complement operations we obtain the equivalent LMI conditions of (2.86). ∇∇∇ 2.8.2 Multiplicative Gain Perturbation we have the following result Theorem 2.31. Consider system (ΣJ ) with multiplicative gain perturbations (2.19). Given matrices 0 < Wj = Wjt , there exists an H∞ controller of the type (2.65) with disturbance attenuation γ > 0 if and only if there exist matrices 0 < Xj = Xjt , Gj , j ∈ IN, V, W and some parameters j > 0, µj > 0, j ∈ IN such that the following LMIs   Γ+ Xj − t W + t t  Ξaj VDoj Φ  V+ t Xj Wj Gj Em Boj Hm Adj   µj Boj Hm Hm t t t  A +W W Foj Φ  oj    • −W − W t  0 0 0 0 0 0     −Rj +  •  • 0 0 0 0 0 t   µj Hm Hm  <0  •  • • −Wj 0 0 0 0    •  • • • − I 0 0 0 j    •  I 0 0 • • • • − j    •  ˆ • • • • • −Wj 0   2  −γ I+  • • • • • • • Φt Φ  −j µj ≥0 (2.88) • −µj are feasible with respect to Xj , W, V. The feedback gain is Ko = W V −1 . 2.8.3 Special Case Starting from either Theorem 2.30 or Theorem 2.31 the following theorem is readily established.

2.8 H∞ Control II

59

Theorem 2.32. Consider system (ΣJ ) without gain perturbations. Given matrices 0 < Wj = Wjt , there exists an H∞ controller with disturbance attenuation γ > 0 if and only if there exist matrices 0 < Xj = Xjt , j ∈ IN, V, W and some parameters j > 0, µj > 0, j ∈ IN such that the following LMIs 

 t Φ Γ + VDoj Xj − V+ t Ξ W X W A aj j j dj t  +W t Foj Φ  Aoj W t    • −W − W t 0  0 0 0    • <0 • −R 0 0 0 j    •  • • −W 0 0 j    •  ˆ • • • −W j 0 2 t • • • • • −γ I + Φ Φ

(2.89)

are feasible with respect to Xj , W, V. The feedback gain is Ko = W V −1 . Proof: Follows from Theorem 2.30 by setting Ea ≡ 0, Ha ≡ 0.

∇∇∇

2.8.4 Results for Delayless Systems For the continuous-time delay-less system (ΣN ), we have the following theorems Theorem 2.33. Consider system (ΣN ) with additive gain perturbations (2.15). There exists an H∞ controller of the type (2.65) with disturbance attenuation γ > 0 if and only if there exist matrices 0 < Xj = Xjt , Lj , j ∈ IN, V, W and some parameters j > 0, µj > 0, j ∈ IN such that the following LMIs   t Γ + VDoj Φ W t+ t t  Ξaj Xj − V + Aoj W µj Boj Ha Hat Lj Ea Boj Ha +W t F t Φ  oj    •  −W − W t 0 0 0 0     −Rj +  •  • 0 0 0 t  <0 µj Ha Ha    •  0 0 • • −j I    •  • • • −j I 0   2  −γ I+  • • • • • Φt Φ  −j µj ≥0 (2.90) • −µj are feasible with respect to Xj , W, V. The feedback gain is Ko = W V −1 . Theorem 2.34. Consider system (ΣN ) with multiplicative gain perturbations (2.19). There exists an H∞ controller of the type (2.65) with disturbance attenuation γ > 0 if and only if there exist matrices 0 < Xj = Xjt , Lj , j ∈ IN, V, W and some parameters j > 0, µj > 0, j ∈ IN such that the following LMIs

60

2 Resilient Control-Continuous Case



 t Γ + VDoj Φ W t+ t t Ξ H X − V + A W G E B oj m aj j oj j t t m  +W t Foj Φ  µj Boj Hm Hm   t  •  −W − W 0 0 0 0     −Rj +  •  • 0 0 0 t  <0 µj Hm Hm    •  • • −j I 0 0    •  0 • • • −j I   2  −γ I+  • • • • • Φt Φ  −j µj ≥0 (2.91) • −µj are feasible with respect to Xj , W, V. The feedback gain is Koj = W V −1 . Theorem 2.35. Consider system (ΣN ) without gain perturbations. There exists an H∞ controller of the type (2.65) with disturbance attenuation γ > 0 if and only if there exist matrices 0 < Xj = Xjt , Lj , j ∈ IN, V, W and some parameters j > 0, µj > 0, j ∈ IN such that the following LMIs   t t Φ Ξaj Xj − V + Aoj W t W t Γ + VDoj Φ + W t Foj  •  −W − W t 0 0  <0 (2.92)  •  0 • −Rj • • • −γ 2 I + Φt Φ are feasible with respect to Xj , W, V. The feedback gain is Ko = W V −1 . 2.8.5 Example 2.6 Consider a polytopic continuous-time system of the type (2.59) with    0 1 0.5 0 0.145 0.75 Ao1 = , Bo1 = , Ad1 = −0.3 −1 0 0.2 0.275 −0.2    0 1 0.1 0 0.115 0.66 Ao2 = , Bo2 = , Ad2 = −0.5 −1.2 0 0.8 0.35 −0.3    0 1 0.2 0 0.1 −1.2 Ao3 = , Bo3 = , Ad3 = −0.1 −0.8 0 1.3 −0.2 −1.1   0.5 1 Ha = , Hm = , Ea = [1 0] , Em = [0 1] , τ + = 0.65 0.7 0.4 A summary of the numerical computations is given in Table 2.7, from which it is clear that the cases without gain perturbations yields less cost bound than those with gain perturbations. Figures 2.4-2.5 provide the response of the controlled system states for cases of additive and multiplicative gain perturbations.

2.9 Dynamic Feedback Control II

61

Table 2.7. Computational Results of Example 2.6 M odel P erturbation Theorem Q (2.1) Additive 2.30 3 0 0 3 (2.1) Multiplicative 2.31 4 0 0 4 (2.1) No 2.32 8 0 0 8 (3.5) Additive 2.33 5 0 0 5 6 0 (3.5) Multiplicative 2.34 0 6 10 0 (3.5) No 2.35 0 10

R Jo W Ko 1.5 1.823 5 0 -9.2556 -7.9873 05 1.5 1.795 6 0 -8.9124 -8.8657 06 3 0.685 6 0 -6.0009 -4.9379 06 1.5 1.561 6 0 -9.5633 -8.0034 06 1.5 1.545 6 0 -9.0124 -9.1746 06 4 0.584 6 0 -4.2845 -5.0507 06

2.9 Dynamic Feedback Control II In the sequel, we examine the problem of designing full-order resilient dynamic output feedback controller such that the resulting closed-loop time-delay system achieves quadratic stability for all admissible uncertainties over the vertex set IN. For this purpose, the controller dynamics have the form: (ΣC ) :

x˙ c (t) = [Ac + ∆Ac (t)]xc (t) + [Bc + ∆Bc (t)]y(t) = A∆c (t)x(t) + B∆c (t)y(t) u(t) = [Kc + ∆Kc (t)]xc (t) = K∆c xc (t)

(2.93)

where xc (t) ∈ IRnc is the controller state vector. One possible representation of the controller perturbations is given by [∆Ac (t) ∆Bc (t)] = Mc ∆c (t)[Nc Nb ] , ||∆c (t)||2 ≤ 1 ∆Kc (t) = Hc ∆c (t)Ec

(2.94)

Now the problem of interest is phrased as follows: Given the uncertain system (ΣJ ) and the dynamic output-feedback controller (ΣC ) with controller perturbations (2.94), it is required to determine the controller matrices Ac , Bc , Kc such that the closedloop controlled system achieve quadratic stability for all admissible uncertainties over the vertex set IN and for all controller perturbations ∆tc (t) ∆c (t) ≤ I. By combining system (ΣJ ) and (ΣC ), the closed-loop time-delay system can be written as (ΣJC ) :

˙ = A∆ξ ξ(t) + A∆dξ ξ(t − τ ) + Γξ w(t) ξ(t) ¯ ∆c ξ(t) u(t) = K z(t) = D∆o ξ(t) + Φ w(t)

(2.95)

62

2 Resilient Control-Continuous Case

Fig. 2.4. Response of First System State x1 : Example 2.6

Fig. 2.5. Response of Second System State x2 : Example 2.6

where





 Γ ∈ IRn+nc , Γξ = , A∆ξ = Aξo + ∆Aξ 0   Ao Bo Kc Ad 0 , A∆dξ = Adξo + ∆Adξ , Adξo = Aξo = Bc Co Ac Bc Cd 0   0 Bo ∆Kc 0 0 ∆Aξ = , ∆Adξ = ∆Bc Co ∆Ac ∆Bc Cd 0 ¯ c + ∆K ¯ c = [0 Kc ] + Hc ∆c E ¯c = [0 Ec ] ¯ ∆c = K ¯c , E K ˆ o + Fˆo ∆c Nc D∆o = [Do Fo Kc ] + [0 Fo Mc ] ∆c Nc = D (2.96) ξ(t) =

x(t) xc (t)

2.9 Dynamic Feedback Control II

63

2.9.1 Guaranteed Cost Design By Definition (2.20), it follows that system (ΣJC ) is quadratically stable for all admissible uncertainties over the vertex set IN if given matrices Qj , Wj , Rj , j ∈ IN the matrix inequality  ¯ t Rj K ¯ ∆c + Qj + Wj Pj A∆dξj Pj A∆ξj + At∆ξj Pj + K ∆c <0 ˆj • −W ˆ j = (1 − τ + )Wj , j ∈ IN W (2.97) is feasible with respect to Pj , Ac , Bc and Kc , where A∆ξj = Aξj + ∆Aξj   Aoj Boj Kc Adj 0 , A∆dξj = Adξj + ∆Adξj , Adξj = Aξj = Bc Cdj 0 Bc Coj Ac   0 Boj ∆Kc 0 0 , ∆Adξj = (2.98) ∆Aξj = ∆Bc Coj ∆Ac ∆Bc Cdj 0 To proceed further, we introduce the following partitioned matrices:    −1 Xsj Ysj Xsj Uj Ysj Vj −1 , Pj = , Sj = , j ∈ IN(2.99) Pj = • Xcj • Ycj 0 Vjt t t t t , 0 < Xcj = Xcj , 0 < Ysj = Ysj , 0 < Ycj = Ycj . By where 0 < Xsj = Xsj −1 considering the fact Pj Pj = I, it can be easily shown that Vj > 0, which would then guarantee that Sj > 0. Now, it follows from [79] and Fact 1 that matrix inequality (2.97) with (2.98) holds if and only if the following inequality  ¯ ct Rj Pj Adξj  Pj Aξj + Atξj Pj + Qj + Wj K  • −Rj 0 + ˆj • • −W   t   t t   t ¯oj ¯ ¯oj ¯ B E B E c c −1        Rj Hc Rj Hc + εj 0 0  + εj 0 0 0 0   t   t t   t ¯ ¯ ¯ cj ¯ cj N N M M bj bj        0 0 −1 +  0 0  + j j 0 0 0 0   t   t t   t ¯ cj ¯ ¯ cj ¯ M N M N c c −1        0 0 0  + σj + σj 0 0 0 0 0     t   t ¯ cj ¯ cj M 0 0 M <0 (2.100) υj−1  0   0  + υj  0   0  ¯t ¯t N N 0 0 d

d

64

2 Resilient Control-Continuous Case

is feasible with respect to Pj for some j , εj , σj , υj , j ∈ IN where   Xsj Boj Hc Uj Mc ¯ ¯ , Mcj = Boj = Ujt Boj Hc Xcj Mc ¯c = [0 Nc ], N ¯d = [Nb Cdj 0] ¯bj = [Nb Coj 0], N N ¯ ¯ ¯ Υj = [Mcj , Mcj , Mcj ] , Λj = diag[j I, σj I, υj I]

(2.101)

By standard Schur complements operations, inequality (2.100) is equivalent to   Pj Aξj + Atξj Pj ¯ ct Rj + K ¯ct E ¯c  +Qj + Wj + εj E Υj Pj Adξj  −1 ¯ t   ¯t N ¯bj + σj N ¯ tN ¯c εj Boj Hc Rj   +j N c bj    < 0 (2.102)  −Rj +   • 0 0 −1 t   εj Rj Hc Hc Rj    0  • • −Λj ˆj • • • −W To cast matrix inequality (2.102) into an LMI format, we perform a congruent transformation  Tj = diag Sj , I, I, I −1 using the fact that Pj−1 Pj = I with Xsj = Zsj and followed by a linearization procedure defined by

ˆc = Zsj Uj Bc , K ˆ c = Kc V t Aˆc = Uj Ac Vjt , B j Ωsj = Zsj Uj Bc Coj Zsj , Φsj =

j Xj−1

(2.103)

, Φcj = j Ysj

t Nbt Nb Coj Ysj j Zsj Coj

, Ψzj = Zsj [Qsj + Wj ]Ysj Ψcj = Lsj = εj Vj , Lcj = σj Vj , Ψsj = Zsj Uj Bc Coj Ysj Ψwj = Ysj [Qsj + Wj ]Ysj , Ωcj = Zsj Uj

(2.104)

The result of the two consecutive transformations is finally expressed by the following LMI:   Πj Πej Ξaj Ξbj Ξcj Ξdj  • −Πdj 0  0 0 0    •  • −Π 0 0 0 hj  <0  •  • • −Λ 0 0 j    •  0 • • • −j t ¯ ˆ ¯ • • • • • − W + υj Nd Nd  −µj εj , j ∈ IN (2.105) • −εj where

2.9 Dynamic Feedback Control II

 Πj =

Πaj Πbj • Πcj



 , Ξbj =

Ωcj Mc Ωcj Mc Ωcj Mc 0 0 0

65



t + Ψsj Πaj = Aoj Zsj + Zsj Atoj + Ωsj + Ωsj ˆ c + Zsj Aˆc + Ψsj + Ψcj + Ψzj Πbj = Aoj Ysj + Zsj Atoj + Boj K ˆ t B t + Ψwj ˆc + K Πcj = Aoj Ysj + Ysj At + Boj K oj

c

oj

t t Πej = [Φsj Coj Nbt , Φcj Coj Nbt , Lsj Ect , Lcj Ect ] , Πhj = Rj − µj Rj Hc Hct Rj Πdj = diag[j I, εj I, σj I]   ˆc 0 Boj Hc Adj + B Ξaj = , Ξdj = Boj Hc 0 Adj  t µj Boj Hc Hc Boj Rj Ξcj = (2.106) ˆ c Rj µj Boj Hc Hct Boj Rj + K

We observe that all the linearization are reversible transformations ensuring that the ensuing resulting conditions are equivalent. In summary, the foregoing analysis has led to the following intermediate result. Lemma 2.36. The nonlinear matrix inequalities (2.97) with (2.98) are feasible with respect to Pj , Ac , Bc , Kc if and only if the LMIs (2.105) with (2.104) and (2.106) are feasible with ˆc , K ˆc respect to the matrices Zsj , Ysj , Ωsj , Ωcj , Ψsj ,Ψcj , Ψzj , Lsj , Lcj , Aˆc , B and the parameters µj , j , εj , σj . The triplet {Ac , Bc , Kc } defining the resilient controller (2.93) is readily derived by inverting (2.103) to yield ˆ c V −t ˆ c , Kc = K Ac = Uj−1 Aˆc Vj−t , Bc = [Zsj Uj ]−1 B j Remark 2.37. Unlike other direct dynamic feedback control schemes [79, 82, 84, 91, 94, 90] using partitioning and Riccati inequalities, the matrix inequality (2.104) offers new potential since it does not involve product of the Lyapunov matrix Pj and system matrices Aξj and Adξj . Essentially, we have adopted a parametrization procedure to cast the nonlinear matrix inequality (2.97) into an alternative LMI characterization which amenable to relaible software computations. 2.9.2 H∞ Design In view of Definition (2.29), it follows for system (ΣJC ) that the control ¯ ∆c ξ(t) is an H∞ controller with disturbance attenuation γ > 0 law u(t) = K if given matrices Wj , j ∈ IN the matrix inequality   Pj A∆ξj +  At∆ξj Pj + Pj A∆dξj Pj Γ + D∆j Φ      < 0, j ∈ IN Wj (2.107)     ˆj 0 • −W • • −γ 2 I + Φt Φ

66

2 Resilient Control-Continuous Case

has a feasible solution Pj , Ac , Bc , Kc for all admissible uncertainties over the vertex IN and controller gain perturbations satisfying (2.94) using the data of (2.98). Proceeding like the resilient GCC case, we exploit the matrices of (2.99) and the results of [79] with Inequality 1.1 of the Appendix. It then follows that matrix inequality (2.106) with (2.98) holds if and only if the matrix inequality   ˆ tΦ Pj Aξj + Atξj Pj + Wj Pj Adj Pj Γξ + D j +  ˆj 0 • −W 2 t • • −γ I + Φ Φ  t   t t  t   ¯ ¯ ¯ ¯ E E Boj Boj c c −1  0   0  + εj  0   0  + εj 0 0 0 0  t   t t  t   ¯ ¯ ¯ cj ¯ cj N N M M bj bj −1        0 0 j + j 0 0  + 0 0 0 0   t   t t   t ¯ ¯ ¯ ¯ Mcj Nc Mcj Nc σj−1  0   0  + σj  0   0  + 0 0 0 0     t   t ¯ cj ¯ cj 0 M 0 M ¯t  N ¯t  + υj−1  0   0  + υj  N d d 0 0 0 0  t   t t    t 0 0 Fˆoj Fˆoj  0   0  + ϑj  0   0  < 0 (2.108) ϑ−1 j Φt Nct Φt Nct 0 0 has a feasible solution. By standard Schur operations, inequality (2.108) is equivalent to   Pj Aξj +   Atξj Pj  Pj Γξ +  ¯ct E ¯c B ¯oj Υj Pj Adξj Fˆ t  +Wj + εj E  oj ˆ tΦ   D j ¯t N ¯bj +  +j N  bj   t ¯ ¯   σj Nc Nc     • −εj I 0 0 0 0     < 0 (2.109) 0 0 0 • • −Λj     ˆ j+ −W   • • • 0 0 t   ¯ N ¯ υj N   d d   • • • • −ϑj I 0   2  −γ I+      Φt Φ+ • • • • • t t ϑj Φ Nc Nc Φ

2.9 Dynamic Feedback Control II

67

In a similar way, we convert matrix inequality (2.109) into an LMI format  by performing the congruent transformation Tj = diag Sj , I, I, I using −1 the fact that Pj−1 Pj = I with Xsj = Zsj and followed by a linearization procedure defined by (2.103)-(2.104). This leads us to:  ¯  t Πj Πej Ξaj Ξbj Ξdj F¯oj Ξtj  • −Πdj 0  0 0 0 0    •  0 0 0 0 • −εj I     ˆ −W +  •  • • 0 0 0 t ¯   < 0, j ∈ IN(2.110) ¯ υj Nd Nd    •  • • • −Λj 0 0    •  • • • • −ϑj 0   2 t  −γ I + Φ Φ  • • • • • • +ϑj Φt Nct Nc Φ where

 ¯ ¯ j = Πaj Πbj , Ψ¯zj = Zsj Wj Ysj Π • Πcj ¯ ˆ c + Zsj Aˆc + Ψsj + Ψcj + Ψ¯zj Πbj = Aoj Ysj + Zsj Atoj + Boj K   t Φ Γ + Zsj Doj 0 t ¯ , Ξtj = (2.111) Foj = t t ˆ tF t Φ Vj Mct Foj Γ + Ysj Doj +K c oj

and the remaining matrices are the same as in (2.106). Finally we provide the following result. Lemma 2.38. The nonlinear matrix inequalities (2.107) are feasible with respect to Pj , Ac , Bc , Kc if and only if the LMIs (2.110) with (2.106) and (2.111) are feasible with ˆc , K ˆc respect to the matrices Zsj , Ysj , Ωsj , Ωcj , Ψsj , Ψcj , Ψzj , Lsj , Lcj , Aˆc , B and the parameters j , εj , σj . Additionally, the triplet {Ac , Bc , Kc } defining the resilient controller (2.53) is readily derived by inverting (2.103). 2.9.3 Special Cases Some relevant special cases are provided Lemma 2.39. Consider system (ΣJ ). There exists a GCC dynamic controller of the type (2.53) given matrices 0 < Qj = Qtj , 0 < Wj = Wjt , 0 < Rj = Rtj , j ∈ IN if and only if the following LMIs 

 ˆ bj Ξdj Πaj Π  • −Πcj 0  < 0 , j ∈ IN ˆ • • −W

(2.112)

68

2 Resilient Control-Continuous Case

are feasible with respect to the matrices ˆc ˆc , K Zsj , Ysj , Ωsj , Ωcj , Ψsj , Ψzj , Aˆc , B ˆ bj = Aoj Ysj + Zsj At + Boj K ˆ c + Zsj Aˆc + Ψsj + Ψzj where Π oj Proof: Follows from Lemma 2.36 by setting Mc ≡ 0, Nb ≡ 0, Nc ≡ 0, Ec ≡ 0, Hc ≡ 0 ∇∇∇ Lemma 2.40. Consider system (ΣJ ). There exists an H∞ GCC dynamic controller of the type (2.93) given matrices 0 < Wj = Wjt , 0 < Rj = Rtj , j ∈ IN if and only if the following LMIs   ˆ bj Ξdj Πaj Π Ξtj  • −Πcj 0  0   < 0 , j ∈ IN (2.113)  •  ˆ • −W 0 • • • −γ 2 I + Φt Φ are feasible with respect to the matrices ˆc , K ˆc Zsj , Ysj , Ωsj , Ωcj , Ψsj , Ψzj , Aˆc , B Proof: Follows from Lemma 2.38 by setting Mc ≡ 0, Nb ≡ 0, Nc ≡ 0, Ec ≡ 0, Hc ≡ 0 ∇∇∇ Lemma 2.41. Consider system (ΣN ). There exists a GCC dynamic controller of the type (2.93) given matrices 0 < Qj = Qtj , 0 < Wj = Wjt , 0 < Rj = Rtj , j ∈ IN if and only if the following LMIs the following LMIs 

Πj Πej Ξaj Ξbj  • −Πdj 0 0   • • −Π 0 hj   • • • −Λj • • • •

 Ξcj  0   −µj εj  0 <0 , , j ∈ IN • −εj 0  −j

are feasible with respect to the matrices ˆc , K ˆc Zsj , Ysj , Ωsj , Ωcj , Ψsj , Ψzj , Aˆc , B and the parameters j , ε j , σ j

(2.114)

2.9 Dynamic Feedback Control II

69

Lemma 2.42. Consider system (ΣN ). There exists an H∞ dynamic controller of the type (2.53) given matrices 0 < Wj = Wjt , j ∈ IN if and only if the following LMIs   ¯ j Πej Ξaj Ξbj F¯ t Π Ξtj oj  • −Πdj 0  0 0 0    •  0 0 • −εj I 0    •  < 0 , j ∈ IN (2.115) 0 • • −Λj 0    •  0 • • • −ϑ j    −γ 2 I + Φt Φ  • • • • • +ϑj Φt Nct Nc Φ are feasible with respect to the matrices ˆc , K ˆc Zsj , Ysj , Ωsj , Ωcj , Ψsj , Ψzj , Aˆc , B and the parameters j , ε j , σ j Lemma 2.43. Consider system (ΣN ). There exists a GCC dynamic controller of the type (2.93) given matrices 0 < Qj = Qtj , 0 < Wj = Wjt , 0 < Rj = Rtj , j ∈ IN if and only if the following LMIs  ˆ bj Πaj Π < 0 , j ∈ IN (2.116) • −Πcj are feasible with respect to the matrices ˆc , K ˆc Zsj , Ysj , Ωsj , Ωcj , Ψsj , Ψzj , Aˆc , B Lemma 2.44. Consider system (ΣN ). There exists an H∞ GCC dynamic controller of the type (2.93) given matrices 0 < Wj = Wjt , 0 < Rj = Rtj , j ∈ IN if and only if the following LMIs   ˆ bj Πaj Π Ξtj  • −Πcj  < 0 , j ∈ IN (2.117) 0 • • −γ 2 I + Φt Φ are feasible with respect to the matrices ˆc , K ˆc Zsj , Ysj , Ωsj , Ωcj , Ψsj , Ψzj , Aˆc , B 2.9.4 Example 2.7 In this section, we present the results of system design and numerical simulation of resilient feedback control applied to a liquid monopropellant rocket motor with a pressure feeding system. A linearized version of the feeding system and combustion chamber equations under the assumption of non steady

70

2 Resilient Control-Continuous Case

flow while taking non-uniform lag into consideration is given by [11, 20]. By considering a lumped approximation of the distributed-lag, the resulting model can be cast into the form (2.59)-(2.61) where the state vector is composed of the relative deviations of the instantaneous combustion champers pressure, the instantaneous mass flow upstream of the capacitance, the instantaneous mass rate of the injected propellant from their steady state value and the deviation of the instantaneous pressure in a special place in the feeding line from its steady-state level relative to the injector pressure drop. The model matrices are given by     0 (γ − 1) 0 0 0    0 0 0 −1/hα   , Bo =  1/hα  Ao =   0   −d/2(1 − α)h 0 −1/(1 − α)h 1/(1 − α)h  0 0 1/ω −1/ω 0     −γβ 0 β 0 0  0 0 0 0 0    Γ =  1  , Ad =  0 0 0 0  0 000 0 Do = [1 0 0 0] , Fo = 1 , Φ = 0.8 Co = [1 0 0 0] , Cd = [0.5 0 0 0] where γ is the pressure exponent of the combustion process, h is the line inertia, α is the fractional length for pressure supply, d is the ratio of steadystate pressure to the steady-state injector pressure drop, β is a corrective factor and ω is the elasticity parameter of the line. The numerical values are h = 2, α = 0.15, d = 0.85, β ∈ [1.5, 3], ω = 1. The uncertainty in the model comes from the parameter γ and is represented by γ ∈ {1−δm , 0, 1+δm }, δm = 0.2 for which the system is considered by to be described by a polytope of three vertices. With δm = 0.2, Qj = I, Wj = 3.5 × I, j = 1, 2, 3, the resilient dynamic controller design is performed for two orders of the controller and the results are summarized in Table 8 for (nc = 2) and in Table 2.9 for nc = 3.

Table 2.8. Results of Dynamic Controllers nc = 2 nc Controller Ac GCC -2.7411 0.0027 2 1.5247 -8.4108 2 H∞ , γ = 15 -3.1681 0.01034 2.2983 -11.8101

Bc Ko -0.3661 -4.2301 0.0858 1.2097 -1.1224 -6.0871 0.1756 0.9006

2.10 Uncertain Model with State-Delays

71

Table 2.9. Results of Dynamic Controllers nc = 3 nc 3

Controller GCC

-3.0819 1.2115 -0.2108 3 H∞ , γ = 12.5 -7.1028 2.1134 1.2664

Ac -0.2665 -0.5128 0.0516 -0.3514 -0.4577 -0.9543

1.0306 0.4557 1.2116 2.1689 1.2182 5.0105

Bc Ko -0.6352 -6.0811 -1.0361 1.1347 1.1668 0.8654 -0.5772 -9.1031 1.1036 -1.2054 3.6861 1.4114

2.10 Uncertain Model with State-Delays Consider the class of dynamical systems with state delays (Σ∆ ) : x(t) ˙ = [Ao + ∆Ao (t)]x(t) + [Ad + ∆Ad (t)]x(t − τ ) + Bo u(t) = A∆o (t)x(t) + A∆d (t)x(t − τ ) + Bo u(t) (2.118) y(t) = Co x(t) + Go u(t)

(2.119)

where x(t) ∈ n is the state vector; y(t) ∈ p is the measured output; u(t) ∈ m is the control input and τ, η are unknown time-varying state delay factors satisfying 0 ≤ τ ≤ τ + , 0 ≤ τ˙ ≤ τ ∗ < 1, where the bounds τ + , τ ∗ are known constants. The matrices Ao , Ad , Co , Go , Bo are known real constant matrices of appropriate dimensions which describe the nominal system of (Σ∆ ). In the sequel, we assume that Co has a full row-rank. The matrices ∆Ao (t) and ∆Ad (t) are real, time-varying matrix functions representing the normbounded parameter uncertainties and represented by: [∆Ao (t) ∆Ad (t)] = Ma ∆p [Na Nd ] ∆

∆p ∈ ∆p (t) = {∆p : ∆tp ∆p ≤ I}

(2.120)

where Ma ∈ n×α , Na ∈ β×n and Nd ∈ β×n are known real constant matrices, with ∆p being unknown, time-varying matrix function where its elements are Lebesgue measurable. The initial condition is specified as x(0), x(s) = xo , φ(s) , where φ(.) ∈ L2 [−τ, 0]. 2.10.1 Resilient Feedback Control A convenient way to control system (Σ∆ ) is by employing a dynamic feedback controller of the form (Σco ) :

x˙ c (t) = Ao xc (t) + Bo u(t) + Ko [y(t) − yc (t)] yc = Co xc (t) + Go u(t) u(t) = Kc xc (t)

(2.121)

72

2 Resilient Control-Continuous Case

where the matrices Ko and Kc are the unknown gain parameters to be determined. In the literature, there have been several methods to cope with this problem [78, 79, 87, 90]. In practice it turns out however that these gains cannot be implemented exactly, leading to the fragility problem and related issues. To address this problem, we consider controller (2.121) to be expressed as (Σc ) :

x˙ c (t) = Ao xc (t) + Bo u(t) + [Ko + ∆Ko (t)][y(t) − yc (t)] yc = Co xc (t) + Go u(t) u(t) = [Kc + ∆Kc (t)]xc (t)

(2.122)

where ∆Ko (t), ∆Kc (t) are additive gain perturbations represented by ∆Ko (t) = Mf ∆f Ng , ∆Kc (t) = Ms ∆d Nh ∆

∆f ∈ ∆f (t) = {∆f : ∆tf ∆f ≤ I} ∆

∆d ∈ ∆d (t) = {∆d : ∆td ∆d ≤ I}

(2.123)

Alternatively for multiplicative gain perturbations, it takes the form (Σc ) :

x˙ c (t) = Ao xc (t) + Bo u(t) + [I + ∆Km (t)]Ko [y(t) − yc (t)] yc = Co xc (t) + Go u(t) u(t) = [I + ∆Kd (t)]Kc xc (t)

(2.124)

where ∆Km (t), ∆Kd (t) are multiplicative gain perturbations represented by ∆Km (t) = Md ∆f Nm , ∆Kd (t) = Mp ∆d Nf

(2.125)

In the sequel, the problem of interest in this part is as follows: Given the uncertain system (Σ∆ ), dynamic-feedback controller (2.122) with additive gain perturbations (2.123), or alternatively dynamic-feedback controller (2.124) with multiplicative gain perturbations (2.125), it is required to determine the feedback gain matrices Ko , Kc such that the closed-loop controlled system achieves quadratic stability for all admissible uncertainties satisfying ∆p ∈ ∆p (t) , ∀t and gain perturbations satisfying ∆f ∈ ∆f (t), ∆d ∈ ∆d (t) , ∀t . Remark 2.45. The norm-bounded uncertainty structure considered here has been widely adopted in robust control. We note that it covers the usual matching condition as a special case by setting N ≡ I. Observe that the unit upperbound condition on ∆(t) does not cause any loss of generality. Indeed ∆p (t) can always be normalized by appropriately selecting the matrices Ma and Na . In the absence of controller perturbations ∆Ko (t) ≡ 0, ∆Kc (t) ≡ 0, then our problem reduces to the well known guaranteed cost control for uncertain continuous-time systems [81].

