Communication and Networking, Part II - FGCN 2011

Communications in Computer and Information Science

266

Tai-hoon Kim Hojjat Adeli Wai-chi Fang Thanos Vasilakos Adrian Stoica Charalampos Z. Patrikakis Gansen Zhao Javier García Villalba Yang Xiao (Eds.)

Communication and Networking International Conference, FGCN 2011 Held as Part of the Future Generation Information Technology Conference, FGIT 2011 in Conjunction with GDC 2011 Jeju Island, Korea, December 8-10, 2011 Proceedings, Part II

13

Volume Editors Tai-hoon Kim Hannam University, Daejeon, Korea E-mail: [email protected] Hojjat Adeli The Ohio State University, Columbus, OH, USA E-mail: [email protected] Wai-chi Fang National Chiao Tung University, Hsinchu, Taiwan, R.O.C. E-mail: [email protected] Thanos Vasilakos University of Western Macedonia, Kozani, Greece E-mail: [email protected] Adrian Stoica Jet Propulsion Laboratory, Pasadena, CA, USA E-mail: [email protected] Charalampos Z. Patrikakis National Technical University of Athens, Greece E-mail: [email protected] Gansen Zhao Sun Yat-sen University, Guangzhou, China E-mail: [email protected] Javier García Villalba Universidad Complutense de Madrid, Spain E-mail: [email protected] Yang Xiao University of Alabama, Tuscaloosa, AL, USA E-mail: [email protected] ISSN 1865-0929 e-ISSN 1865-0937 ISBN 978-3-642-27200-4 e-ISBN 978-3-642-27201-1 DOI 10.1007/978-3-642-27201-1 Springer Heidelberg Dordrecht London New York Library of Congress Control Number: Applied for CR Subject Classification (1998): C.2, H.4, I.2, H.3, D.2, F.1 © Springer-Verlag Berlin Heidelberg 2011 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, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. 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: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)

Foreword

Future generation communication and networking is an area that attract many professionals from academia and industry for research and development. The goal of the FGCN conference is to bring together researchers from academia and industry as well as practitioners to share ideas, problems and solutions relating to the multifaceted aspects of future-generation communication and networking. We would like to express our gratitude to all of the authors of submitted papers and to all attendees for their contributions and participation. We acknowledge the great effort of all the Chairs and the members of Advisory Boards and Program Committees of the above-listed event. Special thanks go to SERSC (Science and Engineering Research Support Society) for supporting this conference. We are grateful in particular to the speakers who kindly accepted our invitation and, in this way, helped to meet the objectives of the conference. December 2011

Chairs of FGCN 2011

Preface

We would like to welcome you to the proceedings of the 2011 International Conference on Future Generation Communication and Networking (FGCN 2011) — one of the partnering events of the Third International Mega-Conference on Future-Generation Information Technology (FGIT 2011) held during December 8–10, 2011, at Jeju Grand Hotel, Jeju Island, Korea. FGCN 2011 focused on various aspects of advances in future-generation communication and networking. It provided a chance for academic and industry professionals to discuss recent progress in the related areas. We expect that the conference and its publications will be a trigger for further related research and technology improvements in this important subject. We would like to acknowledge the great effort of the FGCN 2011 Chairs, International Advisory Board, Committees, Special Session Organizers, as well as all the organizations and individuals who supported the idea of publishing this volume of proceedings, including the SERSC and Springer. We are grateful to the following keynote, plenary and tutorial speakers who kindly accepted our invitation: Hsiao-Hwa Chen (National Cheng Kung University, Taiwan), Hamid R. Arabnia (University of Georgia, USA), Sabah Mohammed (Lakehead University, Canada), Ruay-Shiung Chang (National Dong Hwa University, Taiwan), Lei Li (Hosei University, Japan), Tadashi Dohi (Hiroshima University, Japan), Carlos Ramos (Polytechnic of Porto, Portugal), Marcin Szczuka (The University of Warsaw, Poland), Gerald Schaefer (Loughborough University, UK), Jinan Fiaidhi (Lakehead University, Canada) and Peter L. Stanchev (Kettering University, USA), Shusaku Tsumoto (Shimane University, Japan), Jemal H. Abawajy (Deakin University, Australia). Last but not the least, we give special thanks to Ronnie D. Caytiles and Yvette E. Gelogo of the graduate school of Hannam University, who contributed to the editing process of this volume with great passion. We would like to express our gratitude to all of the authors and reviewers of submitted papers and to all attendees, for their contributions and participation, and for believing in the need to continue this undertaking in the future. December 2011

Tai-hoon Kim Hojjat Adeli Wai-chi Fang Thanos Vasilakos Adrian Stoica Charalampos Z. Patrikakis Gansen Zhao Javier Garc´ıa Villalba Yang Xiao

Organization

Honorary Chair Dae-sik Ko

Mokwon University, Korea

General Co-chairs Wai-chi Fang Thanos Vasilakos Adrian Stoica

National Chiao Tung University, Taiwan University of Western Macedonia, Greece NASA JPL, USA

Program Co-chairs Charalampos Z. Patrikakis Gansen Zhao Javier Garc´ıa Villalba Tai-hoon Kim Yang Xiao

National Technical University of Athens, Greece Sun Yat-sen University, China Universidad Complutense of Madrid, Spain GVSA and University of Tasmania, Australia University of Alabama, USA

Workshop Chair Byungjoo Park

Hannam University, Korea

Publicity Co-chairs Houcine Hassan Damien Sauveron Qun Jin Irfan Awan Muhammad Khurram Khan Yang Xiao J.H. Abawajy

Polytechnic University of Valencia, Spain University of Limoges, France Waseda University, Japan University of Bradford, UK King Saud University, Saudi Arabia The University of Alabama, USA Deakin University, Australia

Publication Chair Maria Lee

Shih Chien University, Taiwan

X

Organization

International Advisory Board Hsiao-Hwa Chen Gansen Zhao Han-Chieh Chao Hamid R. Arabnia Gongzhu Hu Byeong-Ho Kang Aboul Ella Hassanien Tughrul Arslan Jianhua Ma Sankar K. Pal Xiaofeng Song

National Sun Yat-Sen University, Taiwan Sun Yat-sen University, China National Ilan University, Taiwan The University of Georgia, USA Central Michigan University, USA University of Tasmania, Australia Cairo University, Egypt The University of Edinburgh, UK Hosei University, Japan Indian Statistical Institute, India Nanjing University of Aeronautics and Astronautics, China Oslo University College, Norway

Frode Eika Sandnes

Program Committee Aboul Ella Hassanien Ai-Chun Pang Aggeliki Sgora Albert Banchs Andres Iglesias Prieto Andrzej Jajszczyk Antonio Lagana Benahmed Khelifa Bogdan Ghita Chao-Tung Yang Chia-Chen Lin Christophe Fouquer´e Chu-Hsing Lin Clement Leung Damien Sauveron Dimitrios D. Vergados Don-Lin Yang Driss Mammass Farrukh A. Khan Gianluigi Ferrari Hong Sun Hsiang-Cheh Huang

Hsin-Hung Chou Hui Chen Huirong Fu J. Vigo-Aguiar Janusz Szczepanski Jiann-Liang Jieh-Shan George Yeh Jiming Chen Juha Ro”ning Kin Keung Lai Kwok-Yan Lam Li Shijian Luis Javier Garc´ıa Villalba Marc Lacoste Matthias Reuter Michel-Marie Deza Ming-Yen Lin Feng Mohammad Moghal Nashwa El-Bendary Neveen I. Ghalii Nikolaos Pantazis

Special Session Organizers Hong Kook Kim Tae-Young Byun Y. Byun

N. Jaisankar Ning Gui Omar Soluiman P.R. Parthasarathy Ricky Yu-Kwong Kwok Robert Goutte Rui L. Aguiar Shun-Ren Yang Soon Ae Chun Stephen Huang Sun-Yuan Hsieh Tae (Tom) Oh Terence D. Todd Victor C.M. Leung Viktor Yarmolenko Vincent Oria Vincenzo De Florio Weili Han Witold Pedrycz Yung-Hui Li Feng Yvette E. Gelogo Ronnie D. Caytiles

Table of Contents – Part II

Studies on the Key Technologies of Multi-Platform Mobile Thin Client System: Cross-Layer Isolation and Session Allocation . . . . . . . . . . . . . . . . . Biao Song, Wei Tang, Tien-Dung Nguyen, Sang-Ho Na, Jun-Hyung Lee, and Eui-Nam Huh

1

LDPC Equalizer for Compensating the CFO and Phase Noise in OFDM System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Do-Hoon Kim and Heung-Gyoon Ryu

11

TC-HMIPv6: A Study of HMIPV6 Handover Management for Packet Transmission Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sung-Gyu Kim, Farkhod Alisherov, and Byungjoo Park

20

A Multi-hop Communication Scheme for IEEE 802.11p Based V2V Communication Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Woong Cho and Hyun Seo Oh

26

A Political Communication Scheme of Citizen Network System on Disembedding and Embedding Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jang-Mook Kang and Bong-Hwa Hong

34

Web Contents Mining System for Real-Time Monitoring of Opinion Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ho-Bin Song, Moon-Taek Cho, Young-Choon Kim, and Suck-Joo Hong An Energy-Efficient Cluster-Based Routing in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Seongsoo Cho, Bhanu Shrestha, Keuk-Hwan La, Bong-Hwa Hong, and Jongsup Lee A Management of Resource Ontology for Cloud Computing . . . . . . . . . . . Hwa-Young Jeong and Bong-Hwa Hong

43

57

65

Development of an Algorithm for Video Quality Measurement for Broadcasting Communications Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sang-Soo Kim, Hae-Jong Joo, and Euy-Soo Lee

73

An Effective Resource Managements Method Using Cluster-Computing for Cloud Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Seong-Sik Hong and Jin-Mook Kim

83

XII

Table of Contents – Part II

Study on Micro-processing Implementation of USN Environment Data by a Graphic-Based Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Young-Wook Lee

90

Web Based Remote Robot Control for Adjusting Position on Manufacturing System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hwa-Young Jeong and Bong-Hwa Hong

96

Discrimination of Speech Activity and Impact Noise Using an Accelerometer and a Microphone in a Car Environment . . . . . . . . . . . . . . . Seon Man Kim, Hong Kook Kim, Sung Joo Lee, and Yun Keun Lee

104

Crosstalk Cancellation for Spatial Sound Reproduction in Portable Devices with Stereo Loudspeakers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sung Dong Jo, Chan Jun Chun, Hong Kook Kim, Sei-Jin Jang, and Seok-Pil Lee Perceptual Enhancement of Sound Field Reproduction in a Nearly Monaural Sensing System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chan Jun Chun, Hong Kook Kim, Seung Ho Choi, Sei-Jin Jang, and Seok-Pil Lee

114

124

Quality-Aware Loss-Robust Scalable Speech Streaming Based on Speech Quality Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jin Ah Kang, Seung Ho Choi, and Hong Kook Kim

132

Artificial Bandwidth Extension of Narrowband Speech Signals for the Improvement of Perceptual Speech Communication Quality . . . . . . . . . . . Nam In Park, Young Han Lee, and Hong Kook Kim

143

Improvements in Howling Margin Using Phase Dispersion . . . . . . . . . . . . . Jae-Won Lee and Seung Ho Choi Secure Client-Side Digital Watermarking Using Optimal Key Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jing-Jing Jiang and Chi-Man Pun Effective Electronic Advertisement Auction System . . . . . . . . . . . . . . . . . . . Tokuro Matsuo and Satoshi Takahashi

154

162

169

Energy-Efficient Fire Monitoring Protocol for Ubiquitous Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Heemin Kim, Ae-cheoun Eun, Sunyoung Han, and Young-guk Ha

179

Design of Optimal Combination for New and Renewable Hybrid Generation System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kun Hyun Park, Chul Uoong Kang, Gi Min Lee, and Jong Hwan Lim

189

Table of Contents – Part II

Parameter Optimization of UWB Short Range Radar Detector for Velocity Measurement in Automobile Applications . . . . . . . . . . . . . . . . . . . Purushothaman Surendran, Chul-Ung Kang, and Seok-Jun Ko Data Signature-Based Time Series Traffic Analysis on Coarse-Grained NLEX Density Data Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Reynaldo G. Maravilla Jr., Elise A. Tabanda, Jasmine A. Malinao, and Henry N. Adorna Automated Video Surveillance for Monitoring Intrusions Using Intelligent Middleware Based on Neural Network . . . . . . . . . . . . . . . . . . . . . Ana Rhea Pangapalan, Bobby D. Gerardo, Yung-Cheol Byun, Joel T. De Castro, and Francisca D. Osorio

XIII

199

208

220

SMS-Based Automatic Billing System of Household Power Consumption Based on Active Experts Messaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mark Dominic Cabioc, Bobby D. Gerardo, Yung-Cheol Byun

229

Hierarchical Clustering and Association Rule Discovery Process for Efficient Decision Support System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bobby D. Gerardo, Yung-Cheol Byun, and Bartolome Tanguilig III

239

Implementation of Energy Efficient LDPC Code for Wireless Sensor Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sang-Min Choi and Byung-Hyun Moon

248

A Multi-layered Routing Protocol for UWSNs Using Super Nodes . . . . . . Abdul Wahid, Dongkyun Kim, and Kyungshik Lim Experimental Measurement for EVM Performance Enhancement of Wireless Repeater System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Daesik Ko and Hwase Park

258

268

Power Model and Analysis of Wireless Transceiver System . . . . . . . . . . . . Jae-Hoon Choi and Heung-Gyoon Ryu

274

Feedback Scheduling for Realtime Task on Xen Virtual Machine . . . . . . . Byung Ki Kim, Kyung Woo Hur, Jae Hyuck Jang, and Young Woong Ko

283

DTAR: Deduplication TAR Scheme for Data Backup System . . . . . . . . . . Sung Woon Kang, Ho Min Jung, Jung Geun Lee, Jin Haeng Cho, and Young Woong Ko

292

Effect of Maximum Node Velocity on GA-Based QOS Routing Protocol (QOSRGA) for Mobile Ad Hoc Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jiwa Abdullah

301

XIV

Table of Contents – Part II

Application of Wireless Accelerometer System for Evaluating Osteoarthritis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dong Rak Kwon and Ho-Cheol Lee

312

A Performance Evaluation of a Novel Clustering Scheme Considering Local Node Density over WSN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jeong-Sam Kim and Tae-Young Byun

320

Performance Analysis of DRAM-SSD and HDD According to the Each Environment on MYSQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hyun-Ju Song, Young-Hun Lee, and Seung-Kook Cheong

330

Dynamic Channel Adjustable Asynchronous Cognitive Radio MAC Protocol for Wireless Medical Body Area Sensor Networks . . . . . . . . . . . . Byunghwa Lee, Jangkyu Yun, and Kijun Han

338

A Multiple-Metric Routing Scheme for QoS in WMNs Using a System of Active Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jangkyu Yun, Byunghwa Lee, Junhyung Kim, and Kijun Han

346

Implementation of Log Analysis System for Desktop Grids and Its Application to Resource Group-Based Task Scheduling . . . . . . . . . . . . . . . Joon-Min Gil, Mihye Kim, and Ui-Sung Song

354

FM Subcarrier Multiplexing Using Multitone Modulation for Optical Coherent Communications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hae Geun Kim and Ihn-Han Bae

364

An Ontology-Based ADL Recognition Method for Smart Homes . . . . . . . Ihn-Han Bae and Hae Geun Kim

371

Analysis of User Preferences for Menu Composition and Functional Icons of E-Book Readers in a Smartphone Environment . . . . . . . . . . . . . . . Mihye Kim, Joon-Min Gil, and Kwan-Hee Yoo

381

Dynamic Transmission Target Selection Scheme for Load-Balancing in WSN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Seok-Yeol Heo, Wan-Jik Lee, and Won-Yeoul Lee

393

Organizing Virtual Research Groups with Light Path Technology . . . . . . Min-Ki Noh, Won-Hyek Lee, Seung-Hae Kim, and Joon-Min Gil Remote Monitoring Information Management System for Preventing Performance Degradation of Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Myung-Ju Kim, Un-Bai Lee, and Kwang Sik Chung Noise Reduction Scheme for Precise Indoor Localization . . . . . . . . . . . . . . Inseok Moon and Won-Kee Hong

403

412 419

Table of Contents – Part II

Development of a Korean Language-Based Augmentative and Alternative Communication Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chang-Geol Kim, Soo-Won Kwak, Ryu Juang Tak, and Byung-Seop Song

XV

429

Adaptive Power Management for Nanoscale SoC Design . . . . . . . . . . . . . . Jeong-Tak Ryu and Kyung Ki Kim

437

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

447

Table of Contents – Part I

Wireless Multimedia Sensor Networks Testbeds and State-of-the-Art Hardware: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Muhammad Omer Farooq and Thomas Kunz An Energy-Efficient Cluster-Based Routing in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Seongsoo Cho, Bhanu Shrestha, Keuk-Hwan La, Bonghwa Hong, and Jongsup Lee

1

15

Event-Driven Test Script Methodology for SOA System . . . . . . . . . . . . . . . Youngkon Lee

23

Business-Context Based SLA Parameters for SOA Management . . . . . . . . Youngkon Lee

31

Prescription-Level Based Test Assertion Model for SOA . . . . . . . . . . . . . . . Youngkon Lee

39

Multithreaded Power Consumption Scheduler Based on a Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Junghoon Lee, Gyung-Leen Park, and Hye-Jin Kim Design of a Composite Sensor Node in Agricultural Ubiquitous Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Junghoon Lee, Gyung-Leen Park, Hye-Jin Kim, Ho-Young Kwak, Seongjun Lee, Jong-Heon Lee, Bong-Soo Kang, and Yun-Hyuk Kim Millimetric Waves Technologies: Opportunities and Challenges . . . . . . . . . Jahangir Dadkhah Chimeh and Saeed Bashirzadeh Parapari

47

53

59

A Reduced Complexity Subcarrier Switching Scheme for PAPR Reduction in OFDM System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sabbir Ahmed and Makoto Kawai

67

An Indoor Location-Aware System Based on Rotation Sampling in Positioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chu-Hsing Lin, Jung-Chun Liu, Chien-Hsing Lee, and Tang-Wei Wu

77

Research on the ZigBee-Based Indoor Location Estimation Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chu-Hsing Lin, Jung-Chun Liu, Sheng-Hsing Tsai, and Hung-Yan Lin

82

XVIII

Table of Contents – Part I

Visible Watermarking Based on Multi-parameters Adjustable Gamma Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chu-Hsing Lin, Chen-Yu Lee, Tzu-Chien Yang, and Shin-Pin Lai

87

A Tree Overlapping-Based Mesh Routing Protocol for Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Inwhee Joe, Yeonyi Choi, and Dongik Kim

93

Performance Comparison among MIMO Techniques at Different Interference Levels for LTE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mohammad T. Kawser, Md.K. Syfullah, Nawshad U.A. Chowdhury, and Md.T. Hoq Handover Method Considering Power Consumption and Video Quality Satisfaction at the Mobile Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hyun Jong Kim and Seong Gon Choi

103

111

Shape Retrieval Combining Interior and Contour Descriptors . . . . . . . . . . Solima Khanam, Seok-Woo Jang, and Woojin Paik

120

Hardware Architecture of Bilateral Filter to Remove Haze . . . . . . . . . . . . . Eun-Kyoung Kim, Jae-Dong Lee, Byungin Moon, and Yong-Hwan Lee

129

An Efficient Interworking Architecture of a Network Processor for Layer 7 Packet Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kyeong-ryeol Bae, Seung-Ho Ok, Hyeon-Sik Son, Sang Yoon Oh, Yong-Hwan Lee, and Byungin Moon A Rectification Hardware Architecture for an Adaptive Multiple-Baseline Stereo Vision System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hyeon-Sik Son, Kyeong-ryeol Bae, Seung-Ho Ok, Yong-Hwan Lee, and Byungin Moon

136

147

Building Self-organizing Autonomic Agent Based on a Mobile Cell . . . . . Kiwon Yeom

156

Modeling u-Healthcare Frameworks Using Mobile Devices . . . . . . . . . . . . . Haeng-Kon Kim

166

Design and Implementation of Smart Meter Concentration Protocol for AMI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Byung-Seok Park, Cheoul-Shin Kang, and Young-Hun Lee Biologically-Inspired Optimal Video Streaming over Wireless LAN . . . . . Yakubu S. Baguda, Norsheila Fisal, Rozeha A. Rashid, Sharifah K. Yusof, Sharifah H. Syed, and Dahiru S. Shuaibu

179

188

Table of Contents – Part I

Emerging of Mobile Ad-Hoc Networks and New Generation Technology for Best QOS and 5G Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jahangir Khan, Zoran S. Bojkovic, and Muhammad Imran Khan Marwat Intelligent Hybrid Anomaly Network Intrusion Detection System . . . . . . . Heba F. Eid, Ashraf Darwish, Aboul Ella Hassanien, and Tai-hoon Kim Remote Data Acquisition and Touch-Based Control of a Mobile Robot Using a Smart Phone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yong-Ho Seo, Hyo-Young Jung, Chung-Sub Lee, and Tae-Kyu Yang The Performance Analysis of LT Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ling Yang, ShiLi Song, Wei Wei Su, Yi Fan Wang, and Hong Wen Unseen Visible Watermarking for Gray Level Images Based on Gamma Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chu-Hsing Lin, Chen-Yu Lee, Shu-Yuan Lu, and Shih-Pei Chien People Counting Using Object Detection and Grid Size Estimation . . . . . Oliver C. Agustin and Byung-Joo Oh The Research of Serially Concatenated FQPSK Demodulation Based on LDPC Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Gao Yuan Zhang, Hong Wen, Liang Zhou, Ling Yang, and Yi Fan Wang

XIX

198

209

219 227

236 244

254

Design and Implementation of Efficient Reed-Solomon Decoder for Intelligent Home Networking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ik Soo Jin

261

A Cost-Effective Multicasting Using an FP-LD Modulator in a WDM-PON . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hyuek Jae Lee

269

A Wireless CCTV Converter Based on Binary CDMA Technology . . . . . . Yeong-Jin Baek and Sang-Hoon Lee A Soft QoS Supporting Multi-path Routing Scheme for Mobile Nodes in MANETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tz-Heng Hsu, Yi-Da Li, and Meng-Shu Chiang An Adaptive Query Optimization in a Hierarchical Mediator System . . . Nam Hun Park and Kil Hong Joo

277

283 293

XX

Table of Contents – Part I

On Implementation of a KVM IaaS with Monitoring System on Cloud Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chao-Tung Yang, Bo-Han Chen, and Wei-Sheng Chen

300

RFID and Supply Chain Management: Generic and Military Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tae Hwan Oh, Young B. Choi, and Rajath Chouta

310

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

325

Studies on the Key Technologies of Multi-Platform Mobile Thin Client System: Cross-Layer Isolation and Session Allocation Biao Song, Wei Tang, Tien-Dung Nguyen, Sang-Ho Na, Jun-Hyung Lee, and Eui-Nam Huh Department of Computer Engineering Internet Computing and Network Security Lab KyungHee University Global Campus, South Korea {bsong,wtang,junhyung,johnhuh}@khu.ac.kr, {ntiendung,shna}@icns.khu.ac.kr

Abstract. Virtualization was considered as the best way to isolate independent thin client sessions on a physical machine. However, the hypervisor, guest OS and guest remote server not only consumes a considerable amount of memory but also degrades the processing power of CPU. In this paper, we propose a novel cross-layer isolation technology to support independent user sessions with only one OS and one remote server. Furthermore, a session allocation/migration algorithm is introduced in this paper. The algorithm solves the multi-user to multi-machine allocation/migration problem within thin client environment. Keywords: Thin client, Multi-user, Isolation, Allocation, VM Migration.

1

Introduction

Nowadays, the rapid development of network promotes the investigation of thin client (remote display) technology. Using thin client system, users are able to remotely control other computers (servers) and delegate actual information processing to them. Thus, thin client technology provides a powerful way to break the barrier between diverse applications and insufficient local hardware/software environment. For example, a mobile device with thin client system permits its user to use the applications running on different mobile platforms (Android, iOS, and so on) or PC platforms (Windows, Linux, and so on). In our previous paper [1], we introduced Multi-platform Mobile Thin Client architecture and invented the concept of crossplatform application market running on Cloud server where ubiquitous access to individual applications and PaaS/SaaS are enabled through thin client system. The features of the Cloud server, such as virtualization, flexibility, security, and dynamic management can be fully utilized to support a large number of thin client users. However, the previous architecture cannot achieve mass acceptance until we find practical solutions to the following technical challenges. First of all, one physical machine should be able to support multiple thin client users. VM, as a completely isolated operating system installation within your normal T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 1–10, 2011. © Springer-Verlag Berlin Heidelberg 2011

2

B. Song et al.

operating system, can be viewed as a feasible way to achieve isolation for multi-user sessions. However, attempting to create VM for every single user session is not always efficient as VM implementation itself consumes considerable amount of memory, and degrades the capacity of CPU. Most of existing thin client systems focused on supporting collaborative multi-user sessions. This focus has resulted in the majority of the work done in this area to be centered on the cursor management and remote display protocol towards heterogeneous client devices. While these approaches have resulted in improvements for collaborative scenarios, they could not provide an effective solution for independent multi-user sessions running on one physical machine. Secondly, the server side should be capable to manage and allocate a multitude of user sessions using its local resources. Although the task placement problem in Cloud environment has been studied in a number of works [4-6], the problem we are facing is distinct from their problem. For one thing, VM is not necessary but alternative carrier for single user session; for another, the metrics of thin client QoS differs from the tradition QoS in Cloud environment. In this context, we first introduce a novel thin client technology to support independent multi-user sessions running on one physical machine. This work is inspired by both THINC system [2] and VNC system [3] where the former one hooks data from server’s device driver layer and the latter one hooks data from server’s hardware frame buffer layer. Unfortunately, both of them are not directly applicable to our scenario. On the one hand, the device driver hooking requires complex translations among different device drivers as well as the hardware support from client device. Thus, using this model in mobile thin client environment may lose the generality since the hardware (especially the graphic component) of mobile client device are highly heterogeneous and usually insufficient to handle various applications. On the other hand, the hardware frame buffer hooking also has a nontrivial drawback in multi-user session environment, which is the failure of audio isolation. The audio information hooked from audio card frame buffer is a mixed signal including all the audio information from multiple user sessions. To address these shortcomings, we adopt a cross-layer approach, which is an effort to provide light weight isolation of remote display, audio playback and input functionalities for multiple user sessions. Comparing with VM isolation, the cross-layer isolation approach consumes less CPU and memory resources. However, the cross-layer isolation approach also has drawbacks. The input interference is one of the problems that may degrade the QoS. For example, when one user is dragging the cursor, other users’ operations have to be paused until the cursor is released by that user. Besides, VM isolation also benefits workload balancing as VM can be migrated without losing running state. Thus, we consider both isolation approaches are alternative, and apply a QoS based selection approach considering the differences in application requirements. Meanwhile, a user session allocation algorithm is proposed. The two goals of session allocation are achieving better resource utilization and reducing service interruption time caused by input interference or VM migration.

Studies on the Key Technologies of Multi-Platform Mobile Thin Client System

3

The rest of this paper is organized as follows. Section 2 discusses some related work in thin client domain. Section 3 gives a brief overview of multi-platform mobile thin client architecture. A description of cross-layer isolation approach is given in section 4. Session allocation algorithm can be found in section 5. We conclude our work in section 6.

2

Related Work

Many collaborative supports have been done in the field of thin client. We found that the main theme among all of these researches was the need of “floor-control management”, which could be used to handle control of a synchronous task. Boyd introduced “fair dragging” where control of a user gains control of the floor once the mouse is dragged [7]. Another well-known study was collaborative VNC [8]. Collaborative VNC was a patch applied to the TightVNC server and client that provides managed collaborative sessions over the RFB protocol. With collaborative VNC, one user has the “floor” (i.e. controls the desktop) at any given time. Other users have the power to take control from or give control to other users at any time. Every user’s cursor is displayed, with each cursor assigned one of several colors. In [9], a multi-user collaborative support named “THINCing Together” was introduced to extend THINC system. The proposal contains a protocol that allows for asynchronous and synchronous multi-user session. It implemented centralized cursor management to optimize bandwidth usage for multiple users. These existing studies provide good solutions for the user input management in an independent multi-user scenario. However, they are not sufficient in terms of remote display and audio support for independent user sessions since they consider same screen and audio output for all users. The remote display protocols can be categorized into three distinctive groups. At application/OS layer, Remote Desktop Protocol (RDP) is a typical protocol developed by Microsoft, which concerns providing a user with a graphical interface to another computer. RDP clients exist for most versions of Microsoft Windows (including Windows Mobile), Linux, Unix, Mac OS X, Android, and other modern operating systems. At device driver layer, THINC uses its virtual device driver to intercept display output in an application and OS agnostic manner [2]. It efficiently translates high-level application display requests to low-level protocol primitives and achieves efficient network usage. At hardware frame buffer layer, VNC uses the RFB protocol to remotely control another computer [3]. Server only sends the graphical screen updates back to the client. In [10], a hybrid protocol was proposed to handle multimedia streaming and interactive gaming applications. Both VNC RFB-protocol and THINC protocol are alternative in this system. Among these protocols, RDP and THINC require graphical hardware support from client device which is hardly provided by mobile thin client device. VNC RFB-protocol, on the other hand, is more general and flexible since it can fully utilize the GPU of server to support mobile clients. Thus, we choose VNC RFB-protocol with non-overlapping window placement to support remote display for independent multi-user sessions.

4

3

B. Song et al.

Multi-platform Mobile Thin Client Architecture

This section introduces the Multi-Platform Mobile Thin Client Architecture. In Fig.1, we present an overview of the architecture. The architecture is composed by several mobile thin clients and a multi-platform thin client server. To receive services, every mobile device must install a thin client viewer which provides all functionalities of remote access. Traditionally, a thin client ran a full operating system for the purposes of connecting to server. A newer trend is called a zero client, which no longer runs a full operating system: the kernel instead merely initializes the network, begins the networking protocol, and handles display of the server's output. Thus, it is not necessary to have an OS on mobile thin client device.

Fig. 1. Multi-Platform Mobile Thin Client Architecture

As Fig.1 shows, the multi-platform thin client server contains a number of modules. At first, the mobile terminal sends a request to authentication module. The request includes the user’s ID, password and terminal information. Then the user selects applications from application store, and provides necessary information to the payment system. The application request is delivered to the task manager. The task manager receives current resource condition and application profiles from the QoS monitor and the internal data repository, respectively. By analyzing the information, the task manager performs task allocation and real-time migration on local resources (only for VM). The QoS monitor continuously monitors the QoS information of running applications as well as the resource condition of local resources. QoS monitor also detects QoS violation, and send reports to task manager for task migration (only for VM). The remote display module, audio support module and remote input module are the three components of thin client protocol. They provide all functionalities of remote access on the server side. Local resources are composed by physical machines, OSs, hypervisors, VM images, and applications. The actual processing power, platform, and services are provided by local resources. The local resources exchange VM image

Studies on the Key Technologies of Multi-Platform Mobile Thin Client System

5

and application data with the internal data repository in order to save or load user sessions. Meanwhile, the application running on local resources may interact with the external data source at any time. The following section will focus on the isolation technology that allows one physical machine to support multiply independent user sessions without using VM.

4

Cross-Layer Isolation Technology

The two issues to be discussed when designing a multi-user thin client system are how to get graphical/audio updates and how to isolate input/output for multiple user sessions. Fig.2 contains a three-layer interception model illustrating the approaches that allow the graphical/audio updates to be retrieved and redirect to the client. This three layer model was first proposed by R.Baratto in his Ph.D. thesis [2]. The original version demonstrates only the display pipeline while we extend the same concept to the audio/input pipeline as well. There exists three interceptions points in the pipeline: (1) between the applications and the device independent layer, (2) between the device independent layer and the device dependent layer, and (3) between the device dependent layer and the hardware layer. To utilize them, the server side must be able handle the application/OS interfaces, the device drivers and the hardware frame buffer, respectively. The requirement for client side also differs from one interception point to another. The interception between the applications and the device independent layer requires client’s OS support and hardware processing capability. Intercepting between the device independent layer and the device dependent layer needs client’s hardware processing capability. Hardware frame buffer interception merely requires the A/V playback functions on client device.

Fig. 2. Cross-layer isolation vs. VM based isolation

Since the gap between the capacity of mobile graphic card and the requirement of PC applications always exists, using the first or the second interception point to support remote display cannot be the proper choice in our scenario. However, the hardware frame buffer interception also has two disadvantages. First, display updates consisting of raw pixels along are typically too bandwidth-intensive. Second, the intercepted signals including audio and video are mixed already, which means we

6

B. Song et al.

need extra isolation technology to support multiple user sessions. The first problem can be solved by optimizing the remote display protocol and the network condition. A successful example is the latest version of VNC, which is able to provide fluent video display under current network circumstance. To solve the second problem, we invent a new isolation approach named cross-layer isolation. Fig.2 also shows the detail information of cross-layer isolation and the well-known VM based isolation. Unlike the VM based isolation in which the hypervisor takes the responsibility to divide and manage the hardware resources, the cross-layer technique has three isolation components deployed between the user sessions and the remote server. The motivation of our design is from the fact that the VM based isolation consumes considerable amount of resources to maintain VM, guest OS and remote server for every single user. Since the modern OSs have the capability to manage several processes running concurrently, the QoS of user sessions can be guaranteed by OS as long as the server has sufficient resources to handle the tasks. The reason we choose the name “cross-layer” is that the audio isolation, the video isolation and the input isolation are implemented in different layers. To isolate the graphical output of one user session from others, the server assigns a non-overlapping rectangle area for each user session. While hooking the whole screen information from the hardware frame buffer, the server can easily extract any user session from the picture by using the coordinates of the corresponding rectangle. Then using RFB protocol, the updates are distributed to the users continuously. The audio isolation utilizes device driver interception. Before the audio signals from multiple user sessions are mixed, the server hooks them and sends to the corresponding clients. For one thing, it is very hard to extract one session’s audio information from a mixed audio signal retrieved from hardware frame buffer; for the other, the mobile clients are able to provide the hardware processing capability for audio. The RTP (Real-time Transport Protocol) is adopted as the protocol for audio transmission. Unlike video and audio isolation, the input isolation intercepts on the client side and performs the actual input on the server side. The input isolation on server side needs the APIs of server OS. Taking our implementation on Microsoft Window XP as an example, the input isolation module requires the handle of each user session. The user inputs are managed by a multi-queue system. When the server wants to perform a user input, it first activates the corresponding session using the handle, and then simulates the mouse or keyboard event. We adopt Shortest-Remaining-Size-First (SRSF) preemptive scheduler known to be optimal for minimizing mean response time and improving the interactivity of a system [11]. In terms of resource consumption, the cross-layer isolation approach is more efficient than the VM-based isolation approach; nevertheless it can be only an alternative solution since it has two disadvantages. First, input interference may occur. Second, session migration is not possible. In next section, we will discuss about these problems and give a primal solution to choose the isolation technology and to allocate user sessions.

Studies on the Key Technologies of Multi-Platform Mobile Thin Client System

5

User Session Management

5.1

Isolation Selection Approach

7

As we introduced in section four, all the user sessions supported by cross layer isolation are running on same OS. While one session gains the control of mouse/keyboard, the inputs of other sessions cannot be performed until the control is released. Since the control shift among these sessions should be made in a quick and implicit manner, we need to figure out which operation may hinder others from getting the control. Any mouse/keyboard operation can be viewed as one of the following two events: instant event and continuous event. Instant events usually do not interfere with each other whereas the total amount of instant events cannot be larger than that can be handled by the input isolation component. Continuous events, such as mouse dragging, greatly interfere with instant events and themselves. Therefore, we need to profile the averaging number of instant event and the cumulative time of continuous event for each application. Let NI j be the number of instant events generated by the user of application j within one minute,

TC j be the cumulative time of continuous events generated by

the user of application

j within one minute. If the value of NI j or TC j highly

depends on user’s behavior, we need to create profile for each user. Let MAX ( NI ) be the maximum number of instant events that can be handled by the input isolation component within one minute, MAX (TC ) be the maximum time that can be occupied by continuous events within one minute. The value of MAX ( NI ) depends on the hardware and software environment. The value of

MAX (TC ) should guarantee that no user feels the existence of other user. We define the first part of selection approach as follow: given a user session is application j , if NI j > MAX ( NI ) × 50% or TC j > MAX (TC ) × 50% , the cross layer isolation cannot be applied to this session. Using

MAX ( NI ) × 50% and

MAX (TC ) × 50% rather than higher values leaves enough space for allocating other sessions supported by cross-layer isolation. The CPU consumption is also considered as an important factor. Since the crosslayer isolation technology does not support session migration, it is not suitable for the sessions whose CPU consumption fluctuates all the time. Let t be the duration between the time when the CPU is overloaded and the time when QoS monitor triggers a migration. Given an application j , the profiling of its CPU consumption takes time n × t where n should be large enough to present the averaging CPU consumption of application j . By that, we can get n averaging CPU consumption information

{ AC j1 , AC j 2 ,..., AC jn } and an overall averaging CPU consumption

8

B. Song et al.

information

AC j . Let FC be the free CPU capacity to buffer the unexpected

workload burst. We define the second part of selection approach as follow: given a user session is application j , if max{ AC jk | (1 < k < n)} − AC j > FC × 50% , the cross layer isolation cannot be applied to this session. The reason of using FC × 50% is also to facilitate session allocation. 5.2

Session Allocation Approach

Given an application

i and a physical machine j , let cij , mij , gij and bij be the

percentages of resource usage regarding CPU, memory, GPU and network bandwidth, respectively. If the application is supported by VM, the extra resource consumption of VM, OS and remote server should be also included. Let fc j , fm j , fg j and fb j be the percentages of idle CPU, memory, GPU and network bandwidth resources, respectively, on machine j . For any resource, at least 25% free capacity should be kept to buffer the unexpected workload burst, and should not be counted in the available idle resources. The application can be allocated to that physical machine only if the following condition can be fulfilled:

cij <= fc j & mij <= fm j & g ij <= fg j & bij <= fb j ,

(1)

While allocating both cross-layer isolated session and VM isolated session, there are two common objectives: (1) avoiding overuse of any resource, and (2) providing more CPU resource for the session. To explicitly define the objectives, we need new notations. After application i is allocated on physical machine j , the average percentage of free resource

apij is defined as: apij = ( fc′j + fm′j + fg ′j + fb′j ) / 4

fc′j = fc j − cij , fm′j = fm j − mij , fg ′j = fg j − gij , fb′j = fb j − bij .

(2)

j , the resource utilization condition after allocating application i is denoted as RUCij , which is a mean-square value and is given by:

For machine

RUCij = ( fc′j − apij ) 2 + ( fm′j − apij ) 2 + ( fg ′j − apij ) 2 + ( fb′j − apij ) 2

(3)

We assume that the applications running on a single physical machine share all of the CPU capacity in a proportional way. To be more specific, application i to be allocated on physical machine capacity.

j will get (cij / (1 − fc′j )) × 100% of the CPU

Studies on the Key Technologies of Multi-Platform Mobile Thin Client System

Regarding application

9

i , the superiority of machine j is defined as Sij :

Sij = cij / (cij / (1 − fc′j )) = 1 − fc′j

(4)

For cross-layer isolated session, the cumulative number of instant events and the cumulative time of continuous events need to be considered. Let NI ij or TCij be the cumulative number of instant events and the cumulative time of continuous events on machine j after application i is allocated. The conditions NI ij < MAX ( NI ) and

TCij < MAX (TC ) must be satisfied. Based on that, we define the trade-off metric for allocating cross-layer isolated session as follows:

CTij = α × RUCij + β × Sij + ε × ( NI ij / MAX ( NI ) + TCij / MAX (TC )) (5)

α

,

β

and

ε

are the coefficients for adjusting the weight of each objective. A

smaller the value of

CTij represents a better allocation for cross-layer isolated

session. For VM isolated session, input interference problem does not exist. However, any possible migration should be avoided after the VM is allocated. Although the latest VM technology allows a running VM to be migrated within 20 seconds, it still makes user uncomfortable. To solve that problem, using only Sij is not enough since

Sij considers the averaging CPU consumption rather than the burst case. As we have the profile information max{ AC jk | (1 < k < n)} −

AC j to illustrate the stability

of CPU consumption for any application j , we can easily figure out whether the CPU consumption on a physical machine is stable or not. The way is to sum up the stability values retrieved from all the applications running on that machine. We use CS j to denote the summation on machine j , and provide the trade-off metric for allocating VM isolated session as follows:

VTij = α × RUCij + β × Sij + ε × CS j

α, β

and

ε

(6)

are the coefficients for adjusting the weight of each objective. A

smaller the value of

VTij represents a better allocation for VM isolated session.

Based on Eq. (5) and Eq. (6), we adopt a greedy approach to achieve the goal of efficient session allocation in our Multi-Platform Mobile Thin Client system.

6

Conclusions

We presented cross-layer isolation technology, a novel and efficient approach to support multiple thin client sessions using with one OS and one remote server. We introduced how to implement the video isolation, the audio isolation and the input isolation in

10

B. Song et al.

different layer. Meanwhile, we analyzed the advantages and disadvantages of both isolation approaches and provided an effective way to select proper isolation technology for an application using profiling method. A session allocation/migration algorithm was proposed in this paper to allocate multiple sessions on multiple physical machines. Acknowledgement. This research was supported by Next-Generation Information Computing Development Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology (20110020515).

References 1. Tang, W., Lee, J.-H., Song, B., Islam, M.M., Na, S., Huh, E.-N.: Multi-Platform Mobile Thin Client Architecture in Cloud Environment. In: CICC-ITOE 2011, vol. 2, pp. 323–327 (June 2011) 2. Baratto, R.: THINC: A Virtual and Remote Display Architecture for Desktop Computing and Mobile Devices, Ph.D. Thesis, Department of Computer Science, Columbia University (April 2011) 3. Richardson, T., Stafford-Fraser, Q., Wood, K.R., Hopper, A.: Virtual Network Computing. Virtual Network Computing 2(1) (January, February 1998) 4. Wood, T., Shenoy, P., Venkataramani, A., Yousif, M.: Black-box and Gray-box Strategies for Virtual Machine Migration. In: 4th USENIX Symposium on Networked Systems Design & Implementation, pp. 229–242 (2008) 5. Clark, C., Fraser, K., Hand, S., Hansen, J.G., Jul, E., Limpach, C., Pratt, I., Warfield, A.: Live migration of virtual machines. In: Proc. NSDI 2005 (May 2005) 6. Gupta, D., Gardner, R., Cherkasova, L.: Xenmon: Qos monitoring and performance profiling tool. Technical Report HPL-2005-187, HP Labs (2005) 7. Boyd, J.: Floor Control Policies in Multi-User Applications. In: INTERACT 1993 and CHI 1993 Conference Companion on Human Factors in Computing Systems, Amsterdam, The Netherlands, pp. 107–108 (1993) 8. Collaborative, VNC, http://benjie.org/software/linux/vnc-collaborate.html 9. Coulthart, D., Das, S., Kim, L.: THINCing Together: Extending THINC for Multi-User Collaborative Support, December 04. e-Publisher, CiteSeerX Publication (2008) 10. De Winter, D., Simoens, P., Deboosere, L.: A Hybrid Thin-Client protocol for Multimedia Streaming and Interactive Gaming Applications. In: The 16th Annual International Workshop on Network and Operating Systems Support for Digital Audio and Video (2006) 11. Bansal, N., Harchol-Balter, M.: Analysis of SRPT scheduling: investigating unfairness. In: Proceedings of the Joint International Conference on Measurement & Modeling of Computer Systems (SIGMETRICS/Performance), pp. 279–290 (June 2001)

LDPC Equalizer for Compensating the CFO and Phase Noise in OFDM System Do-Hoon Kim and Heung-Gyoon Ryu Department of Electronic Engineering, Chungbuk National University, 12 Gaesin-dong, Cheongju, Korea [email protected], [email protected]

Abstract. In this paper, we propose decision feedback equalizer based on LDPC (Low Density Parity Check) code for fast processing and performance improvement. LDPC code has good error correcting ability and performance approaching Shannon capacity. However, it has more long parity check matrix and the number of iteration. In proposed system, MSE (mean square error) of signal between decision device and decoder is fed back to equalizer. If we use this method, proposed system can improve performance because it corrects estimated channel response more accurately. In addition, proposed system can reduce complexity because it has a lower number of iterations than system that does not include feedback at same performance. We evaluate performance of OFDM system considered CFO and phase noise in multipath channel via simulation. Keywords: Decision feedback equalizer, LDPC, MSE, iteration number, CFO, phase noise, OFDM.