2.10 Uncertain Model with State-Delays

73

2.10.2 Error Dynamics We proceed by defining the state error e(t) = x(t) − xc (t). With focus on the additive gain perturbations, we obtain from (2.118)-(2.119) and (2.122), the error system ˙ = [Ao − Ko Co − ∆Ko (t)Co ]e(t) + ∆Ao (t)x(t) (Σea ) : e(t) + [Ad + ∆Ad (t)]x(t − τ )

(2.126)

By combining systems (Σ∆ ) and (Σea ), we get the augmented system (ΣA ) : where

 ξ(t) = 

x(t) e(t)



˙ = A∆ξ ξ(t) + A∆dξ ξ(t − τ ) ξ(t)

∈ 2n×2n , Aξo =



(2.127)

Ao + Bo Kc −Bo Kc 0 Ao − Kc Co

Ad 0 Ad 0 A∆ξ = Aξo + ∆Aξ , A∆dξ = Aξdo + ∆Aξd   ∆Ao + Bo ∆Kc −Bo ∆Kc ∆Ad 0 , ∆Aξd = ∆Aξ = ∆Ao −∆Ko Co ∆Ad 0



Aξdo =

(2.128)

The deisgn objective is to develop a tractable synthesis procedure for selecting the gain matrices Go and Ko such that for all admissible uncertainties ∆Ao (t), ∆Ad (t) and gain perurbations ∆Ko (t), ∆Kc (t), system (ΣA ) is robustly asymptotically stable. In case of additive gain perturbations, we use (2.120) and (2.123) to obtain   Ma Bo Ms ∆p (t)[Na 0] + ∆d (t)[Nh − Nh ] ∆Aξ = Ma 0  0 + ∆f (t)[0 − Ng Co ] Mf ¯a + M ¯ b ∆d (t)N ¯h + M ¯ f ∆f (t)N ¯g ¯ a ∆p (t)N = M  Ma ¯ a ∆p (t)N ¯d ∆p (t)[Nd 0] = M ∆Aξd = (2.129) Ma Alternatively, we can use (2.120) and (2.125) to get   Ma Bo Mp ∆p (t)[Na 0] + ∆d (t)[Nf Kc ∆Aξ = Ma 0  0 + ∆f (t)[0 − Nm Ko Co ] Md ¯a + M ¯ d ∆d (t)N ¯f + M ¯ f ∆f (t)N ¯m ¯ a ∆p (t)N = M ¯ ¯ ∆Aξd = Ma ∆f (t)Nd

− N f Kc ]

(2.130)

74

2 Resilient Control-Continuous Case

2.10.3 Robust Stability Results To examine the stability of system (ΣA ), we use the Lyapunov-Krasovskii functional

t t V (ξ) = ξ (t)Pξ(t) + ξ t (s)Qξ(s) ds (2.131) t−τ

where



Ps 0 P= 0 Pc





Qs 0 , Q= 0 Qc

, Ps = Pst > 0

Pc = Pct > 0 , Qs = Qts > 0 , Qc = Qtc > 0

(2.132)

The following theorem establishes the first main stability result. Theorem 2.46. Consider system (ΣA ) with additive gain perturbations and given matrices 0 < Qs = Qts ∈ n×n , 0 < Ws = Wst ∈ IRn×n , 0 < Qc = Qtc ∈ n×n , 0 < Wc = Wct ∈ IRn×n . There exist matrices 0 < Xs = Xst ∈ n×n , 0 < Xc = Xct ∈ IRn×n , Ys , Yc , Za , Yd , Yh , Zb , Zc and scalars ε1 > 0, ..., ε4 > 0, such that the LMIs   Ao Xs + Xs Ato Π Π X Q A a s s s d   +Bo Ys + Yst Bot     • −Λs 0 0 0  < 0 (2.133)    0 • • −Λs 0     • • • −Qs 0 t ˆ • • • • −Qs + ε4 Nd Nd   t Ao Xc + Xc Ao t Πc Zc Xs Qs Ad  +Yc + Yct      • −Λ 0 0 0 c   < 0 (2.134)   • • −ε I 0 0 3     • • • −Qc 0 ˆ s + ε4 N t N d • • • • −Q d   Bo Yd + Ydt Bot + 2Ma 2Ma 2Ad   Yh + Yht     ≥ 0 • 2ε I 0 0 (2.135) 1     • • 2ε4 I 0 ˆ s − ε4 N t N d • • • 2Q d have a feasible solution. Moreover, the gain matrices are given by Kc = Ys Xs−1 , Ko = −Yc Xc−1 Co†

1

Co† = Cot [Co Cot ]−1 is the pseudo-inverse of Co .

1

.

2.10 Uncertain Model with State-Delays

75

Proof: Differentiating (2.131) with respect to t and using (2.127) and arranging terms, we obtain   t  PA∆ξ + At∆ξ P + Q PA∆dξ ξ(t) ξ(t) (2.136) V˙ (ξ) = ξ(t − τ ) ξ(t − τ ) • −(1 − τ˙ )Q By applying Inequalities I-II of the Appendix and using (2.128), it follows that: −1 ¯ ¯t ¯ ¯t ¯t ¯ P∆Aξ + ∆Atξ P ≤ ε−1 1 P M a Ma P + ε 1 N a N a + ε 2 P M b Mb P ¯f M ¯ht N ¯h + ε−1 P M ¯ ft P + ε3 N ¯gt N ¯g +ε2 N 3 −1  t ¯ −1 −1 t ˆ ¯ Atξdo P PA∆dξ (1 − τ˙ ) Q A∆dξ P ≤ PAξdo Q − ε4 Nd Nd

¯ ¯t +ε−1 4 P M a Ma P

(2.137)

ˆ = (1 − τ ∗ )Q . By invoking the Schur complements to (2.136) and where Q considering (2.128) and (2.137), we obtain  t ˙ ¯ ¯t ¯t ¯ V (ξ) = ξ (t) PAξo + Atξo P + Q + ε−1 1 P M a Ma P + ε 1 N a N a −1 ¯ ¯t ¯ ¯t ¯t ¯ ¯t ¯ + ε−1 2 P M b Mb P + ε 2 N h N h + ε 3 P M f Mf P + ε 3 N g N g  −1 ˆ − ε4 N ¯ aM ¯dt N ¯d ¯ at P ξ(t) + PAξdo Q Atξdo P + ε−1 P M 4



x(t) = ξ (t) Ξ ξ(t) = e(t) ∆

t

t 

Ξs Ξd • Ξc



x(t) e(t)

(2.138)

From Lyapunov theory, it is known that V˙ (ξ) < 0 is a sufficient condition for stability of system (ΣA ), which corresponds from (2.138) to the inequality Ξ < 0. Expanding the composite matrix Ξ and incorporating the matrices (2.132) with lengthy but standard algebraic manipulations, we get: t Ξs = Ps Ao + Ato Ps + Ps Bo Kc + Kct Bot Ps + Qs + ε−1 1 Ps Ma Ma Ps

−1 t t t t + ε1 Nat Na + ε−1 2 Ps Bo Ms Ms Bo Ps + ε2 Nh Nh + ε4 Ps Ma Ma Ps  ˆ s − ε4 N t Nd −1 At Ps + Ps Ad Q (2.139)

Ξc = + + Ξd = +

d d t t Pc Ao + Ao Pc − Pc Ko Co − Cot Kot Pc + Qc + ε−1 1 Pc Ma Ma Pc −1 −1 t t t t ε2 Nh Nh + ε3 Pc Mf Mf Pc + ε3 Co Ng Ng Co + ε4 Pc Ma Mat Pc  ˆ s − ε4 N t Nd −1 At Pc Pc Ad Q (2.140) d d −1 t t −Ps Bo Kc + ε1 Ps Ma Ma Pc − ε2 Nh Nh  ˆ s − ε4 N t Nd −1 At Pc + ε−1 Ps Ma M t Pc Ps Ad Q (2.141) d d a 4

Introducing the linearization matrices for some ε1 > 0, ε2 > 0, ε3 > 0

76

2 Resilient Control-Continuous Case

Xs = Ps−1 , Xc = Pc−1 , Ys = Kc Ps−1 , Yc = −Ko Co Pc−1 Yd = Kc Xc , Yh = ε2 Ps−1 Nht Nh Pc−1 , Zb = ε2 Nh Pc−1 Zc = ε3 Ng Co Pc−1 , Za = ε1 Na Ps−1

(2.142)

and define the block matrices for some scalars ε1 > 0, ..., ε4 > 0: Λs = [ε1 I, ε2 I] , Πa = [Ma , Bo Ms ] , Πs = [Zat , Zbt ] Λc = [ε1 I, ε3 I, ε4 I] , Πc = [Ma , Mf , Ma ] (2.143) Now, we premultiply (2.139) by Ps−1 and postmultiply the result by Ps−1 and use (2.142) to yield t Ξs = Ao Xs + Xs Ato + Bo Ys + Yst Bot + Xs Qs Xs + ε−1 1 M a Ma −1 −1 t t t t + ε−1 1 Za Za + ε2 Bo Ms Ms Bo + ε2 Zb Zb  ˆ s − ε4 N t Nd −1 At + Ad Q (2.144) d d

Then, we premultiply (2.140) by Pc−1 and postmultiply the result by Pc−1 and use (2.142) again to yield t Ξc = Ao Xc + Xc Ato − Yc − Yct + Xc Qc Xc + ε−1 1 Ma Ma −1 t −1 t t + ε−1 3 Mf Mf + ε3 Zc Zc + ε4 Ma Ma  ˆ s − ε4 N t Nd −1 At c + Ad Q d d

(2.145)

Next, we premultiply (2.141) by Ps−1 and postmultiply the result by Pc−1 and use (2.142) once more to yield t Ξd = −[Bo Kc + ε2 Ps−1 Nht Nh ]Xc + ε−1 1 Ma Ma  ˆ s − ε4 N t Nd −1 At ε−1 Ma M t + Ad Q d d 4 a

(2.146)

Since it is sufficient for the stability requirement Ξ < 0 for system (2.127) that Ξs < 0, Ξd = 0, Ξc < 0, we enforce the latter conditions on (2.144)(2.146) and use the Schur complements with some matrix manipulations, we arrive at the LMIs (2.133)-(2.135). ∇∇∇ Had we considered the multiplicative gain perturbations, we would have obtained from (2.118)-(2.119) and (2.124), the error system (Σem ) : e(t) ˙ = [Ao − Ko Co − ∆Km (t)Ko Co ]e(t) + ∆Ao (t)x(t) + [Ad + ∆Ad (t)]x(t − τ ) (2.147) By combining systems (Σ∆ ) and (Σem ), we get the augmented system (ΣA ) along with (2.125) and (2.130). The following theorem establishes the second main stability result. Theorem 2.47. Consider system (ΣA ) with multiplicative gain perturbations and given matrices 0 < Qs = Qts ∈ n×n , 0 < Ws = Wst ∈ IRn×n , 0 <

2.10 Uncertain Model with State-Delays

77

Qc = Qtc ∈ n×n , 0 < Wc = Wct ∈ IRn×n . There exist matrices 0 < Xs = Xst ∈ n×n , 0 < Xc = Xct ∈ IRn×n , Ys , Yc , Za , Yd , Yh , Zb , Zc and scalars ε1 > 0, ..., ε4 > 0, such that the LMIs   Ao Xs + Xs Ato Π Π X Q A a s s s d   +Bo Ys + Yst Bot     • −Λs 0 0 0    < 0  0 • • −Λs 0 (2.148)     0 • • • −Q s    ˆs+  −Q • • • • ε4 Ndt Nd   Ao Xc + Xc Ato ¯ t ¯ Ad  Πc Zc Xs Qs  +Yc + Yct     • −Λc 0 0 0     < 0 • • −ε I 0 0 (2.149) 3     0 • • • −Q c    ˆs+  −Q • • • • ε4 Ndt Nd   Bo Yd + Ydt Bot  +Yh + Yht 2Ma 2Ma 2Ad      • 2ε1 I 0 0   ≥ 0 (2.150)   I 0 • • 2ε 4    ˆs−  2Q • • • ε4 Ndt Nd have a feasible solution. Moreover, the gain matrices are given by Kc = Ys Xs−1 , Ko = −Yc Xc−1 Co† . ¯c = Proof: Follows by parallel development to Theorem 2.46 using Π [Ma , Md , Ma ] instead of Πc and the linearizations Z¯c = ε3 Nm Ko Co Pc−1 instead of Zc . ∇∇∇ 2.10.4 Special Cases Consider the nominal augmented system (ΣAO ) :

˙ = Aξo ξ(t) + Aξdo ξ(t − τ ) ξ(t)

(2.151)

where Aξo , Aξdo are given by (2.128). The following results hold for additive or multiplicative gain perturbations Corollary 2.48. Consider system (ΣAO ) with additive gain perturbations and given matrices 0 < Qs = Qts ∈ n×n , 0 < Ws = Wst ∈ IRn×n , 0 < Qc = Qtc ∈ n×n , 0 < Wc = Wct ∈ IRn×n . There exist matrices 0 < Xs = Xst ∈ n×n , 0 < Xc = Xct ∈ IRn×n , Ys , Yc , Zs , Yd , Yh , Zb , Zc and scalars ε2 > 0, ε3 > 0, such that the LMIs

78

2 Resilient Control-Continuous Case



 Ao Xs + Xs Ato + Bo Ys + Yst Bot Bo Ms Zb Xs Qs Ad  • −ε2 I 0 0 0      < 0 I 0 0 • • −ε 2    • • • −Qs 0  ˆs • • • • −Q   Ao Xc + Xc Ato + Yc + Yct Mf Zct Xs Qs Ad  0 0  • −ε3 0    • • −ε3 I 0 0    < 0  • • • −Qc 0  ˆs • • • • −Q  Bo Yd + Ydt Bot + Yh + Yht 2Ad ˆs ≥ 0 • 2Q

(2.152)

(2.153)

(2.154)

have a feasible solution. Moreover, the gain matrices are given by Kc = Ys Xs−1 , Ko = −Yc Xc−1 Co† 2 . Proof: Follows from Theorem 2.46 by setting Ma ≡ 0, Na ≡ 0, Nd ≡ 0. ∇∇∇ Corollary 2.49. Consider system (ΣAO ) with multiplicative gain perturbations and given matrices 0 < Qs = Qts ∈ n×n , 0 < Ws = Wst ∈ IRn×n , 0 < Qc = Qtc ∈ n×n , 0 < Wc = Wct ∈ IRn×n . There exist matrices 0 < Xs = Xst ∈ n×n , 0 < Xc = Xct ∈ IRn×n , Ys , Yc , Za , Yd , Yh , Zb , Zc and scalars ε2 > 0, ε3 > 0, such that the LMIs   Ao Xs + Xs Ato + Bo Ys + Yst Bot Bo Ms Zb Xs Qs Ad  • −ε2 I 0 0 0      < 0 I 0 0 • • −ε (2.155) 2    • • • −Qs 0  ˆs • • • • −Q   t t t Ao Xc + Xc Ao + Yc + Yc Md Z¯c Xs Qs Ad  • −Λc 0 0 0      < 0 I 0 0 • • −ε (2.156) 3    • • • −Qc 0  ˆs • • • • −Q  Bo Yd + Ydt Bot + Yh + Yht 2Ad (2.157) ˆs ≥ 0 • 2Q have a feasible solution. Moreover, the gain matrices are given by Kc = Ys Xs−1 , Ko = −Yc Xc−1 Co† . Proof: Follows from Theorem 2.46 by setting Ma ≡ 0, Na ≡ 0, Nd ≡ 0. ∇∇∇ 2

Co† = Cot [Co Cot ]−1 is the pseudo-inverse of Co .

2.10 Uncertain Model with State-Delays

79

In the absence of gain perturbation, we have the following result Corollary 2.50. Consider system (ΣAO ) and given matrices 0 < Qs = Qts ∈ n×n , 0 < Ws = Wst ∈ IRn×n , 0 < Qc = Qtc ∈ n×n , 0 < Wc = Wct ∈ IRn×n . There exist matrices 0 < Xs = Xst ∈ n×n , 0 < Xc = Xct ∈ IRn×n , Ys , Yc , Zs , Yd such that the LMIs   Ao Xs + Xs Ato + Bo Ys + Yst Bot Xs Qs Ad  • −Qs 0  < 0 (2.158) ˆs • • −Q   Ao Xc + Xc Ato + Yc + Yct Xs Qs Ad  • −Qc 0  < 0 (2.159) ˆs • • −Q  Bo Yd + Ydt Bot 2Ad (2.160) ˆs ≥ 0 • 2Q have a feasible solution. Moreover, the gain matrices are given by Kc = Ys Xs−1 , Ko = −Yc Xc−1 Co† . Proof: Follows from Corollary 2.48 or 2.49 by setting Mf ≡ 0, Ng ≡ 0, Nh ≡ 0, Ms ≡ 0, Md ≡ 0. ∇∇∇ Next, consider the state-delay system (Σ0 ) : x(t) ˙ = Ao x(t) + Ad x(t − τ ) + Bo u(t) y(t) = Co x(t) + Go u(t)

(2.161) (2.162)

where the corresponding state-space matrices contain uncertainties represented by a real convex bounded polytopic model of the type   Ao Ad Bo A(oλ) A(dλ) B(oλ) ∆ ∈ Sλ = = Co Go 0 C(oλ) G(oλ) 0   N  Aoj Adj Boj ,λ ∈ Λ (2.163) λj Coj Goj 0 j=1

where Λ is the unit simplex  ∆

Λ=

(λ1 , · · · , λN ) :

N 

 λ j = 1 , λj ≥ 0

(2.164)

j=1

Define the vertex set IN = {1, ..., N }. In the sequel, we use {Ao , · · · , Go } to imply generic system matrices and {Aoj , · · · , Goj , j ∈ IN} to represent the respective values at the vertices. The following two theorem are readily established

80

2 Resilient Control-Continuous Case

Corollary 2.51. Consider system (ΣO ) with additive gain perturbations and given matrices 0 < Qs = Qts ∈ n×n , 0 < Ws = Wst ∈ IRn×n , 0 < Qc = Qtc ∈ n×n , 0 < Wc = Wct ∈ IRn×n . There exist matrices 0 < Xs = Xst ∈ n×n , 0 < Xc = Xct ∈ IRn×n , Ys , Yc , Zs , Yd , Yh , Zb , Zc and scalars ε2 > 0, ε3 > 0, such that the LMIs   t Aoj Xs + Xs Atoj + Boj Ys + Yst Boj Boj Ms Zb Xs Qs Adj  • −ε2 I 0 0 0      < 0 (2.165) • • −ε I 0 0 2    • • • −Qs 0  ˆs • • • • −Q   t t t Aoj Xc + Xc Aoj + Yc + Yc Mf Zc Xs Qs Adj  • −ε3 0 0 0      < 0 I 0 0 • • −ε (2.166) 3    • • • −Qc 0  ˆs • • • • −Q  t Boj Yd + Ydt Boj + Yh + Yht 2Adj (2.167) ˆs ≥ 0 • 2Q have a feasible solution over the vertex set IN. Moreover, the gain matrices are given by Kc = Ys Xs−1 , Ko = −Yc Xc−1 Co† . Theorem 2.52. Consider system (ΣO ) with multiplicative gain perturbations and given matrices 0 < Qs = Qts ∈ n×n , 0 < Ws = Wst ∈ IRn×n , 0 < Qc = Qtc ∈ n×n , 0 < Wc = Wct ∈ IRn×n . There exist matrices 0 < Xs = Xst ∈ n×n , 0 < Xc = Xct ∈ IRn×n , Ys , Yc , Za , Yd , Yh , Zb , Zc and scalars ε2 > 0, ε3 > 0, such that the LMIs   t Aoj Xs + Xs Atoj + Boj Ys + Yst Boj Boj Ms Zb Xs Qs Adj  0 0 0  • −ε2 I    • • −ε2 I 0 0    < 0 (2.168)  • • • −Qs 0  ˆs • • • • −Q   t t t Aoj Xc + Xc Aoj + Yc + Yc Md Z¯c Xs Qs Adj  • −Λc 0 0 0      < 0 • • −ε I 0 0 (2.169) 3    • • • −Qc 0  ˆs • • • • −Q  t Boj Yd + Ydt Boj + Yh + Yht 2Aoj ≥ 0 (2.170) ˆs • 2Q have a feasible solution over the vertex set IN. Moreover, the gain matrices are given by Kc = Ys Xs−1 , Ko = −Yc Xc−1 Co† .

2.10 Uncertain Model with State-Delays

81

Remark 2.53. It is significant to observe that Corollaries 2.48-2.49 for the nominal system and Theorems 2.51-2.52 for the polytopic system new contributions to the area of resilient control. In addition, Corollary 2.50 provides an improved LMI-based conditions for the observer-based control over the results of [79, 87, 94, 90]. 2.10.5 Example 2.8 Consider a time-delay system of the type (2.118)-(2.119) with    1 0.7 0.5 0.5 0.32 Ao = , Bo = , Ad = , Ms = 1.2 0 −2 0.75 0 −0.5 Co = [−1 0] , Go = 1 , Mat = [1 1.4] , Na = [0.35 0.45] Mft = [0.7 0.6] , Mp = 0.3 Ng = 0.5 , Nh = [0.25 0.15] , Nf = 0.8 , Mdt = [0.5 0.5] , Nm = [0.6 0.7] By resorting to MATLAB, the ensuing results are summarized in Table 1 for both types of gain perturbations Table 2.10. Results of Dynamic Controllers: Example 2.8 P erturbatiion Kc Ko Additive -0.0411 -1.6345 -4.4615 -3.7004 Multiplicative -0.1209 -1.4786 -5.3155 -2.9874

2.10.6 Example 2.9 Consider a third-order time-delay system with the following data       −0.1 0.2 0 00 −2 0.95 −0.05 Ao =  0.1 −2 0.03  , Bo =  0 1  , Ad =  0 −0.4 0.1  0 0 −0.6 10 0.1 0 4    01 10 Co =  0 0  , Go = 01 10 Mat = [1 1.4 1.3] , Na = [0.35 0.45 0.55] , Mft = [0.07 0.06 0.05] Mpt = [1 2] , Nf = [0.7 0.3] , Nm = [0.6 0.7 0.4] , Mdt = [0.5 0.5 0.5] Ng = [−2

− 1] , Nh = [−1

−2

− 1] , Mst = [0.7 0.3]

The results are summarized in Table 2.11 for both types of gain perturbations

82

2 Resilient Control-Continuous Case Table 2.11. Results of Dynamic Controllers- Example 2.9 Kc Ko P erturbatiion 0.0155 13.9865 -0.1874 -0.2453 10.1027 Additive -3.0308 45.1042 11.8975 7.0135 -1.0025 12.1253 -2.0145 Multiplicative 0.1035 14.0015 -0.2038 -0.3055 10.2105 -4.0126 46.8865 12.9084 7.1004 -1.0845 11.8954 -1.9035

2.11 Notes and References This Chapter contains the basic methods of resilient feedback control theory applied to classes of time-delay systems. There are a lot of other methods/approaches to be examined or developed to complete the resilient control theory. Some of these methods will be exposed in the coming Chapters. Others are left to the interested readers and researchers. One of the challenging problems is to address the gain perturbation representation. In our view, we felt that the norm-bounded representation is general enough to encompass other related form. However, other forms would deserve further investigations.

3 Resilient Control-Discrete Case

3.1 Introduction A basic ingredient of robust control system design is the employment of a fixed quadratic Lyapunov function in order to guarantee robust stability [74, 105, 126]. In the presence of uncertain parameters, it is often desirable to perform the control system design not only to ensure stability but also to guarantee an adequate level of system performance. This brings about the notion of guaranteed cost control (GCC) [103]. In the course of controller implementation based on different control design methods it has been reported turns that the controllers are very sensitive with respect to errors in the controller coefficients [43]. Broadly speaking, the sources for this include, but not limited to, imprecision in analog-digital conversion, fixed word length, finite resolution instrumentation and numerical roundoff errors and it has been demonstrated that relatively small perturbations in controller parameters could even destabilize the closed loop system. This calls for fine tuning or, loosely speaking, additional redesign. Thus it is considered beneficial, and almost certainly essential, that the designed (nominal) controllers should be capable of tolerating some level of controller gain variations. This has been the heart of the controller fragility problem (see Chapter 1). The objective of this Chapter is to develop resilient controllers for a class of discrete-time systems with norm-bounded uncertainties and controller gain variations. Previous results to the analysis, robustness and design of discretetime systems can be found in [28, 29, 101, 115, 116] and their references. In the sequel, we focus on the development of resilient controllers and extend the results of [118, 121] to uncertain discrete-time systems with both types of gain variations. Necessary and sufficient conditions are established such that the resulting closed-loop feedback control system is quadratically stable for all admissible perturbations and uncertainties. These conditions are conveniently expressed in the form of linear matrix inequalities (LMIs). The feedback stabilization schemes are based on guaranteed cost control and H∞

M.S. Mahmoud: Resilient Control of Uncertain Dynamical Systems, LNCIS 303, pp. 83-98, 2004. © Springer-Verlag Berlin Heidelberg 2004

84

3 Resilient Control-Discrete Case

control approaches. System examples are provided to illustrate the theoretical developments.

3.2 Model of Discrete-Time Systems The class of linear discrete-time systems under consideration is represented by: (Σ∆ ) : xk+1 = [Ao + ∆Ak ]xk + Bo uk + Γ wk = [Ao + M ∆sk N ]xk + Bo uk + Γ wk zk = Co xk + Do uk

(3.1)

where xk ∈ n is the state, uk ∈ p is the control input, wk ∈ r is the external disturbance, zk ∈ q is the controlled output and matrices Ao ∈ n×n , Bo ∈ n×p , Co ∈ q×n , Do ∈ q×p , M ∈ n×α , and N ∈ β×n are real and known constant matrices with ∆sk ∈ α×β being an unknown matrix representing the norm-bounded uncertainties and belong to a bounded set, that is ∆ ∆sk ∈ ∆s = {∆tsk : ∆tsk ∆sk ≤ I , ∀k} The cost function associated with system (Σ∆ ) is given by: J=

∞ 

xtk Qxk + utk Ruk

(3.2)

k=0

where 0 < Q = Qt and 0 < R = Rt are given weighting matrices. 3.2.1 Gain Perturbations For a given state-feedback controller uk = Ko xk with Ko being the nominal controller gain, the actual implemented controller is assumed to be uk = [Ko + ∆Kk ] xk

(3.3)

where ∆Kk represents the gain perturbation, which is assumed to be either of two forms: 1) Additive Type ∆Kk = Ha ∆ck Ea

(3.4)

where Ha and Ea are known constant matrices and ∆ck is the uncertain parameter matrix represented by ∆

∆tck ∈ ∆c = {∆tck : ∆tck ∆ck ≤ I , ∀k

3.2 Model of Discrete-Time Systems

85

2) Multiplicative Type ∆Kk = Hm ∆ck Em Ko

(3.5)

where Hm and Em are known constant matrices. For simplicity in exposition, we define the following matrices: Ao = Ao + Bo Ko , ∆Ak = ∆Ak + Bo ∆Kk , A∆ = Ao + ∆Ak K∆ = Ko + ∆Kk

(3.6)

3.2.2 Quadratic Stability We start by the following definition Definition 3.1. Consider system (Σ∆ ) with cost function (3.2) and wk ≡ 0. The state-feedback control law (3.3) with controller gain perturbations (3.4) or (3.5) is said to be a guaranteed cost control (GCC) with quadratic cost matrix P > 0 if given matrices 0 < Q = Qt and 0 < R = Rt , the following LMI   t −P + Q At∆ P K∆ R  • −P 0  < 0 (3.7) • • −R has a feasible solution with respect to P for all admissible uncertainties ∆Ak and ∆Kk satisfying (3.1) and (3.4) or (3.5) Remark 3.2. The norm-bounded uncertainty structure considered here has been widely adopted in robust control [64, 74, 80, 82, 105, 64]. We note that it covers the usual matching condition as a special case by setting N ≡ I. Observe that the unit upperbound condition on ∆k does not cause any loss of generality. Indeed ∆k can always be normalized by appropriately selecting the matrices M and N . In the absence of controller perturbations ∆Kk ≡ 0, then Definition 3.1 reduces to the well known definition of guaranteed cost control for uncertain discrete-time systems [74]. In this chapter, we closely examine the following problem: Given the uncertain system (Σ∆ ) with cost function (3.2) along with the state-feedback control law (3.3) with controller gain perturbations (3.4) or (3.5), it is required to determine the feedback gain matrix Ko such that the closed-loop controlled system achieve quadratic stability for all admissible uncertainties satisfying ∆tk ∆k ≤ I , ∀k. We will derive expressions for gain matrix Ko using guaranteed cost and H∞ approaches. Considering the guaranteed cost approach, the following theorem provides a link between the notion of quadratic cost matrix, addressed in Definition 3.1, and an upper bound on the cost function J.

86

3 Resilient Control-Discrete Case

Theorem 3.3. Consider system (Σ∆ ) with cost function (3.2). Suppose that the state-feedback control law (3.3) with controller gain perturbations (3.4) or (3.5) is a quadratic GCC with quadratic cost matrix (QCM) P > 0. Then the resulting closed-loop system with wk ≡ 0 xk+1 = A∆ xk

(3.8)

is quadratically stable with the cost function satisfies the bound J ≤ xto P xo

(3.9)

Conversely, if system (Σ∆ ) is quadratically stable then there will be a QCM for this system and cost function (3.2). Proof (=⇒) : Let 0 < P = P t be a quadratic cost matrix for system (Σ∆ ) and cost function (3.2), it follows from (3.7) and the Schur complements that   t t (3.10) x A∆ P A∆ − P x < 0 ∀ x = 0 , ∀ ∆tk ∈ ∆s Hence, system (Σ∆ ) is quadratically stable. Next, introduce the Lyapunov functional V (xk ) = xtk P xk . Along the state sequences of (3.8), we have V (xk+1 ) − V (xk ) = xtk+1 P xk+1 − xtk P xk   = xtk [Ao + ∆Ak ]t P [Ao + ∆Ak ] − P xk   t < − xtk Q + K∆ RK∆ xk

(3.11)

Summing up k over [0, ∞) and using the quadratic stability of (Σ∆ ), we obtain   ∞  t RK∆ xk ≤ V (xo ) = xto P xo xtk Q + K∆ (3.12) J = k=0

since limk→∞ x(k) = 0. (⇐=) : Let system (Σ∆ ) be quadratically stable . If follows that there exists 0 < P = P t such that t RK∆ < 0 [Ao + ∆Ak ]t P [Ao + ∆Ak ] − P + Q + K∆ ∀∆sk ∈ ∆s , ∀∆ck ∈ ∆c

(3.13)

Hence, one can find  > 0 such that the following inequality holds: t RK∆ ] = −1 [Ao + ∆Ak ]t P [Ao + ∆Ak ] − −1 [P + Q + K∆ t¯ t ¯ ¯ ¯ [Ao + ∆Ak ] P [Ao + ∆Ak ] − P + Q + K∆ RK∆ ] < 0 ∀∆sk ∈ ∆s , ∀∆ck ∈ ∆c

(3.14)

3.3 Guaranteed Cost Control Synthesis

87

¯ = −1 Q, R ¯ = −1 R The above inequality implies that there exists matrices Q −1 ¯ such that the matrix P =  P is a QCM for system (Σ∆ ) and the proof is completed ∇∇∇ Having developed the basic stability theory, we proceed to the synthesis problem.

3.3 Guaranteed Cost Control Synthesis In this section, we establish results for GCC synthesis with both types of controller perturbations. We formulate the core of the work in the form of theorems such that their derived conditions will eventually lead to the design procedure. 3.3.1 Additive Gain Perturbations The following theorem provides a necessary and sufficient condition for GCC under additive gain perturbations. Theorem 3.4. Consider system (Σ∆ ) with cost function (3.2). There exist state feedback gain Ko such that the control law (3.3) with additive gain perturbations (3.4) is a GCC with a quadratic cost matrix P > 0 given matrices 0 < Q = Qt and 0 < R = Rt if and only if there exist matrices 0 < X = X t , Y and some parameters ε1 > 0, ε2 > 0 such that the following LMIs   −X XQ XN t XEat XAto + Y t Bot Yt  • −Q 0  0 0 0    • • −ε1 I 0  0 0    • •  0 0 • −ε2 I   < 0 t   −X + ε1 M M  • •  • • 0 t t   +ε2 Bo Ha Ha Bo   −1  −R +  • • • • • ε2 Ha Hat    −X ε1 M ε2 Bo Ha −R−1 ε2 Ha  • −ε1 I  0 < 0, < 0 (3.15) • −ε2 I • • −ε2 I are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 . Proof: By Definition 3.1 and [74], it follows that inequality (3.7) holds if and only if the LMI   −P + Q (Ao + Bo Ko )t P Kot R  • −P 0  + • • −R

88

3 Resilient Control-Discrete Case



 t  t   t t 0 0 N N    0 0  ε1  P M   P M  + ε−1 1 0 0 0 0 t  t t  t t   Ea Ea 0 0    0 0  + ε2  P Bo Ha   P Bo Ha  + ε−1 2 0 0 RHa RHa   −P + Q+ t t Ko R   ε−1 N t N + ε−1 E t Ea [Ao + Bo Ko ] P a 2   1   −P +   t t  = ε P M M P + • R P B H H ε 1 o a a  2    ε2 P Bo Ha Hat Bot P     −R+ • • ε2 RHa Hat R

< 0(3.16)

holds some parameters ε1 > 0, ε2 > 0. Grouping like terms and multiplying every block entry of (3.16) on the left and on the right by P −1 and setting X = P −1 and Y = Ko P −1 with some Schur complement operations we obtain the equivalent conditions of (3.15). ∇∇∇ Remark 3.5. It should be observed that matrix R > 0 is a part of the data and hence its inverse is readily available. However the presence of R−1 render the LMIs formalism ill-posed. To overcome this problem, we introduce a relaxation variable S = R−1 thereby in implementation we seek a feasible solution of the following system of LMIs:   −X XQ XN t XEat XAto + Y t Bot Yt  • −Q 0  0 0 0    • • −ε1 I 0  0 0    • •  • −ε2 I 0 0   < 0 t   −X + ε1 M M  • •  • • 0 t t   +ε2 Bo Ha Bo Ha    −S+  • • • • • ε2 Ha Hat     −X ε1 M ε2 Bo Ha S I −S ε2 Ha  • −ε1 I  0 > 0, ≥ 0 > 0, • −ε2 I •R • • −ε2 I with respect to X, Y and S. In the absence of system uncertainties (M ≡ 0, N ≡ 0) we obtain the nominal system (Σo ) : x(k + 1) = Ao x(k) + Bo u(k) for which implementation of Theorem 3.4 reduces to:

(3.17)

3.3 Guaranteed Cost Control Synthesis

89

Theorem 3.6. Consider system (Σo ) with cost function (3.2). There exist state feedback gain Ko such that the control law (3.3) with additive gain perturbations (3.4) is a GCC with a quadratic cost matrix P > 0 given matrices 0 < Q = Qt and 0 < R = Rt if and only if there exist matrices 0 < X = X t , Y, 0 < S = S t and some parameter ε2 > 0 such that the following LMIs   −X XQ XEat XAto + Y t Bot Yt  • −Q 0  0 0    • • −ε2 I  < 0 0 0   t t  • •  0 • −X + ε2 Bo Ha Bo Ha • • • • −S + ε2 Ha Hat    −X ε2 Bo Ha −S ε2 Ha S I < 0, < 0, ≥ 0 (3.18) •R • −ε2 I • −ε2 I are feasible with respect to X, Y and S. The feedback gain is Ko = Y X −1 . 3.3.2 Multiplicative Gain Perturbations The following theorem provides a necessary and sufficient condition for GCC under multiplicative gain perturbations. Theorem 3.7. Consider system (Σ∆ ) with cost function (3.2). There exist state-feedback gain Ko such that the control law (3.3) with multiplicative gain perturbations (3.5) is a GCC with a quadratic cost matrix P > 0 given matrices 0 < Q = Qt and 0 < R = Rt if and only if there exist matrices 0 < X = X t , Y and some parameters ε1 > 0, ε2 > 0 such that the following LMIs   t −X XQ XN t Y t Em XAto + Y t Bot Yt  • −Q 0  0 0 0    • • −ε1 I 0  0 0    • •  • −I 0 0   < 0 t   −X + ε1 M M  • •  • • 0 t t   +ε2 Bo Hm Bo Hm   −1  −R +  • • • • • t ε 2 Hm Hm    −X ε1 M ε2 Bo Hm −R−1 ε2 Hm  • −ε1 I  0 < 0 (3.19) < 0, • −ε2 I • • −ε2 I are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 . Proof: Following Theorem 3.4, it follows that inequality (3.7) under multiplicative perturbations (3.5) holds if and only if the LMI

90

3 Resilient Control-Discrete Case



 −P + Q+ −1 t   [Ao + Bo Ko ]t P Kot R  −1ε1 N N +   ε K t E t Em K o  o m  2    −P +   t t  ε1 P M M P + ε2 P Bo Hm Hm R  •   t   P ε2 P Bo Hm Bot Hm     −R+ • • t ε2 RHm Hm R

< 0 (3.20)

holds some parameters ε1 > 0, ε2 > 0. Multiplying every block entry of (3.20) on the left and on the right by P −1 and setting X = P −1 and Y = Ko P −1 with Schur complement operations we obtain the equivalent conditions of (3.19). ∇∇∇ In the absence of system uncertainties (M ≡ 0, N ≡ 0) we obtain the nominal system (3.17) for which Theorem 3.4 reduces to: Theorem 3.8. Consider system (Σo ) with cost function (3.2). There exist state-feedback gain Ko such that the control law (3.3) with additive gain perturbations (3.5) is a GCC with a quadratic cost matrix P > 0 given matrices 0 < Q = Qt and 0 < R = Rt if and only if there exist matrices 0 < X = X t , Y, 0 < S = S t and some parameter ε2 > 0 such that the following LMIs   t −X XQ Y t Em XAto + Y t Bot Yt  • −Q 0  0 0    • •  < 0 −I 0 0   t t  • •  0 • −X + ε2 Bo Hm Bo Hm −1 t • • • • −R + ε2 Hm Hm   −X ε2 Bo Hm −R−1 ε2 Hm < 0, (3.21) • −ε2 I • −ε2 I are feasible with respect to X and Y . The feedback gain is Ko = Y X −1 . Remark 3.9. Needless to stress that implementation of Theorems 3.7 and 3.8 is performed in the mannar of Remark 3.5. 3.3.3 Example 3.1 Consider a discrete-time system of the type (3.1) with     0.99 0.01 1 0.4 0.1 0 , Bo = , Γ = , M= , Hm = 0.2 Ao = 0.01 0.51 0 0.5 0 0.1   0.1 0.05 0.2 N = , Co = [0.5 0.5] , Do = 1 , Ha = −0.02 0.1 0.1 Ea = [0.3 0.6] , Em = 0.4 A summary of the numerical computations is given in Table 1, from which it is clear that the case of multiplicative gain perturbations yields less cost bound than additive gain perturbations.

3.4 H∞ Control Synthesis

91

Table 3.1. Computational Results of Example 3.1 P erturbation Theorem Q R J o ε1 ε2 Ko Additive 3.4 1 0 0.75 0.0823 0.01 0.3 -0.9457 -0.1697 0 1 Multiplicative 3.7 1.2 0 0.5 0.0376 0.05 0.7 -1.0612 -0.1925 0 1.2

3.3.4 Example 3.2 Consider the following discrete-time system of the type (3.1) with       0.2113 0.0087 0.4524 0.6135 0.6538 100 Ao =  0.0824 0.8096 0.8075  , Bo =  0.2749 0.4899  , Γ =  0 1 0  0.7599 0.8474 0.4832 0.8807 0.7741 001      0.1 100 1 0 N = [0.05 − 0.02 0.1] , Co =  0 1 0  , Do = , M =  0.2  01 0.3 001   0.15 0 0.6 0 Em = , Hm = , Ea = [0.3 0.6 − 0.1] 0 0.44 0 −0.38  0.2 −0.3 Ha = 0.1 0.2 Observe that the system is open-loop unstable as λ(Ao ) = {0.3827, −0.4919, 1.6133} A summary of the numerical computations is given in Table 2 which shows comparable results in both cases of gain perturbations. Table 3.2. Computational Results of Example 3.2 P erturbation Theorem Additive 3.4 1 0 0 Multiplicative 3.7 1.2 0 0

Q 0 1 0 0 1.2 0

0 0 1 0 0 1.4

R Jo Ko 1 0 5.6532 0.3024 0.3758 0.3114 01 0.2299 0.4852 0.5001 1 0 5.4899 0.3016 0.3809 0.3632 01 0.2304 0.5014 0.4988

3.4 H∞ Control Synthesis In the sequel, we extend the results attained in the forgoing section to the case of H∞ control. For a given nominal state-feedback controller uk = Ko xk , the

92

3 Resilient Control-Discrete Case

closed loop transfer function from wk to zk is given by: Hwz (ζ) = C∆ [ζ I − A∆ ]−1 Γ where

∆

C∆ = Co + Do ∆K ,

(3.22)

Co = Co + Do Ko

This means that given a prespecified H∞ -norm level γ, it is required to develop conditions for the state-feedback controller (3.3) with controller gain perturbations (3.4) or (3.5) to render the closed-loop system (3.6) quadratically stable and ||Hwz ||∞ ≤ γ for all admissible uncertainties ∆tk ∆k ≤ I. Following [110], we have the following definition: Definition 3.10. Consider system (Σ∆ ). The state-feedback control law (3.3) with controller gain perturbations (3.4) or (3.5) is said to be a H∞ with disturbance attenuation γ > 0 if there exists a matrix 0 < P t = P such that the following LMI   t −P At∆ P C∆ At∆ P Γ  • −P 0  0   < 0 (3.23)  •  • −I 0 2 t • • • −γ I + Γ P Γ has a feasible solution for all admissible uncertainties ∆Ak and ∆Kk satisfying (3.1) For simplicity in exposition, we introduce the following vector quantities:     t   t 0 N 0 Ea PM   0   P Bo Ha   0            , σ4 =  0     0 σ1 =     0  , σ2 =  0  , σ3 =     0   0    0  0 0 0 0 0       0 0 0   0   0   0             0 σ5 =  Do Ha  , σ6 =  0  , σ7 =  (3.24)   0   Γ tP M   Γ t P Bo Ha  0 0 0 The following theorems four summarize the corresponding results for controller gain variations of both types. For the additive type, we have Theorem 3.11. Consider system (Σ∆ ) with cost function (3.2). There exist state-feedback gain Ko such that the control law (3.3) with additive gain perturbations (3.4) is a H∞ control with disturbance attenuation γ > 0 if and only if there exist matrices 0 < X = X t , Y and some parameters ε1 > 0, ........., ε7 > 0 such that the following LMIs

3.4 H∞ Control Synthesis



93



¯ t XE ¯ t XAt + Y t B t XC t + Y t Dt XAt + Y t B t 0 −X X N o o o o o a o • −Φ 0 0 0 −XΓ 0   • • −Ψ 0 0 0 0   • • • −Υ1 0 0 0   < 0 0 0  • • • • −Υ2  • • • • • −γ 2 I Γt  • • • • • • −Υ3      −X ε1 M ε2 Bo Ha −X ε4 M ε5 Ea  • −ε1 I  < 0,  • −ε4 I 0  < 0, −I ε3 Do Ha < 0(3.25) 0 • −ε3 I • • −ε2 I • • −ε5 I

        

are feasible with respect to X, Y where    N ¯ = N  , E ¯a = Ea , Υ1 = X − ε1 M M t − ε2 Bo Ha Bot Hat N Ea N Ψ = diag[ε2 I ε3 I] , Υ3 = X − ε4 M M t − ε5 Ea Eat Φ = diag[ε1 I ε4 I ε5 I] , Υ2 = I − ε3 Do Ha Hat Dot

(3.26)

The feedback gain is Ko = Y X −1 . Proof: By Definition 3.1 and Theorem 3.4 with simple Schur operations, it follows that inequality (3.23) holds if and only if the LMI   −P (Ao + Bo Ko )t P (Co + Do Ko )t (Ao + Bo Ko )t P 0  • −P 0 Γ 0     • • −I 0 0     • • • −γ 2 I Γ tP  • • • • −P −1 −1 t t t t t + ε1 σ1 σ1t + ε−1 1 σ 2 σ 2 + ε2 σ 3 σ 3 + ε2 σ 4 σ 4 + ε3 σ 5 σ 5 + ε3 σ 4 σ 4

−1 t t t + ε4 σ6 σ6t + ε−1 4 σ 2 σ 2 + ε5 σ 7 σ 7 + ε5 σ 2 σ 2 < 0

(3.27)

holds for some parameters ε1 > 0, ....., ε7 > 0. Grouping like terms and by the congruent transformation [X X I I X] with X = P −1 , Y = Ko P −1 and some Schur complement operations we obtain the equivalent conditions of (3.25). ∇∇∇ Theorem 3.12. Consider system (Σo ) with cost function (3.2). There exist state-feedback gain Ko such that the control law (3.3) with additive gain perturbations (3.4) is a H∞ control with disturbance attenuation γ > 0 if and only if there exist matrices 0 < X = X t , Y and some parameters ε2 > 0, ε3 > 0 such that the following LMIs

94

3 Resilient Control-Discrete Case



 ¯ t XAt + Y t B t XC t + Y t Dt XAt + Y t B t 0 −X X E o o o o o a o  • −Ψ 0 0 0 0     •  • −Υ 0 0 0 1o   < 0  •  • • −Υ 0 0 2    • Γt  • • • −γ 2 I • • • • • −Υ3o    −X ε2 Bo Ha −X ε5 Ea −I ε3 Do Ha > 0, > 0, > 0 (3.28) • −ε2 I • −ε5 I • −ε3 I are feasible with respect to X, Y where Υ1o = X + ε2 Bo Ha Hat Bot , Υ3o = X − ε5 Ea Eat

(3.29)

The feedback gain is Ko = Y X −1 . For the multiplicative type, we have the following results Theorem 3.13. Consider system (Σ∆ ) with cost function (3.2). There exist state feedback gain Ko such that the control law (3.3) with multiplicative gain perturbations (3.5) is a H∞ control with disturbance attenuation γ > 0 if and only if there exist matrices 0 < X = X t , Y , Z and some parameters ε1 > 0, ....., ε7 > 0 such that the following LMIs   ¯ t Y tE ¯ t XAt + Y t Bot XC t + Y t Dot XAto + Y t Bot 0 −X X N a o o  • −Φ 0 0 0 −XΓ 0     •  • −Ψ 0 0 0 0    •  < 0 • • −Υ 0 0 0 1    •  • • • −Υ 0 0 2    • Γt  • • • • −γ 2 I • • • • • • −Υ33     −X ε1 M ε2 Bo Hm −X ε4 M ε5 Em Y  • −ε1 I <0  < 0 ,  • −ε4 I 0 0 • • −ε2 I • • −ε5 X 2  −I ε3 Do Hm <0 (3.30) • −ε3 I are feasible with respect to X, Y, and S. The feedback gain is Ko = Y X −1 t where Υ33 = X − ε4 M M t − ε5 Em ZZ t Em . Theorem 3.14. Consider system (Σo ) with cost function (3.2). There exist state-feedback gain Ko such that the control law (3.3) with multiplicative gain perturbations (3.5) is a H∞ control with disturbance attenuation γ > 0 if and only if there exist matrices 0 < X = X t , Y and some parameters ε2 > 0, ε3 > 0, ε5 > 0, ε6 > 0, ε7 > 0 such that the following LMIs

3.5 Design Example



95



¯ t XAt + Y t B t XC t + Y t Dt XAt + Y t Bot 0 −X Y t E a o o o o o • −Ψ 0 0 0 0   • • −Υ1 0 0 0   < 0 • • • −Υ2 0 0   Γt  • • • • −γ 2 I • • • • • −Υ33    −X ε2 Bo Hm −I ε3 Do Hm −X ε5 Em Y < 0, < 0, <0 (3.31) • −ε5 X 2 • −ε2 I • −ε3 I

      

are feasible with respect X, Y, Z . The feedback gain is Ko = Y X −1 .

3.5 Design Example In this section, we present the results of system design and numerical simulation of multivariable digital feedback control for regulating vapor compression cycles in two-phase flow heat exchangers with recycling. The original fifthorder state space model and the associated experimental data are described in [39] using the length of the two-phase section, pressure and temperature of tube wall in the evaporator stage in addition to the pressure and temperature of tube wall in the condenser stage as state variables and the compressor speed and valve opening as control variables. In this dynamic system, the output variables are the evaporating temperature and the superheat, and hence the objective is to control indoor room temperature (or evaporating temperature, or cooling capacity) through the compressor speed and to regulate superheat by the expansion valve opening. Using a linearized model around a representative operating point for typical residential air conditioners, the continuous-time model was derived in [39]. By employing a sampling time of 2 sec, the discretized model matrices in the type (3.1) are given by:   0.8955 0.4683 −0.1851 0.0429 0.0481  0.1276 −0.8179 0.2970 0.0144 0.0162     Ao =   0.0740 0.1056 0.9297 0.0084 0.0094   0.0119 0.0170 0.0278 −0.5002 0.5605  0.0016 0.0023 0.0037 0.0668 0.9441     −0.0178 0.0127 0.025 −0.12  −0.0071 0.0013   0.031 −0.044       , Γ =  −0.051 0.007  −0.0041 0.0008 Bo =       0.0377 −0.0148   0.018 −0.018  0.005 −0.002 −0.048 0.009   0.32 −0.17  0.27 −0.31      , Do = 0.2 0 0.41 −0.04 , Ea = [1 2] Cot =    0 0.1  0.09 −0.06  0.07 −0.12

96

3 Resilient Control-Discrete Case

M t = [0.2 0.3 − 0.2 0.1 0.4] , Hat = [0.4 − 0.2] N = [0.1 0.2 0.1 0.2 0.1] In deriving the model, we used about %10 recycling ratio corresponding to building up some minor feed-back loops around the condenser and evaporator stages to allow mixing up the respective variables. By invoking the MATLAB toolbox, the solutions of Theorems 3.4-3.14 are summarized in Tables 3 and 4. Table 3.3. Results of Guaranteed Cost Control P erturbation Theorem Q R Ko Additive 3.4 I5 0.1 ∗ I2 5.3098 0.0987 0.4358 1.9765 Multiplicative 3.7 I6 0.3 ∗ I2 6.1136 0.0744 0.8107 2.4823

Table 3.4. Results of H∞ Control P erturbation Theorem γ Ko Additive 3.11 1.7851 4.3098 0.0055 0.1546 11.1233 Multiplicative 3.13 1.9662 3.2207 0.0715 0.8265 12.8644

We have considered by computer simulation the effects of changing the desired evaporating temperature (first measurable output) and the desired superheat (second measurable output). In Figures 3.1-3.4, simulation results are plotted for the two perturbation patterns and control design methods. As demonstrated, the compressor speed (first control input) and the valve opening (second control input) generated by the developed design methods show that in case of multiplicative perturbations, the response is more effective in regulating the dynamic model of vapor compression cycles in two-phase flow heat exchangers.

3.6 Notes and References We have presented some of the resilient control results for a class of discretetime systems. More results will be presented in Chapter 6 dealing with delaydependent issues. The reported results have been largely based on [28, 29, 74, 101, 115, 116, 118, 121]. There are other works in the open literature that

3.6 Notes and References

97

Fig. 3.1. Behavior of Control Input u1 due to Step Change in Output y1 : GCC Approach

Fig. 3.2. Behavior of Control Input u2 due to Step Change in Output y2 : GCC Approach

can be examined for a redesign viewpoint based on the resilience approach. Needless to stress that more work is needed particularly in case of observerbased resilient control, static output-feedback resilient control and nonlinear resilient control, to name a few.

98

3 Resilient Control-Discrete Case

Fig. 3.3. Behavior of Control Input u1 due to Step Change in Output y1 : H∞ Approach

Fig. 3.4. Behavior of Control Input u2 due to Step Change in Output y2 : H∞ Approach

4 Resilient Adaptive Control

4.1 Introduction This Chapter takes on some steps further along the way towards the controller fragility and performance deterioration issues due to inaccuracies in controller implementation. Interestingly enough, it addresses the problem of resilient adaptive control problem for classes of uncertain continuous-time and discrete-time systems with state-delays against controller gain variations. In the continuous case, design results on both norm-bounded and convexbounded parametric uncertainties are derived. In the discrete-case, design results on norm-bounded parametric uncertainties are developed. Adaptive control schemes are constructed when gain perturbation bounds are known and then extended to accommodate unknown norm-bounded perturbations. To complete the design profile, special cases of delayless continuous and discrete systems are provided. All the developed results are conveniently expressed in LMI feasibility problem. It is fair to mention that almost all the analytical developments hereafter are new and original contributions to the field of adaptive control applied to time-delay systems. Results of several simulation examples are presented to demonstrate the developed theory.

4.2 Continuous-Time Model A schematic of the problem setup is displayed in Figure 4.1 which shows a plant P subjected to uncertainties ∆p and a controller K having gain perturbations ∆c . We consider the plant P to be represented by the following class of time-delay systems: x(t) ˙ = A∆ x(t) + Bo u(t) + A∆d x(t − τ ) y(t) = Co x(t)

(4.1)

where x(t) ∈
M.S. Mahmoud: Resilient Control of Uncertain Dynamical Systems, LNCIS 303, pp. 99-127, 2004. © Springer-Verlag Berlin Heidelberg 2004

100

4 Resilient Adaptive Control

A∆ ∈ n×n , B∆ ∈ n×p and A∆d ∈ n×n , are represented by [A∆ A∆d ] = [Ao Ad ] + M ∆p (t)[Na Nd ]

(4.2)

where Ao ∈ n×n , Bo ∈ n×p , Co ∈ q×n , Ad ∈ n×n , M ∈ n×α , Na ∈ β×n and Nd ∈ β×n , are real and known constant matrices with ∆p (t) is a matrix of parametric uncertainties which belongs to a compact and bounded set in the form ∆

∆p (t) ∈ ∆(t) = {∆p (t) : ∆p (t)∆tp (t) ≤ I ∀ t}

Fig. 4.1. Uncertain Plant with Controller Perturbations

4.2.1 Nominal Model In the absence of uncertainties (∆ ≡ 0), system (4.1) reduces to x(t) ˙ = Ao x(t) + Bo u(t) + Ad x(t − τ ) y(t) = Co x(t)

(4.3)

It is a straightforward task to show by selecting a Lyapunov functional of the type

0 V (x(t)) = xt (t)P x(t) + xt (s)Qx(s) ds −τ

0
n×n

, 0 < P = P t ∈ n×n

that the nominal state-feedback controller

4.3 Adaptive Schemes ∆

u(t) = 1/2 Bot P x(t) = Ko x(t)

101

(4.4)

renders system (4.3) asymptotically stable for arbitrary constant delay τ ∈ [0 → τ ∗ ]. That is, if given a matrix 0 < Q = Qt ∈ n×n there exists a matrix 0 < P = P t ∈ n×n such that the LMI   P Ao + Ato P P Ad P Bo  • −Q 0  < 0 (4.5) • • −I More on the design of feedback controllers can be found in [81]. 4.2.2 Design Objective In practical situations, there are at least two sources of inaccuracies when implementing controller (4.4). The first source is obviously due to the presence of uncertainties in the system matrices and the second source arises from gain perturbations due to various reasons [16, 38]. Therefore, it is naturally to consider, for a given nominal feedback controller u(t) = Ko x(t), that the actual implemented controller is assumed to have two-terms: u(t) = [µ Ko + ∆K(t)] u(t)

(4.6)

where µ is an adjustable gain factor, Ko is the gain matrix to be determined and ∆K(t) represents the gain perturbation, which is assumed to be normbounded of the form: ||∆K(t)|| ≤ β

(4.7)

where β > 0 is an upper bound to be dealt with in the subsequent analysis. The problem we would like to examine in this section is to develop a feedback control scheme that ensures that the closed-loop system of (4.1)(4.3) is asymptotically stable. Among the various possible approaches, we aim at constructing an adaptive scheme to achieve the cited design objective. Needless to stress the salient features of adaptive stabilization methods are well-established [98].

4.3 Adaptive Schemes To achieve our goal, we will proceed in two stages. In the first stage, we attempt to construct an adaptive schemes for the uncertain time-delay system (4.1) assuming that the gain perturbation bound is known. Then in the second stage, we extend the results to accommodate bounded-but-known gain perturbations.

102

4 Resilient Adaptive Control

4.3.1 Known Perturbation Bound When the gain perturbation bound is known, then the purpose of adaptation is to accommodate the uncertainties of system (4.1). The following adaptive scheme is proposed u(t) = [˜ µ Ko + ∆K(t)]x(t) ˙µ ˜−1 Ko x , µ ˜(0) = µ+ , g > 0 ˜ = −g µ ˜ + xt Kot µ

(4.8)

where K, g > 0 represent, respectively, a control gain matrix and a growth factor, both will be determined in the sequel. A convenient Lyapunov functional V (.) is given by

t

Vn (x, µ) = xt (t)P x(t) +

xt (s)Qx(s)ds + 1/2 µ ˜2

(4.9)

t−τ

where 0 < P = P t ∈ IRn×n and 0 < Q = Qt ∈ IRn×n . The following theorem summarizes the first main result: Theorem 4.1. System (4.1) under the adaptive controller (4.8) is asymptotically stable if for a given matrix 0 < Q = Qt and a scalar β > 0, there exist matrices 0 < X = X t , Y, Z and scalars ε > 0,  > 0 such that the LMIs   Ao X + XAto + Z + Z t M M XQ L Y A d  +βBo X + βXBot      • −εI 0 0 0 0 0     • • −I 0 0 0 0     < 0 • • • −Q 0 0 0 (4.10)     • • • • −εI 0 0     • • • • • −I 0    −Q+  • • • • • • Ndt Nd have a feasible solution. Moreover, the feedback gain is Ko = Y X −1 . Proof: Evaluation of the derivative of Vn (x, µ) along the solutions of system (4.1)-(4.2) using adaptive controller (4.8) with some algebraic manipulations yields:  t t ˙ ˜) = x (t) P A∆c + A∆c P + Q x(t) + 2xt P A∆d x(t − τ ) Vn (x, µ ˜µ ˜˙ − xt (t − τ ) Q x(t − τ ) + µ  t t −1 t t = x (t) P A∆c + A∆c P + Q + P A∆d Q A∆d P + Ko Ko x(t)   t t −1 t − x (t − τ ) − x (t)P A∆d Q x(t − τ ) − A∆d P x(t) − g µ ˜2

4.3 Adaptive Schemes

103

 t t −1 t t ≤ x (t) P A∆c + A∆c P + Q + P A∆d Q A∆d P + Ko Ko x(t) ∆

= xt (t) Ξn x(t) A∆c = A∆ + µ ˜ Bo Ko + Bo ∆K(t) = [Ao + µ ˜ Bo Ko ] + ∆A + Bo ∆K(t) = Ac + ∆A + Bo ∆K(t)

(4.11)

(4.12)

From Lyapunov theory, it follows that V˙ n (x, µ ˜) < 0 is guaranteed if Ξn < 0. By [81] with Inequalities I1-I-2 of the Appendix, it is a straightforward task to show that P A∆c + At∆c P ≤ P Ao + Ato P + ε−1 P M M t P + εNat Na P A∆d Q−1 At∆d P ≤ −1 P M M t P + P Ad [Q − Ndt Nd ]−1 Atd P for some scalars ε > 0,  > 0. Using these inequalities and the Schur complements the stability condition holds if the inequality:   P Ao + Ato P +  Q + Kot Ko + εNat Na    P Ad  +βP Bo + βBot P + P M P M    t  µ <0 ˜ Ko P (4.13)  ˜ P Bo Ko + µ    • −εI 0 0     • • −I 0 t • • • −Q + Nd Nd has a feasible solution. Using the congruent transformation T = diag[X I I I], X = P −1 and defining

˜ Bo Ko P −1 , L = εP −1 Y = Ko , Z = µ

it follows that Schur complement operations convert (4.13) to (4.10) and thus the proof is completed. ∇∇∇ Remark 4.2. The dynamical relation of µ ˜ consists of two-terms: one is growth factor and the other is a product of µ ˜ and x so as to preserve inter-coupling between the states and the gain factor. The selection of the growth factor g > 0 guarantees the asymptotic stability of system (4.8) and different values will only affect the speed of convergence. This is illustrated by the following example.

104

4 Resilient Adaptive Control

4.3.2 Example 4.1 This example is motivated by the dynamics of bio-strata in water-quality studies on the river Nile. A pilot-scale model of the form (4.1) is described by:       −0.2 0 0 0 0 0 0.8 0 Ao =  0 −0.9 −0.3  , Ad =  −0.7 0 0  , Bo =  0.2 0.3  0.8 0 −1 0 −0.8 0 0 0.4         −0.1 −0.1 −0.2 10 M =  0.1  , Nat =  0  , Ndt =  0  , Cot =  0 0  0.3 0.1 0.2 01 The feasible solution of LMIs (4.10) is given by   −2.146 0.01 0  K =  0.157 −3.804 −1.626 In Figures 4.2-4.3, the behavior of the output variables and the gain factor µ ˜ are displayed for different values of g, from which it is clear that relatively-high values of g tend to yield effective stabilization.

Fig. 4.2. Plot of Output Response for different values of g

4.3.3 Unknown Gain Perturbation Bound Now we consider the application of controller (4.6) subject to bound (4.7) where β is unknown. The following adaptive scheme is then proposed u(t) = µ ¯ Kx(t) µ ¯˙ = −g µ ¯ + xt K t µ ¯−1 Kx + β ||x||2 , µ ¯(0) = µ∗ β˙ = −h β − µ ¯ ||x||2 + β −1 xt Rx , β(0) = β ∗

(4.14)

4.3 Adaptive Schemes

105

Fig. 4.3. Plot of Gain Factor µ ˜ for different values of g

where h > 0 represents a growth factor. Note that scheme (4.14) is constructed in the same way as scheme (4.8). A convenient Lyapunov functional V (.) is given by

t Vb (x, µ ¯, β) = xt (t)P x(t) + xt (s)Qx(s)ds + 1/2 µ ¯2 + 1/2 β 2 (4.15) t−τ

The following theorem summarizes the desired result: Theorem 4.3. System (4.1) under the adaptive controller (4.14) is asymptotically stable if for given matrices 0 < Q = Qt , 0 < R = Rt there exist matrices 0 < X = X t , Y, L, Z and scalars g > 0, ε > 0,  > 0 such that the LMIs   Ao X + XAto + Z + Z t t M M Na Y XQ XR Ad   +L + Lt     • −εI 0 0 0 0 0 0     • • −I 0 0 0 0 0     • • • −σI 0 0 0 0   < 0 (4.16)   • • • • −I 0 0 0     • • • • • −Q 0 0     • • • • • • −R 0    −Q+  • • • • • • • Ndt Nd have a feasible solution. Moreover, the feedback gain is K = Y X −1 and the adjustable factor µ ¯−1 = Bo Y Z −1 . ¯, β) along the solutions of Proof: An evaluation of the derivative of Vb (x, µ system (4.1)-(4.2) using (4.12) and adaptive controller (4.8), yields:

106

4 Resilient Adaptive Control

 t t ˙ ¯, β) = x (t) P A∆c + A∆c P + Q x(t) + 2xt P A∆d x(t − τ ) Vb (x, µ ¯µ ¯˙ + β β˙ − xt (t − τ ) Q x(t − τ ) + µ  t t −1 t t = x (t) P A∆f + A∆f P + Q + P A∆d Q A∆d P + K K + R x(t)   t t −1 t − x (t − τ ) − x (t)P A∆d Q x(t − τ ) − A∆d P x(t) − gµ ¯2 − hβ 2  ≤ xt (t) P A∆f + At∆f P + Q + P A∆d Q−1 At∆f P + K t K + R x(t) ∆

A∆f

= xt (t) Ξf x(t) ¯ Bo K] + ∆A = Af + ∆A = [Ao + µ

(4.17) (4.18)

Following parallel development to Theorem (4.1), it is readily evident that ˜) < 0 holds if there exist scalers ε > 0,  > 0 the stability condition V˙ n (x, µ P A∆f + At∆f P + Q + P A∆d Q−1 At∆d P + K t K + R + S < 0 =⇒ ¯ P Bo K + µ ¯ K t Bot P P Ao + Ato P + Q + K t K + εNat Na + µ + ε−1 P M M t P + −1 P M M t P + P Ad [Q − Ndt Nd ]−1 Atd P < 0 ⇐⇒   P Ao + Ato P + R+  Q + K t K + εNat Na P M P M  P Ad   t t  +¯  ¯ K Bo P  µ P Bo K + µ <0 (4.19)   • −εI 0 0     • • −I 0 t • • • −Q + Nd Nd Using the congruent transformation T = diag[X I I I], X = P −1 and defining Y = KX, L = µ ¯ Bo KP −1 , it follows that Schur operations converts (4.19) to (4.16) under the constraint ε−1 σ = 1 and hence the proof is completed. ∇∇∇ 4.3.4 Example 4.2 The following example is motivated by the dynamics of machining chatter with the matrices of system (4.1) given by:       0 0 1 0 0 000 0.2  0     0 0 1   , Ad =  0 0 0 0  , Bo =  0  Ao =   −16 10 0  10 0 0 0   0.5  0  −5 −15 0.02 −0.25 0 000 0.8         −0.1 0.1 0 1  0.2        0 0.2 t t t      0 M =  0.4  , Na =  −0.1  , Nd =  0  , Co =  1  −0.5 0 −0.2 0

4.3 Adaptive Schemes

107

The feasible solution of LMI (4.10) is K = [−0.387 1.245

− 0.336

− 0.804] ,

ε = 1.345

The system output y and the control input u are plotted in Figure 4.4 for g = 5, h = 4 while the variation of µ ¯ and β are displayed in Figure 4.5.