1

Introduction

LDPC code is well known as a good performance channel coding method. It was proposed by Robert G. Gallager in 1962[1] and got attention again by Mackay and Neal in 1997[2]. It is error correction code which are the nearest to Shannon’s capacity. Recently, LDPC code is commonly used in high data rate communication such as DVB-S2 (Digital Video Broadcast - Satellite 2) and IEEE 802.16 (WiMax). LDPC code is good error correcting performance and possible fast processing with sparse parity check matrix. However, it has more long sparse parity check matrix and iteration number in multipath channel. In this paper, we propose a decision feedback equalizer (DFE) based on LDPC code. The proposed DFE based on LDPC code is lower complexity than not used feedback system. In addition, it can remove ICI components which inhibit bit error ratio (BER) performance such as CFO and phase noise in OFDM system. So, we consider CFO and phase noise effects in OFDM system for compensating that ICI components. Finally, we analyze LDPC code theory and frequency domain equalizer, and then evaluate performance of the DFE based on LDPC code in OFDM system. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 11–19, 2011. © Springer-Verlag Berlin Heidelberg 2011

12

2

D.-H. Kim and H.-G. Ryu

System Model

Figure 1 shows system model of the LDPC Equalizer based on OFDM system. We use LDPC channel coding method. And get channel characteristic using channel estimation. In this paper, we design and propose feedback block. Then adaptive equalizer is combined with LDPC decoder. The LDPC equalizer system calculates mean square error (MSE) in feedback block.

Fig. 1. Block diagram of OFDM system using proposed DFE based on LDPC

3

Low Density Parity Check Code

(N , K ) . LDPC code has a sparse parity check matrix (H matrix) of ( N − K ) × N size. H LDPC code is one of the block code methods. They can be expressed as matrix is non-systematic sparse parity check matrix. 3.1

LDPC Encoding

The received vector is corrupted by an error vector

e as follows.

r = c ⊕ e = [p m ] ⊕ e .

(1)

Where, p is parity vector and m is message vector. The parity vector and message vector are assumed to be located at the former or latter part of the code vector, respectively. The decoder of the receiver is supposed to apply for this received signal vector to find the syndrome vector as

 H1T  s = rH = ([p m ] ⊕ e )  T  H2  . = pH1T ⊕ mH 2T ⊕ eH T . T

(2)

LDPC Equalizer for Compensating the CFO and Phase Noise in OFDM System

13

Noting that this syndrome vector should be zero for the non-error case of e = 0 .

s = pH1T ⊕ mH 2T = 0 . We can write the parity vector

(3)

p in terms of the message vector m as p = mH 2T H1−T .

(4)

This amounts to having the generator matrix

G =  H 2T H1−T

I 

so that the code vector can be generated by post-multiplying the message vector with the generator matrix G as

c = mG = mH 2T H1−T

m  = [p m ]

(5)

m (6)

Here, the parity-check matrix is the key to the performance of a LDPC code, and therefore the design issue of the parity-check matrix is very important 3.2

LDPC Decoding

LDPC decoder calculates the probability in variable nodes and check nodes, respectively. LDPC decoding algorithm can be divided into sum-product algorithm (SPA) and min-sum algorithm (MSA). SPA has better BER performance, but higher complexity. We use SPA to process LDPC decoding. LDPC decoder process 4 steps. Step 1 is initialized step. Step 2 is updating check nodes step. Step 3 is updating variable nodes step. Step 4 is decision step.

4

Channel Estimation and Equalization

A frequency-domain training sequence is

X (k ) = X R (k ) + jX I (k )

(7)

with the corresponding channel output.

Y ( k ) = YR ( k ) + jYI ( k ) Channel estimation with

(8)

X (k ) and Y (k ) is

Hp =

Y (k ) . X (k )

(9)

14

D.-H. Kim and H.-G. Ryu

The long preamble used for channel estimation is

X (k ) = 1 or −1(with

X I (k ) = 0) .

(10)

The channel estimation signal can be simplified as

Hp =

Y (k ) YR (k ) + jYI (k ) = . X (k ) X R (k )

(11)

Since the long preamble contains two repeated training sequences, the average of the FFTs ( Y1 ( k ) and Y2 ( k ) ) of the channel outputs can be taken for better channel estimation.

Hp =

1 Y1 (k ) + Y2 (k ) . 2 X (k )

(12)

So, estimated channel can equalize the output to compensate the channel effect.

Y′ Xˆ (k ) = . Hˆ p

5

(13)

Proposed DFE Based on LDPC

Figure 2 shows feedback system of proposed DFE based on LDPC in receiver. Figure 2 is subsystem after FFT in OFDM receiver.

Hˆ p Yk ′

Ls CF

Ld

Fig. 2. Block diagram of DFE based on LDPC in receiver

After LDPC decoder, we feed with MSE of Ls and Ld .

CF factor back to compensate channel estimation

Ls = norm( Ls ) .

(14)

LDPC Equalizer for Compensating the CFO and Phase Noise in OFDM System

Lˆk =

 L

k∈sd

CF =

*

s

Ld .

Lˆk L



2

15

(15)

.

(16)

s

k∈sd

H p is the channel characteristic after channel estimator using long preamble.

Hˆ p = H p + CF .

(17)

We improve the performance following equation with compensated channel.

Yk ′ =

Yk . Hˆ

(18)

p

Therefore, the proposed DFE based on LDPC can improve the better performance and complexity than used independently.

6

Simulation Result

Table 1 is the simulation parameters for LDPC Equalizer. In this paper, we consider IEEE 802.11n format. Code rate is 3/4 and parity check matrix size is 720. We consider the AWGN and multipath fading channel. And we also consider normalized CFO component to be 0.03 and phase noise power to be -12dBc. Table 1. The simulation parameters of proposed system Parameters # of data subcarriers # of pilot subcarriers # of padded zeros FFT size # of samplesin a GI Modulation level

Values 52 4 8 64 16 4QAM

Figure 3 shows BER curve according to iterative number of LDPC in AWGN channel. Here, code rate is 3/4 and parity check matrix size is 720. The more iterative number is higher, the more BER performance is improved.

16

D.-H. Kim and H.-G. Ryu 0

10

Iter=3 Iter=5 Iter=8

-1

10

-2

BER

10

-3

10

-4

10

-5

10

-6

10

0

0.5

1

1.5

2

2.5 3 Eb/N0 (dB)

3.5

4

4.5

5

Fig. 3. BER performance according to iterative number in AWGN 0

10

w/o EST. & EQ. DFE based on LDPC -1

10

-2

BER

10

-3

10

-4

10

-5

10

0

0.5

1

1.5

2

2.5 3 E b/N0 (dB)

3.5

4

4.5

5

Fig. 4. Comparison of BER performance of DFE based on LDPC with and without channel estimation and equalization

Figure 4 shows BER performance of DFE based on LDPC with and without channel estimation and equalization. The performance of DFE based on LDPC is much better than without channel estimation and equalization in considerable CFO and phase noise effects. When the BER probability is 10-5, the SNR is less than 5dB. Here, iterative number is 5. So, DFE based on LDPC is possible to compensate the CFO and phase noise. Figure 5 shows BER performance of DFE based on LDPC with and without channel estimation and equalization under considering same iterative number. The performance

LDPC Equalizer for Compensating the CFO and Phase Noise in OFDM System

17

of DFE based on LDPC is better than conventional equalizer. The DFE based on LDPC is better performance about 0.5dB at 10-4. Also, iterative number of LDPC is 5. Figure 6 shows similar BER performance of DFE based on LDPC and conventional equalizer each other iterative number. When SNR is 3.5dB, BER of the DFE based on LDPC and conventional equalizer are exactly same. Iterative number of DFE based on LDPC is 3, but conventional equalizer case is 5. Complexity of system is closely related to iterative number. Therefore, DFE based on LDPC has lower complexity than conventional equalizer. 0

10

DFE based on LDPC w/o feedback

-1

10

-2

BER

10

-3

10

-4

10

-5

10

-6

10

0

0.5

1

1.5

2

2.5 3 Eb/N0 (dB)

3.5

4

4.5

5

Fig. 5. Comparison of BER performance of the DFE based on LDPC 0

10

DFE based on LDPC(Iter.=3) w/o feedback(Iter.=5)

-1

10

-2

BER

10

-3

10

-4

10

-5

10

-6

10

0

0.5

1

1.5

2

2.5 3 Eb/N0 (dB)

3.5

4

4.5

5

Fig. 6. Comparison of BER performance of the DFE based on LDPC

18

7

D.-H. Kim and H.-G. Ryu

Conclusion

We propose LDPC Equalizer in order to compensate the CFO and phase noise in OFDM system. We use long preamble in order to estimate the channel. After LDPC decoder, we calculate MSE with Ls and Ld . And feed MSE factor ( CF ) back to equalizer. Then the equalizer is combined with LDPC decoder. So, the channel estimator is more accurate to equalize channel characteristic. Therefore, the LDPC Equalizer is improved BER performance and getting lower complexity than conventional equalizer. The LDPC Equalizer has very good BER performance with ICI components such as CFO and phase noise. According to Fig. 5, when iterative number of the LDPC Equalizer and conventional equalizer are same, the BER performance of LDPC Equalizer is better than conventional equalizer. Also, according to Fig. 6, when the LDPC Equalizer is similar BER performance to conventional equalizer, the complexity of LDPC Equalizer is lower than conventional equalizer, because complexity is very closely related to iterative number of LDPC decoder. Therefore we compensate the CFO and phase noise by using LDPC Equalizer based on OFDM system. Also, we improve performance and decrease complexity of system. Finally, we look forward to implementation of the LDPC Equalizer for compensating the CFO and phase noise. Acknowledgments. This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology. (No. 2010-0007567).

References 1. Gallager, R.G.: Low Density Parity Check codes. IRT Trans. Inform. Theory IT-8, 21–28 (1962) 2. Mackay, D.J., Neal, R.M.: Neal Shannon limit performance of low density parity check codes. Electronic Letters 45, 457–458 (1997) 3. Mackay, D.J.: Good error correcting codes based on very sparse matrix. IEEE Trans. Inform. Theory 33(6), 399–431 (1999) 4. Chen, J., Dholakia, A., Eleftheriou, E., Fossorier, M.P.C., Hu, X.: Reduced-complexity decoding of LDPC codes. IEEE. Trans. Commun. 53, 1288–1299 (2005) 5. Papaharalabos, S., Sweeney, P., Evans, B.G., Mathiopoulos, P.T., Albertazzi, G., VanelliCoralli, A., Corazza, G.E.: Modified sum-product algorithm for decoding low-density parity check codes. IET. Commun. 1, 294–300 (1955) 6. Glolami, M.R., Nader-Esfahani, S., Eftekhar, A.A.: A new method of phase noise compensation in OFDM. In: ICC 2003, IEEE International Conference on Communications, vol. 5, pp. 3443–3446 (2003) 7. Pollet, T., van Bladel, M., Moeneclaey, M.: BER sensitivity of OFDM systems to carrier freuency offset and Wiener phase noise. IEEE Trans, on Comm. 43(2), 887–895

LDPC Equalizer for Compensating the CFO and Phase Noise in OFDM System

19

8. Moose, P.H.: A technique for OFDM Frequency Offset Correction. IEEE Trans. on Comm. 42(10), 2908–2914 9. Songping, W., Bar-Ness, Y.: A phase noise suppression algorithm for OFDM-based WLANs. IEEE Communications Letters 6, 535–537 10. Ryu, H.G., Li, Y.S.: Phase noise analysis of the OFDM communication system by the standard frequency deviation. IEEE Transactions on Consumer Electronics 49(1), 41–47

TC-HMIPv6: A Study of HMIPV6 Handover Management for Packet Transmission Analysis Sung-Gyu Kim, Farkhod Alisherov, and Byungjoo Park* Department of Multimedia Engineering, Hannam University 133 Ojeong-dong, Daeduk-gu, Daejeon, Korea {sgkim,bjpark}@hnu.kr, [email protected]

Abstract. Recently, the Wireless internet user increased drastically due to the increase of concerns about the performance improvement of the mobile device and Smart phone. Wireless device users required the service which has a higher and the Mobile Communication traffic of the voice was expanded to a various multimedia services which includes various forms of voice data servicecentered including wireless internet, SMS, MMS, and etc. The significant importance of Mobile IPv6 stood out as the user of the wireless internet increased and various of mobile service rapidly developed such as 3G and WIBRO. In Mobile IPv6, when the mobile device moves it possible creates a controversial issue between the base station, the packet loss according to the handover generates a delay time and it reverse the handover of packets and it’s process. In this paper, we present a novel method of solving the problems packet handover. We based our proposed packet handover scheme on the existing handover process. Our proposed scheme has several features such as a buffer installation, Look-Up and Reverse Binding in the beacon Messages, and each Access Point in the existing Hierarchical Mobile IPv6 based was shown. Keywords: MIPv6, HMIPv6, Reverse Binding, Look-Up.

1

Introduction

Recently, Wireless internet user increased drastically due to various mobile device technologies developed such as Smart Phone, Lap-top, PDA and etc. Accordingly, Developed of technology wireless network offer network service in mobile device such as WIBRO, Wireless LAN and 3G system. Standard Mobile IPv6 when mobile node moves to other base station. Standard mobile IPv6 in the event of a handover needs more time to process. For these reason, many problems happen such as Packets Loss and Out-of-sequence. So, wireless network users can’t provide seamless service. Much research is progressing to solve those problems, such as Fast Handover for Mobile IPv6, Hierarchical Mobile IPv6, Proxy Mobile IPv6 and etc. But wireless network still has problems. In this paper a solution for solving problems such as long delay time handover in standard Hierarchical Mobile IPv6. We used Reverse Binding mechanism and LookUp process. *

Correspondent author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 20–25, 2011. © Springer-Verlag Berlin Heidelberg 2011

TC-HMIPv6: A Study of HMIPV6 Handover Management

2

21

Related Work

This section discussed the detailed information about standard Mobile IPv6, Hierarchical Mobile IPv6 operation and Handover procedure.

Fig. 1. HMIPv6 handover procedure scheme

2.1

Standard Mobile IPv6 (MIPv6)

In MIPv6 Mobile Node (MN) the communication will start with the address of its own Home Address (HoA) in a Home Network. However, When the MN moves to a different base station, the MN can’t maintain communication. Therefore, after moving out from the home network, the MN can maintain communication with Care-ofAddress (CoA). CoA can be automatically created by combining router prefix and the MN’s interface address. When MN moves to another base station, it informs own

22

S.-G. Kim, F. Alisherov, and B. Park

moves to New Access Point (NAP) and New Access Router (NAR). After creating New CoA (NCoA), the MN starts Duplicate Address Detection (DAD) process. The MN starts registration of NCoA to Home Agent (HA) and Correspondent Node (CN) through Binding Update (BU) message. 2.2

Standard Hierarchical Mobile IPv6 (HMIPv6)

Figure 1 shows the processing procedure for handover in standard HMIPv6. HMIPv6 uses the top of the hierarchy of router to manage the MN mobility. Mobility Anchor Point (MAP) is the top of router. It is handle process certification and registration. In case when the MN moves in the same MAP domain, the HA and CN is not concerned MN’s mobility.

3

New Proposed Scheme TC-HMIPV6

Each Access Point (AP) in TC-HMIPv6 has buffer. AR’s (AR : Access Router) information periodically stored and updated in AP’s buffer.

Fig. 2. TC-HMIpv6 Handover procedure scheme

TC-HMIPv6: A Study of HMIPV6 Handover Management

3.1

23

Layer 2 Handover

Figure 2 shows the processing procedure for handover in TC-HMIPv6. NAR’s information stored buffer in the NAP. So, the MN can know NAR's information by process that inform own position to the NAP. Router Solicitation message and Router Advertisement message aren't need. When the MN creates the NCoA, Standard MIPv6 uses DAD process to confirm for duplicates of addresses. However, it needs 1000ms to perform a DAD process. It is largest proportion in total handover latency. DAD process time should decrease to reduce total handover delay time. In this paper, DAD process was replaced with a Look-Up Process. Look-Up process examines the Cache if the MN receives NAR's information. If there is a duplicate address, the MN can use address after store to Cache's entry, but, if the isn’t a duplicate address, the MN can use already created address in stored address table. Look-Up process takes or the worst case 5.28 . time to performed process for best case 3.36

㎲

3.2

㎲

Registration to MAP, HA and CN

The MN sends the Local Binding Update message to the MAP2. The MAP2 performs Look-Up process for RCoA. The MAP2 send the Binding Acknowledge message to the MN. MN’s CoA registers to the HA by sending the Binding Update message (BU). The HA sends the Binding Acknowledge message (BA) to the MN at the same time the CN receives the Fast Binding Update message (FBU) from the HA. In standard HMIPv6, the MN registers the the CN after registers the HA. The HA sending the Reverse Binding Update message (RBU) to the MAP2. The MAP2 sending the RBU to the HA. The HA receives the Reverse Binding Acknowledge message (RBA) from the NAR. RBU and RBA have MN’s information such as LCoA, RCoA and received packet information.

4

Performance Analysis

4.1

Handover Latency Analysis

TC-HMIPv6 happens at the same time that Registration messages are sent to HA and CN procedure, we calculated longer time message. Table 1 shows Performance Analysis Parameters. Equation 1 shows the formulation of standard MIPv6. Equation 2 shows the formulation of standard HMIPv6. Equation 3 shows the formulation of proposed TC-HMIPv6. Standard MIPv6 = αtL2 + tCoA + tHA_REG + tCN_REG = α(4t1+2t2) + tCoA + 4t1 + 4t2 + 2t4 + 2t5 + β( t1 + t2 + t5 )

(1)

Standard HMIPv6 = αtL2 + 2tCoA + tMAP_REG + tHA_REG + tCN_REG = α(4t1+2t2) + 2tCoA + 6t1 + 6t2 +6t3 + 2t4 + 2t5 + β( t1 + t2 + t5 )

(2)

24

S.-G. Kim, F. Alisherov, and B. Park

Proposed TC-HMIPv6 = αtL2 + 2tLook-Up + tMAP_REG + tHA_REG + tCN_REG = α2t1 + 2tLook-Up + 4t1 + 4t2 + 4t3 + 4t4 + β( t1 + t2 + t5 ) 4.2

(3)

Performance Comparison

Figure 3 shows a graph comparing the changes in the value of α. In case of same the value of β, proposed TC-HMIPv6 is efficient more than Standard MIPv6 and HMIPv6.

Fig. 3. Handover Latency Comparison at time α

5

Conclusion

In this paper, we Proposed TC-HMIPv6 over Standard HMIPv6. TC-HMIPv6 protocol is efficiently management Handover. Each AP’s buffer stored AR’s information. AR’s information is periodically updated. It is replaced from DAD process to Look-Up process on TC-HMIPv6. If the MN moves to a previous AR, the NAR, MAP2, and HA already know MN’s Information by RBU message. So, Handover delay time is reduced. Acknowledgments. This research was supported by the Security Engineering Research Center, granted by the Korea Ministry of Knowledge Economy (No. 11-8).

TC-HMIPv6: A Study of HMIPV6 Handover Management

25

References 1. Johnson, D., Perkins, C., Arkko, J.: Mobility Support in IPv6. IETF RFC 6275 (July 2011) 2. Soliman, H., Castelluccia, C., El Malki, K., Bellier, L.: Hierarchical Mobile IPv6 Mobility Management (HMIPv6). IETF RFC 4140 (July 2005) 3. Vivaldi, I., Ali, B.M., Habaebi, H., Prakash, V., Salil, A.: Routing Scheme for Macro Mobility Handover in Hierarchical Mobile IPv6 Network. In: 4th National Conference on Telecommunication Technology Proceedings (January 2003) 4. Park, B., Latchman, A.H.: A Fast Neighbor Discovery and DAD Scheme for Fast Handover in Mobile IPv6 Networks. In: ICNICONSMCL, 0-7695-2552-0106 (2006)

A Multi-hop Communication Scheme for IEEE 802.11p Based V2V Communication Systems* Woong Cho and Hyun Seo Oh Electronics and Telecommunications Research Institute (ETRI) Rep. of Korea, 305-700 {woongcho,hsoh5}@etri.re.kr

Abstract. Information and communication technologies (ICT) have been converged to other industrial areas, which creates new applications and supports various services. One of the attractive convergence areas is vehicular communication which combines ICT and vehicle technologies. In this paper, we investigate a multi-hop communication scheme for IEEE 802.11p based communication systems which support both vehicle-to-infrastructure (V2I) and vehicle-to-vehicle (V2V) communications. By building IEEE 802.11p based communication systems, we first measure the performance in terms of communication range, packet error rate (PER), link setup time, and latency by focusing on V2V communication. Then, a multi-hop communication scheme for both broadcast and unicast is discussed. Based on the suggested scheme, practical measurement results are also introduced. Keywords: Vehicular communications, performance, vehicle multi-hop protocol, IEEE 802.11p, WAVE.