Fig. 4.4. Plot of Output Response and Control Input: Example 4.2

In all cases smooth behavior is recorded which supports the flexibility of the developed adaptive control scheme.

Fig. 4.5. Variation of factors µ and β: Example 4.2

108

4 Resilient Adaptive Control

4.3.5 Special Cases In case of delay-less system x(t) ˙ = [Ao + M ∆p (t)Na ]x(t) + Bo u(t) y(t) = Co x(t)

(4.20)

Theorems 4.1-4.3 specialize to the following corollaries Corollary 4.4. System (4.20) under the adaptive controller (4.8) is asymptotically stable if for a given matrix 0 < Q = Qt and a scalar β > 0, there exist matrices 0 < X = X t , Y, Z and a scalar ε > 0 such that the LMIs   Ao X + XAto + Z + Z t M L Y   +βBo X + βXBot    • −εI 0 0  (4.21)   < 0  • • −εI 0  • • • −I have a feasible solution. Moreover, the feedback gain is Ko = Y X −1 . Corollary 4.5. System (4.20) under the adaptive controller (4.14) is asymptotically stable if for given matrices 0 < Q = Qt , 0 < R = Rt there exist matrices 0 < X = X t , Y, L, Z and scalars g > 0, ε > 0 such that the LMIs   Ao X + XAto + Z + Z t t M Na Y XR   +L + Lt    • −εI 0 0 0    < 0 (4.22)  • • −σI 0 0     • • • −I 0  • • • • −R have a feasible solution. Moreover, the feedback gain is K = Y X −1 . Thus far, we have developed LMI-based sufficient conditions for checking the feasibility of resilient feedback controller to stabilize a class of dynamical systems with norm-bounded parametric uncertainties. In the next section, we direction attention to the case of convex-bounded parametric uncertainties.

4.4 Polytopic Model Consider a plant represented by the class of linear continuous-time systems with state-delay: (ΣJ ) :

x(t) ˙ = Ao x(t) + Ad x(t − τ ) + Bo u(t) , y(t) = Co x(t) (4.23)

4.5 Design Results

109

where x ∈ IRn is the system state, u(t) ∈ IRm is the control input, y(t) ∈ IRp is the measured output and τ is a time-delay factor such that 0 < τ ≤ τ ∗ , 0 < τ˙ ≤ τ + < 1 The corresponding state-space matrices contain uncertainties represented by a real convex bounded polytopic model of the type  [Ao ,

Ad ,

∆

Co ] ∈ Sλ =

Bo ,

=

N 

[A(oλ) , A(dλ) , B(oλ) ,

C(oλ) ] 

λj [Aoj , Adj , Boj , Coj ], λ ∈ Λ

(4.24)

j=1

where Λ is the unit simplex  ∆

Λ=

(λ1 , · · · , λN ) :

N 

 λ j = 1 , λj ≥ 0

(4.25)

j=1

Define the vertex set IN = {1, ..., N }. Throughout this section, we use {Ao , · · · , Fo } to imply generic system matrices and {Aoj , · · · , Cdj , j ∈ IN} to represent the respective values at the vertices. The initial condition is specified as x(0), x(s) = xo , φ(s) , where φ(.) ∈ L2 [−τ, 0]. In line of the analytical developments of the foregoing section, we focus hereafter on the design of adaptive feedback control schemes that ensures that the closed-loop system of (4.23) is asymptotically stable. The actual implemented controller is assumed to have two-terms: u(t) = [µ Ko + ∆K(t)] u(t)

(4.26)

where µ is an adjustable gain factor, Ko is the gain matrix to be determined and ∆K(t) represents the gain perturbation, which is assumed to be normbounded of the form: ||∆K(t)|| ≤ β

(4.27)

where β > 0 is an upper bound to be dealt with in the subsequent analysis.

4.5 Design Results The control design proceeds in two stages. In the first stage, we attempt to construct an adaptive schemes for the uncertain time-delay system (4.23) assuming that the gain perturbation bound is known. Then in the second stage, we extend the results to accommodate bounded-but-known gain perturbations.

110

4 Resilient Adaptive Control

4.5.1 Known Perturbation Bound When the gain perturbation bound is known, then the purpose of adaptation is to accommodate the uncertainties of system (4.1). The following adaptive scheme is proposed u(t) = [˜ µ Ko + ∆K(t)]x(t) µ ˜˙ = −g µ ˜ +

N 

λj xt Kot µ ˜−1 Ko x , µ ˜(0) = µ+ , g > 0

(4.28)

j=1

where Ko , g > 0 represent, respectively, a control gain matrix and a growth factor, both will be determined in the sequel. Combining system (ΣJ ) and controller (4.28), we get the closed-loop dynamics ˜ Bo Ko + Bo ∆K(t) x˙ = A∆c x(t) + Ad x(t − τ ) , A∆c = Ao + µ µ ˜˙ = −g µ ˜ +

N 

λj xt Kot µ ˜−1 Ko x , µ ˜(0) = µ+ , g > 0

(4.29)

j=1

A parameter-dependent Lyapunov functional V (.) is given by Vn (x, µ) =

N 



t λj x (t)Pj x(t) +



t t

x (s)Qj x(s)ds

+ 1/2 µ ˜2 (4.30)

t−τ

j=1

where 0 < Pj = Pjt ∈ IRn×n and 0 < Qj = Qtj ∈ IRn×n . The following theorem summarizes the main result: Theorem 4.6. System (4.23) under the adaptive controller (4.28) is asymptotically stable if for a given matrix 0 < Qj = Qtj , j ∈ IN and a scalar β > 0, there exist matrices 0 < Xj = Xjt , Yj , Zj , j ∈ IN such that the LMIs   Aoj Xj + Xj Atoj + Zj + Zjt Xj Qj Yj Adj  t  +βBoj Xj + βXj Boj    • −Qj 0 0  (4.31)  < 0 , j ∈ IN   • • −I 0  ˆj • • • −Q have a feasible solution. Moreover, the feedback gain is Ko = Yj Xj−1 . Proof: Evaluation of the derivative of Vn (x, µ) along the solutions of system (4.29) with some algebraic manipulations yields: V˙ n (x, µ ˜) =

N 

 λj xt (t){Pj A∆cj + At∆cj Pj + Qj }x(t) + 2xt Pj Adj x(t − τ )

j=1

N  2 − (1 − τ˙ )x (t − τ ) Qj x(t − τ ) − g µ ˜ + λj xt Kot Ko x t

j=1

(4.32)

4.5 Design Results

111

where ˜Boj Ko + Boj ∆K = Acj + Boj ∆K A∆cj = Aoj + µ

(4.33)

ˆ j = (1 − τ + )Qj , completing the squares and using (4.27) it follows Letting Q that  N  V˙ n (x, µ ˜) ≤ λj xt (t) Pj [Aoj + µ ˜Boj Ko ] + [Aoj + µ ˜Boj Ko ]t Pj + Qj j=1

t ˆ −1 Atdj Pj + Kot Ko + βPj Boj + βBoj P + Pj Adj Q j x(t) j ∆

=

N 

λj xt (t) Ξnj x(t)

(4.34)

j=1

From Lyapunov theory, it follows that V˙ n (x, µ ˜) < 0 is guaranteed if Ξnj < 0. By the Schur complements, it follows that Ξnj < 0 holds if and only if the matrix inequality :   t Pj Aoj + Atoj Pj + µ ˜Pj Boj Ko + µ ˜Kot Boj Pj A P j dj  t  Pj +Qj + Kot Ko + βPj Boj + βBoj <0 (4.35) ˆj • Q Using the congruent transformation T = diag[Xj I], Xj = Pj−1 and introducing the linearization variables Yj = Ko Xj , Zj = µ ˜ Boj Ko Pj−1 , it follows that standard Schur complement operations convert (4.35) to (4.31) and thus the proof is completed. ∇∇∇ Remark 4.7. The dynamical relation of µ ˜ consists again of two-terms: one is growth factor and the other is a product of µ ˜ and x so as to preserve inter-coupling between the states and the gain factor. The selection of the growth factor g > 0 guarantees the asymptotic stability of system (4.28) and different values will only affect the speed of convergence. This is illustrated by the following example. 4.5.2 Example 4.3 This example is motivated by the dynamics of bio-strata in water-quality studies on the river Nile. A pilot-scale model of the form (4.23) with two modes of operation is described by:       −0.2 0 0 0 0 0 0.8 0 Ao1 =  0 −0.9 −0.3  , Ad1 =  −0.7 0 0  , Bo1 =  0.2 0.3  0.8 0 −1 0 −0.8 0 0 0.4       −0.6 0 0.1 −0.1 0 0 0.3 0 Ao2 =  0 −0.8 0.1  , Ad2 =  0.5 0 −0.2  , Bo2 =  0.4 0.4  0.6 0 −1.2 0 0.4 −0.5 0 0.9

112

4 Resilient Adaptive Control

The feasible solution of LMIs (4.31) is given by   −2.3015 0.0104  0 K =  0.1682 −2.9704 −1.8246 In Figure 4.6, the behavior of the output variables is displayed for different values of g, from which it is clear that relatively-high values of g tend to yield effective stabilization.

Fig. 4.6. Plot of Output Response for different values of g

4.5.3 Unknown Gain Perturbation Bound Now we consider the application of controller (4.26) subject to bound (4.27) where β is unknown. The following adaptive scheme is then proposed u(t) = µ ¯ Kx(t) µ ¯˙ = −g µ ¯ +

N 

λj xt K t µ ¯−1 Kx + β ||x||2 , µ ¯(0) = µ∗

j=1

β˙ = −d β − µ ¯ ||x||2 +

N 

λj β −1 xt Rx , β(0) = β ∗

(4.36)

j=1

where d > 0 represents a growth factor. Note that scheme (4.36) is constructed in the same way as scheme (4.28). A parameter-dependent Lyapunov functional V (.) is given by

4.5 Design Results

Vb (x, µ ¯, β) =

N 



λj xt (t)Pj x(t) +

j=1

113



t t

x (s)Qj x(s)ds

t−τ

+ 1/2 µ ¯2 + 1/2 β 2

(4.37)

The following theorem summarizes the desired result: Theorem 4.8. System (4.23) under the adaptive controller (4.36) is asymptotically stable if for given matrices 0 < Q = Qt , 0 < R = Rt there exist matrices 0 < X = X t , Y, L, Z and scalars g > 0 such that the LMIs   Aoj Xj + Xj Atoj Yj Xj Qj Xj Rj Adj   +Lj + Ltj    • −I 0 0 0    < 0 (4.38)  0 0  • • −Qj    • • • −Rj 0  ˆ • • • • −Q have a feasible solution. Moreover, the feedback gain is K = Yj Xj−1 . Proof: An evaluation of the derivative of Vb (x, µ ¯, β) along the solutions of system (4.23) using (4.32) and adaptive controller (4.28), yields: ¯, β) = V˙ b (x, µ

N 

 λj xt (t){Pj [Aoj + µ ¯Boj K] + [Aoj + µ ¯Boj K]t Pj + Qj }x(t)

j=1

+ 2xt Pj Adj x(t − τ ) − (1 − τ˙ )xt (t − τ ) Qj x(t − τ ) +µ ˜µ ˜˙ + β β˙

(4.39)

Proceeding like Theorem 4.6, it follows that  ¯, β) = xt (t) Pj [Aoj + µ ¯Boj K] + [Aoj + µ ¯Boj K]t Pj + Qj V˙ b (x, µ t x(t) + K t K + Rj Pj Adj (1 − τ˙ )Q−1 A P j dj j   t t −1 t − x (t − τ ) − x (t)P A∆d Q x(t − τ ) − A∆d P x(t) − g µ ¯2 − dβ 2  t ≤ x (t) Pj [Aoj + µ ¯Boj K] + [Aoj + µ ¯Boj K]t Pj + Qj ∆ −1 t t ˆ + K K + Rj Pj Adj Qj Adj Pj x(t) = xt (t) Ξf x(t) (4.40) It is readily evident that the stability condition V˙ n (x, µ ˜) < 0 holds if Ξf < 0 and by the Schur complements this corresponds to

114

4 Resilient Adaptive Control



Pj Aoj + +Atoj Pj t µ ¯Pj KBoj  ¯Pj Boj K + µ  •   •   • •

Kt Q −I • • •

0 −Q • •

 R P Adj   0 0  <0 0 0   −R 0  ˆj • −Q

(4.41)

Using the congruent transformation T = diag[Xj , I, I, I, I], Xj = Pj−1 and defining Yj = KXj , Lj = µ ¯ Boj KPj−1 , it follows that Schur operations converts (4.41) to (4.38) and hence the proof is completed. ∇∇∇ 4.5.4 Example 4.4 The following example deals with an alternative dynamic model of machining chatter with the matrices of system (4.23) given by:       0 0 1 0 0 000 0.2  0     0 0 1   , Ad1 =  0 0 0 0  , Bo1 =  0  Ao1 =   −16 10 0     0 10 0 0 0 0.5  −5 −15 0.02 −0.25 0 000 0.8       0 0 1 0 0 000 0.3  0     0 0 1   , Ad2 =  0 0 0 0  , Bo2 =  0  Ao2 =   −17 9     0 0 12 0 0 0 0.4  −4.8 −16 0.03 −0.35 0 000 0.7       0 0 1 0 0 000 0.4  0     0 0 1   , Ad3 =  0 0 0 0  , Bo3 =  0  Ao3 =   −15 12 0     0 11 0 0 0 0.6  −5.2 −17 0.04 −0.45 0 000 0.6 The feasible solution of LMI (4.38) is K = [−0.5647 1.5124

− 0.6044

− 0.7824]

4.5.5 Model Reference State Regulation We now direct attention to an alternative approach to resilient adaptive control. This approach belongs to the class of model-reference adaptive techniques which employs both feedforward and feedback gains [98]. Our effort concentrates hereafter on regulating the system state to the origin leaving other techniques like state-tracking, output regulation or output tracking to later investigations. Following [63], we start by selecting output transformation matrix Lo ∈ p×p , and invertible feedforward matrix 0 < Ho ∈ m×m such that the reference model has the form ˙ = Do ξ(t) + Fo υ(t) , ω(t) = Go ξ(t) ξ(t) Do = Ao − Fo Ko , λ(Do ) < 0 , Fo = Bo Ho , Lo Go = Co (4.42)

4.5 Design Results

115

where υ(t) ∈ m is a bounded, piecewise continuous vector function of time and ω(t) ∈ p is the reference output vector. It should be observed that the matrices Lo , Ho can be selected by the designer from past experience and/or acquaintance of the desired system behavior and the nominal gain Ko is selected to meet some stability consideration. To accommodate gain perturbations ∆K(t), we let the gain matrix Ko be replaced by Ko + ∆K(t) → K in the model (4.42) and proceed like the foregoing section to determine K to ensure the asymptotic stability of the closed-loop model. Now given the objective of adaptive state regulation, we construct the resilient adaptive controller as u(t) = −Ho K x(t) − µm Γ x(t)

(4.43)

which consists of two terms: the first term −Ho K x(t) is a stabilizing part and the second term µm Γ x(t) is an adaptive part with µm being a scalar adaptation gain adjusted according to µ˙m (t) = −σ µm +

N 

λj xt (t)Γ t Ho−1 Γ x(t)

j=1

µm (0) = µmo , σ > 0

(4.44)

Figure 4.7 shows the structure of the controlled system. Over the vertex set IN, the matrices Do , Fo , Ho have the values Doj , Foj , Hoj , j ∈ IN. It follows on combining (4.23) and (4.42)-(4.44) with K re placing Ko that the closed-loop dynamics takes the form −1 x(t) ˙ = Doj x(t) + µm Foj Hoj Γj x(t) + Adj x(t − τ )

µ˙m (t) = −σ µm +

N 

−1 λj xt (t)Γjt Hoj Γj x(t)

(4.45)

j=1

In this respect, a convenient Lyapunov functional Va (.) is given by 

t N  λj xt (t)Pj x(t) + xt (s)Qj x(s)ds + µ2m (4.46) Va (x, µm ) = j=1

t−τ

where 0 < Pj = Pjt ∈ IRn×n and 0 < Qj = Qtj ∈ IRn×n . In the sequel we t select the gain matrix Γj = Foj Pj . The following theorem summarizes the third main result: Theorem 4.9. System (4.1) with the reference model (4.42) under the adaptive controller (4.43)-(4.44) is asymptotically stable if for a given matrix 0 < Qj = Qtj , j ∈ IN there exist matrices 0 < Xj = Xjt , Yj , j ∈ IN such that the LMIs   t Aoj Xj + Xj Atoj + Foj Yj + Yjt Foj Xj Qj Adj  • −Qj 0  < 0 , j ∈ IN (4.47) ˆj • • −Q have a feasible solution. Moreover, the feedback gain is K = −Yj Xj−1 .

116

4 Resilient Adaptive Control

Proof: Evaluation of the derivative of Vn (x, µ) along the solutions of system (4.45) with some algebraic manipulations in the manner of Theorem 4.6, it yields: ˜) = V˙ a (x, µ

N 

 −1 t t λj xt (t){Pj Doj + Doj Pj − 2µm Pj Foj Hoj Foj Pj + Qj }x(t)

j=1

+ 2x Pj Adj x(t − τ ) − (1 − τ˙ )x (t − τ ) Qj x(t − τ ) t

t

− 2σ µ2m + 2

N 

−1 t Foj Pj x(t) λj xt (t)µm Pj Foj Hoj

j=1

≤

N 

 t Pj λj x (t) Pj Aoj + Atoj Pj − Pj Foj Kj + Qj + Kjt Foj t

j=1

t ˆ + Pj Adj Qj Aoj x(t) ∆

=

N 

λj xt (t)Ξaj x(t)

(4.48)

j=1

By the Schur complements, the stability condition Ξaj < 0 using Xj = Pj−1 and introducing the linearization variables Yj = −KXj can be readily written into the LMI (4.47) and thus the proof is completed. ∇∇∇ In the single-input single-output case, we use the following change of variables Hoj −→ hj , Goj −→ gj , Loj −→ rj Foj −→ fj = hj boj , rj gjt −→ ctj , Γj −→ γj = hj Pj boj (4.49) A little algebra yields the following result Corollary 4.10. Consider system (4.23) with single input and single output (m = 1, p = 1). The reference model (4.42) under the adaptive controller (4.43)-(4.44) renders the closed-loop system asymptotically stable if for a given matrix 0 < Qj = Qtj , j ∈ IN there exist matrices 0 < Xj = Xjt ∈ n×n , yj ∈ n×1 , j ∈ IN such that the LMIs   Aoj Xj + Xj Atoj + hj boj yjt + hj yj btoj Xj Qj Adj  • −Qj 0  < 0 , j ∈ IN (4.50) ˆj • • −Q have a feasible solution. Moreover, the feedback gain is k = −Xj−1 yj .

4.5 Design Results

117

Fig. 4.7. Structure of the controlled system

4.5.6 Example 4.5 In order to demonstrate the model-reference state regulation, we consider a system of the type (4.23) with the following data       11 0.01 0 0 −1 0.5 −0.2 Ao1 =  0 −0.5 −2  , Ad1 =  0 −0.02 0  , Bo1 =  0 1  10 0 0 0.03 0.15 0 −2       0 −0.1 0 10 −0.8 0 0.2 Ao2 =  −0.5 2 0.1  , Ad2 =  0.2 0 0  , Bo2 =  1 1  0 −0.1 0 01 0.2 0 −2.5     10 01 t t Co1 =  1 0  , Co2 = 1 0 01 01 Let the invertible feedforward matrix 0 < Ho be selected as     1.3 0.1 0.7 −0.2 20 40 Ho1 = , Ho2 = , ; Lo1 = , Lo2 = −0.1 0.8 0.2 1.1 02 04 Using Q1 = 0.3 × I3 , Q2 = 1.2 × I3 , the feasible solution of LMI (4.47) is given by  1.0102 0.9945 −0.0015 K= −0.0013 0.0004 2.9675

118

4 Resilient Adaptive Control

4.5.7 Special Cases Once again, in case of the delay-less system (ΣF ) :

x(t) ˙ = Ao x(t) + Bo u(t) , y(t) = Co x(t)

(4.51)

where the matrices Ao , Bo , Co are represented by (4.24). The following corollaries can be easily established. Corollary 4.11. System (4.51) under the adaptive controller (4.28) is asymptotically stable if for a given scalar β > 0, there exist matrices 0 < Xj = Xjt , Yj , Zj , j ∈ IN such that the LMIs 

 Aoj Xj + Xj Atoj + Zj + Zjt Y j  t  +βBoj Xj + βXj Boj < 0 , • −I

j ∈ IN

(4.52)

have a feasible solution. Moreover, the feedback gain is Ko = Yj Xj−1 . Corollary 4.12. System (4.51) under the adaptive controller (4.36) is asymptotically stable if for a given matrix 0 < R = Rt there exist matrices 0 < X = X t , Y, L, Z and a scalar g > 0 such that the LMIs   Aoj Xj + Xj Atoj Y X R j j j  +Lj + Ltj   < 0 (4.53)  • −I 0  • • −Rj have a feasible solution. Moreover, the feedback gain is K = Yj Xj−1 . Corollary 4.13. System (4.51) with the reference model (4.42) under the adaptive controller (4.43)-(4.44) is asymptotically stable if there exist matrices 0 < Xj = Xjt , Yj , j ∈ IN such that the LMIs t < 0 , Aoj Xj + Xj Atoj + Foj Yj + Yjt Foj

j ∈ IN

(4.54)

have a feasible solution. Moreover, the feedback gain is K = −Yj Xj−1 . We now proceed to consider uncertain discrete-time systems

4.6 Discrete-Time Model In this section, we focus on the development of resilient adaptive controllers for a class of linear discrete-time systems with norm-bounded parametric uncertainties and bounded controller gain variations. In the absence of controller gain variations, adaptive control schemes for uncertain time-delay systems are developed in [96, 21, 63, 72, 81]. We extend these results to a class of

4.6 Discrete-Time Model

119

uncertain discrete-time systems with state-delay. In this regard, the results compiled hereafter are new contributions to the research investigations on resilient control of uncertain time-delay systems. Towards our goal, we consider the following class of time-delay systems with parametric uncertainties: xk+1 = A∆ xk + Bo uk + A∆d xk−d yk = Co xk

(4.55)

where xk ∈ n is the state vector; uk ∈ p is the control input, y(t) ∈ q is the controlled output, d is a constant time-delay and the uncertain matrices A∆ ∈ n×n and A∆d ∈ n×n , are represented by [A∆ A∆d ] = [Ao Ad ] + M ∆k [Na Nd ]

(4.56)

where Ao ∈ n×n , Bo ∈ n×p , Co ∈ q×n , Ad ∈ n×n , M ∈ n×α , Na ∈ β×n and Nd ∈ β×n , are real and known constant matrices with ∆k is a bounded matrix of uncertainties such that ∆

∆k ∈ ∆k = {∆tk : ∆k ∆k < I , ∀ k} Remark 4.14. The class of systems (4.55) emerges in many areas dealing with functional difference equations or delay-difference equations [35]. On the application side, these systems appear in cold rolling mills [51] and decision-making of manufacturing systems [112]. Our purpose now is to proceed to design a resilient feedback controller for system (4.55) in the sense that it should be capable of accommodating the norm-bounded parametric uncertainties and guaranteeing desirable behavior against controller gain perturbations. The following definition is recalled [81]. Definition 4.15. Given matrix 0 < Q = Qt ∈ n×n . System (4.55) is said to robustly quadratically stable (RQS) if and only if the following inequality  −1 t At∆ P A∆ − P + At∆ P A∆d Q − At∆d P A∆d A∆d P A∆ + Q < 0 ∆

∆k ∈ ∆k = {∆tk : ∆k ∆k < I , ∀ k}

(4.57)

has a feasible solution 0 < P = P t ∈ n×n . 4.6.1 Stability and Stabilization Results To proceed further, we introduce the Lyapunov function k−1 

Vk = xtk P xk +

j=k−d

A preliminary stability result is provided

xtj Qxj

(4.58)

120

4 Resilient Adaptive Control

Theorem 4.16. Given a matrix 0 < Q = Qt ∈ n×n , system (4.55) with uk ≡ 0 is RQS if and only if the linear matrix inequality (LMI)   −P + Q + Nat Na Nat Nd Ato P 0  • −Q + Ndt Nd Atd P 0    < 0 (4.59)  • • −P P M  • • • −I has a feasible solution 0 < P = P t ∈ n×n . Proof (Sufficiency): Suppose that 0 < P = P t ∈ n×n is a feasible solution of LMI (4.59) for all admissible uncertainties ∆k ∈ ∆k . Given Vk of (4.58), the corresponding Lyapunov difference ∆Vk = Vk+1 − Vk along the solutions of (4.55) has the form: ∆Vk = xtk [At∆ P A∆ − P + Q]xk + xtk At∆ P A∆d xk−d + xtk−d At∆d P A∆ xk − xtk−d Qxk−d t  t   A∆ P A∆ − P + Q At∆ P A∆d xk xk = xk−d • −Q xk−d ∆

= ηkt Ξk ηk , ηk = [xtk

xtk−d ]t

(4.60)

By Schur complements operation and in view of Definition 4.15, it follows that ∆Vk < 0 for ηk = 0 and hence system (4.55) with uk ≡ 0 is RQS. (Necessity): Suppose that system (4.55) with uk ≡ 0 is RQS and given a matrix 0 < W = W t . It follows from Definition 4.15 using (4.56) and Schur complements that there exists a matrix 0 < S = S t such that     −S + W 0 At∆ S −S + W 0 Ato S  • −W At∆d S  =  • −W Atd S  + • • −S • • −S    t 0 Na  0  ∆k [Na Nd 0] +  Ndt  ∆tk [0 0 S t M t ] < 0 SM 0 ∀ ∆k ∈ ∆k Therefore, given any ∆k ∈ ∆k we get   −S + Q 0 Ato S ∆ • −W Atd S  < Υ =  • • −S    t 0 Na  0  ∆k [Na Nd 0] +  Ndt  ∆tk [0 0 S t M t ] SM 0 Letting σ = [0 0 S t M t ]t , ψ = [Na Nd 0], we have the inequality

(4.61)

(4.62)

4.6 Discrete-Time Model

ξ t Υ ξ < −2ξ t σ ∆k ψξ ∀ ∆k ∈ ∆k , ξ = 0

121

(4.63)

From which it follows that given ξ = 0 ξ t Υ ξ < −2 max {ξ t σ ∆k ψξ : ∆k ∈ ∆k } ≤ 0

(4.64)

Hence, given ξ = 0 [ξ t Υ ξ]2 > 4 max {[ξ t σ ∆k ψξ]2 : ∆k ∈ ∆k }

(4.65)

By Lemma 8.3 of the Appendix, we get [ξ t Υ ξ]2 > 4 ξ t σσ t ξ ξ t ψ t ψξ , ξ = 0

(4.66)

Application of Lemma 8.4 of the Appendix implies that there exists a α > 0 such that α2 σσ t + α Υ + ψ t ψ < 0

(4.67)

∆

With P = αS, W = αQ, inequality (4.67) reduces to the LMI   −P + Q + Nat Na Nat Nd Ato P   < 0 • −Q + Ndt Nd Atd P • • −P + P M M t P

(4.68)

Finally, a simple Schur complements operation turns (4.68) to (4.59) and therefore the proof is completed. ∇∇∇ Theorem 4.17. Given a matrix 0 < Q = Qt ∈ n×n , system (4.55) is robustly quadratically stabilizable by the feedback control uk = Ko xk if and only if the LMI   −X XNat Nd XAto + Y t Bot XQ XNat  • −Q + N t Nd Atd P 0 0  d   t  • 0 0  • −X + M M (4.69)   < 0  • • • −Q 0  • • • • −I has a feasible solution 0 < X = X t ∈ n×n , Y ∈ m×n . Moreover the gain matrix is given by Ko = Y X −1 . Proof: Follows from Theorem 4.16 on re placing Ao by Ao + Bo Ko and employing the congruent transformation diag[X I I] with X = P −1 and ∇∇∇ Y = Ko P −1 . As we learned before in the continuous-time model, there are at least two sources of inaccuracies when implementing the state-feedback controller uk = Ko xk beyond the availability of state measurements. The first source is obviously due to the presence of uncertainties in the system matrices and the

122

4 Resilient Adaptive Control

second source arises from gain perturbations due to various reasons [16, 38]. Therefore, it is naturally to consider uk = Ko xk as a nominal feedback controller but the actual implemented controller is assumed to have two-terms: uk = [µ Ko + ∆Ko ] xk

(4.70)

where µ is an adjustable gain factor, Ko is the gain matrix to be determined and ∆Ko represents the gain perturbation, which is assumed to be belong to a compact set of the form: ∆

∆Ko ∈ ∆k = {∆Ko : ∆Ko = β Imn }

(4.71)

where Imn is the unit matrix of order m × n and β > 0 is an upper bound to be dealt with in the subsequent analysis. The problem of interest in the sequel is to develop a feedback control scheme that ensures that the closed-loop system of (4.55)-(4.56) and (4.70)is robustly quadratically stable. Among the various possible approaches, we aim at constructing an adaptive scheme to achieve the cited design objective.

4.7 Adaptive Stabilization Methods We will proceed along two consecutive stages. In the first stage, we attempt to construct an adaptive stabilization method for the uncertain time-delay system (4.55) assuming that the gain perturbation bound β is known. Then in the second stage, we extend the results to accommodate bounded-butknown gain perturbations thereby establishing another . adaptive stabilization method 4.7.1 Known Perturbation Bound When the gain perturbation bound is known, then the purpose of adaptation is to accommodate the uncertainties of system (4.1). The following adaptive scheme is proposed uk = [Ko + ∆Ko ]xk µ ˜k+1 = g µ ˜k + αt Ko xk , µ ˜0 = µ+ , |g| < 1

(4.72)

where Ko ∈ m×n , α ∈ m×1 represent, respectively, a control gain matrix and an adjustable factor, both will be determined in the sequel. A convenient Lyapunov functional Vk is given by Vk = xtk P xk +

k−1 

xtj Qxj + µ ˜2k

(4.73)

j=k−d

where 0 < P = P t ∈ IRn×n and 0 < Q = Qt ∈ IRn×n . The following theorem summarizes the first main result:

4.7 Adaptive Stabilization Methods

123

Theorem 4.18. Consider system (4.55) under the adaptive controller (4.72) and given matrices 0 < Q = Qt , α and a scalar β . The closed-loop system is RQS if and only if the fo1lowing LMI   XAto + Y t Bot t t t Z  −X XNa Nd +βXInm Bot XQ XNa Y α      −Q+ t  •  XAd 0 0 0 0 t   N N d d     −X+  •  < 0, • 0 0 0 0 t   MM    •  • • −Q 0 0 0    •  • • • −I 0 0    •  • • • • −I 0 • • • • • • −1 + g 2 |g| < 1

(4.74)

have a feasible solution 0 < X = X t , Y, Z, g. Moreover the gain matrix is given by Ko = Y X −1 . Proof: By similarity to Theorem 4.17, we replace Ao by Ao +Bo Ko +Bo ∆Ko and apply the congruent transformation diag[X I I] to (4.68) with X = P −1 , Y = Ko P −1 and the linearization Z = gP −1 Kot α. Then performing Schur complements, we obtain LMI (4.74). ∇∇∇ 4.7.2 Unknown Perturbation Bound When the gain perturbation bound is unknown, then the adaptive scheme should be constructed to accommodate the uncertainties of system (4.55) and robustify the closed-loop system. For this purpose, the following adaptive scheme is provided uk = [Ko + βk Imn Ko ]xk µ ¯k+1 = g µ ¯k + αt Ko xk , |g| < 1, µ ¯0 = µ+ βk+1 = d βk + ψ t Ko xk , |d| < 1, β0 = β +

(4.75)

where α ∈ m×1 , ψ ∈ m×1 represent adjustable factors, both will be determined in the sequel. In this case, a convenient Lyapunov functional Vk is given by Vk = xtk P xk +

k−1 

xtj Qxj + µ ¯2k + βk2

(4.76)

j=k−d

where 0 < P = P t ∈ IRn×n and 0 < Q = Qt ∈ IRn×n . The following theorem summarizes the first main result:

124

4 Resilient Adaptive Control

Theorem 4.19. Consider system (4.55) under the adaptive controller (4.75) and given matrices 0 < Q = Qt , α, ψ and a scalar β. The closed-loop system is RQS if and only if the fo1lowing LMI   XAto + Y t Bot t t t t XQ XNa Y α W ψ Z R  −X XNa Nd  +W t Bot     −Q+ t  •  0 0 0 0 0 0 XAd t   N N d d   t  •  • −X + M M 0 0 0 0 0 0    •  < 0 • • −Q 0 0 0 0 0    •  • • • −I 0 0 0 0    •  • • • • −I 0 0 0    •  • • • • • −I 0 0   2  •  • • • • • • −1 + g 0 • • • • • • • • −1 + d2 |g| < 1

,

|d| < 1

(4.77)

have a feasible solution 0 < X = X t , Y, Z, W, R, g, d. Moreover the gain matrix is given by Ko = Y X −1 . Proof: By similarity to Theorem 4.16, we replace Ao by Ao + Bo Ko + Bo βImn Ko and apply the congruent transformation diag[X I I] to (4.75) with X = P −1 , Y = Ko P −1 and the linearizations Z = gP −1 Kot α, W = βImn P −1 , R = dP −1 Kot ψ. Performing Schur complements lead to the LMIs (4.77). ∇∇∇ Remark 4.20. In implementing the LMI (4.74) or LMIs (4.77), we start by selecting Q as diagonal matrix of weighting factors and α (same for ψ) as row vector of equal adjustable elements and update the selection till reaching a feasible solution. The availability of convenient software based on interiorpoint minimization [26] helps much in performing computer simulation.

4.8 Examples Some examples will be solved to demonstrate the theoretical developments 4.8.1 Example 4.6 In terms of system (4.55)-(4.56), the data are:     −2 0 0 0.1 1 0.7 Ao = , Ad = , Bo = , M= 0 1 0.1 0.5 1 0.8    0.4 0 0.7 Nat = , Γ = , Ndt = , Cot = [0.5 0.5] 0.6 1 0.8

4.8 Examples

125

Observe that the system is open-loop unstable as λ(Ao ) = {−2, 1}. It took several computations to reach at appropriate weighting matrices. A summary of the feasible solutions is given in Table 4.1. It is quite clear that the selection of input data Q, β, α, ψ plays a crucial role in the computational burden. However, the resulting gain Ko changes slightly. A typical state response of the closed-loop system is depicted in Figures 4.8 and 4.9.

Fig. 4.8. Closed-loop State Response-x1: Example 4.6

Fig. 4.9. Closed-loop State Response-x2: Example 4.6

126

4 Resilient Adaptive Control

4.8.2 Example 4.7 Consider the following discrete-time system of the type (4.55)-(4.56) with     0.5 0.2 0.1 0.2 0.1 1 −0.6 Ao = , Bo = , M= , Ad = 0.6 0.4 0 0.1 0.1 0.4 0.5    0.02 2 0.02 , Γ = , Ndt = , Cot = [1 0] Nat = 0.01 1.5 0.03 A summary of the numerical computations is given in Table 4.2 which shows comparable results in both cases of gain perturbations. Table 4.1. Computational Results of Example 4.6 Theorem Q 4.17 2 0 0 2 4.18 5 0 0 5 4.19 7 0 0 3.5 4.17 3 0 0 6 4.18 8 0 0 8 4.19 10 0 0 5 4.17 4 0 0 8 4.18 10 0 0 10 12 0 4.19 0 15

β

α

1.4 2.3

ψ

Ko -0.7514 0.5496 -0.7386 0.4068

1.9 3.25 1.75 -0.8025 0.3576 -0.8223 0.6755 2.7 3.7

-0.7669 0.6407

3.1 5.4 2.66 -0.8787 0.5617 -0.8345 0.9116 3.6 6.1

-0.8785 0.8237

4.9 7.3 7.5 -0.8066 0.7446

Table 4.2. Computational Results of Example 4.7 Theorem Q β α ψ 10 2 4.17 2 14 4.18 15 1 3.5 6.15 1.45 1 5 4.19 8 3 2.7 7.44 2.37 6.75 1.56 3 16

Ko 4.6818 -0.4981 -6.2863 0.6743 -3.9785 0.4068 7.1235 -1.0578 -5.2775 1.1616 6.1765 -0.6452

4.9 Notes and References

127

4.9 Notes and References Most of the results reported in this Chapter are new. As such, it paves the way to more extension by exploiting other ideas and methods of adaptive control theory. This is equally true for continuous and discrete-time systems. Of particular interest is the sliding-mode and variable-structure control methods.

5 Resilient Linear Filtering

5.1 Introduction An integral part of deterministic and stochastic dynamical models describing physical and man-made systems can be conveniently cast into the framework of time-delay systems TDs, that is a wide class of retarded or neutral differential equations containing state and/or input delays [81, 93]. It becomes increasingly apparent that delays occur due to various reasons including finite capabilities of information processing among different parts of the system, inherent phenomena like mass transport flow and recycling and/or by product of computational delays [81]. Considerable discussions on delays and their stabilization/destabilization effects in control systems have commanded the interests of numerous investigators in recent years, see [22, 52, 78, 79, 86, 87, 89, 90, 95] and the references cited therein. On another direction of research, robust filtering arose out of the desire to determine estimates of non-measurable state variables for dynamical systems with uncertain parameters. From this perspective, robust filtering can be viewed as an extension of the celebrated Kalman filter to uncertain dynamical systems. The past decade has witnessed major developments in robust filtering problem using various approaches [6, 24, 33, 97, 104]. In the course of filter implementation based on different design algorithms, it turns out that the filters are very sensitive with respect to errors in the filter coefficients [16, 122]. The sources for this include, but not limited to, imprecision in analog-digital conversion, fixed word length, finite resolution instrumentation and numerical roundoff errors. By means of several examples, it is demonstrated in the control design formalism [43] that relatively small perturbations in controller parameters could even destabilize the closed loop system. Such controllers are often termed ”fragile”. Hence, it is considered beneficial that the designed (nominal) controllers should be capable of tolerating some level of controller gain variations. This illuminates the controller fragility problem for which some relevant results are available in [16, 38]. Extension of this effort to the

M.S. Mahmoud: Resilient Control of Uncertain Dynamical Systems, LNCIS 303, pp. 129-169, 2004. © Springer-Verlag Berlin Heidelberg 2004

130

5 Resilient Linear Filtering

problem of robust nonfragile Kalman filtering with respect to estimator gain perturbations was considered in [122]. The objective of this Chapter is to contribute to the further development of resilient linear filtering for a class of uncertain systems. Initially, we consider the robust linear filtering problem and extend the results of [122] to a class of continuous-time systems with norm-bounded uncertainties. We start by describing a class of nonfragile linear filters with additive filter gain perturbations then we formulate the design problem of robust nonfragile filtering as a convex optimization over linear matrix inequalities (LMIs). Further, we establish an important special case which recovers, in the nominal case, previously known results. By a limiting approach, we derive an LMI result on robust nonfragile Kalman filter and finally we show that the case of multiplicative filter gain perturbations can be developed conveniently as an extension of the forgoing results. Later on, we look into the class of time-delay systems with parametric uncertainties belong to a convex bounded polytopic domain. In particular, we focus attention to the resilient L2 − L∞ filter and establish complete results under additive and multiplicative gain perturbations. Specifically, we tackle both the filter design problem from a general viewpoint that conveniently yield both full-order and reduced-order filters. We formulate the design problem of linear resilient filtering as a convex optimization over linear matrix inequalities (LMIs). In the absence of gain perturbations, we develop tractable synthesis conditions. Initially, the delay-independent case is considered. Then we proceed to develop delay-dependent case using the descriptor and extended Newton-Leibniz approaches. Simulation results show that the descriptor approach yields less conservative results for both full-order and reduced-order filters.

5.2 A Class of Uncertain Systems We consider a class of uncertain systems described by: Σ∆ :

x(t) ˙ = [A + ∆A(t)]x(t) + Bw(t) y(t) = [C + ∆C(t)]x(t) + Dw(t)

(5.1) (5.2)

z(t) = Lx(t)

(5.3)

where x(t) ∈ IRn is the system state, y(t) ∈ IRm is the measurement output, z(t) ∈ IRp is the vector of state combination to be estimated, w(t) ∈ IRm is a zero-mean white noise input with unity power spectrum density matrix. The initial condition x(0) is a zero-mean random variable uncorrelated with the input noise w(t) for all t ≥ 0. The matrices A, B, C, D, L are known real constants of appropriate dimensions which describe the nominal system whereas ∆A(t) and ∆C(t) are real, time-varying matrix functions representing the norm-bounded parameter uncertainties where

5.3 The Resilient Filter



∆A(t) ∆C(t)



 =

Ma Na

131

∆a (t) Ha , ∆ta (t) ∆a (t) ≤ I , ∀ t

(5.4)

where Ma ∈ IRn×α , Na ∈ IRp×α and Ha ∈ IRβ×n are known real constant matrices and ∆a (t) ∈ IRα×β is an unknown matrix with Lebesgue measurable elements. The problem of robust linear filtering considered hereafter is to design an estimate zf of z given by zˆ = F · y, where F is a linear operator, such that as t → ∞ it minimizes an upper bound σ(F) of the estimation error variance for all admissible parameter uncertainty. Define the estimation error ∆ e(t) = z(t) − zf (t), then the problem of interest becomes min σ(F) ,

F ∈Cf

sup ||∆||2 ≤I

IE[et (t)e(t)] ≤ σ(F)

(5.5)

where the feasible set Cf represents the set of all linear operators with minimum state-space realization of the form: ˆ(0) = 0 x˙ f (t) = Af xf (t) + Bf w(t) , x zf (t) = Lf xf (t) nf ×nf

nf ×m

(5.6) (5.7)

, Bf ∈ IR and Lf ∈ IR and the where the matrices Af ∈ IR scalar nf > 0 are the design parameters. We will focus on the stationary case where all matrices of the uncertain system as well as the filter are timeinvariant. The solution to the robust linear filtering cited here before is known [24, 33, 81, 97, 104] using different algorithms. One of our prime interests in this Chapter is the issue of filter realization or implementation. It turns out that implementing the designed filter (5.6)(5.7), much in line with the controller design counterpart, one might encounter inaccuracies or parameter perturbations which eventually leads to the fragility problem [16]. In this regard, this problem has been overlooked in the literature. The only exception to this is the problem of robust nonfragile Kalman filtering with respect to estimator gain perturbations that was considered in [122]. In this part, we generalize this result in different ways by considering the robust nonfragile linear filtering with respect to perturbations in both the estimator gain and dynamics matrices. p×nf

5.3 The Resilient Filter This section contains the analytical developments and main results. Initially, we describe a class of linear filters with additive gain perturbations then we formulate the design problem of robust nonfragile filter as a convex optimization over LMIs. Further, we establish an important special case which recovers, in the nominal case, previously known results. By a limiting approach, we derive an LMI result on robust nonfragile Kalman filter and finally we show that the case of multiplicative filter gain perturbations can be developed conveniently as an extension of the forgoing results.

132

5 Resilient Linear Filtering

5.3.1 A Class of Linear Filters In the sequel, to cope with the filtering fragility problem we consider that the feasible set Cf now represents the set of all linear operators with minimum state-space realization of the form: Σf :

x˙ f (t) = [Af + ∆Af (t)]xf (t) + Bf w(t) , zf (t) = [Lf + ∆Lf (t)]xf (t)

xf (0) = 0

where the additive gain perturbations are represented by   ∆Af (t) Mf ∆f (t) Hf , ∆tf (t) ∆f (t) ≤ I , ∀ t = Nf ∆Lf (t)

(5.8) (5.9)

(5.10)

Consider the case of full-order filtering, n = nf , and connecting the filter (5.8)-(5.9) to the system (5.1)-(5.3), we can write: Σ∆f :

where

 ∆

x ˆ(t) =

x(t) xf (t)



ˆx(t) + Bw(t) ˆ x ˆ˙ (t) = Aˆ ˆ x(t) e(t) = Lˆ

ˆ= ∈ 2n , B



B Bf D

(5.11)



ˆ = [L − [Lf + ∆Lf (t)] = [L − Lf ] − Nf ∆f (t)Hf L   A 0 A + ∆A(t) 0 = Aˆ = Bf C Af Bf [C + ∆C(t)] Af + ∆Af (t)    Ma 0 ∆a (t) 0 Ha 0 + Bf Na Mf 0 ∆f (t) 0 Hf

(5.12)

Therefore, the estimation error variance as t → ∞ satisfies ˆ L ˆt] IE[et (t) e(t)] = T r[LIP

(5.13)

for all admissible uncertainties satisfying (5.4) where 0 < IP = IPt ∈ 2n×2n is the solution of the Lyapunov equation ˆ +B ˆB ˆt = 0 IPAˆt + AIP

(5.14)

5.3.2 LMI-Based Design Conditions It follows from [24, 97], using monotonicity property of the Lyapunov equation and invoking Schur complements, that the design problem of robust linear filtering can be cast as:

5.3 The Resilient Filter

min T r[W ]  ˆ + IPAˆt B ˆ AIP < 0 ˆt −I B  ˆt IP IPL > 0 ˆ LIP W

133

(5.15) (5.16)

for all admissible uncertainties satisfying (5.4). Our immediate goal is to transform inequalities (5.15) and (5.16) into LMIs which implies that the foregoing design problem is converted into a convex optimization problem. To achieve our goal, we start by expressing matrix IP in the form   Y R Px Uc −1 , IP = (5.17) IP = Uct Pf Rt Qf where Px , Pf , Y, Qf ∈ n×n are all symmetric and positive definite matrices. Among several relations one can derive from (5.17) and the fact IPIP−1 = I are Px Y +Uc Rt = I and Uct Y +Pf Rt = 0. It then follows from these conditions that Px = Y −1 + Uc Pf Uct from which we observe that for any given symmetric and positive definite matrices Px > Y −1 > 0 and Uc nonsingular, then R must nonsingular and by the Schur complements it is always possible to find Pf > 0. In turn, this assures that IP > 0. This guarantees that the matrix   −1 Px Y 0 (5.18) T =  0 Rt 0  0 0 I is nonsingular. On multiplying (5.15) to the right by T and to the left by T t then overbounding the uncertainties using Fact 1, we get for some , µ > 0.   X A + At X Υa XB X Ma 0  • Υb YB + GD 0 YMa + GNa      < 0 • • −I 0 0     • • • −µI 0 • • • • −µI  1 ≥ 0 (5.19) •µ where Υa = X A + At Y + C t G t + Γ t + (X Ma Mat Y + X Ma Nat G t ) + µHat Ha Υb = YA + At Y + GC + C t G t + µHat Ha (5.20) and X = Px−1 , G = RBf , Γ = RAf Uct X are relaxation variables introduced to insure the linearity of inequality (5.19). The second inequality in (5.19) is incorporated to guarantee well-posedness of the matrix inequalities [25]. In a

134

5 Resilient Linear Filtering

similar way, performing same congruent transformation to inequality (5.16), we obtain for some ε > 0 the following linear matrix inequality:   X X −Z + Lt Rt • Y 0  Lt   (5.21)  • • W − εNf Nft 0  > 0 • • • −εI where Z = Lf Uct X , R = Hf Uc X are relaxation variables introduced again to insure the linearity of inequality (5.21). Therefore, the foregoing analysis leads to the following general result: the robust nonfragile linear filtering design problem of (5.15)-(5.16) with additive gain perturbations is readily equivalent to the following convex programming problem over LMIs: min

W, X , Y, G, Z, R, µ, ε,

T r[W ]

subject to LM Is (5.19) − (5.21) In view of the invertibility of the linear transformation, we immediately obtain the nonfragile filter gain matrices Af = Rc−1 Γ [Uct X ]−1 , Bf = Rc−1 G , Lf = Z[Uct X ]−1

(5.22)

Remark 5.1. One should observe that a necessary condition for the feasibility of inequality (5.19) is that X A + At X < 0 which means that A is a stable matrix. This is an implicit essential requirement of Kalman filtering. Remark 5.2. It is readily seen that in the absence of system uncertainties ∆a ≡ 0, the nonfragile linear filtering design problem reduces to the following convex programming problem over LMIs: min

W,X ,Y,G,Z,R,ε

T r[W ]

   X X −Z t + Lt Rt XB X A + At X Υ¯a • Y Lt 0    • Υ¯b YB + GD  < 0 ,   • • W − εNf Nft 0  > 0 • • −I • • • −εI t t t t t ¯ ¯ Υa = X A + A Y + C G + Γ , Υb = YA + A Y + GC + C t G t 

for which the filter gain matrices are given by (5.22).

∇∇∇

5.3 The Resilient Filter

135

5.3.3 Special Case The next theorem summarizes an important special case: Theorem 5.3. Consider system (5.2)-(5.3) with nonfragile linear filter (5.8)(5.9) subject to uncertainties (5.4) and additive gain perturbations (5.10). The gain matrices of the minimum error variance filter F ∈ Cf defined by Lf = L , Af = −Y

−1

 −1 −1 Γ I −Y X , Bf = −Y −1 G

(5.23)

where 0 < W = W t , 0 < X = X t , 0 < Y = Y t , Γ, G, R, ε, µ,  are the optimal feasible solution of the following convex programming problem over LMIs: min T r[W ]   X X Rt t • Y 0  > 0 , Y L > 0 • W • • −εI   X A + At X Υ a XB  • Υb YB + GD  < 0 • • −I 

(5.24)

Proof: We start by inequality (5.21) and performing some Schur operations, it follows that:     X X −Z t + Lt Rt X −Z t + Lt X Rt t • Y   Lt 0    > 0 ⇐⇒  • W − εNf Nf L 0  > 0 t  • • W − εNf Nf 0  • • Y 0  • • • −εI • • • −εI  −1 −1 t −1 t t t X − X Y X + ε R R −X Y L − Z + L > 0 ⇐⇒ • W − εNf Nft − LY −1 Lt    X X Rt t • Y 0  > 0 , Y L > 0 , Z = L (I − Y −1 X ) (5.25) • W • • −εI On making the selection Uc = Uct = X −1 − Y −1 and using Lf from (5.22) along with Z from (5.25), we get  −1  Lf = Z[Uct X ]−1 = L I − Y −1 X X −1 X −1 − Y −1 = L

(5.26)

136

5 Resilient Linear Filtering

which proves the first part of (5.23). From (5.17), we have Px Y + Uc Rt = I =⇒ R = Rt = −Y which eventually leads to the remaining two parts of (5.23).

∇∇∇

Remark 5.4. It is interesting to observe that Theorem 5.3 has a fewer variables to handle and a simpler LMIs to manipulate than the case of general result derived earlier. This took place on the expense of constraining the choice of the optimal output filter matrix Lf to be Lf = L. 5.3.4 A Limiting Approach to Kalman Filtering Following Remark 5.1, we introduce a matrix 0 < Xˆ = Xˆ t such that Xˆ = ω X , ω > 0 is an arbitrary small scalar and note that inequality (5.19) remains feasible as long as ω → 0+ . Bearing in mind that in Kalman filtering, the gain matrix satisfies [6]: Af = A − Bf C In the present analysis and recalling the results of Theorem 5.3, this requires in the limiting case that  Γ = − YA + GC + +µHat Ha holds for some µ > 0. This leads to the following result: the robust nonfragile Kalman filtering design problem of (5.15)-(5.16) is readily equivalent to the following convex programming problem with respect to Y, G, µ over LMIs: min T r[W ] YA + At Y + GC + C t G t + µHat Ha YB + GD < 0 • −I  Y Lt > 0 • W



(5.27)

The nonfragile filter gain matrices are given by Af = A − Bf C , Bf = Y −1 G , Lf = L

(5.28)

Remark 5.5. The significance of the foregoing result stems from the fact that by enforcing the regularity conditions B Dt = 0, D Dt = I and further ∆ substituting S = Y −1 and G = −C t , the robust nonfragile Kalman filtering problem after some Schur complements operations becomes min



S>0,µ>0

T r[L S Lt ]

 SAt + AS + µHat Ha SC t B  • I 0  < 0 • • −I

(5.29)

5.3 The Resilient Filter

137

The nonfragile filter gain matrices are now given by Af = A − Bf C , Bf = S C t , Lf = L

(5.30)

5.3.5 Multiplicative Gain Perturbations Here we consider the feasible set Cf to represent the set of all linear operators with minimum state-space realization of the form with multiplicative gain perturbations: x˙ f (t) = Af [I + ∆Af (t)]xf (t) + Bf w(t) , zf (t) = Lf [I + ∆Lf (t)]xf (t)

xf (0) = 0

(5.31) (5.32)

where the matrices Af ∈ IRnf ×nf , Bf ∈ IRnf ×m and Lf ∈ IRp×nf and the scalar nf > 0 are the design parameters and ∆Af (t), ∆Lf (t) are given by (5.10). The results pertaining to this case can be derived from the foregoing ¯f = Lf Nf ¯ f = Af Mf , N design results by simply defining two matrices M to replace Mf and Nf , respectively. Therefore, we summarize the following results: the robust nonfragile linear filtering design problem of (5.15)-(5.16) with multiplicative gain perturbations corresponds to the following convex programming problem over LMIs: min



W, X , Y, G, Z, R, K, µ, ε,

X A + At X Υ a  • Υb   • •   • • • •  X X −Z + Lt • Y Lt   • • W − εK • • •

T r[W ]

 XB X Ma 0  YB + GD 0 YMa + GNa    < 0 , 1 −I 0 0 ≥ 0  •µ  • −µI 0 • • −µI  t R 0   > 0 0  −εI

¯f N ¯ t is a linearizing matrix variable. The filter gain matrices where K = N f are given by (5.22). In a similar way, in the absence of system uncertainties ∆a ≡ 0, the nonfragile filtering design problem with multiplicative gain perturbations reduces to: min    X X A + At X Υ¯a XB •  • Υ¯b YB + GD  < 0 ,  • • • −I •

W,X ,Y,G,Z,R,K,ε

T r[W ]

 X −Z t + Lt Rt Y Lt 0   > 0 • W − εK 0  • • −εI

138

5 Resilient Linear Filtering

5.3.6 Example 5.1 In order to illustrate the theoretical results of this paper, we consider a pilotscale model of a stirred-tank heater which can fall into the type (5.1)-(5.3) with the following data at a particular operating point:    −0.4 0.3 10 0.15 0 A= , C= , B= , Ha = [0.3 0.3] 3 −4.5 01 0 0.25    0.2 0.3 0.6 0 Ma = , Na = , L= 0.1 −0.2 0 0.4 In the implementation of the linear filter, we consider that the gain perturbations have   0.6 −0.3 Mf = , Nf = , Hf = [0.2 0.4] 0.4 0.5 Numerical computations of the robust nonfragile linear filter show that    0 1 0.201 0 0.598 0.01 Af = , Bf = , Lf = −0.524 −5.132 −0.02 0.245 −0.2 0.4 T r[W ] = 3.1246 , ε = 1.2097 , µ = 2.3416 ,  = 0.8412 On the other hand, application of Theorem 5.3 yields   0 1 0.223 −0.03 Af = , Bf = −0.511 −5.723 0.01 0.264 T r[W ] = 5.0634 , ε = 3.4066 , µ = 3.4615 ,  = 0.4436 The higher value of the minimum upper bound to the error variance is due to the constrained choice of Lf = L. 5.3.7 Example 5.2 As another example, we consider a single-reach model for stream water-quality which can fall into the type (5.1)-(5.3) with the following parameters:     0 1.01 0.49 1 A =  −5.01 −0.02 0  , C t =  1  1.51 0 −0.1 −2   00 B =  1 0  , Ha = [0.3 0.3 0.5] 10   0.2 Ma =  0.1  , Na = 0.6 0.7   1.2 Lt =  0.8  −2

5.4 Continuous Polytopic Systems

139

In the implementation of the linear filter, we consider that the gain perturbations have     0.6 −0.3 Mf =  0.4  , Nf =  0.5  , Hf = [0.2 0.4 0.4] 0.2 −0.2 Numerical computations of the robust nonfragile linear filter show that     0 0.1 1 0.201 0 0 0.65  , Bf =  −0.02 0.245  Af =  −0.1 −1.337 −1.196 −6.025 0 0.311   0.911 Ltf =  −0.04  0.63 T r[W ] = 2.9835 , ε = 3.7625 , µ = 4.6115 ,  = 1.0245 On the other hand, application of Theorem 5.3 yields     0 1 −0.05 0.311 0 0 −0.2 0.75  , Bf =  −0.03 0.264  Af =  −1.4011 −1.201 −5.986 0.01 0.401 T r[W ] = 4.9711 , ε = 2.6576 , µ = 3.6588 ,  = 1.4016 Once again, the higher value of the minimum upper bound to the error variance is due to the constrained choice of Lf = L. In the remaining part of the Chapter, we shift attention to the resilient linear filtering problem for a class of uncertain time-delay systems with parametric uncertainties belong to a convex bounded polytopic domain. In particular, we focus attention to the resilient L2 − L∞ filter and establish complete results under additive and multiplicative gain perturbations.

5.4 Continuous Polytopic Systems Consider a class of uncertain continuous time-delay systems described by (ΣJ ) :

x(t) ˙ = Ao x(t) + Ad x(t − τ ) + Γ w(t) y(t) = Co x(t) + Φw(t) z(t) = Do x(t)

(5.33) (5.34) (5.35)

where x ∈ IRn is the system state, y(t) ∈ IRp is the measured output, z(t) ∈ IRr is the controlled output and τ is a delay factor which is an unknown and bounded time-varying function such that 0 < τ ≤ β , 0 < τ˙ ≤ α < 1. The corresponding state-space matrices contain uncertainties represented by a real convex bounded polytopic model of the type

140

5 Resilient Linear Filtering

[Ao , =

N 

Ad ,

∆

Co , Do ] ∈ Sλ = [A(oλ) ,

λj [Aoj ,

A(dλ) , C(oλ) ,

Adj , Coj , Doj ], λ ∈ Λ

D(oλ) ] (5.36)

j=1

where Λ is the unit simplex   N  ∆ Λ = (λ1 , · · · , λN ) : λ j = 1 , λj ≥ 0

(5.37)

j=1

Define the vertex set IN = {1, ..., N }. Throughout the paper, we use {Ao ,· · ·, Fo } to imply generic system matrices and {Aoj , · · · , Foj , j ∈ IN} to represent the respective values at the vertices. The initial condition is specified as x(0), x(s) = xo , φ(s) , where φ(.) ∈ L2 [−τ, 0]. Consider the problem of designing a filter of order k x˙ f (t) = Af xf (t) + Bf y(t) zf = Cf xf (t)

(Σf o ) :

(5.38)

where the matrices Af ∈ k×k , Bf ∈ k×p and Cf ∈ r×k are the unknown (nominal) filter parameters to be determined subject to some prescribed criteria. Note that model (5.38) is quite general in the sense that we consider k = n for full-order and 1 ≤ k ≤ n for reduced-order filter. In the literature, there have been several methods to cope with this problem [78, 79, 86, 87, 88, 89, 90]. In practice it turns out however that these gains cannot be implemented exactly or can be tolerated for fine tuning purposes thereby leading to the so-called fragility (resilience) problem [16, 43]. It is well known that this problem is closely related to the issue of performance deterioration. In the remainder of this paper, we closely investigate the resilience problem in filtering design of time-delay systems

5.5 The Resilient Filtering Problem To address the resilience problem in filtering design and related issues, we consider the filter (5.38) to be expressed as (Σf a ) :

x˙ f a (t) = [Af + ∆Af a (t)] xf a (t) + [Bf + ∆Bf a (t)] y(t) zf a = [Cf + ∆Cf a (t)] xf a (t) (5.39)

where ∆Af a (t), ∆Bf a (t), ∆Cf a (t) are additive gain perturbations represented by ∆Af a (t) = Ma ∆f a (t)Na , ∆Bf a (t) = Ma ∆f a (t)Nb ∆Cf a (t) = Mc ∆f a (t)Nc ∆

(5.40)

∆f a (t) ∈ ∆FA (t) = {∆f a (t) : ∆tf a (t)∆f a (t) ≤ I} ⊂ α×β (5.41)

5.5 The Resilient Filtering Problem

141

Fig. 5.1. The resilience filtering system

A schematic diagram of the resilience filtering problem is depicted in Figure 5.1. Now by combining systems (Σ∆ ) and (Σf a ), we get the augmented system ˙ = Aξ∆ ξ(t) + Aξd ξ(t − τ ) + Γξ w(t) (ΣAa ) : ξ(t) ∆

¯ ξ(t) zξ (t) = ze (t) = L where

 ξ(t) = 

x(t) xf a (t)

(5.42)

∈ n+k×n+k , Aξ∆ = Aξo + ∆Aξ

Ao 0 Bf Co Af   Ad Ad 0 [I 0] = A¯d E = Aξd = 0 0 0  0 0 ¯ = [Do − Cf − ∆Cf a ] = L ¯ o + ∆L ¯ ∆Aξ = , L ∆Bf Co ∆Af ¯ = [0 − ∆Cf a ] ¯ o = [Do − Cf ] , ∆L L   Γ Γ Γξ = = Bf Φ (Bf + ∆Bf a )Φ  0 + = Γo + ∆Γ (5.43) ∆Bf a Φ Aξo =

The design objective is to develop a tractable synthesis procedure for selecting the gain matrices Af , Bf and Cf such that for all admissible uncertainties over the vertex IN and gain perurbations ∆Af (t), ∆Bf (t), ∆Cf (t) system (ΣA ) is robustly exponentially stable and ||zξ ||22 < γ 2 ||w||22 .