1

Introduction

Vehicular communications have been considered as one of the attractive research areas by applying communication technologies to the road. In the form of vehicle-to-vehicle (V2V) or vehicle-to-infrastructure (V2I) communications, vehicular communications provide numerous applications such as traffic information service, anti-collision warning, accident alarming, internet service, multimedia downloading, and etc. Among them, the most promising application is the public safety [3,5,7]. Existing literature considers some specific safety applications. Cooperative collision is studied in [12] with focus on congestion control. Reliability analysis using measurement data and numerical analysis of latency for reliable rear-end collision avoidance is introduced in [2] and [9], respectively. To characterize the wireless channels in vehicular communications, practical data are collected in various vehicular environments [1,4,8]. Throughput and frame error rate of IEEE 802.11a/b/g based system are investigated in *

This research was supported by a grant from Construction Technology Innovation Program (CTIP) funded by Ministry of Land, Transportation and Maritime Affairs (MLTM) of Korean government (SMART Highway Project (07 Technology Innovation A01)).

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 26–33, 2011. © Springer-Verlag Berlin Heidelberg 2011

A Multi-hop Communication Scheme for IEEE 802.11p

27

[11] and IEEE 802.11p, which is commonly referred as wireless access in vehicular environments (WAVE), based prototype is introduced in [10]. On the other hand, standardizations of vehicular communication have been developed in several institutions such as IEEE, ETSI and ISO. IEEE 802.11p specifies MAC/PHY layer standard in 5.9GHz frequency band [6] and its basic MAC/PHY structures are widely accepted in vehicular communications. In this paper, we first build the IEEE 802.11p based communication prototype. Then, practical measurements are carried out in freeway environments with 5.9GHz frequency band. The various performance factors, i.e., communication range, packet error rate (PER), link setup time, and latency, are tested with V2V communication. After showing measurement results, a vehicle multi-hop protocol (VMP) is introduced. The rest of this paper is organized as follows. The overall system architecture of the vehicular communication system is introduced in Section 2. In Section 3, the measurement results of communication system are presented. The VMP is discussed in Section 4 including some measurement results, and concluding remarks are given in Section 5.

2

System Specification

Fig. 1 depicts the overall system architecture. We consider the communication system which consists of antenna, on-board-unit (OBU), vehicle terminal, road-sideequipment (RSE), and ITS center/server. Both OBU and RSE have the communication module which is equipped with Modem, MAC and RF module. The OBU provides both V2V and V2I communications, and multi-hop communication is supported. Our system is design to support the IEEE 802.11p based system. The general features satisfy the following features:  RF frequency: 5.8GHz band (5.835-5.925GHz)  Channel bandwidth: 10MHz  Modulation: OFDM (BPSK, QPSK, 16QAM)  MAC protocol: CSMA/CA, EDCA  Routing: Position based

Fig. 1. Overall system architecture

28

3

W. Cho and H.S. Oh

Performance Measurements

In this section, we present the experimental measurement results of our system. For performance measurement, we use 5.85 GHz center frequency with QPSK signaling in freeway environment. With 8dBi gain of omni-antenna for both transmitter and receiver, the transmitter uses approximately 20dBm EIRP. We consider various packet lengths (512, 1024 and 1518 bytes) and vehicle speeds (20km/h, 60km/h, 100km/h, 120km/h, 160km/h and 180km/h). 3.1

Communication Range

Measurement Setup: One vehicle (vehicle 1) is located at the fixed point. To assure a proper distance for reaching the target vehicle speed, vehicle 2 is located approximately 3km away from the vehicle 1. The vehicle 1 continuously sends packets, and vehicle 2 moves towards the vehicle 1. On receiving the packet at the vehicle 2, whether the received packet is error or not. If the vehicle 2 indicates “no error”', that distance is recorded as the communication range. Otherwise, the vehicle is moving about 50m from that distance, and the packet error is checked again. Results: The correct reception starts when the vehicle 2 passes approximately 1km distance from the vehicle 1 regardless of the vehicle speed and packet size. 3.2

PER

Measurement Setup: Two vehicles are located approximately 3km away from each other. Two vehicles move towards incoming direction and vehicle 1 sends 2000 packets when the distance between the vehicles is 500m. Then, the received packets are recorded. For PER, we use the maximum vehicle speed of 120km/h, which corresponds 240km/h in relative speed. Results: For unicast, regardless of vehicle speed and packet size, the worst PER is 0.15%, which corresponds to 1997 reception out of 2000 transmissions. However, this result does not count the retransmission in unicasting. For broadcast, the worst PER is approximately 5% with 1518 bytes of packet size, and it is observed that the PER decreases for long packet in general. 3.3

Link Setup Time

Measurement Setup: Initial setup is the same as the case of measuring communication range. The vehicle 2 moves towards the vehicle 1, and the vehicle 2 sends a packet to the vehicle 1 when the distance between the vehicles is 500m. Then, the vehicle 1 returns the acknowledgement signal. When the vehicle 2 receives the acknowledgement signal, the response time is recorded. Results: The link setup time is approximately 1ms regardless of the vehicle speed and packet size.

A Multi-hop Communication Scheme for IEEE 802.11p

3.4

29

Latency

Measurement Setup: The setup for measuring latency is the same as the link setup time measurement. The only difference is in the measured layer. Link setup time is measured between the MAC layers of the transmitter and receiver, whereas the latency is measured between the application layers. Results: Our measurement results reveal that the latency is different depending on the packet size, and the vehicle speed does not affect the latency. The average latency is approximately 3ms, 5.2ms, and 7.4ms for 512 bytes, 1024 bytes, and 1518 bytes, respectively. In this section, we represented the practical measurement results of communication system. Summarizing, the results indicate that it is possible to support reliable communication link with low link setup time, low latency, and less than 10\% PER with high mobility, i.e., 240km/h.

4

Vehicle Multi-hop Communications

In this section, we introduce VMP including measurement results. 4.1

VMP-BROADCAST

VMP-BROADCAST is used for broadcasting the data to all nodes in the specific area. If the designated nodes are located within the communication range, the source node simply transmits data by broadcasting. Otherwise, the data is transmitted by using multi-hop communication of the forward nodes (forwarder) where the forwarder is predetermined by the source node. The destination node of VMPBROADCAST is based on the location not identification number. Therefore, all nodes located in the designated area are regarded as the target nodes and receives the data. The nodes check the location information using the header information. Depending on the location information, each node set the node as the target node or deletes the received information. Fig. 2 represents the data flow of VMP-BROADCAST. Before transmitting VMPBROADCAST message, each node, i.e., a vehicle, broadcasts “hello” message to all nodes. The “hello” message is transmitted every 100msec. Then, each vehicle knows the location, moving direction and speed of neighbor nodes using “hello” message. The source node can predetermine the destination as the designated area or the number of hop. In Fig. 2, the vehicle 1 broadcasts the data to the designated area. Since the vehicle 1 has the information of neighbor vehicles, vehicle 1 can select the forwarder for multi-hop communication; the vehicle 3 is selected as the forwarder in Fig. 2. When the vehicle 1 broadcasts the data, both the vehicle 2 and 3 receive the data, and the vehicle 2 deletes the received data, whereas the vehicle 3 forwards the received data. The vehicle 3 can also decide the next target nodes. In our example, the vehicle 4 is the destination node, and vehicle 3 broadcasts the data to the vehicle 4.

30

W. Cho and H.S. Oh

Fig. 2. Data flow of VMP-BROADCAST

4.2

VMP UNICAST

VMP-UNICAST is used for transmitting the data where the geographical information of specific node is determined. The data flow of VMP-UNICAST is depicted in Fig. 3. When the target node is the one-hop neighbor node, the source transmits the data to the target node directly by unicast. When the location of target nodes is undetermined, first the VMP-location request (VMP-LREQ) message is flooded to all networks by broadcast to detect the current location of the specific node. Notice that the information of neighbor nodes is already known to the source node by using “hello” message as the same as the case of broadcast. When VMP-location reply (VMPLREP) is received, the location of target node is determined. Then, the closest node between the neighbor node and the target node is regarded as the forwarder. Finally, the data is transmitted to the target node by the multi-hop unicast of forwarders. 4.3

Experimental Measurements

In this section, we represent experimental measurement results of VMP. The performance of VMP is measured with respect to packet delivery ratio and throughput. Due to the long communication range, it is hard to establish the measuring environment for broadcast. Therefore, our measurement is carried out using unicast. Measurement Setup: We use 5 OBUs and measure the performance up to 4 hops. To measure the performance, 5 OBUs are located sequentially, and each OBU is set up to know only the information of adjacent OBU(s). Then, the source node generates and transmits the data using Smart bit instrument.

A Multi-hop Communication Scheme for IEEE 802.11p

Packet delivery ratio (%)

Fig. 3. Data flow of VMP-UNICAST

Fig. 4. Packet delivery ratio depending on the number of hop

31

32

W. Cho and H.S. Oh

5 4 3 2 1 0 1 hop

2 hop

3 hop

4 hop

Fig. 5. Throughput depending on the number of hop

Results: For packet delivery ratio, the length of 64 bytes and 1024 bytes frame is used with 16 QAM. For each length, 100000 frames are transmitted. Fig. 4 represents the measurement results of packet delivery ratio depending on the number of hop. The results show that the packet delivery ratio decreases as the number of hop increases and the packet length is long. This is due to the increment of collisions between the nodes. For throughput measurement, we use 1024 bytes with 16 QAM. Fig. 5 depicts the throughput depending on the number of hops. Similar to the Fig. 4, the throughput decreases as the number of hop increases. This is also due to the nature of CSMA/CA. For increasing the packet delivery ratio and throughput, an advanced MAC may be required, which is directly related to the collision issue in IEEE 802.11 based standard.

5

Conclusions

In this paper, we investigated the multi-hop communication schemes for IEEE 802.11p based communication systems. We first introduced the general system architecture and showed some basic measurement results of V2V communications. Then, a multi-hop scheme for broadcast and unicast is studied. Practical measurement results showed that the overall performance of multi-hop communication decreases as the number of hop increases, and it may be required a new MAC to guarantee the reliable communication links.

References 1. Acosta-Marum, G., Ingram, M.A.: Six time-and frequency-selective empirical channel models for vehicular wireless LANs. IEEE Vehicular Technology Magazine 2(4), 4–11 (2007)

A Multi-hop Communication Scheme for IEEE 802.11p

33

2. Bai, F., Krishnan, H.: Reliability analysis of DSRC wireless communication for vehicle safety applications. In: Proc. of Intelligent Transportation Systems Conference, pp. 355– 362 (2006) 3. Biswar, S., Tatchikou, R., Dion, F.: Vehicle-to-Vehicle wireless communication protocols for enhancing highwary traffic safety. IEEE Commun. Mag. 44(1), 74–82 (2006) 4. Cheng, L., Henty, B.E., Cooper, R., Stancil, D.D., Bai, F.: A measurement study of timescaled 802.11a waveforms over the mobile-to-mobile vehicular channel at 5.9GHz. IEEE Communicaions Magzine 46(5), 84–91 (2008) 5. The CAMP Vehicle Safety Communications Consortium, Vehicle Safety Communications Project Task 3 Final Report, U.S. Department of Transportation (2005) 6. IEEE Std 802.11p, IEEE standard for information technology-telecommunications and information exchange between systems-local and metropolitan area networks-specific requirements, Part 11, Amendment 6: Wireless Access in Vehicular Environments (2010) 7. Jiang, D., Taliwal, V., Meier, A., Holfelder, W., Herrtwich, R.: Design of 5.9GHz DSRCbased vehicular safety communication. IEEE Wireless Commun. 13(5), 36–43 (2006) 8. Kukshya, V., Krishnan, H.: Experimental measurement and modeling for vehicle-tovehicle dedicated short range communication (DSRC) wireless channels. In: Proc. of Vehicle Technology Conference-Fall, pp. 1–5 (2006) 9. Nekovee, M.: Quantifying performance requirements of vehicle-to-vehicle communication protocols for rear-end collision avoidance. In: Proc. of Vehicle Technology ConferenceSpring, pp. 1–5 (2009) 10. Xiang, W., Haung, Y., Majhi, S.: The design of a wireless access for vehicular environment (WAVE) prototype for intelligent transportation system (ITS) and vehicular infrastructure integration (VII). In: Proc. of Vehicle Technology Conference-Fall, pp. 1–2 (2008) 11. Wellens, M., Westphal, B., Mähönen, R.: Performance evaluation of IEEE 802.11-based WLANs in Vehicular scenarios. In: Proc. of Vehicle Technology Conference-Spring, pp. 1167–1171 (2007) 12. Yang, X., Liu, J., Zhao, F., Vaidya, N.H.: A vehicle-to-vehicle communication protocol for cooperative collision warning. In: Proc. of International Conference on Mobile and Ubiquitous Systems: Networking and Services, pp. 114–123 (2004)

A Political Communication Scheme of Citizen Network System on Disembedding and Embedding Principle Jang-Mook Kang1 and Bong-Hwa Hong2,* 1

Electronic Commerce Research Institute, Dongguk University, 707 Seokjang-dong, Gyeongju, Gyeongsangbuk-do, 780-714, Korea

[email protected] 2

Dept. of Information and Communication, Kyunghee Cyber University, Hoegi-dong, Seoul, 130-701, Korea

[email protected] Abstract. This study looks into disembedding and embedding, which could occur when there is a restructuring from the conventional hierarchical order into network principle brought about by mobile network age. By nature, the advent of smart device transforms mechanism that is dichotomous, hierarchical, vertical and central into one that is dispersive, horizontal, relational and so on. It can also encourage collective collaboration and self-purification of creating a virtuous circle of ‘participation-sharing-openness’ during circulation of political information. This study will propose citizen network system that delivers communication of desirable political information through disembedding and embedding principle, of which the system will hopefully evolve into a scheme supporting political communication. Keywords: Political Network, Smartphone, SNS (social network service), Mobile-Web, participation, shares, openness

1

Introduction

The emergence of ubiquitous computing was made possible by sensor performance and network expansion. Sensor precision, network bandwidth expansion and computers’ upgraded processing capacity benefit humans as they are applied to diverse services for citizens. Specifically, web 1.0 put computers at the center of information processing and put humans at the periphery by clustering them. Web 2.0, on the other hand, is a more individual oriented trend, which put interaction between computer processing and network on the periphery and hence delivers individualized services in a distributed environment. Such introduction of services optimized for individuals, not groups brought about faster Internet, faster reaction rate of computers and development of mobile devices. *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 34–42, 2011. © Springer-Verlag Berlin Heidelberg 2011

A Political Communication Scheme of Citizen Network System

35

In particular, Internet environment has changed and developed drastically, so few users now access it via modems, which were widely used a few years ago [1]. As a result, the traditional Virtual Terminal (VT) environment has been replaced by the Graphic User Interface (GUI), and text‐based simple HTML (Hypertext Markup Language) service has become a multimedia based service [1]. Graphic processing optimized to users and seamless Internet that allows access to the network anytime, anywhere propelled huge political, social and cultural changes. This study will observe the impact mobile network and the subsequent changes in political network have on citizen participation and propose ideal communication model.

2

Related Works

2.1

IT-Driven Mass Participation

New terms and technologies such as cloud computing, smartphone, mesh network and social network service are given different approaches. For example, they are dubbed as new media in journalism and mass communication and Internet politics or network politics in political science. ICT (information and communication technology) is used as means to explore new government in public administration. Early on, Manuel Castells said the application of new technologies in social science will have tremendous ripple effects. If information technology is better theorized and incorporated into the central social science theories that guide thinking about how government works, researchers will possess more powerful tools for explanation and prediction [2]. Fig. 1 shows Technology Enactment Framework. Technologies will not grow in the market and community unless technologies become part of the framework and system. The community will not advance either if it fails to properly absorb new technologies as part of its system and culture. As indicated in Fig. 1, citizens leverage instituted technologies to engage themselves. Prerequisites for enacted technology are trust, social capital and interoperability. The government also needs to commit itself to hierarchy, jurisdiction, standardization, rules, files and stability to protect transparency and efficiency. By learning how the government, market, individuals and technologies structure themselves and interact with one another, we can figure out how technologies get through for proper use by citizens. This study gives an observation on citizenengaging platform suggested as a concrete service in a network structure, which boosts efficiency of political information and accordingly looks into the environment on which the system operates. Fig. 1 provides a general glimpse on operating principles and background for ‘participation, sharing, openness’ required to engage citizens to politics.

36

J.-M. Kang and B.-H. Hong

Fig. 1. The Technology Enactment Framework [3]

2.2

Social Network Service (SNS)

Me2day (http://www.me2day.net) by Naver and yozm (http://yozm.daum.net) by Daum are hot social network services in Korea these days. Google’s Google Plus (https://plus.google.com), Twitter (http://twitter.com) and Facebook (http://www.facebook.com) are also popular social network services. Although there is a slight difference from one service to the other, all of them share ‘openness’ in their network and put no restrictions on moving or connecting contents.

A Political Communication Scheme of Citizen Network System

37

Fig. 2. The Technology Enactment Framework [4]

Being such, users have the freedom to edit or give away contents no matter their subject, theme, preference, published date or producer. For instance, there are already 200 million Twitter accounts and 460,000 new accounts are added every day. 97% of users have less than 100 followers and 81% of users follow less than 100. 70% of Twitter accounts were opened outside of the U.S. and 75% of Twitter traffic comes from outside Twitter. 40% of Twitt is sent from mobile device and there are one million applications related to Twitter. Twitter, me2day, Facebook and yozm broaden open network sphere by competing and cooperating at the same time and the evolution of such social network service is introducing a new form of communication. 2.3

Long Tail and Political Information

Chris Anderson of Wired Magazine described a new business model called “The Long Tail”, which is defined as ‘on-line retailers are finding that even the most obscure content sells at an acceptable level on line’. The following Fig. 3 is the long tail for digital sound.

38

J.-M. Kang and B.-H. Hong

Fig. 3. Monthly Download Performance of Rhapsody-Source-Wired Magazine [5]

If long tail as described in Fig. 3 can be likened to politics, it is high-profile politicians, influential officials and headliners who consumed political information in the age of web 1.0 but it is ‘the small stories of ordinary citizens are newly emerging as political information’ in the age of web 2.0. For example, President Barack Obama raised $319.9 million from 3.1 million supporters during his presidential campaign in 2009 and 50% of them donated less than $200. Their average donation was $86 and most of them went online for donation. As in Obama’s case, long tail revolutionized big-sized fundraising by just a few into donation of small amounts by many citizens. iPhone application YouTube (http://www.youtube.com), Obama’08 (http://www.barackobama.com/iphone-demo), MyBo (http://MyBarackObama.com), Obama blog, Obama homepage (www.barackobama.com) and Facebook (www.facebook.com/barackobama) joined forces to engage citizens and help them voluntarily raise funds for which they support.

A Political Communication Scheme of Citizen Network System

39

Long tail is a new theoretical framework applicable to CD, digital sound all the way to politics and the community. In-depth discussions on its technological potential and its institution into a system make it more concrete. 2.4

Other Research

Aside from research on technology and social science point of view, latest research on mobile phone focused on smart phone. The following technologies are also

available for mobile devices: First, there is the micro-grid technology and clouding platform. With this diversification of distribution systems, distributed generators used in existing systems can be subdivided into smaller units, called micro-grids [6]. The application of smart grid technology has turned the Micro-Grid into a system that can digitize in real-time all the processes of power generation, distribution and demand chain [7]. Second, recent advances in wireless communications and electronics have enabled the development of low-cost, low-power, multi-functional sensor nodes [8]. These sensor nodes leverage the idea of sensor networks [9]. A Ubiquitous Sensor Network [10] is a wireless network which consists of a large number of lightweight, lowpowered sensor nodes. Such sensor nodes consist of sensing, data processing and communicating components. Sensor networks are drawing a lot of attention as a way of realizing a ubiquitous society [8]. These researches can be interpreted as the attempts to apply various sensors to many fields of the society (education, medical service, government sector, commerce, etc.) [11].

3

A Political Communication Scheme of Citizen Network System

3.1

Overview and Problems of the Proposed System

This study is to propose the architecture design and method of A Political Communication Scheme of Citizen Network System. This system addresses the limitations and problems of the real world as described below, and is proposed as a new model for political communication. Hardly anyone rushing to his/her own business on the streets gives their ears to the suffering who has been treated unfairly. Just a little engagement and interest is enough but taking time to listen to one’s frustrations or consume political information entails cost. For sure, communication of political information will become much more active if such transaction cost can be minimized. The divergent stories made by citizens and individuals, in particular, are cases in point of disembedding that had been adjusted and negotiated through social

40

J.-M. Kang and B.-H. Hong

institutions and laws. If such settlement failed to be effective, citizens hold illegal rallies or violent demonstrations to draw people’s attention. This study proposes a model that addresses such disembedding with communication technology to search for embedding in political information. 3.2

System Configuration

The following Fig. 4 is a system aspired by this study to solve political information containing citizens’ voices and conflicts through a social method characterized by ‘sharing-participation-openness.’ Social network system and basic technologies for service are required to deliver this technology.