142

5 Resilient Linear Filtering

5.5.1 Stability Results To study the stability of the foregoing augmented system, we introduce the Lyapunov-Krasovskii functional

t t V (ξ) = ξ (t)Pξ(t) + ξ t (s)E t QEξ(s) ds (5.44) t−τ

where 0 < P = P t , 0 < Q = Qt are weighting matrices. Evaluation of V˙ (ξ, j) along the solutions of (5.42) with (w ≡ 0) and completing the squares, it yields:   V˙ (ξ) = ξ t (t) Atξ∆ P + PAξ∆ + E t QE ξ(t) + 2ξ t P A¯d Eξ(t − τ )  ¯ − ξ t (t − τ )(1 − τ˙ )E t QEξ(t − τ ) ≤ ξ t (t) Atξ∆ P + PAξ∆ + Q  ∆ ˆ −1 A¯td P ξ(t) = + P A¯d Q ξ t (t)Ξξ(t) (5.45) ˆ = (1 − α)Q, Q ¯ = E t QE. where Q The following theorem establishes an LMI-based sufficient condition for the robust exponential stability of the closed-loop system (ΣAa ). Theorem 5.6. Consider system (ΣAa ) and given matrices 0 < Q = Qt ∈ n×n . If there exists a matrix 0 < P = P t ∈ n+k×n+k such that the inequalities  ¯ P A¯d PΓξ  PAξ∆ + Atξ∆ P + Q  ˆ 0  < 0 (5.46) • −Q • • −I  t ¯ P L > 0 (5.47) • γ2I has a feasible solution for all admissible uncertainties over the vertex IN and for all gain perturbations ∆f a (t) ∈ ∆FA (t), then the filtering-error system (5.42) is exponentially stable with an L2 − L∞ disturbance attenuation level γ>0 Proof: Given the Lyapunov functional V (ξ) and its derivative V˙ (ξ), it follows that from (5.46) and the fact that PΓξt Γξ P > 0, it follows that Ξ < 0 is guaranteed and we conclude that V˙ (ξ) < 0 for all ξ = 0 and V˙ (ξ) ≤ 0 for all ξ. Since ||ξ(t+β)|| ≤ ϕ||ξ(t)|, ∀β ∈ [−τ, 0] and some ϕ > 0 [55], it follows from  (5.44) that V (ξ) ≤ ξ t (t)Pξ(t) + µ||ξ||2 where µ = ϕ τ λM [P] + λM [E t QE] . Therefore, for all ξ = 0, we have   ξ t Ξξ V˙ (ξ) λm [−Ξ] ∆ ≤ t (5.48) ≤ − σ = − min V (ξ) ξ Pξ + µ||ξ||2 j∈IN λM [P] + µ

5.5 The Resilient Filtering Problem

143

It is readily seen from (5.48) that σ > 0 and hence we get V˙ (ξ) ≤ −σ V (ξ), which implies that V (ξ) ≤ e−σ t V (ξo )

(5.49)

Therefore system (5.39) is exponentially stable at a rate σ > 0. To establish the L2 −L∞ performance we consider that ξo ≡ ξ(t = 0) = 0 thereby leading to V (ξo ) = 0 . Now, consider the performance index

t ∆ J = V (ξ) − wt (φ) w(φ) d φ 0

For any 0 = w(t) ∈ L2 [0, ∞), t ≥ 0, it follows that

t wt (φ) w(φ) d φ J = V (ξ) − V (ξo ) − 0

t

t  t ˙ V (ξ, j) − w (φ) w(φ) d φ ≤ ϑt (φ) Υ ϑ(φ) d φ = 0

where

(5.50)

0



 ξ t (φ) ϑ(φ) =  ξ t (φ − τ )  w(φ)   PA∆ξ + At∆ξ P + E t QE P A¯d PΓξ Υ = • −Q 0  • • −I

(5.51)

Obviously (5.46) guarantees that J < 0 over the entire domain Sλ . Furthermore, we have

t ξ t (t) P ξ(t) ≤ V (ξ) < wt (φ) w(φ) d φ (5.52) 0

¯tL ¯ < γ 2 P which in turn, for all t ≥ 0 In view of (5.47), it is readily seen that L leads to ¯tL ¯ ξ(t) < γ 2 ξ t (t)P ξ(t) zet (t) ze (t) = ξ t (t) L

t

wt (φ) w(φ) d φ ≤ γ 2 < γ2 0

∞

wt (φ) w(φ) d φ (5.53)

0

||ze (t)||2∞

< γ 2 ||w(t)||2∞ ∀ 0 = By taking supremum over t ≥ 0, we have ∇∇∇ w(t) ∈ L2 [0, ∞) and therefore the proof is completed. 5.5.2 Delay-Independent Filter Synthesis On considering (5.46)-(5.47) in view of (5.43) and applying Fact 1, it follows that

144

5 Resilient Linear Filtering

¯ P A¯d PΓξ   PAξo + At P + Q ¯ P A¯d PΓo  PAξ∆ + Atξ∆ P + Q ξo  ˆ 0 = ˆ 0 + • −Q • −Q • • −I • • −I     t ¯ Π P Ad PΓo P∆Aξ + ∆Aξ P 0 P∆Γ   < 0 (5.54) ˆ • 0 0  ≤  • −Q 0 t t • • 0 • • −I + ε2 Φ Nb Nb Φ 

and in a similar way   ¯t ¯t ¯ ¯ to P L L P − ε−1 3 Nc Nc ≥ ¯ cM ¯t > 0 • γ2I • γ 2 I − ε3 M c

(5.55)

where ¯ + ε−1 P M¯a M ¯bt N ¯b ¯ at P + ε−1 P M¯a M ¯ at P + ε1 N Π = PAξo + Atξo P + Q 2 1 ¯b = [Nb Co Na ] ¯ at = [0 Mat ] , N M ¯ ¯ Nc = [0 Nc ] , Mct = [0 Mct ] (5.56) For simplicity in exposition, we introduce the following matrices   W RΛ I 0 n×n k×k P= , W∈ , V =∈  , T= • V 0 V −1 Λt Ga = RΛV −1 Λt , Gc = ΛMa Gb = Na VΛt , Λ ∈ k×k Rt = [Ik 0k×(n−k) ] ∈ IRk×n , Θa = [RGc , RGc , RGc , RGc ] Θb = [ε1 I, ε2 I, ε1 I, ε2 I] (5.57) The existence of resilient L2 −L∞ filter is established by the following theorem. Theorem 5.7. Consider system (ΣAa ) and given a matrix 0 < Q = Qt ∈ n×n and a scalar γ > 0. A kth−order resilient L2 − L∞ filter exists if there ˆf , Cˆf , Ga , Gb , Gc , exist matrices 0 < W = W t , 0 < V = V t , 0 < Λ, Aˆf , B Θa , Θb , and scalars ε1 > 0, ε2 > 0, ε3 > 0,  > 0, µ > 0 such that the LMIs   WAoj + Atoj W t ˆ RAf + Aoj Ga   ˆf Coj + WΓ + +RB   t ˆ WAdj Θ +Coj Bf + a   t ˆt t ˆf Φ RB Coj Bf R   t t ε C N G   1 b t oj b Nbt Nb Coj + Q   +ε1 Coj     Aˆtf + Aˆf t   G Γ + a t t  < 0(5.58)  • G A 0 G G + +ε dj 1 b a b ˆf Φ   B t t   Gc Gc + µGc Gc   ˆ   • • − Q 0 0     −I+  0  • • •   ε2 Nb Nbt • • • • −Θb

5.5 The Resilient Filtering Problem

 

µ 1 • ε1



 ≥ 0 ,

 1 • ε2

145

≥ 0

 t  W − ε−1 Dot 3 Nc Nc R Λ W Nct   > 0 , • V −Cft > 0 • ε3 I • • γ 2 I − ε3 Mct Mc

(5.59)

(5.60)

have a feasible solution for all admissible uncertainties over the vertex IN and for all gain perturbations ∆f a (t) ∈ ∆FA (t). Moreover, the filter matrices are given ˆf , Cf = Cˆf Af = Λ−1 Aˆf Λ−t V , Bf = Λ−1 B

(5.61)

Proof: It follows from (5.60) and the Schur complements for some ε3 > 0 that 0 < W = W t ∈ IRn×n , 0 < V = V t ∈ IRk×k and more importantly t −1 t t W − ε−1 Λ R > 0. Observe that P as constructed in (5.57) is 3 Nc Nc − RΛV positive-definite since W > 0, V > 0 and W − RΛV −1 Λt Rt > 0. Introducing the following change of variables:   −1   −t ˆf Aˆf B Af Bf Λ 0 Λ V0 (5.62) = Cf 0 0 I 0 I Cˆf 0 On performing the congruent transformation C I a = [T, I, I], C I b = [T, I] to the last inequalities of (5.54) and (5.55), respectively, and using (5.56), we obtain   Πs Π d WAd WΓ + RΛBf Φ  • Πf ΛV −1 Λt Rt Ad ΛV −1 Λt Rt Γ + ΛBf Φ    < 0 (5.63)  • •  ˆ −Q 0 • • • −I + ε2 Nbt Nb   t W − ε−1 Dot 3 Nc Nc R Λ   < 0 • −V −Cft (5.64) 2 t • • γ I − ε 3 Mc M c where for j ∈ IN t Πs = WAoj + Atoj W + RΛBf Coj + Coj Bft Λt Rt + Q −1 t t t t t t t + ε1 Coj Nbt Nb Coj + ε−1 1 RΛMa Ma Λ R + ε2 RΛMa Ma Λ R −1 t t t t Πf = ΛAf V −1 Λt + ΛV −1 Atf Λt + ε−1 1 ΛMa Ma Λ + ε2 ΛMa Ma Λ

+ ε1 ΛV −1 Nat Na V −1 Λt t t Πd = RΛAf V −1 Λt + Atoj RΛV −1 Λt + Coj Bft Λt + ε1 Coj Nbt Na V −1 Λt −1 t t t t + ε−1 1 RΛMa Ma Λ + ε2 RΛMa Ma Λ

(5.65)

By using the linearizations (5.57), invoking Schur complements and performing standard matrix manipulations, we cast the matrix inequalities (5.63)(5.64) with (5.65) into the LMIs (5.58) and (5.60) subject to the constraints

146

5 Resilient Linear Filtering

 ε2 = 1, µ ε1 = 1. This is readily expressed by (5.59). Therefore, in view of Theorem 5.6, we conclude that the filter (5.39)-(5.41) with realization defined by (5.61) guarantees that the filtering error system (5.42) has an L2 − L∞ disturbance attenuation level bounded by γ. This completes the proof of the theorem. ∇∇∇ Interestingly enough, in the absence of gain perturbations (∆Af a (t) ≡ 0, ∆Bf a (t) ≡ 0, ∆Cf a (t) ≡ 0), we have the nominal augmented system (ΣAao ) :

˙ = Aξ0 ξ(t) + A¯d Eξ(t − τ ) + Γo w(t) ξ(t) ∆ ¯ o ξ(t) zξ (t) = ze (t) = L

(5.66)

for which the following result holds: Corollary 5.8. Consider the nominal augmented system (ΣAao ) and given a matrix 0 < Q = Qt ∈ n×n and a scalar γ > 0. A kth−order robust L2 − L∞ ¯f , C¯f , , Ga filter exists if there exist matrices 0 < W = W t , 0 < V, A¯f , B such that the LMIs   WAoj + Atoj W ¯f + At Ga R A WΓ +  oj ¯f Coj + C t B ¯t t  +RB WAdj oj f R t ˆ ¯f Φ   R B +C B f oj   +Q   t   < 0 G Γ + (5.67) t t a ¯ ¯   Ga Adj ¯ • A f + Af   B Φ f    ˆ • • −Q 0  • • • −I   t  W R Λ Do W Nct  • V −Cft  > 0 , > 0 (5.68) • ε3 I • • γ2I have a feasible solution for all admissible uncertainties over the vertex IN. Moreover, the filter matrices are given by ¯f , Cf = C¯f Af = Λ−1 A¯f Λ−t V , Bf = Λ−1 B

(5.69)

5.5.3 Multiplicative Gain Perturbations Had we considered the filter perturbation matrices to be gain dependent, an alternative representation of the resilient filter would then be (Σf m ) :

x˙ f m (t) = [I + ∆Af m (t)]Af xf m (t) + [I + ∆Bf m (t)]Bf y(t) zf m = [I + ∆Cf m (t)]Cf xf m (t) (5.70)

where ∆Af m (t), ∆Bf m (t), ∆Cf m (t) are multiplicative gain perturbations represented by

5.5 The Resilient Filtering Problem

147

∆Af m (t) = Md ∆f m (t)Ne , ∆Bf m (t) = Md ∆f m (t)Nf ∆Cf m (t) = Mh ∆f m (t)Ng ∆

∆f m (t) ∈ ∆FM (t) = {∆f m (t) : ∆tf m (t)∆f m (t) ≤ I} ⊂ α×β (5.71) Augmenting systems (Σ∆ ) and (Σf m ) yields: (ΣAm ) :

˙ = Aζ∆ ζ(t) + Aζd ζ(t − τ ) + Γζ w(t) ζ(t) ∆ ¯ ζ(t) zζ (t) = ze (t) = H

where



x(t) xf m (t)

(5.72)



∈ 2n×2n , Aζ∆ = Aζo + ∆Aζ    Ad Ao 0 Ad 0 , Aζd = = [I 0] = A¯d E Aζo = Bf Co Af 0 0 0   0 0 Γ ∆Aζ = , Γζ = = Γζo + ∆Γ ∆Bf m Bf Co ∆Af m Af (I + ∆Bf m )Bf Φ   0 Γ Γζo = , ∆Γ = ∆Bf m Bf Φ Bf Φ ¯ o + ∆H ¯ ¯ = [Do − [I + ∆Cf m ]Cf ] = H H ¯ = [0 − ∆Cf m Cf ] ¯ o = [Do − Cf ] , ∆H (5.73) H ζ(t) =

By similarity to the case of system (ΣAa ), the following results are easily established for system (ΣAm ): Theorem 5.9. Consider system (ΣAm ) and given matrices 0 < Q = Qt ∈ n×n . If there exists a matrix 0 < P = P t ∈ n+k×n+k such that the inequality  ¯ P A¯d PΓζ  PAζ∆ + Atζ∆ P + Q  ˆ 0  < 0 (5.74) • −Q • • −I  t ¯ P H > 0 (5.75) • γ2I has a feasible solution for all admissible uncertainties over the vertex IN and for all gain perturbations ∆f a (t) ∈ ∆FM (t), then the filtering-error system (5.72) is exponentially stable with an L2 − L∞ disturbance attenuation level γ>0 Likewise, the existence of resilient L2 − L∞ filter in case of multiplicative gain perturbations is established by the following theorem. Theorem 5.10. Consider system (ΣAm ) and given a matrix 0 < Q = Qt ∈ n×n and a scalar γ > 0. A kth− order resilient L2 − L∞ filter exists if

148

5 Resilient Linear Filtering

there exist matrices Θa , Θb , Ψa , Ψb , Ψc , such that the LMIs  Σm Σa WAdj   • Σb Gta Adj    • • −Q ˆ    • • • 

0 < W = W t , 0 < V = V t , Aˆf , , Cˆf , Λ, Ga , Ψe and scalars ε1 > 0, ε2 > 0, ε3 > 0,  > 0, µ > 0

 Θa Σc  Gta Γ + 0  ˆ  Bf Φ  0 0   < 0  −I+ 0  t  ε2 Ψb Ψb • • • • −Θb   µ 1  1 ≥ 0 , ≥ 0 • ε1 • ε2   t  W − ε−1 Dot 3 Ψc Ψc R Λ W Ψct   > 0 , • V −Cft > 0 • ε3 I • • γ 2 I − ε3 Mht Mh

(5.76)

(5.77)

(5.78)

have a feasible solution for all admissible uncertainties over the vertex IN and for all gain perturbations ∆f a (t) ∈ ∆FM (t). Moreover, the filter matrices are given ˆf , Cˆf = Cf Af = Λ−1 Aˆf Λ−t V , Bf = Λ−1 B

(5.79)

where ˆf Coj + C t B ˆt t Σm = WAoj + Atoj W + RB oj f R t + ε1 Coj Ψbt Ψb Coj + Q t ˆ t Σa = RAˆf + Atoj Ga + Coj Ψet Bf + ε1 Coj ˆf Φ Σc = WΓ + RB

Σb = Aˆtf + Aˆf + ε1 Gtb Gb + Gd Gtd + µGd Gtd

(5.80)

Proof: Follows from a parallel development of Theorem 5.10 using the linearizations Gd = Λ Md , Ψa = Bft Nft Ne Af VΛt , Ψb = Nf Bf , Ψc = Ng Cf , Ψe = ∇∇∇ ΛVNat Nf Bf . Remark 5.11. It is crucial to note that the approach developed here before provides a tractable procedure of resilient filter design for polytopic state-delay systems since the case of resilient full-order filter can be readily derived by setting R ≡ In . Interestingly enough, this has been enabled by the significant employment of linearization procedures in Theorem 5.7 to Theorem 5.10 and carried out in the same way in all subsequent theorems. By adopting the approach of [90], the ensuing resilient results can be extended to multiple statedelays. In the absence of gain perturbations, it yields a filter structure, which is basically quite different from the results of [27].

5.6 Resilient Delay-Dependent Filtering

149

Remark 5.12. When solving for the resilient filter matrices in the case of additive gain perturbations, it is possible to cast the scalar γ as one of the LMI variables in which case we have to solve the following convex optimization problem: M inimize π = γ 2 ˆf , Cˆf , Ga , Gb , Gc , Θa , Θb over W, V, Λ, Aˆf , B subject to(5.58) and (5.60) √ The minimum attainable value will be γ ∗ = π ∗ . A similar convex optimization problem can be also formulated for multiplicative gain perturbations case.

5.6 Resilient Delay-Dependent Filtering In order to exhibit the delay-dependence dynamics in the filter design procedure, one can utilize different methods [81], each of which has merits and demerits. In the sequel, we are going to develop two methods to resilient delaydependent filtering. The first method is based on the descriptor transformation approach [22] and the resulting design conditions belong to the class of strong delay-dependent [93] since they require the availability of the time-delay factor at every time instant. The second uses the conventional Newton-Leibniz method in addition to some appropriate relaxation variables [95] and since the results require bounds on the time-delay factor and its derivative, the filter design is weak delay-dependent [93]. 5.6.1 A Descriptor Approach Unlike several transformations [93] for time-delay systems, it has been reported that the descriptor approach leads to a system which is equivalent to the original time-delay system, it does not essentially depend on additional assumptions on stability of the transformed system and requires bounding fewer cross-terms. Thus it is expected that it yields least conservative results [22]. In the discussions to follow, we initially focus on the additive gain perturbations case. Thus, we rewrite system (ΣAa ) into the form: ˙ = σ(t) (ΣD ) : ξ(t) 0 = −σ(t) + [Aξ∆ + Aξd ]ξ(t) + Γξ w(t)

t σ(s)ds − Aξd

(5.81)

t−τ (t)

∆ ¯ ξ(t) zξ (t) = ze (t) = L

(5.82)

where the variable σ(t) is treated as fast variable, that is σ˙ = 0. Proceeding further, we introduce the matrices

150

5 Resilient Linear Filtering

   0 0 I0 IPξ 0 ˜ ¯ , Q= P = U IP , U = ; IP = 0 Qσ IPd IPσ 00   I 0 ∆ ¯ ˜= E1 = , E2 = , L [L 0] 0 I

(5.83)

Let the Lyapunov-Krasovskii functional for system (ΣDa ) be selected as: ∆

Vt (η) = Vξ (ξ) + Vσ (ξ)



Vξ (ξ) = ξ t (t)IPξ ξ(t) , Vσ (ξ) =



0

−τ (t)

t

σ t (s) Qσ σ(s) ds d θ (5.84) t+θ

Using (5.82) and (5.83), we get:  ˙ ˙ = 2ξ t (t)IPt ξ(t) ˙ = 2η t IPt ξ(t) ∂Vξ /∂ξ ξ(t) ξ 0  σ(t) = 2[ξ t (t) σ t (t)] IPt −σ(t) + [Aξ∆ + Aξd ]ξ(t)  0 t t t  + 2[ξ (t) σ (t)] IP t −Aξd t−τ σ(s)ds Simple manipulations using (5.82) yield:  σ(t) 2[ξ t (t) σ t (t)] IPt = −σ(t) + [Aξ∆ + Aξd ]ξ(t)   ξ(t) [ξ t (t) σ t (t)] IPt [Aξ∆ + Aξd ] + [Aξ∆ + Aξd ]t IP σ(t) Using Fact 1, it follows that  t t t 2[ξ (t) σ (t)] IP

t−τ

(5.86)

σ(s)ds

t

[ξ t (t) σ t (t)] IPt A¯d σ(s)ds

=2 t−τ



≤ τ

−Aξd

0 t

(5.85)

ξ(t) σ(t)

t

 t ξ(t) −1 ¯t ¯ IP Ad Qσ Ad IP + σ t (s)Qσ σ(s) ds (5.87) σ(t) t−τ t

Therefore from (5.85)-(5.87) we get:  t   ξ(t) ξ(t) t t ¯ t −1 ¯t ˙ IP [Aξ∆ + Aξd ] + [Aξ∆ + Aξd ] IP + τ IP Ad Qσ Ad IP Vξ ≤ σ(t) σ(t)

t σ t (s)Qσ σ(s) ds (5.88) + t−τ

It is straightforward to show that,

5.6 Resilient Delay-Dependent Filtering

V˙ σ = τ σ t (t)Qσ σ(t) −

0

−τ

σ t (t + θ)Qσ σ(t + θ) dθ

151

(5.89)

By using (5.82)-(5.84) and (5.88), it follows that V˙ t = V˙ ξ + V˙ σ  t  x(t) IPt (Aξ∆ + Aξd ) + (Aξ∆ + Aξd )t IP = σ(t)  x(t) t ¯ −1 ¯t + τ IP Ad Qσ Ad IP + τ Q σ(t)

(5.90)

Letting Aξ∆d = Aξ∆ + Aξd   0 0 Ao + Ad 0 + = ∆Bf Co ∆Af Bf Co Af

(5.91)

Then in line of Theorem 5.6, we have the following stability result Theorem 5.13. Consider system (ΣD ) subject to additive gain perturbations (5.41) and given a matrix 0 < Qσ = Qtσ ∈ n×n . If there exists a matrix 0 < IP = IPt ∈ n+k×n+k and scalars ε1 > 0, ε2 > 0, ε3 > 0 such that the inequality   ˜ τ IPt A¯d ΠD + Q IPt Γξ   < 0 (5.92) • −τ Qσ 0 • • −I + ε2 Φt Nbt Nb Φ  t ¯t ¯ ˜t IP − ε−1 L 3 Nc Nc (5.93) ¯ cM ¯ ct > 0 • γ 2 I − ε3 M t ¯ ¯t ΠD = IPt [Ao + Ad ] + [Ao + Ad ]t IP + ε−1 1 IP Ma Ma IP + t ¯ ¯t ¯t ¯ ε−1 2 IP Ma Ma IP + ε1 Nb Nb

(5.94)

has a feasible solution for all admissible uncertainties over the vertex IN and for all gain perturbations ∆f a (t) ∈ ∆FA (t), then the filtering-error system (5.81)-(5.82) is exponentially stable with an L2 − L∞ disturbance attenuation level γ > 0 Proof: Inequality (5.92) is obtained from applying the Schur complements to the sufficient stability condition V˙ t < 0 followed by bounding the result via inequality I.1 of the Appendix. Inequality (5.94) results from the L2 − L∞ performance consideration followed by bounding the result via inequality I.1 as well. ∇∇∇ The existence of resilient delay-dependent L2 − L∞ filter with additive gain perturbations is established by the following theorem.

152

5 Resilient Linear Filtering

Theorem 5.14. Consider system (ΣD ) subject to additive gain perturbations (5.41) and given a matrix 0 < Qσ = Qtσ ∈ n×n and a scalar γ > 0. A kth−order resilient L2 − L∞ filter exists if there exist matrices 0 < IPξ = ˜f , C˜f , Z, Y and scalars IPtξ ∈ n×n , 0 < IPσ = IPtσ ∈ k×k , IPd , A˜f , B ε1 > 0, ε2 > 0, ε3 > 0 such that the LMIs   Ωmj Ωaj IPtd Ma IPtd Ma IPtξ Adj IPtξ Γ + YΦ   A˜tf + A˜f  •  ˜t Φ 0 0 0 B f t   +ε1 Na Na + Qσ    •  0 0 0 • −ε1 I   < 0 (5.95)  •  0 0 • • −ε2 I    •  0 • • • −Qσ    −I+  • • • 0 0 ε2 Nb Nbt   t  IPξ − ε−1 Dot 3 Nc Nc IPd IPξ Nct t   • IPσ −Cf > 0 , > 0 (5.96) • ε3 I • • γ 2 I − ε3 Mct Mc have a feasible solution for all admissible uncertainties over the vertex IN. Moreover, the filter matrices are given by −t ˜ ˜ ˜ (5.97) Af = IP−t σ Af , Bf = IPσ Bf , Cf = Cf where Ωmj = IPtξ (Aoj + Adj ) + (Atoj + Atdj )IPξ + Cot Y t Ωaj

t + YCo + ε1 Coj Nbt Nb Coj t ˜ t = Coj Nbt Na Bf + Z + ε1 Coj

(5.98)

Proof: Follows directly from Theorem 5.13 using (5.83) and applying the ∇∇∇ linearizations Y = IPtd Bf , Z = IPtd Af . Had we considered the case of multiplicative gain perturbations, we would had arrived at the following results Theorem 5.15. Consider system (ΣD ) subject to multiplicative gain perturbations (5.71) and given a matrix 0 < Qσ = Qtσ ∈ n×n . If there exists a matrix 0 < IP = IPt ∈ n+k×n+k and scalars ε1 > 0, ε2 > 0, ε3 > 0 such that the inequality   ˜ τ IPt A¯d IPt Γξ ΠE + Q  < 0  (5.99) • −τ Qσ 0 t t t • • −I + ε2 Φ Bf Nf Nf Bf Φ  ˜t ˜ ˜t IPξ − ε−1 L 3 Nc Nc (5.100) ˜ hM ˜t > 0 • γ 2 I − ε3 M h

t ¯ ¯t ΠE = IPt [Ao + Ad ] + [Ao + Ad ]t IP + ε−1 1 IP Md Md IP t ¯ ¯t ˜t ˜ +ε−1 2 IP Md Md IP + ε1 Nb Nb ˜c = [0 Ng Cf ] , M ˜ t = [0 M t ] (5.101) ˜b = [Nf Bf Co Ne Af ] , N N h h

5.6 Resilient Delay-Dependent Filtering

153

has a feasible solution for all admissible uncertainties over the vertex IN and for all gain perturbations ∆f a (t) ∈ ∆FM (t), then the filtering-error system (5.81)-(5.82) is exponentially stable with an L2 − L∞ disturbance attenuation level γ > 0 Proof: Follows from parallel development to Theorem 5.13. ∇∇∇ The existence of resilient delay-dependent L2 − L∞ filter is established by the following theorem. Theorem 5.16. Consider system (ΣD ) subject to multiplicative gain perturbations (5.71) and given a matrix 0 < Qσ = Qtσ ∈ n×n and a scalar γ > 0. A kth−order resilient L2 − L∞ filter exists if there exist matrices 0 < IPξ = ˜f , C˜f , Y, Z, Bj , D, K, N IPtξ ∈ n×n , 0 < IPσ = IPtσ ∈ k×k , IPd , A˜f , B and scalars ε1 > 0, ε2 > 0, ε3 > 0 such that the LMIs   Θaj IPtd Md IPtd Md IPtξ Adj IPtξ Γ + YΦ Θmj   A˜tf + A˜f   • ˜t Φ IPtσ Md 0 0 B f t   +ε D D + Q 1 σ     • I 0 0 0 • −ε 1  < 0  (5.102)   • • • −ε I 0 0 2     • • • • −Qσ 0    −I+  • • • 0 0 ε2 KKt   t  IPξ − ε−1 Dot 3 N N IPd IPξ N t   > 0 , • IPσ −Cft > 0 (5.103) • ε3 I 2 t • • γ I − ε 3 Mh M h have a feasible solution for all admissible uncertainties over the vertex IN. Moreover, the filter matrices are given by −t ˜ ˜ ˜ Af = IP−t σ Af , Bf = IPσ Bf , Cf = Cf

(5.104)

where Θmj = IPtξ (Aoj + Adj ) + (Atoj + Atdj )IPξ + Cot Y t + YCo Θaj

+ ε1 Bjt Bj t ˜ = Coj Bf + Z + ε1 Bjt Na

(5.105)

Proof: Follows directly from Theorem 5.15 using (5.83) and applying the linearizations Y = IPtd Bf , Z = IPtd Af , Bj = Nb Bf Coj , D = Ne Af , K = Nf Bf , N = Ng Cf ∇∇∇ As a final closure to this part of our work, in the absence of gain perturbations (∆Af a (t) ≡ 0, ∆Bf a (t) ≡ 0, ∆Cf a (t) ≡ 0), we have the nominal augmented system (ΣAao ) for which the following result holds:

154

5 Resilient Linear Filtering

Corollary 5.17. Consider the nominal augmented system (ΣAao ) and given a matrix 0 < Q = Qt ∈ n×n and a scalar γ > 0. A kth−order robust L2 − L∞ filter exist if there exist matrices 0 < IPξ = IPtξ ∈ n×n , 0 < IPσ = IPtσ ∈ ˜f , C˜f such that the LMIs k×k , IPd , A˜f , B 

 IPtξ (Aoj + Adj )+ t ˜  (Atoj + Atdj )IPξ + Coj Bf + Z IPtξ Adj IPtξ Γ + YΦ    t t  Co Y + YCo      < 0 ˜t + A˜f A t f   ˜ • 0 Bf Φ   +Q σ     0 • • −Qσ • • • −I   IPξ IPd Dot  • IPσ −Cft  > 0 • • γ2I

(5.106)

(5.107)

have a feasible solution for all admissible uncertainties over the vertex IN. Moreover, the filter matrices are given by −t ˜ ˜ ˜ Af = IP−t σ Af , Bf = IPσ Bf , Cf = Cf

(5.108)

Remark 5.18. Similarly, when solving for the resilient filter matrices, it is possible to cast the scalar γ as one of the LMI variables in which case we have to solve the following convex optimization problem: M inimize π = γ 2 ˆf , Cˆf , Ga , Gb , Gc , Θa , Θb , over W, V, Λ, Aˆf , B Y, Z, Bj , D, K, N subject to(5.102) and (5.103) √ The minimum attainable value will be γ ∗ = π ∗ . 5.6.2 Extended Newton-Leibniz Approach All Newton-Leibniz methods have the starting point of using the classical rule

t ˙ ξ(t − τ ) = ξ(t) − ξ(s) ds t−τ

which eventually transforms (5.42) into the form:

(ΣN ) :

t

˙ = (Aξ∆ + Aξd ) ξ(t) − A¯d E ξ(t)

˙ ξ(s) ds + Γξ w(t) t−τ

∆

¯ ξ(t) zξ (t) = ze (t) = L

(5.109)

5.6 Resilient Delay-Dependent Filtering

155

In preparation for the stability analysis based on Lyapunov theory, there are several Lyapunov-Krasovskii functionals that have been used [81, 93]. Hereafter, we introduce the following Lyapunov-Krasovskii functional

t t Vs (ξ) = ξ (t)Pξ(t) + ξ t (s)E t QEξ(s) ds + (1 − α)−1



0

−τ

t−τ

t

˙ dsdφ ξ˙t (s)E t RE ξ(s)

(5.110)

t+φ

where 0 < P = P t , 0 < Q = Qt , 0 < R = R are weighting matrices. The following theorem establishes the main result: Theorem 5.19. Consider system (ΣN ) and given matrices 0 < Q = Qt ∈ n×n . If there exists a matrix 0 < P = P t ∈ n+k×n+k such that the inequality  t  Aξ∆ P + PAξ∆ + E t QE+ ¯ t t P A − Y PΓ A E R d ξ α ξ∆   βX + E t Yt + YE   t t  ˆ ¯ • −Q 0 Ad E Rα    (5.111)  • • −I Γξt E t Rα  • • • −Rα  t ¯ P L > 0 (5.112) • γ2I has a feasible solution for all admissible uncertainties over the vertex IN and for all gain perturbations ∆f a (t) ∈ ∆FA (t) with Rα = β/(1 − α) R such that  XY ≥ 0 (5.113) •R then the filtering-error system (5.42) is exponentially stable with an L2 − L∞ disturbance attenuation level γ > 0 Proof: Evaluation of V˙ s (ξ) along the solutions of (5.109) and bounding some terms yields:   V˙ s (ξ) ≤ ξ t (t) (Aξ∆ + Aξd )t P + P(Aξ∆ + Aξd ) ξ(t) + 2ξ t (t)PΓξ w(t)

t

˙ ξ t (s)P A¯d E ξ(s) ds + ξ(t)E t QEξ(t)

−2 t−τ

− (1 − α)ξ t (t − τ )E t QEξ(t − τ )

t β t t ˙ ˙ ˙ ˙ RE ξ(s) ds ξ(s)E + ξ(t)E RE ξ(t) − (1 − α) t−τ   ≤ ξ t (t) (Aξ∆ + Aξd )t P + P(Aξ∆ + Aξd ) ξ(t) + 2ξ t (t)PΓξ w(t)

156

5 Resilient Linear Filtering



t

˙ ξ t (s)P A¯d E ξ(s) ds + ξ(t)E t QEξ(t)

−2 t−τ

− (1 − α)ξ t (t − τ )E t QEξ(t − τ )   ¯ + Aξ∆ ξ(t) + Ad Eξ(t − τ ) + Γξ w(t)   t ¯ E Rα E Aξ∆ ξ(t) + Ad Eξ(t − τ ) + Γξ w(t)

t t ˙ ˙ RE ξ(s) ds ξ(s)E

−

(5.114)

t−τ ∆ ∆ ∆ ∆ ˙ Using Lemma 8.5 with h = ξ(t), m = E ξ(t), IM = P A¯d , Z = R, we get:

t ˙ ξ t (s)P A¯d E ξ(s) ds −2 t−τ  t t t t ¯ ¯ ≤ ξ (t) βX + E Y − E Ad P + YE − P Ad E ξ(t) 

t t ˙ ˙ RE ξ(s) ds (5.115) ξ(s)E − 2ξ t (t) Y − P A¯d Eξ(t − τ ) + t−τ

such that X, Y, R satisfy (5.113). Setting w ≡ 0, it follows from (5.114) and (5.115) that   t t ˙ Vs (ξ) ≤ ξ (t) (Aξ∆ + Aξd ) P + P(Aξ∆ + Aξd ) ξ(t) + 2ξ t (t)PΓξ w(t)  t ¯ − 2ξ (t) N − P Ad Eξ(t − τ ) + ξ(t)E t QEξ(t) − (1 − α)ξ t (t − τ )E t QEξ(t − τ )  t + Aξ∆ ξ(t) + Aξd ξ(t − τ ) + Γξ w(t) E t Rα E   Aξ∆ ξ(t) + Aξd ξ(t − τ ) + Γξ w(t)  t t t t ¯ ¯ + ξ (t) βX + E Y − E Ad P + YE − P Ad E ξ(t)  = where

ξ(t) ξ(t − τ )E t

t

 Ξ

ξ(t) ξ(t − τ )E t

(5.116)

 Atξ∆ P + PAξ∆ + E t QE+  βX + E t Yt + YE P A¯d − Y + Atξ∆ E t Rα EAξd   (5.117) Ξ= t t   +Aξ∆ E Rα EAξ∆ t t ˆ • −Q + Aξd E Rα EAξd 

5.6 Resilient Delay-Dependent Filtering

157

By a parallel development to Theorem 5.6, the internal exponential stability of the filtering-error system (5.42) with w ≡ 0 is guaranteed by (5.111) and the Schur complements which in turn ensure that Ξ < 0 over the vertex set IN. t ∆ By considering the performance measure J = Vs (ξ) − 0 wt (s)w(s)ds for 0 = w(t) ∈ L2 [0, ∞) at zero initial condition and noting that internal stability implies Vs (ξ) → 0 as t → 0, it follows that  t  

t

t  ξ(t) ξ(t)  ξ(t − τ )E t  Ξ¯  ξ(t − τ )E t  V˙ s (ξ) − wt (s)w(s) ds ≤ J = 0 0 w(t) w(t) where

 ¯ Ξ¯c Ξ¯d Ξa ˆ + A¯t E t Rα E A¯d Ξ¯ =  • −Q A¯td E t Rα EΓξ  d • • −I + Γξt E t Rα EΓξ Ξ¯a = At P + PAξ∆ + E t QE + βX ξ∆

+ E t Yt + YE + Atξ∆ E t Rα EAξ∆ Ξ¯c = P A¯d − Y + Atξ∆ E t Rα E A¯d Ξ¯d = PΓξ + Atξ∆ E t Rα EΓξ

(5.118)

It is readily evident once again on using the Schur complements that (5.111) guarantees Ξ¯ < 0 over the vertex set IN and therefore J < 0, which in turn t leads to ξ t (t)Pξ(t) ≤ Vs (ξ) < 0 wt (s)w(s)ds. Moreover, by Theorem 5.6, inequality (5.53) can be easily derived and hence the proof is completed. ∇∇∇ Finally, on considering (5.118) in view of (5.43) and applying Inequalities 1-2 of the Appendix, it follows that  t  Aξo P + PAξo + E t QE+ ¯ P A − Y PΓ d o   βX + E t Yt + YE  Ξ¯ =   ˆ • −Q 0  • • −I    t   t t t Aξ∆ E t Aξ∆ E t ∆Aξ P + P∆Aξ 0 P∆Γ + • 0 0  +  A¯td E t  Rα  A¯td E t  Γξt E t Γξt E t • • 0   Atξo P + PAξo + E t QE+ ¯ tN ¯   βX + E t Yt + YE + ε1 N P A¯d − Y b b+  −1 PΓo + Atξo R−1 Γo  ρ ¯ ¯ aM ¯ aM ¯ at P + ε−1 P M ¯ at P +Atξo R−1   ε1 P M ρ Ad 2   −1 t t t −1   ≤  + E Na Na E + Aξo Rρ Aξo  t −1 ¯ t −1   ˆ ¯ ¯ • −Q + Ad Rρ Ad Ad Rρ Γo   t −1  −I + Γo Rρ Γo  • • +ε2 Φt Nbt Nb Φ

158

5 Resilient Linear Filtering

= Ξ+ < 0 ∆

t , Rρ = R−1 α − ψMa Ma ∆

With reference to (5.113), we let  X1 0 X= 0 X2

(5.119) 

,

Y1 Y= Y2

(5.120)

then the existence of resilient L2 − L∞ filter is established by the following theorem. Theorem 5.20. Consider system (ΣN ) subject to additive gain perturbations (5.41) and given a matrix 0 < Q = Qt and a scalar γ > 0. A kth−order resilient L2 − L∞ filter exists if there exist matrices 0 < W = W t , 0 < V = ¯ 2 , Y2 and scalars ¯a , Θ ¯ b , X1 , Y1 , X ˆf , Cˆf , Ga , Gb , Gc , Θ V t , 0 < Λ, Aˆf , B ε1 > 0, ε2 > 0, ε3 > 0,  > 0, µ > 0, ψ > 0 such that the LMIs 

 WΓ + t ¯a Υ Υ WA A Θ m d dj oj ˆf Φ   RB   t   t ˆt  • Υn Gta Adj Ga Γ + Coj  0 B f   ˆ Bf Φ    • • −Q t ˆ (5.121) 0 Adj 0    < 0   −I+  • • • Γot 0    ε2 Nb Nbt    • • • • −Rρ 0  ¯b • • • • • −Θ   µ 1  1 ≥ 0 , ≥ 0 (5.122) • ε1 • ε2   t  W − ε−1 Dot 3 Nc Nc R Λ W Nct   > 0 , • V −Cft > 0 (5.123) • ε3 I • • γ 2 I − ε3 Mct Mc have a feasible solution for all admissible uncertainties over the vertex IN and for all gain perturbations ∆f a (t) ∈ ∆FA (t) where t t ˆf Coj + C t B ˆt t Υm = WAoj + Atoj W + RB oj f R + ε1 Coj Nb Nb Coj

+ Q + αX1 + Y1 + Y1t t ˆ t Υd = RAˆf + Atoj Ga + Coj Nbt Gb + Y2t Bf + ε1 Coj Υn = Aˆtf + Aˆf + ε1 Gtb Gb + αV −1 Λt X2 Λ + Gc Gtc + µGc Gtc ¯ b = [ε1 I, ε2 I, ε1 I, ε2 I, ψI] ¯ a = [RGc , RGc , RGc , RGc , Nat ] , Θ Θ −1 t ¯ 2 = V Λ X2 Λ (5.124) X Moreover, the filter matrices are given ˆf , Cf = Cˆf Af = Λ−1 Aˆf Λ−t V , Bf = Λ−1 B

(5.125)

5.6 Resilient Delay-Dependent Filtering

159

Proof: Follows parallel developments to Theorem 5.7 applied to Ξ¯ of (5.118). ∇∇∇ The case of multiplicative gain perturbations can be readily stated Theorem 5.21. Consider system (ΣN ) subject to multiplicative gain perturbations (5.41) and given a matrix 0 < Q = Qt ∈ n×n and a scalar γ > 0. A kth−order resilient L2 − L∞ filter exists if there exist matriˆf , Cˆf , Ga , Gb , Gc , ces 0 < W = W t , 0 < V = V t , 0 < Λ, Aˆf , B ¯ 2 , Y2 Ψb and scalars ε1 > 0, ε2 > 0, ε3 > 0,  > 0, µ > ¯a , Θ ¯ b , X1 , Y1 , X Θ 0, ψ > 0 such that the LMIs 

¯ ¯  Υm Υd WAdj    • Υn Gta Adj    • • −Q ˆ    • • •    • • • • • •  µ 1 ≥ 0 , • ε1  t W − ε−1 3 Ψc Ψc  • •

 WΓ + t ¯a A Θ oj ˆf Φ  RB  t  Ga Γ + t ˆt  ˆf Φ Coj Bf 0  B  0 Atdj 0   < 0  −I+ Γot 0   t ε2 N b N b  • −Rρ 0  ¯b • • −Θ   1 ≥ 0 • ε2   RΛ Dot W Ψct t  V −Cf > 0 , > 0 • ε3 I • γ 2 I − ε3 Mht Mh

(5.126)

(5.127)

(5.128)

have a feasible solution for all admissible uncertainties over the vertex IN and for all gain perturbations ∆f a (t) ∈ ∆FM (t) where t t ˆf Coj + C t B ˆt t Υ¯m = WAoj + Atoj W + RB oj f R + ε1 Coj Ψb Ψb Coj

+ Q + αX1 + Y1 + Y1t t ˆ t Ψet + Y2t Υ¯d = RAˆf + Atoj Ga + Coj Bf + ε1 Coj

(5.129)

Moreover, the filter matrices are given ˆf , Cf = Cˆf Af = Λ−1 Aˆf Λ−t V , Bf = Λ−1 B

(5.130)

We end this part of our work by considering the nominal augmented system (ΣAao ), we have for which the following result holds: Corollary 5.22. Consider the nominal augmented system (ΣAao ) and given a matrix 0 < Q = Qt ∈ n×n and a scalar γ > 0. A kth−order robust L2 − L∞ filter exists if there exist matrices 0 < W = W t , 0 < V = V t , 0 < ¯ 2 , Y2 such that the LMIs ˆf , Cˆf , Ga , Gb , Gc , Θ ¯a , Θ ¯ b , X1 , Y1 , X Λ, Aˆf , B

160

5 Resilient Linear Filtering



 ˆm Υˆd WAdj WΓ + At Υ oj  ˆf Φ  RB   t   t ˆt   • Υn Gta Adj Ga Γ + Coj B f < 0  ˆ (5.131) Bf Φ    • • −Q  t ˆ 0 A  dj   • • • −I Γot  • • • • −Rρ   µ 1  1 ≥ 0 , ≥ 0 (5.132) • ε1 • ε2   t  W − ε−1 Dot 3 Nc Nc R Λ W Nct   > 0 , • V −Cft > 0 (5.133) • ε3 I • • γ 2 I − ε3 Mct Mc have a feasible solution for all admissible uncertainties over the vertex IN where t ˆt t ˆf Coj + Coj Bf R + Q + αX1 + Y1 + Y1t Υˆm = WAoj + Atoj W + RB t ˆ Bf + Y2t Υˆd = RAˆf + Atoj Ga + Coj (5.134) Υˆn = Aˆt + Aˆf + αV −1 Λt X2 Λ f

Moreover, the filter matrices are given by −t ˜ ˜ ˜ Af = IP−t σ Af , Bf = IPσ Bf , Cf = Cf

(5.135)

Remark 5.23. Likewise, when solving for the resilient filter matrices, it is possible to cast the scalar γ as one of the LMI variables in which case we have to solve the following convex optimization problem: M inimize π = γ 2 ˆf , Cˆf , Ga , Gb , Gc , over W, V, Λ, Aˆf , B ¯ 2 , Y2 Ψb ¯a , Θ ¯ b , X1 , Y1 , X Θ subject to(5.126) and (5.128) √ The minimum attainable value will be γ ∗ = π ∗ . Remark 5.24. A simple comparison between the resilient delay-dependent filtering approaches shows that the descriptor approach depends on a larger information set than the extended Newton-Leibniz approach since the former approach demands instantaneous value of the delay τ, whereas the latter approach needs only the bounds α, β . This suggests that the descriptor approach belongs to the family of strong delay-dependent methods and extended NewtonLeibniz approach is a member of the weak delay-dependent methods.

5.7 Examples

161

5.7 Examples In the sequel, we provide some examples to shed some light on the theoretical development of this part of the Chapter. 5.7.1 Example 5.3 We consider a pilot-scale single-reach water quality system which can fall into the type (5.33)-(5.35) with the following associated date :      0.9 δd 0 1 + δa 2δa −0.2 1 + δd  , Γ =  −0.7  0 0 Ao =  0 1.5δa 3δa  , Ad =  0.1 0.5 −3 −7 −5 −0.5 −0.2 −0.4 + 2δd 

Co = [2, 0, 1] , Do = [1, 3, 1] , Φ = 0.65 τ = 0.6 , −0.4 ≤ δa ≤ 0.4 , −0.2 ≤ δd ≤ 0.2 , α = 0.6 , β = 0.5 Invoking the MATLAB software environment, we solve the convex optimization problem outlined in Remark 5.12 using   200 Qσ =  0 5 0  008 Considering the delay-independent resilient filter synthesis, the feasible solutions for both full-order and reduced-order resilient filters with additive and multiplicative gain perturbations are summarized in Table 5.1. It is quite Table 5.1. Delay-Independent Filter Design Results: Example 5.3 T ype k Additive 2 0.5174 0.9175 -7.1156 Additive 3 0.7236 0.8535 -11.6015 Multipl- 2 0.1764 icative 2.7885 -12.1586 Multipl- 3 1.1764 icative 3.4175 -14.1673

Af 11.0175 2.4567 -12.6457 12.2011 3.0517 -17.5337 10.1856 3.7464 -17.6557 1.7775 1.4557 -12.4587

-2.5466 -0.4176 -8.6953 -3.1633 -0.1616 -12.9045 -2.6236 -2.1796 -38.5553 -8.5626 -3.4146 -28.6924

Bf 0.7846 -0.9464 1.0314 0.8655 -1.1244 1.2254 0.4995 -2.6333 1.0544 1.4356 -1.4694 5.0304

Cf γ∗ -1.0482 -23.7784 6.3714 1.8765

-1.2015 -19.5466 7.9805 1.6224

-1.2357 -22.4788 5.7815 2.0134

-1.4882 -17.7894 11.3144 1.7988

evident that minimum disturbance-attenuation level bound γ ∗ = 1.6624 corresponding to full-order with additive gain perturbations. Next we examine the variation of the resultant filtering-error system over the vertex set

162

5 Resilient Linear Filtering ∆

IN = {(δa , δd ) : −0.4 ≤ δa ≤ 0.4 , −0.2 ≤ δd ≤ 0.2} through computer simulation of Theorem 5.6. This is displayed in Figures 2 and 3 from which it is clear that the maximum L2 − L∞ gain bound over the set IN is γ ∗ = 1.4932 for full-order filter and is attained at the point {δa = 0.4 , δd = −0.2} and is γ ∗ = 1.5158 for reduced-order filter and is attained at the point {δa = 0.4 , δd = 0.2}. In turn, this illuminates the effectiveness of the resilient filter design.

Fig. 5.2. Resilient L2 − L∞ Gain Bound: full-order delay-independent filter

Turning to the delay-dependent approaches, the results related to the feasible solutions are included in Table 5.2. As expected, we record two important observations: 1) the descriptor approach yields less conservative results that the extended Newton-Leibniz approach and 2) the delay-dependent resilient filter design outperforms the delay-independent design. For the purpose of completeness, the variations of the resultant filtering-error system over the vertex set IN are is displayed in Figures 4 and 5 for full-order resilient filter design with additive gain perturbations using both delay-dependent approaches.

5.7 Examples

163

Fig. 5.3. Resilient L2 − L∞ Gain Bound: reduced-order delay-independent filter

Fig. 5.4. Resilient L2 − L∞ Gain Bound: descriptor delay-dependent filter

164

5 Resilient Linear Filtering Table 5.2. Delay-Dependent Filter Design Results: Example 5.3 Approach Descriptor

T ype Additive Multiplicative Descriptor Additive Multiplicative Newton-Leibniz Additive Multiplicative Newton-Leibniz Additive Multiplicative

k 2 2 3 3 2 2 3 3

γ∗ 1.4985 1.5026 1.4213 1.4355 1.5546 1.5787 1.5136 1.5248

Fig. 5.5. Resilient L2 − L∞ Gain Bound: Newton-Leibniz delay-dependent filter

5.7.2 Example 5.4 Consider a linear dynamical system with polytopic uncertainties and described by     0 −3 1 + δa −0.2 0.1 + δd 0.5 + δd  , Ad =  0.5 1 −1 −0.7 + δd  Ao =  0.3 −2.5 + 0.5δa 0 1 + δd −2.5 + δd −0.1 0.3 + δa −4 + 0.6δa   1 Γ =  0  , Cot = [0.8, 0.3, 0.1] 1

5.7 Examples

165

Do = [0.2, −0.3, −0.6] , Φ = 0.5 , α = 0.6 , β = 0.5 τ = 0.7 , −0.3 ≤ δa ≤ 0.3 , −0.1 ≤ δd ≤ 0.1 By resorting the MATLAB software environment, we solve the convex optimization problem outlined in Remark 5.12 using   4 0 0 Qσ =  0 12 0  0 0 7 Considering the delay-independent resilient filter synthesis, the feasible solutions for both full-order and reduced-order resilient filters with additive and multiplicative gain perturbations are summarized in Table 5.3. It is quite Table 5.3. Delay-Independent Filter Design Results: Example 5.4 T ype k Additive 2 -6.1654 1.9105 0.7256 Additive 3 -5.12567 1.6255 0.5045 Multipl- 2 -7.1764 icative 1.7895 0.8156 Multipl- 3 -5.5006 icative 1.6715 0.7113

Af 2.0195 -2.5967 -8.1247 1.8767 -3.7907 -7.3007 2.1856 -3.4694 -7.6597 1.7687 -3.8037 -6.8237

2.4646 -2.4336 -3.6773 2.3722 -1.5305 -3.4895 2.6446 -2.2796 -3.8853 2.4572 -1.5605 -4.2774

Bf 0.4776 2.7484 3.9314 -1.8995 3.4224 4.9854 0.4785 2.6773 4.0344 -1.9685 3.6105 5.0314

Cf γ∗ -1.0204 -2.4101 2.7144 2.1135

-1.0105 -1.6026 2.4305 1.7014

-1.0317 -2.4258 2.7801 2.3224

-1.1123 -1.5928 2.5105 1.8108

evident that minimum disturbance-attenuation level bound γ ∗ = 1.7014 corresponding to full-order with additive gain perturbations. 5.7.3 Example 5.5 Consider another linear dynamical system with polytopic uncertainties and described by     −0.1 −2.5 0.5 + δa 0 1 + δd 0.6 + δd  , Ad =  0.1 0.5 −0.1 + δd  0.3 Ao =  0.1 −3.5 + 0.1δa −0.1 0.8 + δa −1.6 + 0.4δa −0.6 1 + δd −0.8 + δd   −0.6 Γ =  0.5  , Cot = [−0.5, 0.2, 0.3] 0 Do = [0, −0.6, 0.2] , Φ = 0.8 , α = 0.8 , β = 0.6 τ = 0.7 , −0.2 ≤ δa ≤ 0.2 , −0.6 ≤ δd ≤ 0.6

166

5 Resilient Linear Filtering

Observe that the variation in the delay matrix takes on wider range. By resorting the MATLAB software environment, we solve the convex optimization problem outlined in Remark 5.12 using   14 0 0 Qσ =  0 4 0  0 0 10 Considering the delay-independent resilient filter synthesis, the feasible solutions for both full-order and reduced-order resilient filters with additive and multiplicative gain perturbations are summarized in Table 5.4. It is quite Table 5.4. Delay-Independent Filter Design Results: Example 5.5 T ype k Additive 2 -6.1654 1.9105 0.7256 Additive 3 -8.4813 5.2715 -4.5005 Multipl- 2 -7.1764 icative 1.7895 0.8156 Multipl- 3 -8.9863 icative 5.4705 -4.4945

Af 2.0195 -2.5967 -8.1247 1.8007 -4.4107 2.1601 2.1856 -3.4694 -7.6597 1.8101 -4.4231 2.1691

2.4646 -2.4336 -3.6773 -1.6322 1.5905 -1.9235 2.6446 -2.2796 -3.8853 -1.6412 1.6005 -1.9246

Bf 0.4776 2.7484 3.9314 -2.4469 1.6644 -2.3140 0.4785 2.6773 4.0344 -2.4519 1.6714 -2.3240

Cf γ∗ -1.0204 -2.4101 2.7144 3.9345

-1.1340 -1.0927 2.3105 3.0184

-1.0317 -2.4258 2.7801 4.1246

-1.2340 -1.1325 2.5555 3.2209

evident that minimum disturbance-attenuation level bound γ ∗ = 3.0184 corresponding to full-order with additive gain perturbations. 5.7.4 Example 5.6 Consider a second-order linear dynamical system described by    0 0 0 11 −0.5 0.7 + δd , Ad = , Γ = Ao = 1 + δc 0 −11 −2.2 + δa −0.4 + δd −0.6 Co = [0, 1] , Do = [1, 0] , Φ = [0, 1 + δc ] τ = 0.8 , −1 ≤ δa ≤ 1 , −0.7 ≤ δd ≤ 0.7 α = 0.8 , β = 0.6 , 0 ≤ δc ≤ 1 Observe that the variation in the input and delay matrices take on wider range and the amplitude of the bounded energy input can vary. By resorting the MATLAB software environment, we solve the convex optimization problem outlined in Remark 5.12 using

5.7 Examples

 Qσ =

10 0 0 6

167



Considering the delay-independent resilient filter synthesis, the feasible solution of full-order resilient filter with additive and multiplicative gain perturbations are summarized in Table 5.5. Table 5.5. Filter Design Results: Example 5.6 T ype k Af Additive 2 -0.2675 10.9903 -11.0013 -1.3055 Multipl- 2 -0.4604 11.1053 icative -11.1155 -2.1525

Bf Cf γ∗ 0.0091 0.5705 -0.0084 1.2154 0.6525 0.0105 0.8205 -0.0205 1.5325 0.4275

5.7.5 Example 5.7 Consider a fourth-order uncertain system of the form   1 −1.5 −0.7 δa  0 δa 1 1.3   , Co = [1, δo , 0, 0] Ao =   −0.1 0 0  δa −0.4 −2.5 −3.5 −4.7   0 0 −0.1 + δd 0.05 + δd  −0.1 0 −0.5 −0.1 + δd   , Φ=1 Ad =   0 0 −0.6 0.1 + δd  0 −0.9 0 −0.5 + δd Γ t = [0, 0, 0, δc ] , Do = [1, 1, 0, 0] τ = 1 , 0.9 ≤ δa ≤ 1 , 0.6 ≤ δd ≤ 0.7 , α = 0.8 β = 0.6 , 0 ≤ δo ≤ 1 , 0.08 ≤ δc ≤ 0.09 Observe that there are four parameters in the model matrices. By resorting the MATLAB software environment, we solve the convex optimization problem outlined in Remark 5.12 using   14 0 0 0 0 9 0 0  Qσ =   0 0 12 0  0 0 0 20 Considering the delay-independent resilient filter synthesis, the feasible solutions for full-order k = 4 resilient filters with additive and multiplicative gain perturbations are summarized in Table 5.6. With focus on the resilient

168

5 Resilient Linear Filtering Table 5.6. Filter Design Results: Example 5.7

T ype Additive -1.1585 1.0845 0 -0.0445 Multipl- -1.2565 icative 2.0510 0.0155 -0.4455

Af 1.3928 -6.9155 -1.8718 3.1555 -0.0728 -4.5775 1.1775 2.4675 1.8115 -8.5775 -1.2718 3.0055 -1.7208 -1.7575 3.0075 2.7465

4.5705 2.6683 0.1185 -2.7783 11.0544 4.8663 0.0186 -2.0183

Bf Cf γ∗ 0.8764 1.1678 -0.8577 -10.4456 -2.4582 24.5445 0.2757 0 0.2015 1.6054 1.7698 -1.5787 -10.7756 -5.5812 30.4872 0.4757 0.0025 0.4115

Table 5.7. Computational Results: Example 5.7

γ

γ

γ

γ

V ertix1 δa = 0.9 δd = 0.6 δo = 0 δc = 0.08 2.3456 V ertix5 δa = 0.9 δd = 0.7 δo = 0 δc = 0.08 3.6986 V ertix9 δa = 1.0 δd = 0.6 δo = 0 δc = 0.08 5.9855 V ertix13 δa = 1.0 δd = 0.7 δo = 0 δc = 0.08 7.2345

V ertix2 δa = 0.9 δa = 0.6 δa = 0 δa = 0.09 2.5789 V ertix6 δa = 0.9 δa = 0.7 δa = 0 δa = 0.09 3.5789 V ertix10 δa = 1.0 δa = 0.6 δa = 0 δa = 0.09 5.8779 V ertix14 δa = 1.0 δa = 0.7 δa = 0 δa = 0.09 7.7895

V ertix3 δa = 0.9 δa = 0.6 δa = 1 δa = 0.08 2.9675 V ertix7 δa = 0.9 δa = 0.7 δa = 1 δa = 0.08 4.5045 V ertix11 δa = 1.0 δa = 0.6 δa = 1 δa = 0.08 5.9685 V ertix15 δa = 1.0 δa = 0.7 δa = 1 δa = 0.08 8.6755

V ertix4 δa = 0.9 δa = 0.6 δa = 1 δa = 0.09 3.4136 V ertix8 δa = 0.9 δa = 0.7 δa = 1 δa = 0.09 2.6456 V ertix12 δa = 1.0 δa = 0.6 δa = 1 δa = 0.09 6.1365 V ertix16 δa = 1.0 δa = 0.7 δa = 1 δa = 0.09 9.4155

filter with full-order subject to additive gain perturbations, a summary of the variations of the resultant filtering-error system over the vertex set IN is summarized in Table 5.7. It is quite evident that minimum disturbanceattenuation level bound γ ∗ = 9.4155 and occurs at the utmost corner of the vertex set.

5.8 Notes and References

169

5.8 Notes and References This Chapter has established the basic foundations for resilient filtering of uncertain time-delay systems. It should open the door for possible build up and extensions along the results of [33, 34, 97, 100, 101, 102]. The topic is exciting and more results should clarify questions like what is the best (least conservative) delay-dependent approach to resilient filtering? Is possible to integrate the delay-factor as one the perturbation parameters. Indeed, the topic of resilient filtering of uncertain discrete-time delay systems is still fresh.

6 Resilient Delay-Dependent Control

6.1 Introduction It becomes increasingly apparent that delays occur in physical and man-made systems due to various reasons including finite capabilities of information processing among different parts of the system, inherent phenomena like mass transport flow and recycling and/or by product of computational delays [81]. Considerable discussions on delays and their stabilization/destabilization effects in control systems have commanded the interests of numerous investigators in recent years, see [81, 8, 22] and their references. In control design, it turns out that the design objectives have to incorporate the impact of parameter shifting, component and interconnection failures which are frequently occurring in practical situations. Robust control theory provides tractable design tools using the time domain and the frequency domain when the plant modeling uncertainty or external disturbance uncertainty is of major concern in control systems. In this Chapter, we direct attention to delay-dependent methodologies for continuous-time and discrete-time systems by exhibiting the delay-dependence dynamics in the design procedure. For the continuous-time case, we develop a descriptor approach to simultaneous H2 /H∞ control design for a class of time-delay systems with additive and multiplicative controller gain perturbations. The delay factor is considered unknown-but-bounded constant or timevarying. The cases of norm-bounded and convex polytopic uncertainties are treated. In the discrete-time case, delay-dependent stability and feedback stabilization methods are developed for the nominal, robust and resilience cases. Seeking computational convenience, all the developed results are cast in the format of linear matrix inequalities (LMIs) and several numerical examples are presented.

M.S. Mahmoud: Resilient Control of Uncertain Dynamical Systems, LNCIS 303, pp. 171-205, 2004. © Springer-Verlag Berlin Heidelberg 2004

172

6 Resilient Delay-Dependent Control

6.2 Continuous-Time Systems We consider a class of time-delay systems by: (ΣJ ) :

x(t) ˙ = Ao x(t) + Ad x(t − τ (t)) + Bo u(t) + Γ w(t) x(s) = φ(s), t ∈ [−τ, 0] y(t) = Co x(t) + Do u(t) z(t) = Go x(t) + Fo u(t) + Φw(t)

(6.1) (6.2) (6.3)

where x(t) ∈ n is the state vector; u(t) ∈ m is the control input; w(t) ∈ q is the disturbance input which belongs to L2 [0, T ]; y(t) ∈ p , z(t) ∈ p is the controlled output, τ ∈ [τ0 , τ ∗ ] is a constant delay factor and Ao , Bo , Ad , Co , Do , Go , Fo , Γ, Φ are known real constant matrices of appropriate dimensions which describe the nominal system of (ΣJ ). The initial condition is specified as x(0), x(s) = xo , φ(s) , where φ(.) ∈ L2 [−τ, 0]. Our purpose in the sequel is to develop methods for determining delaydependent stabilizing controllers which minimizes the upper bound of an H2 performance measure while guaranteeing that a prescribed upper bound on an H∞ performance is attained for all possible w(t) ∈ L2 [0, ∞]. It will be shown that this problem is to the existence of a feasible solution of linear matrix inequalities (LMIs) thereby providing a clear key to designing the state feedback controller. Note that system ΣJ is continuous-time system with point-delay. In the sequel, the main thrust is to transform this system to an appropriate form in order to exhibit its delay dependence behavior. There are several methods to achieve this [81] each of which has merits and demerits. We will accomplish this by the following method which is known to be less conservative than other methods. 6.2.1 Descriptor Transformation We employ the descriptor system approach [22] to exhibit the delay-dependence dynamics in the design procedure. Thus, we rewrite system (ΣJ ) into the form: (ΣD ) : x(t) ˙ = σ(t) 0 = −σ(t) + [Ao + Ad ]x(t) + Bo u(t)

t σ(s)ds + Γ w(t) − Ad

(6.4)

t−τ (t)

y(t) = Co x(t) + Do u(t),

(6.5)

z(t) = Go x(t) + Fo u(t) + Φw(t),

(6.6)

Under the state-feedback control law u(t) = Ko x(t)

(6.7)

6.3 Feedback Control Design

173

into system (ΣDo ), it becomes: (ΣDK ) :

x(t) ˙ = σ(t)



0 = −σ(t) + AKd x(t) − Ad

t

σ(s)ds + Γ w(t) (6.8) t−τ (t)

y(t) = CK x(t)

(6.9)

z(t) = GK x(t) + Φw(t)

(6.10)

where CK = Co + Do Ko , GK = Go + Fo Ko , Aod = Ao + Ad AKd = Ao + Bo Ko + Ad = Aod + Bo Ko

(6.11)

In the discussions to follow, we develop tractable synthesis procedures to compute the feedback gain Ko .

6.3 Feedback Control Design We focus attention on three design methodologies based on H2 , H∞ and simultaneous H2 /H∞ approaches, respectively. For each approach, we divide our effort into two schemes: a nominal scheme where the gain Ko is assumed to be exactly implemented and a perturbed scheme that takes into consideration various sources of inaccuracies. 6.3.1 Nominal H2 Design In this regard, we introduce the H2 performance measure  ∞ ∆ t J2 = y (s)y(s) ds

(6.12)

0

The problem of H2 state-feedback control could be phrased as follows: Given system ΣJ determine the control law (6.7) which achieves the minimal value of H2 performance measure (6.12) Since our objective is to develop tractable synthesis to determine the feedback gain Ko , we introduce, for convenience, the following matrices   I0 Px 0 P¯ = U IP ; U = ; IP = 00 P d Pσ    I 0 Yx 0 E1 = (6.13) , E2 = , Y = IP−1 = Yd Yσ 0 I Yx = Px−1 , Yσ = Pσ−1 , Yd = −Yx Pd Yσ   0 I 0 ¯ , C¯K = [CK 0] , Ad = AoKd = AKd −I Ad and provide the following stability result.