Fig. 4. Social embedding system to address citizens’ disembedding

A Political Communication Scheme of Citizen Network System

41

Fig. 4 is a diagram on an effective sharing system module that leverages augmented reality technology to share, expand, reproduce or transmit civil movements or citizens’ voices with others. Its development can be described by each stage of system realization as follows: First, store location information in a smart mobile device via the built-in location information receiving function.. Second, send received location information to management server together with current (at the time of transmission) time information. Third, search civil movements matching with the foregoing information location and time information in DB. Fourth, send AR images on civil movements or individual voices to smart mobile device. Fifth, display the foregoing smart mobile device, which will display an image combining the real world reflected in the displayers via built-in camera and the foregoing AR image.

4

Discussion and Conclusion

This study observes disembedding and embedding, which could occur when the advent of mobile network age realigns the existing hierarchical order into network principles. It starts with a different viewpoint as the proposed system does not view voices of citizens or NGOs as conflicts or disembedding. In other words, political information can enter into a virtuous circle only when self-defensive arguments raised by civil movements and individuals are perceived as contents for communication. And this system can maximize its impact in an environment upholding virtuous circulation of political information. The emergence of smart devices inevitably restructures dichotomous, hierarchical, vertical and central mechanism into distributed, horizontal and relational principles. It also has the power to encourage collective collaboration and self-purification capable of creating a virtuous circle of ‘participation-sharing-openness’ in circulating political information. This study proposes a political communication scheme of citizen network system on disembedding and embedding principle. It is particularly designed in a network structure to support connection with social network service. An open network structure is also a condition to turn information of disembedding into one of embedding through the system. Acknowledgment. This work was supported by the National Research Foundation of Korea Grant funded by the Korean Government (NRF-2011-330-B00017). This study revised part of ‘Social Network System for Citizens Campaign’s ShareParticipation-Open’ by inventor Kang, J.-M and Yun, S.-H, patent application number (10-2011-0046410) of the Republic of Korea in thesis format. This study is a restructure of presentation on [Network Political Theory and Smart Technology; social citizen network service and augmented reality system connecting

42

J.-M. Kang and B.-H. Hong

disembedding and embedding] on Apr. 30th, 2011 at Kyunghee University, Hoegidong, Seoul, Republic of Korea into a thesis. I would like to express my fullest appreciation to presenters Dr. Lee, H.-C. and Dr. Han, J.-H. for making good points during seminar hosted by Social Sciences Korea in April.

References [1] Hong, S.-R., Na, W.-S., Kang, J.-M.: A QoS Scheme for a Congestion Core Network Based on Dissimilar QoS Structures in Smart-Phone Environments. Sensors 10, 1424 (2010) [2] Castells, M.: The Network Society From Knowledge to Policy, Center for transatlantic relations, p. 179 (August 31, 2011), http://www.umass.edu/digitalcenter/research/pdfs/ JF_NetworkSociety.pdf#page=28 [3] Fountain, J.E.: Building the virtual state: Information Technology and Institutional Change, p. 91. Brookings Institution Press, Washington, D.C (2001) [4] http://goo.gl/JlpS4 (August 31, 2011) [5] Wired magazine [6] An, J.-B., Yoo, D.-W., Park, J.-H.: KERI; Busan university ministry of education science and technology. In: Development of Autonomous Demand-Management Type MicroGrid; KERI: Changwon, Korea, pp. 20–45 (2006) [7] Moon, H.-H., Lee, J.-J., Choi, S.-Y., Cha, J.-S., Kang, J.-M., Kim, J.-T., Shin, M.-C.: A Study Using a Monte Carlo Method of the Optimal Configuration of a Distribution Network in Terms of Power Loss Sensing. Sensors 11, 7824 (2011) [8] Lee, W.-J., Kim, J.-I., Kang, J.-M.: Automated Construction of Node Software Using Attributes in a Ubiquitous Sensor Network Environment. Sensors 10, 8664 (2010) [9] Akyildiz, I.-F., Su, W.-L., Sankarasubramaniam, Y., Cayirci, E.: A survey on sensor networks. IEEE Commun. Mag. 40, 102–114 (2002) [10] Chong, C.-Y., Kumar, S.-P.: Sensor networks: Evolution, opportunities and challenges, vol. 91, pp. 1247–1256. IEEE (2003) [11] Kang, J.-M., Hong, B.-H.: A Study on the SNS (Social Network Service) Based on Location Model Combining Mobile Context-Awareness and Real-Time AR (Augmented Reality) via Smartphone Communications in Computer and Information Science. 184, Part 2, 302–304 (2011)

Web Contents Mining System for Real-Time Monitoring of Opinion Information Ho-Bin Song1, Moon-Taek Cho1, Young-Choon Kim2,*, and Suck-Joo Hong3 1

Dept. of Electrical & Electronic Engineering, Daewon University, 599 Sinwol-dong, Jecheon, Chungbuk, 380-702, Korea [email protected], [email protected] 2 Dept. of Car Engineering, Kongju National University, 275 Budae-dong, Seobuk-gu, Cheonam-si, Chungnam, 331-717, Korea [email protected] 3 Dept. of Information and Telecommunication, Kyung Hee Cyber University, 1 Hoegi-Dong, dongdaemun-Gu, Seoul, 130-701, Korea [email protected]

Abstract. As the use of the Internet has recently increased, the demand for opinion information posted on the Internet has grown. However, such resources only exist on the website. People who want to search for information on the Internet find it inconvenient to visit each website. This paper focuses on the opinion information extraction and analysis system through Web mining that is based on statistics collected from Web contents. That is, users' opinion information which is scattered across several websites can be automatically analyzed and extracted. The system provides the opinion information search service that enables users to search for real-time positive and negative opinions and check their statistics. Also, users can do real-time search and monitoring about other opinion information by putting keywords in the system. Keywords: Motoring Search System, Opinion Information Automatic Extraction, Web Contents Mining, Opinion Information Monitoring.

1

Introductions

As the use of the internet gradually gets active lately, many people tend to express their opinions on the internet through the media such as Blog and Wiki[1]. And, in evaluating the value of the specific information, such demand for referring the opinion information other people put online is increasing. However, such opinions existing on the internet exist only at individual web site, and the user is to search all of such individual web sites manually in order to use such opinion informations. To settle this problem, technology for extracting the opinion of the user is actively being studied in home and foreign academics, and various technologies are being studied in *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 43–56, 2011. © Springer-Verlag Berlin Heidelberg 2011

44

H.-B. Song et al.

a field of information retrieval by great improvement from early 2000[1,3,5]. But the existing information retrieval technology is simply providing retrieval based on information having a keyword, and not being able to provide high-dimensional retrieval based on contents rated positively / negatively in document and sentence which each keyword appears. Recently an attempt to apply the technology for extracting the opinion of the user on information retrieval is in progress, but is in level of simply separating positive, negative document yet. This thesis suggests web contents mining system for real-time monitoring of opinion information to settle this problem. Suggesting system provides opinion information information retrieval service able of retrieval and statistics in each positive/negative opinion by automatically extracting and analyzing opinion information of user from web contents scattered in many websites existing on internet. As a result, opinion retrieval users easily can use the system searching and monitoring opinion information of other users on the specific keyword readily at eye, and the function of automatically extracting and analyzing opinion information real-time in web contents is provided. Construction of this thesis is as follows. In chapter 2, existing web mining technique, opinion extracting technique and theoletical background of multilingual linguistic dictionary are examined and the problem is examined for theoletical inquiry of this thesis. In chapter 3, plan and design method of web contents mining system able to collect and analyze opinion information on the internet is suggested. Finally in chapter 5, conclusion Is formed. 2

Relevant Study

2.1

Web Mining Technique

Web Mining is aimed at all data originated on the web or existing on the internet, and indicates the process of extracting and analyzing useful information by applying data mining technique based on such data. That is the application which data mining technique is applied to the web, massive data assembly [2,4,5]. Such web mining uses data mining technique to find and extract information automatically from web document and service. So, it can be defined as the process of finding useful information and knowledge not known previously from web data. Field of study of web mining is holding in common much parts studied in the field of Information Retrieval or Information Extraction [3,4,5]. 2.2

Opinion Extracting Technique

This is the study of opinion classification in which unit of document and sentence is classified more in detail into unit of phrase and word. Study of classifying opinion at the unit of phrase and word is studied by the method based on rule at the beginning, and the machine learning method of studying information around phrase and word and deciding polarity of phrase and word is studied afterward [7,8].

Web Contents Mining System for Real-Time Monitoring of Opinion Information

45

2.2.1 Rule-Based Model Study of Nasukawa and Zhongchao Fei is done as the study of extracting opinion in unit of phrase based on rule-based method [9,10]. Part of speech information and polarity information of each word are put together with that word and tag is attached [11,12]. Polarity information is determined to three, good, bad, neutralm, and word part of speech being object is determined to adjective, noun, adverb and verb. Tag is attached on positive verb and negative verb. At this point, tag is attached on one pertinent word or also on phrase forming one expression with that word and showing specific polarity. 2.2.2 Machine Learning Based Model It was mainly focused on the method of manually constructing opinion expressing resource at previous rule-based method. Machine learning is carried out using Corpus tagged to positive or negative expression part in sentence on the machine learning method. After automatic construction of tagged Corpus, the machine learning for opinion classification of word / phrase unit using this Corpus is done. HMM (Hidden Markov Model) is used as machine learning for opinion classification [16].

Fig. 1. Probabilistic Parameters at HMM

2.2.3 Pros and Cons of Prior Method Prior rule-based method has merit of showing high accuracy on corresponding pattern. So it can be the good method to approach if many opinion patterns could be constructed delicately. But, as the problem the rule-based method was constantly being pointed out, in case of rule-based method, there is weak point of which reproducibility drops sharply in case pattern already constructed come out in transformed form, and there is limit of not being able to reflect surrounding context information to studying. Constructing and maintaining this opinion pattern on all other domains and linguistic range is the hard work necessary of much manpower and time. There is difficult which opinion resource is to be constructed manually in both rule-based method and machine learning based method for opinion classification of word / phrase. For the opinion classifying of word / phrase unit is possible if such opinion resource is constructed, one of the most important problems in opinion classifying of word / phrase unit could be seen as construction of Corpus which opinion pattern construction or opinion expression is tagged.

46

2.3

H.-B. Song et al.

Multilingual Linguistic Dictionary

2.3.1 Foreign Words Automatic Transcribing Model Study on automatically constructing translation knowledge used in application of natural language such as machine translation and information retrieval of cross lanugage actively has been progressed [6,7]. Phonetic translation means of generally transcribing English word in language of non-English-area based on pronunciation. Many of the words phonetic translated are not registered in dictionary for many of them are coined words showing new ideas. Therefore automatically obtaining translation knowledge of phonetic translation is very important to build effective translation knowledge. We have automatic phonetic translation and phonetic translation interlinear pair extraction, etc. as the study of obtaining phonetic translation interlinear words on given English words. Automatic phonetic translation is the technique of phonetic translating given English word into a word of non-English-area [7]. Phonetic translation interlinear pair extraction, the study of automatically extracting English and phonetic translated word corresponding to English from the form bilingual corpora to widen applicative range of translation dictionary, and the translation knowledge is limited to phonetic translation interlinear pair. Automatic phonetic translation and phonetic translation interlinear pair extraction are actively being progressed as the method of handling phonetic translated word, but the study of integratively using these two methods is not thoroughgoing enough. 2.3.2 Statistics Based Phonetic Translation Model Generally Roman notation of Chinese characters or Pinyin is used in comparing with English in Chinese phonetic translation interlinear pair extraction [15]. With assumption of E as English, C as Chinese, TU (Translation Unit) as phonetic translation unit in statistics based phonetic translation model, conditional probability P(C|E) is substituted with P(Chinese|English) and can be converted to the problem of seeking P(C|E) probability. And, Unigram, Bigram, Trigram for English, and Pinyin's first syllable, last syllable or the whole Pinyin for Chinese are used as TU. Method of automatically presuming parameter by applying EM(Expectation Maximization) algorithm [13] without pronouncing dictionary is used, and match type information is added to phonetic translation model. With adding match type (M) to P(C|E) formula, we have formula 1. P(C|E)

≒ max P(C|M,E)P(M|E) ≒ max P(C|M,E)P(M)

(formula 1)

3

Web Contents Mining System for Real-Time Monitoring of Opinion Information

3.1

Outline of System

Web contents mining system for real-time monitoring of opinion information is the system to automatically extract and analyze opinion information in web contents, and

Web Contents Mining System for Real-Time Monitoring of Opinion Information

47

that platform is as figure 2. Proposed system is the system that provides opinion information searching service able to check search and statistics in each positive/negative opinion by automatically extracting and analyzing user opinion information from web documents scattered over many websites existing on the internet. Positive opinion and negative opinion are automatically extracted. Proposed system of figure 2 is formed by including data collection processing, opinion/non-opinion automatic construction, Opinion information resources, indexing transaction, opinion indexing information resources, opinion expression machine learning, multilingual linguistic dictionary automatic registration, multilingual opinion information resources, opinion search transaction and user terminal, etc. are included in forming.

Fig. 2. Opinion information automatic extraction web mining system platform

3.2

Data Collection Processing

Data collection processing performs function of collecting various web contents existing on the internet. So, data collection processing downloads HTML(Hyper Text Markup Language) information of each Web Site existing on the internet in real time. Moreover, data collection processing can extract information data of at least one among necessary information such as text, image or video, etc. from web contents downloaded as above and store in separate data storing module. Data collection processing can sort and collect web contents including opinion information data (that is general sentence/document data and information data with positive/negative valuation on it) as table 1. Object data collected through data collection processing as shown in table 1 is the opinion information data that is the general sentence/document data and the information data with positive/negative

48

H.-B. Song et al.

valuation on it. At this timee, the above positive/negative valuation can be expressedd in point within fixed range or o variously valued using the asterisk( ) or other marrks. Positive/negative valuation n so expressed in various methods are all recalculatedd in same point range and used in i this thesis.

★

Table 1. Opinion Information Data

Expresssion

★★★★★ ★★★★★ ★★★★☆ ★★★★★ ★★★★☆

Point 10 10 8 9 8

★★★★★

10

★★★★★

10

★★★★★

10

★★★☆☆

6

★★☆☆☆

5

Opinion Contents Interesting. statement Story of ‘smart’ people. statement Wise people mending ordinary life! statement Fascinated by uncle ... statement Story of ordinary people. statement heart warming love story with spelendid acting and interesting story. statement Really touching story. statement Heart warming movie. Interesting also, statement Warm and funny ... Could be longer ... statement palpable story after all. statement

∼

To explain this in the con ncrete, if the point range of collected data is c d when the point range used in workin ng example of this thesis is a b, the pertinent collecttion point x is changed as formu ula 2.

∼

(formula a 2)

∼

For instance, in case this th hesis uses the point between 1 10 point (positive as clooser to 10 point), and the collectted data uses the point between 1 5 point, if the colleccted data is 2 point, it is calculatted as formula 3.

∼

(formula a 3)

Web Contents Mining System for Real-Time Monitoring of Opinion Information

49

Data collected by data collection processing is stored in opinion information resources data structure like table 2 to express in the set of opinion point {(data, point), (data, point), ... (data, point)} used in this thesis like chart 1. Table 2 is showing field name of opinion information resources data structure, data type on pertinent field and explanation on field. Table 2. Opinion Information Resources Data Structure

id

user_id

date

topic

sentence

polarity

.....

bigserial (PK)

char varying(200)

bigint

char varying(200)

char varying(1000)

char varying(10)

.....

This thesis has automatically building study corpus which opinion information "word/phrase" is tagged as an object. Thus opinion information "word/phrase" is automatically classified through machine learning method using corpus automatically built. At this time, data collection processing collects data which positive/negative opinion information in blocks of sentence easily sought in the internet is expressed using opinion information data structure of formula 2 and table 2 to automatically built corpus that opinion information "word/phrase" is tagged. 3.3

Opinion/Non-opinion Automatic Construction

Method of simply using the number of times appealing in positive/negative document based on rule like formula 2 is inaccurate on data in form of point like 1~10 point. And in case of using the absolute number of times appealed when the number of positive document and negative document are different, there is the problem of having the point leaned to the pertinent data set collection of bigger size. Method of automatically constructing opinion/non-opinion information resources in this clause has feature of automatically seeking through interpolation the positive/negative probability and the probability of appearing in opinion sentence of candidate word to be used after generating possible opinion information expressions by the whole dictionary based N-Gram analyzers. Opinion intensity of various point collection like 1~10 point is reflected in process of seeking positive/negative probability and probability of appealing in opinion sentence, and normalization is also proposed to settle the problem of point leaning of data size itself of specific point collection has grown. 3.3.1 Word Point Calculating Method of Proposed Method Data which opinion is indicated in sentences is used in this clause to automatically construct opinion information word resource. After that, the point on N-Gram of each morpheme is sought after dividing sentence into morphemes.

50

H.-B. Song et al.

shows the number n of times the word Wj appeals in the point collecttion Si. So supposing the point of 10 appeared 9 times and the point of 1 1 time for the word "movie", the extreme pooint using the frequency on "m movie" in positive sentence and negative sentence iss as formula 4.

(formula a

4)

Formula 4 is the examplle of which problem arises when former method of claause 2.2 is directly applied. It is used under the condition in which the sizes of poositive/negative data were sam me at the former study. Problem arises if the size of eeach data is calculated as formulla 4. In case the word “movie” appeared 9 times at 10 pooint collection and 1 time at 1 point p collection, very positive point of 1.6 is won by sim mply calculating with the above numerical formula. But as we see the example above, we can see that each word ap ppears more at 10 point collection for 10 point collecttion itself is big. So the problem m arises in case size of each point collection varies in seeeking polarity point of the word. As the example above, the t word "movie" is the word that appears a lot in all pooint collections in common at the t movie review. At this time, the problem of which the point is excessively biased to big point in case the absolute value simply is used in the condition which the absolu ute size of 10 points itself is big. Therefore, normalizattion without simply taking averaage of point is needed in case size of each point collecttion varies. Formula 5 is the numerical formula for converting absolute frequency to reelative probability. (formulaa 5) On formula 5, the point collection Si,

indicates the number of times the word Wj appearrs at means the value that added the numbeer of

times the word Wk appearss at all point collections, after all the number of times Wk appears in the whole data.. So is the value which absolute frequencyy is converted to relative probaability. Normalizing absolute frequency to relative vaalue having relative probability based on this comes to formula 6.

(formulaa 6)

Web Contents Mining Sy ystem for Real-Time Monitoring of Opinion Information

51

Normalized opinion poiint not leaned to specific point collection is obtainedd by changing the part which haad the average of absolute value formerly to the averagee of normalized values accordin ng to formula 6. 3.3.2 Calculating Meth hod Using Subjective Point Subjectivity of word as well as polarity point should be calculated in constructting word resources. At this timee, the element having an influence on subjectivity as callculating polarity point is to be b the generative probability of part of speech informattion of that word rather than thee generative probability of that word itself. It is becausee of being weak to various proper noun, naturalized word and other words not in stuudy data in case of generative probability p of word itself, and that part os speech inform mation can be usefully used fo or the specific part of speech combination exists. Subjecctivity of word can equally be calculated using the method of calculating point of w word previously proposed, but th he subjectivity point of that word can be sought by calcuulating in seeing calculating sub bject data as the part of speech data of opinion data and the part of speech data of non n-opinion data as positive(10 point) and negative(1 pooint) data each. Formula 7 is the method d of calculating subjective point in which indicaates part os speech information of word , and indicates each point collection. is seen as the data collection n not including opinion and is seen as the data colllection including opinion in caalculating subjective point.

(formulaa 7) Formula 8 is the propossal method automatically constructing opinion informattion word of this thesis which seeks the Opinion Score of word using two points, polaarity point and subjective point, and using interpolation method. This Opinion Score is the point showing how much opinion word that word is, thus the word under speccific point is not used as the opin nion word.

(formulaa 8) max(S) of the above numerrical formula means the maxium point collection from the point collection. Reason of having absolute value on part and Poolarity Score part is to have neg gative words heighten opinion point also at Polarity Scorre.

52

3.4

H.-B. Song et al.

Opinion Expression Machine Study

After opinion/non-opinion tagging corpus is automatically constructed, the machine learning for opinion classification in word/phrase is done. HMM described in relevant study of chapter 2 is used as machine learning for opinion classification. But there are evaluation problem, decoding problem and estimation problem to be settled for HMM to be actually applied [16]. Evaluation problem to be settled first is the problem of how to seek the probability observed in the model when sequence O=O1O2...Or and P ( O | λ ) of data O model λ = ( A , B , π ) of the observed symbol is given. Decoding problem to be settled second is the problem of what the optimum state transition sequence Q = q 1 q 2 ... q t is when sequence

O=O1O2...Or and model λ = ( A , B , π ) of observed symbol is. Estimation problem to be settled third is the problem of deciding model parameter λ = ( A , B , π ) showing the biggest π = { π i } . Three problems above can be settled by Forward Algorithm, Viterbi Algorithm and Baum-Welch Algorithm each. In this thesis, state transition probability, observation probability and initial state probability which is the model parameters are obtained from the tagging corpus collected from opinion/non-opinion automatic construction module. 3.5

Indexing Transaction

Indexing transaction performs the function of indexing for the opinion informations of pertinent web contents to be stored in opinion indexing information resource in linguistic qualities of opinion sentence sorted from opinion/non-opinion automatic construction. Opinion indexing information resource here performs the function of summary information of relevant opinion sentence of linguistic qualities of each opinion sentence indexed through indexing transaction and basic and opinion informations of relevant web contents to be stored as database. Table 3. Opinion Indexing Information Resource Data Structure

id

comm entct

date

snippet

data

pola rity

topic

url

...

big serial (PK)

char varying (50)

bigint

char varying (200)

char varying (2000)

char varying (10)

char varying (10)

char varying (200)

...

Opinion indexing information resource which opinion informations of pertinent web contents are to store in linguistic qualities classified from opinion/non-opinion automatic construction is stored in data structure like table 3. Table 3 shows the explanation on field name of opinion indexing information resource data structure, data type on pertinent field and field.

Web Contents Mining System for Real-Time Monitoring of Opinion Information

53

Fig. 3. Large Classification Indexing

Indexing process is constituted of the large classification indexing process to improve search speed and the detailed classification indexing to use in actual information search process by indexing guide word and contents information of each document. Large classification indexing shows the document including technical terms. Figure 3 below shows composition of large classification index.

Fig. 4. Detailed Classification Indexing

Detailed classification indexing is the process of forming document table to search actual documents including the keyword user presented. Information such as subject information, File name (storing channel), document contents and major related keyword included in document contents are stored in document table. Movie related keywords are extracted and stored through inquiring movie review sentence in analyzing morpheme here. Figure 4 below shows composition of detailed classification indexing.

54

H.-B. Song et al.

3.6

Multilingual Linguistics Dictionary Automatic Registration

In this clause, Effect of drastically reducing human power compared to existing passive linguistics dictionary constructing method is obtained by proposing the method of automatically constructing multilingual linguistics dictionary at double language on the internet using statistics based phonetic translation model. This clause suggests the phonetic translation interlinear pair extraction method at mass comparative corpus using phonetic translation frequency and phonetic similarity based on dynamic window and tokenizer technique applied to parallel corpus.

Fig. 5. Phonetic Translation Interlinear Pair Extraction Process in English-Chinese Corpus

English-Chinese phonetic translation automatic extraction model proposed in this clause first extracted proper noun applying proper noun recognizing module on English sentence of English-Chinese parallel corpus, chose only English words to be phonetic translated among them and extracted phonetic translation word from corresponding Chinese sentence. Figure 5 shows the process of extracting phonetic translation interlinear pair from English-Chinese parallel corpus, and the parallel corpus data structure is as table 4. Table 4. Multilingual Opinion Information Resource Data Structure

id bigserial (PK)

from language char varying(50)

to language char varying(55)

from_text

to_text

...

char varying(2000)

char varying(2000)

...

Web Contents Mining System for Real-Time Monitoring of Opinion Information

55

As observed in related study of chapter 2, error frequently occurs in extracting phonetic translation interlinear pair applying statistics based phonetic translation model if Chinese string similar in pronunciation with English word given in one sentence exist a lot. This thesis proposes dynamic window technique and tokenizer technique, non phonetic translation technique using entropy, and phonetic translation extraction technique using similarity and frequency of voice to settle error. 3.7

Opinion Information Search Transaction

Opinion information search transaction provided of specific opinion information or type information of user transmitted through web server and linked with indexing transaction or opinion indexing information storing resource, carries out function of searching indexing informations related with specific opinion search keyword or type information, and forwarding to web server to be transmitted to pertinent user.

4

Conclusion

Opinions existing on the internet exist only in individual web sites, so the user has to search such individual web sites one by one manually in case of using such opinion informations. Web contents mining system for real-time monitoring of opinion information is proposed in this thesis to settle such problems. Proposed system provides opinion searching service that can check retrieval and statistics in positive/negative opinions by automatically extracting and analyzing user opinion informations from web contents scattered in several web sites existing on the internet. Expected effect of the proposing system is that the users are able of easy and at a glance searching and monitoring of opinion information of other uses on the specific keyword and the time spent for searching opinion of other uses can be greatly shortened by automatically extracting and analyzing user opinion informations scattered in several web sites existing on the internet and providing opinion searching service to be able to check searching and statistics in positive/negative opinions. As the tasks to be solved, the web contents opinion searching system for the complete monitoring search engine is to be made by adding multilingual (Korean, Chinese, Japanese, English) search and machine translation function to solve language barrier on the internet and to be able of monitoring foreign informations in native language, and the reliability on opinion monitoring is to be verified.

References [1] Joo, H., Park, Y.: Design of Web Contents Mining System for Monitoring Search Engine. In: KICS, vol. 34(2), pp. 53–60 (February 2009) [2] Jang, N., Hong, S., Jang, J.: Data Mining. DaeChung, 32–56 (2007) [3] Anand, S., Bell, D., Hughes, J.: The Role of Domain Knowledge in Data Mining. In: CIKM (1995)

56

H.-B. Song et al.

[4] Anand, S., Hughes, J.: Hybrid Data Mining Systems: The Next Generation. In: PAKDD 1998, pp. 13–24 (1998) [5] Adriaans, P., Zantinge, D.: Data Mining. Addison Wesley Longman, England (1996) [6] Berry, J., Linoff, G.: Data Mining Techniques: For Marketing, Sales, and Customer Support. John Wiley & Sons (1997) [7] Kosala, R., Blockeel, H.: Web Mining Research: A Survey. In: ACM SIGKDD (July 2000) [8] Lee, C.H., Yang, H.C.: A Web Text Mining Approach Base on Self-Organizing Map. In: Proceedings of the 2nd International Workshop on Web Information and Data Management, WIDM 1999, Kansas City, MO, USA, pp. 59–62 (1999) [9] Mulvenna, M., Anand, S., Büchner, A.: Personalization on the Net using Web Mining. Communications of the ACM 43(8) ( August 2000) [10] Dagan, I., Church, K.W., Gale, W.A.: Robust bilingual word alignment for machine aided translation. In: Proceedings of the Workshop on Very Large Corpora, pp. 1–8 (1993) [11] Lee, J.S., Choi, K.S.: Enflish to Korean Statistical transliteration for information retrieval. Journal og Computer Processing of Oriental Languages 12(1), 17–37 (1998) [12] Kang, B.J., Choi, K.-S.: Automatic Transliteration and Back-transliteration by Decision Tree Learning. In: Proceedings of LREC (2000) [13] Goto, I., Kato, N., Uratani, N., Ehara, T.: Transliteration Considering Context Information Based on the Maximum Entropy Method. In: Proceedings of MT-Summit IX (2003) [14] Yan, Q., Grefenstette, G., Evans, D.A.: Automatic transliteration for Japanese-to-English text retrieval. In: Proceedings of ACM SIGIR 2003, pp. 353–360 (2003) [15] Paola, V., Paola, V., Khudanpur: Transliteration of Proper Names in Cross-Lingual Information Retrieval. In: ACL 2003 Workshop on Multilingual and Mixed-language Named Entity Recognition (2003) [16] Dorre, J., Gerstl, P., Seiffert, R.: Text Minin g:Finding Nuggets in Mountains of Textual Data. In: Dorre, J., Gerstl, P., Seiffert, R. (eds.) Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (1999) [17] Yan, L.H.: Text Mining-Knowledge Discovery from Text. In: Trend in Knowledge Discovery from Databases ( June 29, 1999) [18] Fayyad, U.M., Piatetsky-Shapiro, G., Smyth, P.: From Data Mining to Knowledge Discovery. In: Advances In Knowledge Discovery and Data Mining, pp. 1–34. AAAI Press/MIT Press, CA (1996) [19] Sproat, R., Tao, T., Zhai, C.: Named Entity Tranliteration with Comparable Corpora. In: Proceddings of the 21st International Conference on Computational Linguistics (2006) [20] Oh, J.-H., Bae, S.-M., Choi, K.-S.: An Algorithm for extracting Engilish-Korean Transliterationpairs using Automatic E-K Transliteration. In: Proceedings of Korean Information Science Society (2004) [21] Lee, C.J., Chang, J.S., Jang, J.S.: Extraction of transliteration pairs form parallel corpora using a statistical transliteration model. Information Science 176, 67–90 (2006) [22] Lee, C.-J., Chang, J.S., Roger Jang, J.-S.: Alignment of bilingual named entities in parallel corpora using statistical models and multiple knowledge sources. ACM Trans. Asian Lang. Inf. Process 5(2), 121–145 (2006) [23] Satish, L., Gururaj, B.: Use of hidden Markov models for partial discharge pattern classification. IEEE Transactions on Dielectrics and Electrical Insulation (April 2003)

An Energy-Efficient Cluster-Based Routing in Wireless Sensor Networks Seongsoo Cho1, Bhanu Shrestha1, Keuk-Hwan La1, Bong-Hwa Hong2, and Jongsup Lee3 1

Department of Electronic Engineering, Kwangwoon University, 26 Kwangwoon–gil, Nowon-gu, Seoul, 139-701, Korea {css,bnu,Khra}@kw.ac.kr 2 Department of Information Communication, Kyunghee Cyber University, Dongdaemun-gu, Seoul, 130-701, Korea [email protected] 3 SK C&C, SK u-Tow4er, 25-1, Jeongia-dong, Bundang-gu, Seongnam-si, 463-844, Korea [email protected]

Abstract. In Wireless Sensor Networks (WSNs), sensor nodes depend on batteries for energy source. The ability to use limited energy efficiently is the key to determining the lifetime of networks and the amount of information transmitted. Low Energy Adaptive Clustering Hierarchy (LEACH) is a representative cluster-based routing protocol designed to ensure energy use efficiency. In this paper, a protocol scheme was proposed wherein member nodes (lower-level nodes) are designed to compare the currently sensed data with the previously sensed one and to switch to sleep mode when a match is achieved. The design is to help improve the transmission energy efficiency. The proposed scheme was tested via simulations and was compared with two existing cluster-based algorithms, i.e., LEACH and Threshold Sensitive Energy Efficient Sensor Network Protocol (TEEN). Performance evaluation was conducted based on the number of surviving nodes in each of the three networks (i.e., LEACH, TEEN, and proposed scheme) over time. The results indicated that the scheme contributed to greater energy efficiency by helping to increase the lifetime of the LEACH network by a maximum of 27%. Keywords: WSN, LEACH, TEEN, Clustering Algorithm, Energy Efficiency.

1

Introduction

A Wireless Sensor Network (WSN) consists of sensor nodes, each of which includes a microcontroller, a radio transceiver, and a sensing module. Data collected by the sensor nodes are sent to the sink node (the data aggregator) mostly through the multihop wireless mesh network. The WSN, deploying a large number of sensor nodes in a specific area, had first started out as the monitoring and patrol application as well as military application for missions involving areas to which humans have limited access. Since then, the network has expanded its application gradually, which now includes environmental monitoring, building hazard diagnosis, and patient monitoring T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 57–64, 2011. © Springer-Verlag Berlin Heidelberg 2011

58

S. Cho et al.

as well as health care service applications [1]. In the sensor networks, MANET (Mobile Ad Hoc Network)-like environment wherein no AP (Access Point) or other fixed infrastructure exists is used, wherein a relatively large number of sensor nodes are deployed throughout a sensor field covering a large area, resulting in various, dynamic topology. Also, autonomous and independent networks are formed between the sensor nodes [2-3]. One of the most important issues about WSN is ensuring the efficiency in energy use, which arises from the energy resource constraints imposed on sensor nodes. For sensor nodes, batteries are the main source of energy, but their operation characteristics do not allow the replacement or charge of batteries. There are three items to evaluate the performance of WSNs: energy efficiency; accuracy of the data; and service quality. Among these items, energy efficiency is the most important one. It acts as a factor that indicates the lifetime of the sensors, as they function, consume energy, and wear out over time [4]. The sensor nodes comprising a WSN are often small and have limited battery capacity. Moreover, the batteries, in most cases, are not replaceable or chargeable due to the characteristics of their operating environment. It is thus crucial that the sensor nodes be designed in such a way that they maximize the efficiency of their energy use and that the process involved be run effectively. One of the solutions to the energy problem is clustering [3]. Networks adopting clustering are hierarchical, and consist of upper-level and lower-level nodes. Because forwarding of data to a remote base station (BS) requires a tremendous amount of energy, transmitting is carried out by only a few upper-level nodes (cluster heads) selected from all the cluster nodes. A big energy spender, upper-level nodes are designed to make sure that each node have an equal chance of becoming a cluster head according to the set probability and transmit data. Taking turns helps extend the lifetime of the sensor network [4]. Low Energy Adaptive Clustering Hierarchy (LEACH), a representative cluster-based routing protocol, assumes that each lower-level node always has data to transmit [5]. The assumption means that in some cases data that were sensed previously may also be sensed afresh, which is a drawback. Under LEACH, nodes react immediately to changes in the sensed data. Threshold Sensitive Energy Efficient Sensor Network Protocol (TEEN), on the other hand, uses threshold values which help prohibit data sensed previously from being sensed afresh or from being sent to upper-level nodes [6]. But this feature can be a drawback to TEEN, since the sensed data failing to meet the thresholds will not be sent to upper-level nodes. Therefore, in this paper, an energy-efficient clustering technique taking into account the way lower-level nodes collect data was proposed.

2

Related Studies

2.1

Overview of Routing Protocols for Wireless Sensor Networks

A WSN is comprised of nodes that are connected to sensors detecting changes in the data under monitoring. Figure 1 illustrates the architecture of the network.

An Energy-Efficient Cluster-Based Routing in Wireless Sensor Networks

59

Fig. 1. Wireless sensor network architecture

Depending on network configuration, routing protocols for WSNs are divided into flat routing protocol and cluster-based hierarchical routing protocol. In flat routing protocol, the entire network is considered a unit, and all the nodes therein participate in routing function with equal probability. In comparison, a cluster-based hierarchical routing protocol uses various clustering processes to divide the network into a number of clusters (units) which group the nodes into a hierarchy according to their roles [4]. In this protocol, lower-level nodes collect sensed changes in data and send them to upper-level nodes. The upper-level nodes then aggregate the data and forward them to BS. Some of the best-known cluster-based hierarchical routing protocols are LEACH, LEACH-C (LEACH-Centralized), and TEEN [7]. The routing protocols can also be divided into proactive network protocol and reactive network protocol depending on the network’s mode of functioning and type of target application. In the proactive network protocol, the nodes in the field periodically switch on their sensors and transmitters but do so only during the time slots assigned to them. The nodes sense changes in the data and transmit them to the upper-level nodes. The proactive network protocol is suitable for applications that require periodic monitoring of data. LEACH and LEACH-C are among the examples of this protocol. In the reactive network protocol, all the nodes in the field are engaged constantly in sensing changes in the data. The nodes react immediately to the changes, and the data are transmitted immediately to the upper-level nodes. It is better suited for time-critical applications. TEEN is included in this protocol. 2.2

LEACH

A WSN using LEACH is built with several clusters, each of which has a cluster head (CH) and non cluster heads (Non-CHs). The CH controls all the sensor nodes within the cluster, fuses data sent by the sensor nodes, and forwards them to BS. Non-CHs, on the other hand, collect data and send them to the CH. Since it is in charge of aggregating data transmitted from Non-CHs and forwarding them to a remote BS, the CH consumes a lot of energy. Thus, the CH is selected from all the nodes at the beginning of a new round, according to the set probability. This taking turns allows each node to have an equal opportunity to become the CH [5].

60

S. Cho et al.

Fig. 2. Timeline showing operation of LEACH

As shown in Figure 2, the operation and structure of LEACH schedule are based on rounds. At each round, Non-CH nodes select their respective CHs (i.e., the set-up stage); data are then transmitted from Non-CHs to the CHs and to BS (i.e., the steadystate stage).[5] Unless it is their time slot, Non-CH nodes remain in sleep mode to save energy. Upon the completion of the current round, a new round begins, with new CHs selected, repeating the aforesaid process all over again. Under this protocol, Non-CH nodes send sensed data to the CHs even if they were the same as the ones sensed previously. In other words, the nodes transmit unnecessary data while consuming energy of the member nodes. Also, in LEACH, CHs are selected based on probability; and clusters are formed based on the location of the selected CHs. This system, therefore, could lead to clusters with non-favorable topology. 2.3

TEEN

Though its cluster structure is basically the same as that of LEACH (proactive network protocol), TEEN is a reactive network protocol where all nodes sense data continuously by reacting immediately to the changes occurring in the data. The reactive network protocol uses two types of threshold in sensing data: hard threshold (HT) and soft threshold (ST).[6] HT is the absolute value of the sensed data. When the value of the data collected by Non-CH nodes is either the same as or less than HT, the data get to be transmitted to the CHs. ST, on the other hand, is a small change in the value of the sensed data. When this value matches or exceeds data collected by Non-CH nodes, the data will get to be sent to the CHs. Once the network starts operating and when the sensed data reach their HT value, Non-CH nodes send the data to the CHs. (The value of the sensed data is stored in the Non-CH nodes.) Next, at the current cluster period, nodes will transmit data only when the current value of the sensed data exceeds HT and at the same time matches or exceeds ST. The purpose of using HT then is to have nodes transmit only the data of importance and to help reduce the number of transmissions conducted by them. In comparison, ST is set to detect minor changes in data when the value of the sensed data is greater than HT. A drawback to TEEN is that the use of thresholds leads to other problem, i.e., Unless the sensed data reach thresholds, communication with upper-level nodes will never occur.

An Energy-Efficient Cluster-Based Routing in Wireless Sensor Networks

3

61

An Energy-Efficient Clustering Technique Taking into Account Data Collection Practice

LEACH assumes that each lower-level node always has data to transmit. When each new cluster is created, Non-CH nodes select the closest CH based on the strength of the ADV (Advertisement) signal information transmitted by the CH. By selecting the closest CH, Non-CH nodes are expected to consume the least amount of energy during the transmission. Nevertheless, the consumed energy is still large compared to the one spent during the nodes’ sleep mode. Moreover, non-favorable cluster topology, where the distance between Non-CHs and their CHs is far, can result from the LEACH protocol’s use of probability in selecting CHs. Previous research on LEACH focuses extensively on how to improve the energy efficiency of CHs, the biggest energy spender, based on the aforesaid assumption of the protocol. In this paper, a method for extending the lifetime of a WSN was explored by focusing on how to reduce the energy consumption at Non-CH nodes, the majority that makes up the cluster. The proposed scheme offers an alternative to help improve the energy efficiency of the entire network while taking into account the data collection occurring at Non-CH level in cluster-based routing protocols. When applied to the current LEACH protocol, the proposed scheme undergoes the process illustrated in Figure 3. In the process, the set-up stage, during which the cluster is created, ends and then triggers a new period, during which Non-CH nodes at once sense data in the environment and compare them with the ones stored in their internal memory during their time slot. In the initial period, no previously collected data exist, and hence no match between data sets. The non-match will next allow the data to be stored in the nodes’ internal memory and at the same time to be transmitted to the CHs. Upon the completion of the transmission, Non-CH nodes switch to sleep mode, just as their counterparts in the current LEACH protocol do. When the second period begins, Non-CH nodes come back from sleep mode and start sensing the environment and collecting data during the second time slot assigned

Fig. 3. Operation of the proposed scheme applied to LEACH

62

S. Cho et al.

to them, according to the TDMA (Time Division Multiple Access) schedule created by the CH of their respective cluster. The data collected this time will be compared with the one stored previously in the internal memory of the nodes, in the same manner adopted during the first time slot. If the previous and current data achieve a match, the nodes will switch to sleep mode to save energy. If the data do not match, the nodes will store the current data in their internal memory and then transmit them to the CHs. This process will be repeated over and over again until the lifetime of the network expires. Note that with the proposed scheme, the advantage of using thresholds in TEEN (i.e., selective collection of data of interest from all the data sensed in the network field) is compromised to a certain degree, because Non-CH nodes in the proposed scheme simply compare the previously sensed data with the currently sensed one and either store and transmit or switch to sleep mode. But the proposed scheme will help solve the big problem in TEEN, i.e., the data collected by Non-CH nodes that fail to reach the thresholds will never get to be transmitted to upper-level nodes. In other words, the proposed scheme will: (a) maintain the characteristics of reactive network protocol by having Non-CH nodes continuously sense the data in the field, react immediately to changes in the data, and transmit them to upper-level nodes; and (b) simultaneously help solve the aforesaid problem of TEEN.

4

Performance Evaluation

4.1

Simulation Environment

For the simulations, it was made sure that the WSN had 100 nodes and a fixed BS. The nodes were randomly assigned to one of the three network algorithms, i.e., LEACH, TEEN, or the proposed scheme. For clustering and other issues, the parameters used in LEACH are listed in Table 1 [5, 8]. To help ease the simulation of the proposed scheme, an environmental scenario was created wherein the temperature was manipulated to fluctuate randomly between 0°C and 200°C at five-second intervals — as applied in TEEN. Regarding energy consumption, incorporating the proposed scheme into an existing routing algorithm for real-life simulation would result in energy consumption that is required for the data comparison process at NonCH level. But the comparison concerned involves only simple comparison between the previously stored data and the currently sensed one. Thus, the amount of energy consumed for the comparison was judged to be sufficiently small, and it was ignored during the simulation of the proposed scheme. Table 1. Simulation parameters for the proposed scheme

Parameter

Network Base station Number of E elec grid coordinate sensor nodes

Value

From (0,0) to (50, 175) (100,100)

100

50nJ / bit

Eamp

Initial energy/ k node

Eda

2J

5nJ / 10 pJ / bit bit / m2

100 pJ / bit / m2

5

E fs

Emp

0.0031 pJ / bit / m 4

An Energy-Efficient Cluster-Based Routing in Wireless Sensor Networks

63

Another issue to consider was the variability in chances of having the same data in the provided simulation environment wherein temperature changed randomly. Thus, the validity of simulation results needed to be improved. This was achieved by conducting as many simulations as possible (i.e., at least 10 times) and combining the data to help reduce the standard deviation. 4.2

Results and Analysis

Simulations were carried out to compare LEACH and the proposed scheme, and the number of surviving nodes in each protocol was compared over time. As shown in Figure 4, the existing LEACH protocol had nodes that survived a maximum of 630 seconds, whereas the proposed scheme contributed to the survival of nodes for up to 800 seconds. In terms of the maximum lifetime of the network, the proposed scheme increased it by approximately 27% compared to the LEACH model. The increased lifetime is mostly contributable to the reduced energy consumption at Non-CH level. Under the proposed scheme, the CH (upper-level node) consumed less energy, too, because there were less data transmitted from the Non-CH nodes in its cluster, and hence less information to fuse and less energy to spend. This indicates that increasing energy efficiency at Non-CH level is just as important as reducing energy consumption at CH level, i.e., the focus of the majority of the previous studies. Next, simulations were conducted to compare TEEN against the proposed scheme, i.e., to compare the number of surviving nodes over time while implementing TEEN’s hard mode (HT) as opposed to implementing the proposed scheme. Under TEEN, the simulation set-up including cluster configuration was mostly similar to that of LEACH. However, the radio electronics model used in LEACH had to be altered so that the model in TEEN would represent both the idle time power dissipation and the sensing power consumption. This was necessary because in TEEN algorithm all the nodes sense the environment continuously.