(6.14)

174

6 Resilient Delay-Dependent Control

Theorem 6.1. In the absence of input disturbance w(t) ≡ 0, controller (6.7) is an H2 -optimal controller for system (ΣDK ) minimizing the H2 -performance measure (6.12) if, given matrix sequence Qσ = Qtσ > 0, i ∈ S there exist matrices Px = Pxt > 0, Pσ = Pσt > 0, Pd = Pdt > 0, i ∈ S, satisfying the system of LMIs   t CK Ω2 τ Pdt Ad Ω1 + CK ∆ • −Ω3 τ Pσt Ad  < 0 (6.15) Πt =  • • −Qσ where Ω1 = Pdt AKd + AtKd Pd , Ω2 = Pxt − Pdt + AtKd Pσ Ω3 = Pσt + Pσ − τ Qσ

(6.16)

An upper bound on the H2 performance measure is given by 

∆ J2 ≤ J+ = xt (0)Px x(0) + τo

1/2

0

−τo

σ t (s) Qσ σ(s) ds

(6.17)

Proof: Let the Lyapunov-Krasovskii functional V (·) of the transformed system (ΣDKo ) be selected as: ∆

Vt (x) = Vx (x) + Vσ (x)



Vx (x) = xt (t)Px x(t) , Vσ (x, i) =



0

−τ (t)

t

σ t (s) Qσ σ(s) ds d θ (6.18) t+θ

Using (6.8) and (6.18), we get: ∂Vx /∂x x(t) ˙ = 2xt (t)Pxt x(t) ˙

x(t) ˙ = 2[x (t) σ (t)]IP 0  σ(t) t t t = 2[x (t) σ (t)] IP −σ(t) + AoKd x(t)  0 t + 2[xt (t) σ t (t)] IPt −Ad t−τ σ(s)ds t

t



t

Simple manipulations using (6.8) yield:  σ(t) 2[xt (t) σ t (t)] IPt = −σ(t) + AoKd x(t)   x(t) [xt (t) σ t (t)] IPt AoKd + AtoKd IP σ(t) Using Inequality I.1 of the Appendix, it follows that

(6.19)

(6.20)

6.3 Feedback Control Design

 2[xt (t) σ t (t)] IPt

−Ad

t



0

t−τ

175

σ(s)ds

t

[xt (t) σ t (t)] IPt A¯d σ(s)ds

=2 t−τ



≤ τ

x(t) σ(t)

t

¯t IPt A¯d Q−1 σ Ad IP



x(t) σ(t)



t

σ t (s)Qσ σ(s) ds

+

(6.21)

t−τ

Therefore from (6.19)-(6.21) we get:  t   x(t) x(t) t ¯ V˙ x ≤ IPt AoKd + AtoKd IP + τ IPt A¯d Q−1 IP A d σ σ(t) σ(t)

t σ t (s)Qσ σ(s) ds (6.22) + t−τ

It is straightforward to show that, V˙ σ = τ σ t (t)Qσ σ(t) −



0

−τ

σ t (t + θ)Qσ σ(t + θ) dθ

(6.23)

By using (6.13)-(6.14) and (6.22), it follows that V˙ t = V˙ x + V˙ σ   t  x(t) x(t) t t ¯ IPt AoKd + AtoKd IP + τ IPt A¯d Q−1 = IP + τ E Q E A 2 σ σ d 2 σ(t) σ(t)   t x(t) x(t) ∆ (6.24) Πo = σ(t) σ(t) Let χ(t) = [xt (t) σ t (t)]t . Considering Πo and using (6.13)-(6.14), some matrix t manipulations convert it to LMIs (6.15). In view of the fact that CK CK > 0, ˙ it follows that Πo < 0 is guaranteed and we conclude that Vt < 0 for all χ(t) = 0 and V˙ t ≤ 0 for all χ(t). Since ||χ(t + β)|| ≤ ϕ||χ(t)|, ∀β ∈ [−τ, 0] t 2 ¯ and some ϕ > 0 [53], it follows from (6.18) that  Vt (.) ≤ χ (t)P χ(t) + µ||χ|| t ¯ where µ = ϕ τ maxi λM [P ] + τ λM [E2 Qσ E2 ] . Therefore, for all χ = 0, we have   χt Πo χ V˙ t (.) λm [−Πo ] ∆ (6.25) ≤ t¯ ≤ −ξ = − Vt (.) λM [P¯ ] + µ χ P χ + µ||χ||2 ∆

It is readily seen from (6.25) that ξ > 0 and hence, on letting Vt (0) = Vt (x = xo ) we get V˙ t ≤ −ξ Vt , which implies Vt (.) ≤ e−ξ t Vt (0)

(6.26)

Therefore system (ΣDK ) is exponentially stable at a rate ξ > 0 which completes the proof. ∇∇∇

176

6 Resilient Delay-Dependent Control

Remark 6.2. It should be noted that Theorem 6.1 establishes an LMI-based sufficient condition for the existence of state-feedback controller (6.7) and hence it depends on Ko . The following theorem provides a method for computing the feedback gains Theorem 6.3. The feedback gain associated with the H2 -optimal controller for system (ΣDK ) is given by Ko = Z Yx−1 where the matrices Yx = Yxt > 0, Yσ = Yσt > 0, Yd = Ydt > 0, satisfy the system of LMIs   Λ1 Λ2 τ Ydt Yxt Cot + Z t Dot  • −Λ3 τ Yσt  0   < 0 (6.27)  • • −τ Qσ  0 • • • −I where Λ1 = Ydt + Yd , Λ2 = Yσ − Ydt + Yxt Atod + Z t Bot Λ3 = Yσt + Yσ − τ Ad Q−1 σ Ad

(6.28)

Proof: By the Schur complements, condition (6.24) is equivalent to t ¯ t ¯t CK + τ IPt A¯d Q−1 AtoKd IP + IPt AoKd + C¯K σ Ad IP + τ E2 Qσ E2 < 0 (6.29)

Using (6.13)-(6.14), premultiplying (6.29) by Y t = IP−t and postmultiplying the result by Y , we obtain: t ¯ t t ¯t CK Y + τ A¯d Q−1 Y t AtoKd + AoKd Y + Y t C¯K σ Ad + τ Y E2 Qσ E2 Y < 0 (6.30)

Expanding inequality (6.30) with the help of (6.11) and using Ko Yx = Z, LMIs (6.27) subject to (6.28) follow immediately. ∇∇∇ 6.3.2 Resilient H2 Design In the course of controller implementation, it turns out that the controllers are very sensitive with respect to errors in the controller coefficients [43]. The sources for this include, but not limited to, imprecision in analog-digital conversion, fixed word length, finite resolution instrumentation and numerical roundoff errors. By means of several examples, it is demonstrated [43] that relatively small perturbations in controller parameters could even destabilize the closed loop system. Such controllers are often termed ”fragile”. Hence, it is considered beneficial that the designed (nominal) controllers should be capable of tolerating some level of controller gain variations [16]. This illuminates the controller fragility problem for which some relevant results are available in [38] and further effort to alleviate this problem can also be found in [118, 121]. Looked at in this light, we consider the actual (implemented) control law be u(t) = [Ko + ∆K] x(t) , ∆K = H∆c (t)E

(6.31)

6.3 Feedback Control Design

177

where ∆c (t) is uncertainty matrix represented by ∆

∆c (t) ∈ ∆C (t) = {∆c (t) : ∆tc (t)∆c (t) ≤ I}

(6.32)

The closed-loop system becomes: (ΣDK ) : x(t) ˙ = σ(t) 0 = −σ(t) + A∆Kd x(t) − Ad



t

σ(s)ds + Γ w(t) (6.33) t−τ (t)

y(t) = C¯K x(t)

(6.34) (6.35)

where A∆Kd = Ao + Bo Ko + Ad + Bo H∆c E = Aod + Bo Ko + Bo H∆c E CˆK = Co + Do Ko + Do H∆c E (6.36) Extending on Theorem 6.3 for system (ΣDK ) and by the Schur complements, it follows that a sufficient condition for exponential stability can be expressed t ˆ t ¯t At∆Kd IP + IPt A∆Kd + CˆK CK + τ IPt A¯d Q−1 σ Ad IP + τ E2 Qσ E2 < 0 (6.37)

Inequality (6.37) holds if and only if the following inequality t ¯t At IP + IPt AtoKd + τ IPt A¯d Q−1 σ Ad IP + τ E2 Qσ E2   oKd −1 −1 t (ε +  )E E+ 0 t <0 [I − Do HH t Dot ]CK +  CK • εBo HH t Bot

(6.38)

holds for some scalars ε > 0,  > 0. Using (6.13)-(6.14), premultiplying (6.29) by Y t = IP−t and postmultiplying the result by Y , we obtain: t t ¯t Y t AtoKd + AoKd Y + τ A¯d Q−1 σ Ad + τ Y E2 Qσ E2 Y   (ε−1 + −1 )E t E+ 0 t t Y < 0 CK [I − Do HH t Dot ]CK +Y t t • εBo HH Bo

(6.39)

The following summarizes the main result. Theorem 6.4. The perturbed feedback gain (6.31) associated with the H2 optimal controller for system (ΣDK ) is given by Ko = Z Yx−1 for some scalars ε > 0,  > 0. where the matrices Yx = Yxt > 0, Yσ = Yσt > 0, Yd = Ydt > 0, satisfy the system of LMIs

178

6 Resilient Delay-Dependent Control



Yxt Cot t t t εYd Bo H Y  Λ1 Λ2 Do +Z   • −Λ3 τ Yσt 0 εYσt Bo H 0   • • −τ Qσ 0 0 0   • • • −I 0 0   • • • • −εI 0   • • • • • −Λ e   • • • • • • τ Ydt

Yxt Cot +Z t Dot 0 0 0 0 0 −I+ Do HH t Dot

        < 0      

(6.40)

for all admissible gain perturbations, where Λe = [εI,

I] , Y = [Yxt E t ,

Yxt E t ]

Proof: Upon expanding inequality (6.39) with the help of (6.11) and applying Inequality I.3 of the Appendix, then we use Ko Yx = Z. Thus LMIs (6.40) follows immediately. ∇∇∇ 6.3.3 Nominal H∞ Design In the sequel, for a prescribed γ > 0, we introduce the performance measure  ∞ ∆ J∞ = {z t (s)z(s) − γ 2 wt (s)w(s)} ds (6.41) 0

The problem of H∞ state-feedback control could be phrased as follows: Given system ΣJ determine the control law (6.7) which guarantees that H∞ performance measure (6.41) is bounded by γ for all w(t) ∈ L2 [0, ∞] The following two theorems summarize the main results pertaining to the H∞ performance: Theorem 6.5. Given a prescribed scalar γ > 0. State feedback controller (6.7) renders system (ΣDK ) exponentially stable with a disturbance attenuation level γ for all w(t) ∈ L2 [0, ∞) if, given matrix sequence Qx = Qtx > 0 there exist matrices Px = Pxt > 0, Pσ = Pσt > 0, Pd = Pdt > 0 satisfying the system of LMIs   ¯t Φ Ω1 + GtK GK Ω2 τ Pdt Ad IPt Γ¯ + G K  • −Ω3 τ Pσt Ad 0 ∆   < 0 (6.42) Πh =    0 • • −Qσ • • • −γ 2 I + Φt Φ Moreover

 ||z(t)|| ≤ γ 2 ||w(t)||2 + xt (0)Px (ηo )x(0)

+ τo

0

−τo

1/2 σ t (s)Qσ (ηo )σ(s)ds

(6.43)

6.3 Feedback Control Design

179

Proof: The exponential stability follows as a result of Theorem 6.1. To show that system (ΣDKo ) has a disturbance attenuation γ, we let the Lyapunov functional V (t) be given by (6.18). Using (6.13)-(6.14) and (6.24), we get   t t ¯ t t −1 ¯t t ˙ Vh = χ (t) IP AoKd + AoKd IP + τ IP Ad Qσ Ad IP + τ E2 Qσ E2 χ(t) + χt (t)IPΓ¯ w(t) + wt (t)Γ¯ t IPχ(t)

(6.44)

 Γ ¯ ¯ K = [GK 0] Γ = , G 0

where

It follows from (6.44) that: z t (t)z(t) − γ 2 wt (t)w(t) + V˙ h     t IPt AoKd + AtoKd IP t ¯ ¯ χ(t) χ(t)  IP Γ + G Φ t ¯ −1 t t K ¯  = +τ IP Ad Qσ Ad IP + τ E2 Qσ E2 w(t) w(t) • −γ 2 I + Φt Φ   t χ(t) χ(t) ∆ (6.45) = Πe w(t) w(t) By the Schur complements, it is easy to show from (6.42) that Πe < 0. Therefore from (6.41), we have: J∞ ≤ Vh (x0 ) =⇒ 2

2

2

||z(t)|| − γ ||w(t)|| ≤

xt0 Px x0

0

+τ −τ

φt (β)Qx φ(β)dβ =⇒

||z(t)||2 ≤ γ 2 ||w(t)||2 + xt0 Px x0

0 +τ φt (β)Qx φ(β)dβ

(6.46)

−τ

which completes the proof.

∇∇∇

Theorem 6.6. Given a prescribed constant γ > 0. The feedback gain associated with the H∞ -controller for system (ΣDK ) is given by Ko = Z Yx−1 where the matrices Yx = Yxt > 0, Yσ = Yσt > 0, Yd = Ydt > 0 , satisfy the system of LMIs   Λ1 Λ2 τ Ydt Yxt Gto + Z t Fot Γ + Y t Gto Φ  • −Λ3 τ Yσt  0 0    • • −τ Qσ  < 0 0 0 (6.47)    • •  • −I 0 • • • • −γ 2 I + Φt Φ Proof: Follows from parallel development to Theorem 6.3.

∇∇∇

180

6 Resilient Delay-Dependent Control

6.3.4 Resilient H∞ Design Following the procedure of resilient H2 design, we develop hereafter a tractable method for resilient H∞ design. The following summarizes the main result. Theorem 6.7. Given a prescribed scalar γ > 0, the perturbed feedback gain (6.31) associated with the H∞ -optimal controller for system (ΣDK ) is given by Ko = Z Yx−1 for some scalars ε > 0,  > 0. where the matrices Yx = Yxt > 0, Yσ = Yσt > 0, Yd = Ydt > 0, satisfy the system of LMIs   t t Γ+ Yxt Cot + t Yx Go + t εYd Bo H Y   Λ1 Λ2 τ Yd Z t Dot Z t Fot Y t Gto Φ   t t   • −Λ3 τ Yσ B H 0 0 0 0 εY σ o     • • −τ Qσ 0 0 0 0 0     • • • −I 0 0 0 0   2  < 0 (6.48)  −γ I+   • • • • 0 0 0 t   Φ Φ     • • • • • −εI 0 0     • • • • • • −Λ 0 e     −I+ • • • • • • • Do HH t Dot for all admissible gain perturbations Proof: Upon expanding inequality (6.47) with the help of (6.11) and (6.32)(6.34) in the manner of Theorem 6.4 and using Ko Yx = Z, LMI (6.48) follows immediately. ∇∇∇ 6.3.5 Simultaneous Nominal H2 /H∞ Design Having developed two separate nominal design methods, one is obviously tempted to combine both methods in one way or another. This can successfully be achieved by a simultaneous H2 /H∞ design approach. Adopting one approach, the problem of simultaneous H2 /H∞ state-feedback control could be phrased as follows: Given system ΣJ determine a control law of the type (6.7) which achieves the minimal value of H2 performance measure while guaranteeing that H∞ performance measure is bounded by γ for all w(t) ∈ L2 [0, ∞]. Remark 6.8. It should be observed that the objective of the simultaneous H2 /H∞ feedback control under consideration is to minimize the energy of the output y(t) while simultaneously satisfying the prescribed H∞ -norm bound of the controlled system w(t) → z(t). By exploiting the benefits of the foregoing results, the solution to the simultaneous nominal H2 /H∞ control problem is established by the following theorem:

6.3 Feedback Control Design

181

Theorem 6.9. Given a prescribed scalar γ > 0. The feedback controller (6.7) with gain Ko = Z Yx−1 is a simultaneous H2 /H∞ controller satisfying the performance measure (6.41) for system (ΣDK ) if there exist matrices Y = Y t > 0, Z, L = Lt > 0, W = W t > 0, such the system of generalized eigenvalue problems  min λ + T r(W ) 

−L I I −Qx

subject to (6.27), (6.47) and   −λ φt (0) −W X t <0, <0, <0 φ(0) −Yx X −L

(6.49)

has a feasible solution. Proof: On observing that xt (0)Px x(0) = λ =⇒ −λ + φt (0)Yx−1 φ(0) < 0 ∆

(6.50)

and in similar way using the cyclic properties of matrix trace [9]  

0

0 xt (s)Q(ηo )x(s)ds = T r xt (s)Q(ηo )x(s) ds = T r XX t L−1 −τo −τo  t −1 (6.51) = T r X L X < T r(W ) =⇒ −W + X t L−1 X < 0 0 where XX t = −τo xt (s)x(s)ds. Achieving the objective of simultaneous H2 /H∞ control leads to the convex minimization in (6.49) subject to (6.27)(6.47). ∇∇∇ 6.3.6 Simultaneous Resilient H2 /H∞ Design Taking gain perturbations into account, we have the following result Theorem 6.10. Given a prescribed scalar γ > 0. The feedback controller (6.31) with gain Ko = Z Yx−1 is a simultaneous resilient H2 /H∞ controller satisfying the performance measure (6.41) for system (ΣDK ) if there exist matrices Y = Y t > 0, Z, L = Lt > 0, W = W t > 0 such that the system of generalized eigenvalue problems  min λ + T r(W ) 

−L I I −Qx

subject to (6.40), (6.48) and   −λ φt (0) −W X t <0, <0, <0 φ(0) −Yx X −L

has a feasible solution.

(6.52)

182

6 Resilient Delay-Dependent Control

6.3.7 Example 6.1 In order to illustrate Theorem (6.9), we consider a pilot-scale single-reach water quality system which can fall into the type (6.1)-(6.3) with the following associated date : 

  −0.2 0 −0.1 0 1 , Ad = , Γ = 0 −0.1 −0.1 −0.1 2    20 10 10 Bo = , Do = , Co = 02 01 01 Ao =

Go = [0.2 0.1] , Φ = 0.5 , Fo = [2 1] Invoking the software environment [26], we solve the system of LMIs (6.49) using   0.4 t 20 e Qσ = , φ = 0.4 t , τ = 0.6 e 02 The feasible solution is given by:    0.9765 0.2240 2.0065 1.1213 2.2354 1.1532 Yx = , Yd = , Yσ = 0.2240 3.1367 1.1213 3.2524 1.1532 4.0422   1.0124 −0.2341 5.2054 2.1443 Z = , W = −0.4125 1.0332 2.1443 4.7805 λ = 1.2154 , γ = 2.1275 The feedback gain and the associated performance bound are  0.5661 −0.2194 K= , J = 11.2013 −0.3710 0.3614 This show the resilient controllers renders the water quality model asymptotically with a disturbance attenuation level of γ = 2.1275. 6.3.8 Example 6.2 Here we try another operating point of the model examined in Example 6.1. It is described by the following information : 

 −2 −1 0 Ao = , Ad = 0 −2 1   10 10 Bo = , Co = 01 02

 2 , Γ = 1  10 , Do = 01

1 0



Go = [0.1 0.1] , Φ = 0.3 , Fo = [1 2]

6.3 Feedback Control Design

183

Similarly, we solve the system of LMIs (6.49) using   0.4 t 10 e , φ = 0.4 t , τ = 0.65 Qσ = = e 02 This yields the feasible solution:   1.3264 0.7443 5.2064 1.4423 Yx = , Yσ = 0.7443 2.8987 1.4423 3.8605  1.6554 1.4142 Yd = 1.4142 6.8055   1.4564 −0.6214 5.2054 2.1443 Z = , W = −0.7250 1.2112 2.1443 4.7805 λ = 1.3461 , γ = 3.6345 The feedback gains and the associated performance bound are  0.3618 −0.2961 K= , J = 13.0542 −0.2533 0.4080 6.3.9 Example 6.3 Examining a different operating point of pilot-scale single-reach water quality system, we consider the date: 

  −1.9 0 −0.9 0 1 Ao = , Ad = , Γ = 0 −1 −1 −1.1 1    20 20 10 Bo = , Co = , Do = 01 02 01 Go = [0.2 0.2] , Φ = 0.2 , Fo = [1 1] We solve the system of LMIs (6.49) using   0.4 t 20 e , φ = 0.4 t , τ = 0.7 Qσ = e 01 The feasible solution is given by:   7.8095 2.4729 2.1134 1.7944 Yx = , Yσ = 2.4729 6.1265 1.7944 5.2634  1.7791 0.7988 Yd = 0.7988 2.275   2.1004 −0.8333 5.2054 2.1443 Z = Z(3) = , W = −0.2975 1.2332 2.1443 4.7805 λ = 2.0825 , γ = 1.8954

184

6 Resilient Delay-Dependent Control

The feedback gains and the associated performance bound are  0.3577 −0.2804 K= , J = 10.8375 −0.1168 0.2484

6.4 Continuous Polytopic Systems In this section, we consider the robust counterparts of Theorems 6.1-6.9 for the system (ΣP ) (ΣP ) :

x(t) ˙ = Ao x(t) + Ad x(t − τ ) + Bo u(t) + Γ w(t) y(t) = Co x(t) + Cd x(t − τ )

(6.53)

z(t) = Do x(t) + Fo u(t) + Φw(t)

(6.55)

(6.54)

where x ∈ IRn is the system state, u(t) ∈ IRp is the control input, y(t) ∈ IRp is the measured output, z(t) ∈ IRr is the controlled output and τ is a time-delay factor such that 0 < τ ≤ β , 0 < τ˙ ≤ α < 1. The corresponding state-space matrices contain uncertainties represented by a real convex bounded polytopic model of the type    A Ao Ad Bo (oλ) A(dλ) B(oλ) ∆  C(oλ) C(dλ) 0  =  Co Cd 0  ∈ Sλ = Do 0 Fo D(oλ) 0 F(oλ)    N Aoj Adj Boj  (6.56) λj  Coj Cdj 0  , λ ∈ Λ j=1 Doj 0 Foj where Λ is the unit simplex   N  ∆ Λ = (λ1 , · · · , λN ) : λ j = 1 , λj ≥ 0

(6.57)

j=1

Define the vertex set IN = {1, ..., N }. Throughout this section, we use {Ao , · · · , Fo } to imply generic system matrices and {Aoj , · · · , Foj , j ∈ IN} to represent the respective values at the vertices. The initial condition is specified as x(0), x(s) = xo , φ(s) , where φ(.) ∈ L2 [−τ, 0]. Under the control law (6.7) and invoking the descriptor transformation, system (ΣP ) becomes: (ΣP K ) : x(t) ˙ = σ(t) 0 = −σ(t) + AP Kd x(t) − AP d



t

σ(s)ds t−τ (t)

+ Γ w(t) y(t) = CP K x(t)

(6.58) (6.59)

z(t) = GP K x(t) + Φw(t)

(6.60)

6.4 Continuous Polytopic Systems

185

where AP Kd = Aoj + Boj Ko + Adj = Aodj + Boj Ko ¯ P K = [GP K 0] CP K = Coj + Doj Ko , GP K = Goj + Foj Ko , G   0 I 0 AP Kd = , C¯P K = [CP K 0] , AP d = (6.61) AP Kd −I Adj In line of the nominal design methods, we have the corresponding results which are listed without proof: 6.4.1 Polytopic H2 Design Much in line with the analytical development in section 5.3.2, the corresponding design results are summarized by the following theorems Theorem 6.11. In the absence of input disturbance w(t) ≡ 0, controller (6.7) is an H2 -optimal controller for the polytopic system (ΣP K ) minimizing the H2 -performance measure (6.12) if, given matrix sequence Qσ = Qtσ > 0, there exist matrices Px = Pxt > 0, Pσ = Pσt > 0, Pd = Pdt > 0, satisfying the system of LMIs   Ω1j + CPt K CP K Ω2j τ Pdt Adj ∆ • −Ω3j τ Pσt Adj  < 0 , j ∈ IN (6.62) Πt =  • • −Qσ where Ω1j = Pdt AP Kd + AtP Kd Pd , Ω2j = Pxt − Pdt + AtP Kd Pσ Ω3j = Pσt + Pσ − τ Qσ

(6.63)

An upper bound on the H2 performance measure is given by 

∆ J2 ≤ J+ = xt (0)Px x(0) + τo

0

−τo

1/2 σ t (s) Qσ σ(s) ds

(6.64)

Theorem 6.12. The feedback gain associated with the H2 -optimal controller for the polytopic system (ΣP K ) is given by Ko = Z Yx−1 where the matrices Yx = Yxt > 0, Yσ = Yσt > 0, Yd = Ydt > 0, satisfy the system of LMIs   t t Λ1 Λ2j τ Ydt Yxt Coj + Z t Doj  • −Λ3j τ Y t  0 σ   < 0 , j ∈ IN (6.65)  •  0 • −τ Qσ • • • −I where t Λ2j = Yσ − Ydt + Yxt Atodj + Z t Boj , Λ3j = Yσt + Yσ − τ Adj Q−1 σ Adj (6.66)

186

6 Resilient Delay-Dependent Control

Theorem 6.13. The perturbed feedback gain (6.31) associated with the H2 optimal controller for the polytopic system (ΣP K ) is given by Ko = Z Yx−1 for some scalars ε > 0,  > 0. where the matrices Yx = Yxt > 0, Yσ = Yσt > 0, Yd = Ydt > 0, satisfy the system of LMIs   t t Yxt Coj Yxt Coj t t t εYd Boj H Y t  Λ1 Λ2j τ Yd +Z t Doj  +Z t Doj   t  • −Λ3j τ Y t  0 εY B H 0 0 σ σ oj    •  • −τ Q 0 0 0 0 σ    •  < 0 (6.67) • • −I 0 0 0    •  • • • −εI 0 0    •  • • • • −Λ 0 e     −I+ • • • • • • t Doj HH t Doj for all admissible gain perturbations 6.4.2 Polytopic H∞ Design The main design results are summarized by the following theorems: Theorem 6.14. Given a prescribed scalar γ > 0. State feedback controller (6.7) renders the polytopic system (ΣP K ) exponentially stable with a disturbance attenuation level γ for all w(t) ∈ L2 [0, ∞) if, given matrix sequence Qx = Qtx > 0 there exist matrices Px = Pxt > 0, Pσ = Pσt > 0, Pd = Pdt > 0 satisfying the system of LMIs   ¯t Φ Ω1j + GtP K GP K Ω2j τ Pdt Adj IPt Γ¯ + G PK  • −Ω3j τ Pσt Adj 0 ∆   < 0 (6.68) Πh =    0 • • −Qσ • • • −γ 2 I + Φt Φ Moreover

 ||z(t)|| ≤ γ 2 ||w(t)||2 + xt (0)Px (ηo )x(0)

+ τo

0

−τo

1/2 t

σ (s)Qσ (ηo )σ(s)ds

(6.69)

Theorem 6.15. Given a prescribed constant γ > 0. The feedback gain associated with the H∞ -controller for system (ΣP K ) is given by Ko = Z Yx−1 where the matrices Yx = Yxt > 0, Yσ = Yσt > 0, Yd = Ydt > 0 , satisfy the system of LMIs   t Γ + Y t Gtoj Φ Λ1 Λ2j τ Ydt Yxt Gtoj + Z t Foj  • −Λ3j τ Y t  0 0 σ    •  < 0 (6.70) • −τ Qσ 0 0    •  • • −I 0 • • • • −γ 2 I + Φt Φ

6.4 Continuous Polytopic Systems

187

Theorem 6.16. Given a prescribed scalar γ > 0, the perturbed feedback gain (6.31) associated with the H∞ -optimal controller for system (ΣP K ) is given by Ko = Z Yx−1 for some scalars ε > 0,  > 0. where the matrices Yx = Yxt > 0, Yσ = Yσt > 0, Yd = Ydt > 0, satisfy the system of LMIs   t Yxt Coj Yxt Gtoj Γ+ t t t   Λ1 Λ2j τ Yd +Z t F t oj Y t Gt ojΦ εYd Boj H Y +Z t Doj   t   • −Λ3j τ Y t 0 0 εY B H 0 0 σ σ oj     • • −τ Q 0 0 0 0 0 σ     • • • −I 0 0 0 0   2 t   • • • • −γ I + Φ Φ 0 0 0     • • • • • −εI 0 0     • 0 • • • • • −Λ e     −I+ • • • • • • • t Doj HH t Doj (6.71) ;< 0

(6.72)

for all admissible gain perturbations 6.4.3 Simultaneous Polytopic H2 /H∞ Design By similarity to the nominal design methods, we provide below the polytopic counterpart of the simultaneous H2 /H∞ design approach. Keeping the same pace, the main results are summarized by the following theorems: Theorem 6.17. Given a prescribed scalar γ > 0. The feedback controller (6.7) with gain Ko = Z Yx−1 is a simultaneous H2 /H∞ controller satisfying the performance measure (6.41) for system (ΣP K ) if there exist matrices Y = Y t > 0, Z, L = Lt > 0, W = W t > 0, such the system of generalized eigenvalue problems  min λ + T r(W ) 

subject to (6.65), (6.70) and   −λ φt (0) −W X t −L I <0, <0, <0 I −Qx X −L φ(0) −Yx

(6.73)

has a feasible solution. Theorem 6.18. Given a prescribed scalar γ > 0. The feedback controller (6.31) with gain Ko = Z Yx−1 is a simultaneous resilient H2 /H∞ controller satisfying the performance measure (6.41) for system (ΣP K ) if there exist matrices Y = Y t > 0, Z, L = Lt > 0, W = W t > 0 such that the system of generalized eigenvalue problems

188

6 Resilient Delay-Dependent Control

 min λ + T r(W ) 

subject to (6.67), (6.72) and   −L I −λ φt (0) −W X t <0, <0, <0 I −Qx φ(0) −Yx X −L

(6.74)

has a feasible solution. Remark 6.19. It should be noted that Theorems 6.9 and 6.17, for the nominal and uncertain models respectively, present design procedures to compute the simultaneous H2 /H∞ controller as the solution of a convex minimization problem which can conveniently solved by the MATLAB software system [26]. 6.4.4 Example 6.4 In order to illustrate Theorem 6.18, we consider a second order system with the following data    0.5 −1 1 + δs −0.1 0 Ao = , Γ = , Ad = 0.5 −3 −2 0.3 −0.2 + δd    0.9 0 10 0.5 0 , Co = , Do = Bo = 0 0.9 01 0 0.5 Go = [0.3 0.2] , Φ = 0.4 , Fo = [2 2] , τ = 0.55 −0.4 ≤ δs ≤ 0.4 , −0.2 ≤ δd ≤ 0.2 In simulations, the vertex set Λ is formed as Λ = {j, k} : (j, k) ∈ [−0.4, −0.3, ..., 0.3, 0.4] ∩ [−0.2, −0.1, ..., 0.1, 0.2] The feasible solution is summarized by   0.8532 0.9260 4.9154 1.3553 K = , W = −1.4317 −1.2628 1.3553 4.1105 λ = 2.7154 , γ = 1.35 , J = 11.7413

6.5 Discrete-Time Systems In this section, we focus attention on the development of resilient delaydependent design results for a class of discrete-time systems. 6.5.1 System Description Consider the following class of discrete-time state-delay systems with parametric uncertainties:

6.5 Discrete-Time Systems

xk+1 = A∆o xk + Bo uk + A∆d xk−dk , x0 = ψ0

189

(6.75)

where xk ∈ IRn is the state vector; uk ∈ IRp is the control input, dk is a positive number representing the delay such that dk ≤ d+ and the uncertain matrices A∆ ∈ IRn×n and A∆d ∈ IRn×n , are represented by [A∆o A∆d ] = [Ao Ad ] + M ∆k [Na Nd ]

(6.76)

where Ao ∈ IRn×n , Bo ∈ IRn×p , Co ∈ IRq×n , Ad ∈ IRn×n , M ∈ IRn×α , Na ∈ IRβ×n and Nd ∈ IRβ×n , are real and known constant matrices with ∆k ∆ is a bounded matrix of uncertainties such that ∆k ∈ ∆k = {∆tk : ∆k ∆k < I , ∀ k}. Remark 6.20. The class of systems (6.77) emerges in many areas dealing with functional difference equations or delay-difference equations [35]. On the application side, these systems appear in cold rolling mills [51] and decision-making of manufacturing systems [112]. It should be stressed that we consider only the case of single-delay since multiple-delay systems can be easily handled using the augmentation procedure