Fig. 4. Comparison of the no. of surviving nodes over time (LEACH vs. proposed scheme)

64

5

S. Cho et al.

Conclusion

LEACH, a representative cluster-based routing algorithm, assumes that each lowerlevel node always has data to transmit. The assumption means that in some cases data that were sensed previously and thus need not be sent to upper-level nodes may well be being transmitted afresh. The redundancy problem was what inspired this study. In this paper, a protocol scheme was proposed to help save energy at lower-level nodes that compare the currently sensed data with the previously sensed one during their time slot and decide whether or not to send the data to upper-level nodes (CHs). The proposed scheme was compared with LEACH and TEEN through simulations. The results indicate that the use of the proposed scheme helped increase the energy efficiency of lower-level nodes as well as decrease the energy consumption required during the data aggregation by upper-level nodes (CHs). As a result, the lifetime of the entire LEACH network increased by 26% at the maximum. Compared with TEEN, the proposed scheme performed slightly better TEEN and helped solve the biggest problem it has, i.e., the backfiring of TEEN’s use of thresholds to prevent redundancy in data transmission which in some cases acts to prevent communication with upper-level nodes entirely.

References 1. Ian, F., Akyildiz, F., Su, W., Sankarsubramaniam, Y., Cayirci, E.: A Survey on Sensor Networks. IEEE Communications Magazine, 102–114 ( August 2002) 2. Hac, A.: Wireless Sensor Network Designs. John Wiley & Sons, Ltd. (2003) 3. Santi, P.: Topology control in wireless ad hoc and sensor networks. ACM Computing Surveys (CSUR) 37(2) (2005) 4. Heinzelman, W.B., Chandrakasan, A., Balakrishnan, H.: Energy-Efficient Communication Protocol for Wireless Microsensor Networks. In: Proc. of the Hawaii International Conference on System Science, pp. 3005–3014 (January 2000) 5. Chandrakasan, P., Heinzelman, W.B.: Application-Specific Protocol Architectures for Wireless Networks. IEEE Transactions on Wireless Communications 1, 660–670 (2000) 6. Manjeshwar, A., Agrawal, D.P.: TEEN: A Protocol for Enhanced Efficiency in Wireless Sensor Networks. In: Proc. of 1st Intl. Workshop on Parallel and Distributed Computing Issues in Wireless Networks and Mobile Computing, San Francisco (April 2001) 7. Heinzelman, W.B., Chandrakasan, A., Balakrishnan, H.: An Application-Specific Protocol Architecture for Wireless Microsensor Networks. IEEE Transactions on Wireless Communications 1(4), 660–670 (2002) 8. Heinzelman, W.R., Sinha, A., Wang, A., Chandrakasan, A.P.: Energy-scalable algorithms and protocols for wireless microsensor networks. In: Proc. IEEE Acoustics, Speech and Signal Processing Conf., vol. 6, pp. 3722–3725 (June 2000)

A Management of Resource Ontology for Cloud Computing Hwa-Young Jeong1 and Bong-Hwa Hong2,* 1

Humanitas College of Kyunghee University, Hoegi-dong, Seoul, 130-701, Korea [email protected] 2 Dept. of Information and Communication, Kyunghee Cyber University Hoegi-dong, Seoul, 130-701, Korea [email protected]

Abstract. Cloud computing is a new trend in web based service. It could be use in service based application such as Infrastructure as a Service(IaaS) or Service as a Service(SaaS). However cloud computing has more wide meaning than service or application. In cloud computing environment, we don’t have to have any platform to use computing resources. All the computing-source and resources is in the web and can be manage them for user. Therefore the efficient method that is to control and manage the resources on the web is necessary. In this paper, we proposed management method to control and interface cloud computing resources. For this purpose, we made ontology for cloud computing resources and agent model to interface between the resources by the ontology. Keywords: Cloud computing, resource ontology, ontology management, cloud-sourcing, cloud computing service.

1

Introduction

Computing is being transformed to a model consisting of services that are commoditized and delivered in a manner similar to traditional utilities. In such a model, users access services based on their requirements without regard to where the services are hosted or how they are delivered. Several computing paradigms have promised to deliver this utility computing vision and these include cluster computing, Grid computing, and more recently Cloud computing. The latter term denotes the infrastructure as a “Cloud” from which businesses and users are able to access applications from anywhere in the world on demand [1]. Cloud computing uses the Internet and central remote servers to maintain data and applications. It is broken down into three segments: “applications”, “platforms”, and “infrastructure”. The cloud is the term for networked computers that distribute processing power, applications, and large systems among many machines. Applications like Flickr, Google, YouTube, and many others use the cloud as their platform, in the way that programs on a desktop computer use that single computer as a platform [2]. Actually, *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 65–72, 2011. © Springer-Verlag Berlin Heidelberg 2011

66

H.-Y. Jeong and B.-H. Hong

cloud is not a new concept. Anyone who has a web-based email account such as ‘Hotmail’ or ‘Gmail’, is using a simple form of cloud-computing [3]. Although not the first to propose a utility computing model, Amazon EC2 contributed to the popularization of the Infrastructure as a Service (IaaS) paradigm, which became closely tied to the notion of cloud computing. An IaaS cloud enables on-demand provisioning of computational resources, in the form of VMs deployed in a cloud provider’s datacenter (such as Amazon’s), minimizing or even eliminating associated capital costs for cloud consumers, allowing capacity to be added or removed from their IT infrastructure in order to meet peak or fluctuating service demands, while only paying for the actual capacity used. In general, an IaaS cloud consists of three main components, namely: a virtualization layer on top of the physical resources including network, storage and compute; the virtual infrastructure manager (VIM) that control and monitor the VMs over the distributed set of physical resources; and a cloud interface that provides the users with a simple abstraction to manage VMs. Cloud computing has emerged as a very promising paradigm to simplify and improve the management of current IT infrastructures of any kind and, in particular, grid ones. Clouds, in their IaaS form, have opened up avenues in this area to ease the maintenance, operation and use of grid sites, and to explore new resource sharing models that could simplify in some cases the porting and development of grid applications [4]. Therefore the method which is to control, manage, and share the various resources is necessary to operate process and interface in cloud computing environment efficiently. In this paper, we proposed management method to control and interface cloud computing resources. And we also made ontology for cloud computing resources and agent model to interface between the resources. This paper is organized as follows. Section 2 describes the environment of cloud computing service. The agent for ontology of cloud computing resources is described in section 3. Benefits of proposed method are discussed in Section 4, and conclusions can be found in Section 5.

2

Cloud Computing Service

The cloud computing service has been implemented as a multi-agent system. A schema of the system is depicted in Figure 1. First of all, a distributed and parallel system has to define how tasks are assigned to an agent. In order to minimize response and execution time, the implemented system uses a load balancing technique for distributing tasks between agents. Each agent has a local task queue, but a central information agent. Benefits of using load balancing algorithms for multi-agent, parallel and grid systems have been extensively studied [4, 5, 6]. Load Balancing uses three principal functionalities: (i)

Service advertisement and discovery is used by different agents to locate where other agents can be accessed and to publish their own location. (ii) Performance prediction is used to assign new tasks to a determined agent in order to equilibrate execution time for different agents. Each agent registers its

A Management of Resource Ontology for Cloud Computing

67

system performance for each type of operation (division, storage, and morphological operations). Using task parameters (for example, number of regions or untreated pixels), execution time is predicted using a logarithmic regression function. And the task is assigned to the agent that has less estimated pending processing time. (iii) Queues of tasks are employed for scheduling purposes. When a new task is assigned to an agent, maybe other tasks are being executed by this agent. Each agent queues the tasks pending execution in a local queue. An important issue in multi-agent systems is the role assignment of the agents that exist in the system. As Figure 1 shows, there are three kinds of agents in the system:

Fig. 1. Cloud Computing Service Schema

• The service access point agent is a unique agent in the system. This agent is the entrance point for final users. It offers the methods for storing and applying operations of the cloud computing service. • The resource index agent is also a unique agent in the system. It implements the service directory. This agent collects information about the working agents in the system. The collected information contains the location of each agent, its current load, the estimation function for performance prediction, and the image processing status. • The working agents are the main agent role in the system. Each working agent manages a relational database containing the complete data-structure. They have all necessary functionalities for storing and analyzing images on its own. So,

68

H.-Y. Jeong and B.-H. Hong

every working agent is capable of (a) dividing an input image (when it is too large), (b) extracting image regions and descriptors, (c) storing image information into the data structure, and (d) applying certain filters and graphbased operations on stored images in their own database [7].

3

An Agent Model for Ontology of Cloud Computing Resources

3.1

Cloud-Sourcing

There are three main types of cloud-sourcing products widely available: [3] • Cloud storage : Cloud storage is typically where a business stores and retrieves data from a data storage facility via the Internet. Storing data in this way offers near unlimited storage and can provide significant cost savings as there is no need for the business to buy, run, upgrade or maintain data storage systems with unused spare capacity.

• Cloud service : This is also known as ‘‘Software as a Service’’ or ‘‘SaaS’’. In SaaS, software applications are run on a SaaS provider’s system and accessed by a customer usually through a webbrowser via the Internet. This means that the software application itself is not hosted on the user’s PC or within a business’s servers but within the SaaS provider’s facilities.

• Cloud infrastructure/platform : In a cloud infrastructure or platform arrangement the provider operates the whole computing platform or operating system for the customer which is accessed via the Internet. Applications can then be run on the cloud platform/operating system in conjunction with utilising cloud storage. All that is required for the user to access the service is a computer with Internet access but little else in the way of computing hardware or ancillary support, maintenance or operating efforts.

A Management of Resource Ontology for Cloud Computing

3.2

69

The Ontology for Cloud Computing Resources

Lamia, Maria, and Dilma [8] proposed cloud computing ontology as Figure 2. In their research, the model depicted as five layers, with three constituents to the cloud infrastructure layer. The layered figure represented the inter-dependency and composability between the different layers in the cloud.

Fig. 2. Cloud computing ontology by Lamia, Maria, and Dilma [8]

Our proposed ontology model is similar than their model as shown in Figure 3.

Fig. 3. Proposed cloud computing ontology

70

H.-Y. Jeong and B.-H. Hong

The cloud application is to the users of the cloud. Normally, the users access the services provided through web-portals, and are sometimes required to pay fees to use them. And the users of the cloud software environment are cloud applications’ developers, implementing their applications for and deploying them on the cloud. The providers of the cloud software environments supply the developers with a programming-language-level environment with a set of well-defined APIs to facilitate the interaction between the environments and the cloud applications, as well as to accelerate the deployment and support the scalability needed of those cloud applications. And the cloud software infrastructure provides fundamental resources to other higher-level, which in turn can be used to construct new cloud software environments or cloud applications. Software kernels at this level can be implemented as an OS kernel, hypervisor, and virtual machine monitor and/or clustering middleware. Hardware and Firmware is the actual physical hardware and switches that form the backbone of the cloud [8]. 3.3

Agent Model for Cloud Computing

In this research, we considered an agent to process and handle each of resources in cloud computing and the ontology. Figure 4 depict the proposed agent model for cloud computing.

Fig. 4. Proposed agent model for cloud computing

The proposed model consists of five layers; Physical machines layer, Cloud resource layer, Resource management layer, Agent layer, and User layer. The Cloud resource layer supports resources which adapted to Cloud by using ontology. And the

A Management of Resource Ontology for Cloud Computing

71

Resource management layer is able to perform management ontology, resource, job, and data in cloud computing. The Agent layer has three types of agent; Service agent, Data management agent, and Resource management agent. Each agent deals with process between the user and system resources.

4

Benefits of Proposed Method

Our proposed agent model with ontology illustrates that cloud applications can be developed on the cloud software environments or infrastructure components. In addition, cloud applications can be perform efficiently as a service by use agent. It also can manage cloud resources by ontology. For example, using the agent with ontology can analyze cloud resources more easy and efficient than the environment without using agent and ontology for cloud computing. Because there are lots of resources for cloud computing, and therefore it need a management method for the resources. In this model, agent cloud perform the handling such as a service in cloud computing. Moreover using an ontology model for cloud computing could manage heterogeneous resource for cloud computing by analysis the ontology analysis and merge.

5

Conclusion

In this research, we proposed management method for cloud computing resources. For this process, we also made ontology model for manage the cloud computing sourcing and resource. Proposed cloud computing ontology has four steps; Firmware/Hardware step, Software kernel step, Cloud software infrastructure, and Cloud application step. To interact each step, we construct architecture of agent with ontology as five layers; Physical machines layer, Cloud resource layer, Resource management layer, Agent layer, and User layer. Finally, using this method has some benefits such as easily develop the cloud software environment and use the application.

References 1. Buyya, R., Yeo, C.S., Venugopal, S., Broberg, J., Brandic, I.: Cloud computing and emerging IT platforms: Vision, hype, and reality for delivering computing as the 5th utility. Future Generation Computer Systems 25 (2009) 2. Ivanova, M., Ivanov, G.: Cloud computing for authoring process automation. Procedia Social and Behavioral Sciences 2 (2010) 3. Joint, A., Baker, E., Eccles, E.: Hey, you, get off of that cloud? Computer Law and Security Review 25 (2009) 4. Leinberger, W., Karypis, G., Kumar, V., Biswas, R.: Load balancing across nearhomogeneous multi-resource servers. In: Proceedings of 9th Heterogeneous Computing Workshop, HCW 2000 (2000)

72

H.-Y. Jeong and B.-H. Hong

5. Junwei Cao, S.A.J., Junwei Cao, S.A.J., Spooner, D.P., Nudd, G.R.: Grid load balancing using intelligent agents. Future Generation Computer Systems 21(1) (2005) 6. Yagoubi, B., Slimani, Y.: Task load balancing strategy for grid computing. Journal of Computer Sciences 3(3) (2007) 7. Alonso-Calvo, R., Crespo, J., García-Remesal, M., Anguita, A., Maojo, V.: On distributing load in cloud computing: A real application for very-large image datasets. Procedia Computer Science 1 (2010) 8. Youseff, L., Butrico, M., Da Silva, D.: Towards a Unified Ontology of Cloud Computing. In: Grid Computing Environments Workshop, GCE 2008 (2008)

Development of an Algorithm for Video Quality Measurement for Broadcasting Communications Services Sang-Soo Kim1, Hae-Jong Joo2,*, and Euy-Soo Lee3 1

Dept. of Electrical & Electronic Engineering, Daewon University, 599 Sinwol-dong, Jecheon, Chungbuk, 380-702, Korea [email protected] 2 Dept. of HUNIC, Dongguk University, 82-1 Pil-dong 2-ga, Jung-gu, Seoul, 100-272, Korea [email protected] 3 Dept. of Chemical and Biochemical Engineering, Dongguk University, 26 Pil-dong 3-ga, Jung-gu, Seoul, 100-715 [email protected]

Abstract. The algorithm that represents the overall video quality must express a user’s evaluation pattern on a video that was watched through an objective index, and QoE measurement index factors that represent the video quality must be developed. Therefore, in this study, a video quality measurement algorithm was developed in terms of user experience of the video media of the IPTV service. Keywords: QoE, Quality of Experience, QoE measurement algorithm, video QoE algorithm.

1

Introduction

For the continuous performance improvement of video quality management systems for broadcasting communication services, there is a need to study the overall video quality index that can represent the quality of the video-related user experience [1]. For the various video services, including the IPTV service, the video QoE (quality of experience) measurement method, which represents the quality of video-related user experience, has been suggested by the ITU-T standardization organization. For the video QoE quality measurement method, which is suggested in the standard document, four methods are offered [1, 2], as shown in Fig. 1. The algorithm that represents the overall video quality must express a user’s evaluation pattern on a video that was watched through an objective index, and QoE measurement index factors that represent the video quality must be developed. Therefore, in this study, a video quality measurement algorithm was developed in terms of user experience of the video media of the IPTV service. *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 73–82, 2011. © Springer-Verlag Berlin Heidelberg 2011

74

S.-S. Kim, H.-J. Joo, and E.-S. Lee

Fig. 1. Video QoE quality measurement methods

2

Related Studies

2.1

Concept of QoE

For the video media of the IPTV service, a user watches the media through a terminal (set-top box) after the IPTV service is transferred to the house via an IP network. QoE is the quality of experience that the user perceives at the moment that the service is provided [3]. ITU-T TR-NGN.QoS explains the relationship between NP, QoS, and QoE, as shown in Fig. 2. NP (network performance) is the measurement of network performance on data transfer, QoS is the measurement of the transfer quality of these data to the terminal, and QoE is the measurement of the quality of the service from the perspective of the user [3, 4].

Fig. 2. Relationship between NP, QoS, and QoE

2.2

Video Quality Measurement Method

The development of a QoE measurement algorithm for the IPTV service must refer not only to the IPTV service but also to the methods recommended by the standardization organizations involved in any video-related service. The standard documentations [6] on multimedia and IPTV services recommended by the standardization organizations are summarized in Table 1.

Development of an Algorithm for Video Quality Measurement

75

Table 1. List of Standard Documents related to the Video Quality Measurement Method Org.

Document

Remark

ITU-T G.IPTV-PMMM:TD0179 IPTV-performance-monitoring measurement model ITU-T G.IPTV-PMPD:TD0178

IPTV-performance-monitoring parameter definition suggested by IPTV-GSI

ITU-T G.IPTV-PMR:C54

IPTV performance measurement and report

ITU-T G.RQAM:TD18

Reference guide for the QoE evaluation method

ITU-T G.IPTV-QMA:C175

Quality management assessment (QMA) for GIPTV Non-intrusive parametric model for securing multimedia stream

ITU-T P.NAMS:TD67 performance ITU-T P.NAMS:TD108

P.NAMS development participation call: operation mode and input/output parameters definition Suggestions, including packet loss and error correction/hiding, in

ITU-T P.NAMS:C48 P.NAMS ITU-T P.NAMS:C80

P.NAMS input parameter suggestion

ITU-T P.NAMS:C128

P.NAMS draft recommendation Suggestion of the P.NAMS parametric model’s area definition of the

ITU-T P.NAMS:C129 bitstream, hybrid, and perceptual models ITU-T P.NBAMS:TD71

Areas of P.NBAMS and J.bitvqm

ITU-T P.NBAMS:C185

QMA-process-related phase

ITU-T P.NBAMS:C186

Multimedia video quality: VQEG multimedia test results

ITU-T J.bitvqm:TD566

J.bitvqm as a new draft rec. for the hybrid perceptual/bitstream model for objective video quality measurement

ITU-T J.bitvqm:C79

Hybrid perceptual/bitstream model for imbedded-video-quality degree

ITU-T J.bitvqm:C99

Hybrid perceptual/bitstream NR model performance report

ITU-T J.bitvqm:C100

Hybrid perceptual/bitstream model for imbedded-video-quality degree Perceptual video quality measurement technology for digital cable TV

ITU-T J.noref:TD95 through the NR method ITU-T Y.1540/1541

Network performance index

ITU-T E.800

Relationship between service quality and network performance

ITU-T E.1000

QoS from the perspectives of the user and supplier

ITU-T G.1000

QoS parameters and four QoS viewpoints’ summary

ITU-T G.1010

Performance objectives by user service performance pattern

ETSI

EG 202 057

QoS parameters

ETSI

TR101 290

Digital video broadcasting (DVB): DVB system measurement guidelines

76

2.3

S.-S. Kim, H.-J. Joo, and E.-S. Lee

Video Quality Measurement Algorithm Verification Method

The verification of the QoE measurement algorithm for video services is as important as the algorithm development because it determines the accuracy of the algorithm. Methods for the verification of the QoE algorithm for video services are recommended in ITU-T P.910/911 and ITU-R BT.500. The international standard documentations [11] related to the verification methods of the video quality measurement algorithm are listed in Table 2.

Table 2. List of Standard Documents related to Video Quality Measurement Algorithm Verification Org.

Document

Remark Introduction of the subjective quality measurement method ACR-HR (absolute category rating with hidden reference)

ITU-T

P.910/911

DCR (degradation category rating) PC (pair comparison method) Introduction of the subjective quality measurement method

ITU-R

BT.500

DSIS (double-stimulus impairment scale) DSCQS (double-stimulus continuous-quality scale) SDSCE (simultaneous double-stimulus for continuous evaluation)

3

Video QoE Algorithm Derivation

3.1

Algorithm Derivation Procedure

The derivation and verification procedures of the video QoE algorithm for broadcasting communications services based on the bitstream layer method consist of the following three steps: (1) Setting up a video quality criterion To develop a video service quality measurement algorithm, the target quality that the algorithm is supposed to measure must be defined. The target quality for video media must be the quality experienced by the user, which can be measured by carrying out the subjective quality test procedure, in which the quality of the user’s experience is measured using a variety of produced videos. In this study, the subjective quality test was performed using the DSCQS method.

Development of an Algorithm for Video Quality Measurement

77

(2) Development of a video QoE algorithm To develop a video QoE algorithm, the parameters that were to be used in the algorithm were selected after analyzing the various elements that could affect the video quality, and their effects on the user’s experience. Moreover, the weight of each parameter was determined according to the degree of its effect on the user. Based on these parameters and weights, an algorithm that best represents the quality of the given video was developed. (3) Verification of the video QoE algorithm To verify the video QoE measurement algorithm that was developed in this study, a test video was prepared, and the video qualities based on the users’ subjective experiences and the developed algorithm were measured. The accuracy of the algorithm was verified by analyzing the correlation of these two results. 3.2

Development of a Video Quality Measurement Algorithm

For the development of a video quality measurement algorithm, the applied degradation patterns were analyzed, and a test video was produced based on the given degradations. The video quality evaluation results from the users were compared and analyzed to identify the elements that had greater effects on the video quality. These elements were then used as the parameters of the video quality measurement algorithm, and the weight of each parameter was determined according to the degree of its effect on the video quality. The procedure for the development of a video quality measurement algorithm largely consists of the steps of finding the parameters after analyzing the degradation elements applied to the test video and the users’ quality evaluation patterns, determining the weights of each parameter, and developing an algorithm that will represent the video quality. (1) Parameter classification and selection In this study, the parameters for the video quality measurement algorithm were determined, as shown in Table 3, and the parameters that were applied based on the characteristics were selected, as shown in Table 4. (2) Weight (W1- W16) calculation The weight of each parameter for the video quality measurement algorithm was calculated based on the correlation between the users’ experienced qualities and the parameter values derived in this study. The following Pearson’s correlation coefficient formula was applied as a correlation formula:

78

S.-S. Kim, H.-J. Joo, and E.-S. Lee Table 3. Parameter Classification and Characteristics

Parameter Characteristics Classification Including the video frame impairment level characteristics

Frame loss

Application of burst level characteristics according to the frame loss degradation injection method

Packet loss

Including the video frame impairment level characteristics The packet loss characteristics are identical with the frame error

(frame error) QP

characteristics. Parameter for displaying the original video quality

Table 4. Parameter Selection Classification

Parameter

Frame loss

Parameter 1

Remark Loss of one frame in the GOP

Frame loss

Parameter 2

Loss of two consecutive frames, including the I-frame

Frame loss

Parameter 3

Loss of two consecutive frames, not including the I-frame

Frame loss

Parameter 4

Loss of three consecutive frames, including the I-frame

Frame loss

Parameter 5

Loss of three consecutive frames, not including the I-frame

Frame loss

Parameter 6

Loss of four consecutive frames, including the I-frame

Frame loss

Parameter 7

Loss of four consecutive frames, not including the I-frame

Frame loss

Parameter 8

Loss of two or more frames in the GOP on a random basis

Frame loss

Parameter 9

Loss of five consecutive frames, including the I-frame

Frame loss

Parameter 10

Loss of five consecutive frames, not including the I-frame

Frame loss

Parameter 11

Subliminal effect applied, the effect of the previous GOP quality on the current GOP quality by frame loss

Packet loss (frame error)

Parameter 12

Loss of packet in the frame header

Packet loss (frame error)

Parameter 13

Loss of packet in the data of the video information area, except for the header at each frame of the GOP

Packet loss (frame error)

Parameter 14

Loss of packet in the GOP in a burst manner, or random loss of a packet or two

Development of an Algorithm for Video Quality Measurement

79

Table 4. (continued) Packet loss (frame error)

Parameter 15

Random loss of three packets or more at each frame of the GOP

Packet loss (frame error)

Parameter 16

Subliminal effect applied, the effect of the previous GOP quality on the current GOP quality by packet loss

QP

Parameter 17

Quantization parameter (quality representative of the original video)

Table 5. Weight by Parameter

Parameter

Correlation

Parameter

Correlation

1

0.914127

9

0.997582

2

0.942992

10

0.911979

3

0.998209

11

0.708549

4

0.953097

12

0.573887

5

0.979551

13

0.900335

6

0.989555

14

0.820754

7

0.940671

15

0.960323

8

0.930665

16

0.911231

(3) Measurement algorithm derivation Broadcasting communications services transfer the original video through a network. Therefore, degradations in quality can arise in the network or at each data transfer step. These degradations by the external effects must be detected and expressed as quality by the algorithm. The video quality measurement algorithm suggested in this study consists of the degraded qualities (parameters 1-16) based on the original quality (parameter 17). The video quality measurement algorithm for broadcasting communications services was derived through the following procedure.

80

S.-S. Kim, H.-J. Joo, and E.-S. Lee Table 6. Derivation of the Video Quality Measurement Algorithm

Classification

Step

Remark/Formula Derivation of basic value for use in the weight value

Step 1

The coefficient value between the user evaluation result and the parameter value is set as a default value for weight calculation. The value is designated as the refresh weight (refW).

Weight

Step 2

Weight is applied to the parameter with value (degraded). The weight is an action weight (aW).

derivation Action weight (aW) derivation step

Step 3

sumWeight : sum of the weights when a parameter value exists aWi = refWi/sumWeight The V-value algorithm derivation step depends on the existence of a parameter with the value seen below. if (any parameter with value exists)

V-value Step 4 derivation

ĺ if (no parameter with value exists) ĺ

3.3

Accuracy Comparison with the Previous Algorithm

Conventionally, the following video quality measurement algorithms are recognized as accurate (in descending order): the human video perceptual model (FR>RR>NR), the bitstream layer model, and the packet layer model. (The hybrid method has been excluded because there is no example of its development.) The accuracy of V-value, developed in this study as a bitstream layer model method, is approximately 0.745, and the comparable data are the correlation data obtained using the FR/RR/NR method of the human video perceptual model, submitted for ITU-T J.247 standardization, as shown in Fig. 3. The accuracy of the 0.745 V-value algorithm was found to be higher than those of the FR quality measurement algorithm of the PSNR method of the human video perceptual model and of the RR quality measurement algorithm of the PSNR method. It was also found to be higher than those of the Psytechnics and SwissCom algorithms that were developed using the NR model method.

Development of an Algorithm for Video Quality Measurement

81

Fig. 3. Correlation data submitted for ITU-T J.247 standardization

4

Conclusion

The development of a video media quality measurement algorithm for broadcasting communications services from the perspective of the user has been required by the international standard organizations, but the research and development on it is not well advanced. It is believed that the video quality (V-value) algorithm developed in this study will protect the rights of the IPTV service users, will provide a foundation for high-definition-quality services, and will serve as a guide for the development of a video quality measurement algorithm. V-value, a video quality measurement algorithm developed using the bitstream method, has about 0.745 accuracy. It thus has medium-level accuracy based on the current technology. For the improvement of the algorithm’s accuracy, the future project will include finding STB’s Jitter Buffer approach method through the approach method for the development of a V-value algorithm, and developing techniques for measuring the more diverse parameters that were excluded from this study, such as TS quality, PTS/DTS quality, PCR quality, client information, and decoding information. In addition, to improve the subjective quality test, more video materials (only three types of video images were tested in this study) and more users (20 users per test

82

S.-S. Kim, H.-J. Joo, and E.-S. Lee

video were included in this study, for a total of 60 users) will be included to more accurately determine the user quality measurement pattern, which will lead to a more accurate target quality. V-value has enough room for accuracy enhancement through these additional works. Finally, the future project of this study is to combine V-value, which was developed in this study, and VMOS to develop a hybrid perceptual/bitstream model, as the international standards related to video quality measurement, or the algorithm developers, are heading towards the hybrid perceptual/bitstream model.

References [1] ATIS-0800004, A Framework for QoS Metrics and Measurements Supporting IPTV Services (December 2006) [2] Multimedia Research Group IPTV Video Quality: QoS & QoE (Quarterly Technology & Content Report) (February 2007) [3] DSL Forum Technical Report TR-126, Triple-play Services Quality of Experience (QoE) Requirements (December 2006) [4] TMF506 v1.5, Service Quality Management Business Agreement [5] TR 101 290, Digital Video Broadcasting (DVB) Measurement Guidelines for DVB systems (May 2001) [6] ITU-T Recommendation I.350, General aspects of quality of service and network performance in digital networks, including ISDNs (March 1993) [7] ITU-T Recommendation Y.1541, Network Performance Objectives for IP-Based Services (February 2003) [8] ITU-T Recommendation Y.1543, Measurements in IP networks for inter-domain performance assessment (November 2007) [9] ITU-T FG IPTV-C-0411, IPTV QoS/QoE Metrics (January 2007) [10] ITU-T FG IPTV-C-0354, A proposal on QoE of EPG, Metadata and Browser (January 2007) [11] ITU-T FG IPTV-C-0507, Suggested addition to Quality of Experience requirements for IPTV (May 2007) [12] ITU-T FG IPTV-C-0127, Classification of IPTV services based on network QoS requirements (October 2006) [13] ITU-T FG IPTV-C-0210, Quality of experience (QoE), an overview (October 2006) [14] ITU-T FG IPTV-C-0184, Quality of experience Requirements for IPTV Services (December 2007) [15] ITU-T FG IPTV-DOC-0147, IPTV Service Requirements (October 2007) [16] ITU-T FG IPTV-DOC-0187, Performance monitoring for IPTV (December 2007)

An Effective Resource Managements Method Using Cluster-Computing for Cloud Services Seong-Sik Hong1 and Jin-Mook Kim2 1

Department of Ineternet Security, Hyejeon College [email protected] 2 Department of IT Education, Sunmoon University [email protected]

Abstract. The interest of Cloud service is rising recently. According to Gartner group's investigation, development view of Cloud service is very hopeful hereafter. However, skeptical sight is presented about Cloud service actually. If all-in-one about existent IT Resources and effective supervision are not gone ahead, because cannot secure the hopeful future about Cloud service. So, I wish to propose effective resource management method applying existent Clustercomputing for Cloud services. I show that EMRCC’s simulation result. It can be manage existent IT resources effectively about Cloud service using Hadoop. Keywords: Cluster computing, Cloud service, Hadoop, Resource managements.

1

Introduction

The cloud computing is main topic of IT industry between the recently 2 ~ 3 year. it is receiving many interests so much that is selected series 2 years in world teens IT strategy achievement according to Gartner Group's examines. The cloud computing is proposed by Google's a researcher in 2006. It is known as that is begun in effort that wish to integrate concept of existent grid computing, cluster computing, virtualization computing, integration computing etc. from service provider's viewpoint. In this way, it must be core business by 3 government bodies that is knowledge economics department, broadcast communication. Because the interest about latest cloud computing is rising. So, they make "Total government cloud activation all-out plan" in domestic. And establishing jointly and try to arrive to world market share 10 until 2014. However, people who have skeptical sight against exist cloud computing exist silver lining. They are deciding that they are not more in effort that cloud computing wishes to integrate existent various kinds computing technologies from service provider's viewpoint. And are possible though smooth integration for necessarily existent IT resources and authoritativeness about cloud service. So, we wish to examine about two problems that should be solved necessarily to provide cloud service in treatise that see hereupon, and propose effective cloud service resource management method to take advantage of cluster computing technology that T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 83–89, 2011. © Springer-Verlag Berlin Heidelberg 2011

84

S.-S. Hong and J.-M. Kim

can solve this. And we experimentalize to that through some simulation scenarios using Hadoop simulator. It can act in Apache web server to show feasibility. Composition of this treatise is as following. We explained curtly about Cloud computing, cluster computing, Hadoop in related research in chapter 2. And explained about ERMCC's characteristic and action surrounding and experiment scenario that wish to propose in treatise that see in 3 chapter. And in 4 chapter, we examine about propriety for experiment scenario for ERMCC. Finally, chapter 5 described conclusion and future works.

2

Related Researches

In this chapter, we describe about cloud computing and cluster computing, and Hadoop to use to experiment ERMCC that propose in treatise curtly. 2.1

Cloud Computing

Cloud computing that proposed by Google's a researcher. It integrates IT resources that exist in existing in geographical, systematic dimension, and it is basis purpose that wish to offer and offer work that user wants conveniently to service concept to user. To make this, integrate existing IT resources in geographical, systematic dimension beforehand. Cloud computing concept displayed in figure 1.

Fig. 1. Concept of Cloud computing

So that we can integrate application programs as well as thing that bind integrating systems that is away geographically as is appearing in figure 1 simply by one and improve of resources usability and integration ability. Therefore, concept of cluster computing is required necessarily to explain next section.

An Effective Resource Managements Method Using Cluster-Computing

2.2

85

Cluster Computing

Cluster computing is research for integrate computer hardware resources that exist. And it can offer super computing system that can offer more excellent performance. In Example, Eucalyptus cluster computing research explains that how can constructs existing IT resources.

Fig. 2. Concept of Cluster computing on Eucalyptus

It connects existent computing systems by network protocol such as TCP/IP. Figure 2 is displaying concept of cluster computing on Eucalyptus. In 2009 to example about cluster computing North Carolina college's Frank. Muller professor showed example that bind device 8 that is PS3 that it is known as game mourning and offer first scientific cluster. Thus identical computer device or similar devices bind that wish to offer high computing power. 2.3

Hadoop

It is research about simulator that can provide cloud service that is based on web service by one during Apache project. This did to introduce cluster computing concept and provide cloud service after connect electronic computing systems that have equal or resembling performance that exist in existing using TCP/IP communication protocol. Cluster composition way that Hadoop supports is 3 types as following. This is single composition way, imagination breakup way, perfection breakup way. Single composition way means cluster that is consisted of computer resources that have equal specification. After imagination breakup way executes Java imagination machine by second, it is method to bind various hardware resources by cluster. Finally, perfection breakup way is method to bind various kind computer resources by cluster. Hadoop Distributed File System is representative perfection breakup way cluster method. This is displaying in figure 3.

86

S.-S. Hong and J.-M. Kim

Fig. 3. Concept of Hadoop Distributed File System

3

ERMCC

We need to necessarily existing IT resources and form cluster to provide cloud service. So we propose about ERMCC (Effective Resource Management method for Cloud services using Cluster computing) that apply clustering techniques to provide cloud service that can manage existent IT resources effectively. 3.1

Compose of ERMCC

Compose of ERMCC in this paper, it wishes to use cluster computing technology. We are going to use Hadoop that is proposed in Apache project. And this treatise selected doing to provide cloud service effectively using perfection distribution way. Because of most cloud service need perfection distribution way actually. ERMCC's composition that propose shows figure 4.

Fig. 4. Composition of ERMCC

An Effective Resource Managements Method Using Cluster-Computing

3.2

87

Performance Assessments Scenarios

We wish to measure data processing time through simulation to confirm ERMCC's efficiency. Firstly, we measured data processing time in CM (Cluster Manager) and Node's computer hardware Spec is identical environment. We measure times that are transmits and writes to CM by each node. Packets consist of 100 slice dividing data of 64 MB. And we measure 5 times average. Result of simulation shows Table 1. Result of simulation shows Table 1. Table 1. Result of simulation about identical environments Device Node 1 Node 2 Node 3 Node 4 Node 5

Spec. Process Time(sec) CPU : IBM compatible Dual 420 Core 2Ghz 418 Memory : 2GB 421 416 419

We know that same spec. node have equivalent process times. Next, we measure process time for two cluster group that computer hardware Spec. is consisted of other nodes by second. In this time CM spec. is same first simulation. Result of simulation shows Table 2. Table 2. Result of simulation about different environments Device Node 1 Node 2 Node 3 Node 4 Node 5

Spec. CPU: Dual core 2Ghz, Memory : 1GB CPU: Dual core 1.66Ghz, Memory : 1GB CPU: Dual core 2.4Ghz, Memory : 2GB CPU: Dual core 2.6Ghz, Memory : 3GB CPU: Quad core 2.4Ghz, Memory : 4GB

Process Time(sec) 468 482 328 320 216

We know that device spec. is different than process time is too different. Because process time fellows device computing power. Table 2 shows that higher device process time is short. But lower computing power has device’s result is slowly about double. We describe more performance assessment in next section.

88

4

S.-S. Hong and J.-M. Kim

Performance Assessment

As refer before, conducted simulation according to two scenarios. Prerequisite supposed that CM's Performance makes use of all identical systems in two simulations. First measure cost time transmit each by node and store data of 64 MB size by CM in first simulation scenario. Because hardware specification is same by device, difference did not appear greatly at process time. If provide cloud service composing devices that have equal hardware specification through this by cluster, the efficiency is thought to be very high. However, is going to have much manufacturing drug to compose cluster that have equal hardware specification actually in this way. Because burden which throw away all hardware devices that have already to compose cluster that have equal hardware specification and is expense enemy who buy hardware devices newly may breed. Second simulation carried out because consider realistic hardware configuration and hardware specification of nodes supposes case that is all. Could know that process time according to hardware specification appears until maximum 2 double as result that examine different view result. We must be afflicted that hardware specification must do how to compose other devices by cluster actually thus. In this paper, We suggest that hardware specification is thought to have more excellent process time if compose resembling devices small scale cluster and do this to process work by other serve cluster and cooperation through CM of super ordinate concept hierarchically more.

5

Conclusion

The interest about cloud service is decayed rawly. Because this is the best solution that cloud service can solve problem about green IT technology, ashes for physical devices - use problem solving, integration of system units etc. However, many difficulties follow actually to provide perfect cloud service. It is because realistic restrictions are many to integrate perfectly various PUs among them. In this paper, we setup cluster group by perfection distribution way selecting Hadoop that is existing cluster computing technology and this did simulation. Process time of cluster that have equal hardware specification if examine simulation result appeared by thing which is more fast. It may be best that compose devices cluster of equal specification and connect these through network to provide effective service when put off deviation. However, this is impossible actually. Because exist IT resources might be consisted of different specification by all necessities. IT resources that have different hardware specification hereupon composing cluster effectively to provide better cloud service more researches to need think.

References 1. Prigge, M.: Confession of a Cloud Skeptic, Inforworld (June 21, 2010) 2. Barnatt, C.: Cloud Computing: A Brief Guide to Cloud Computing (2011) 3. Dean, J., Ghemawat, S.: Mapreduce: Simplified Data Processing on Large Clusters. Communications of the ACM (2001)

An Effective Resource Managements Method Using Cluster-Computing

89

4. Barhan, P., Dragovic, B., Fraser, K., Hand, S., Harris, T., Ho, A., Neugebauer, R., Pratt, I., Warfield, A.: Xen and the art of virtualization. In: Proceeding of the 19th ACM Symposium on Operating Systems Principles (2003) 5. Chase, J., Irwin, D., Grit, L., Moore, J., Sprenkle, S.: Dynamic virtual clusters in a grid site manager. In: Proceedings of 12th IEEE International Symposium on High Performance Distributed Computing 2003, pp. 90–100 (2003) 6. McNett, M., Gupta, D., Vahdat, A., Voelker, G.M.: Usher: Aen Extensible Framework for Managing Clusters of Virtual Machines. In: Proceeding of the 211st Large Installation System Administration Conference, LISA (December 2007) 7. Hadoop (2009), http://hadoop.apache.org 8. IBM Blue Cloud project (2009), http://www.ibm.com/jct03001c/press/us/en/appengine 9. Amazon Elastic Compute Cloud (2007), http://awa.amazon.com/ec2 10. Eucalyptus Public Cloud(EPC) (2006), http://eucalyptus.cs.ucsb.edu/wiki/EucalyptusPublicCloud/ 11. http://www.gartner.com/it/page.jsp?id=1454221 12. http://www.gartner.com/it/page.jsp?id=1210613 13. http://www.gartner.com/it/page.jsp?id=777212 14. , , " 3 “, 23 9 (June 5, 2011)

이호현 강홍렬 클라우드 서비스의 가지 본질적 속성 정보통신정책 동향분석 제 권 호

Study on Micro-processing Implementation of USN Environment Data by a Graphic-Based Programming Young-Wook Lee Dept. of Computer Science of Semyung University [email protected]

Abstract. Recently USN computing technique is able to acquire and process environments data by using the USN small sensors. Especially such data as temperature, humidity, intensity of illumination, positioning information of GPS, CO2 gas density etc. at construction site are able to be effectively made use of design modification and improvement according to construction process. A TinyOS-based USN technique using wireless communication sensors is applied for the construction site which needs those environments data. In addition, efforts to visit the construction site and the allotted period to check corresponding construction process, are able to be reduced. Such environmental data aquired from the USN sensors are processed by the graghic-based programming in this study and showed the data and graghical results on a PC monitor in real time or on a mobile LCD. Keywords: USN, Envorinmental Data, Construction, Data Acquisition, Micro- Processing.

1

Introduction

Recently the USN(Ubiquitous Sensor Network) techniques for goods circulation, information of ecology and environmental control of greenhouse which is able to compute, control and communicate, has been implemented by using wireless small sensors[2],[4],[7]. A computer (PC or notebook) senses any information that is needed for us and provides us with such information through the appropriate interfaces. Such technique coupled by a processor, sensors and communication etc. needs to cognate and determinate the related information for oneself. A sensor is one of ubiquitous input devices which detect the change of external environments. A processor is a device for analyzing and determining the obtained data from the ubiquitous sensors. In addition, the function of communication with such sensors is also required. In this study, recently a developed wireless USN (Ubiquitous Sensor Network) and the graphic-based programming technique (using a LabView) are applied for acquisition and process of the environmental information at the construction site. A microprocessor system is interfaced with a PC (Personal Computer) and USN (Ubiquitous Sensor Network) sensors. The environmental information at the construction site is obtained from the interfaced microprocessor system with USN T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 90–95, 2011. © Springer-Verlag Berlin Heidelberg 2011

Study on Micro-processing Implementation of USN Environment Data

91

sensors and is displayed on a PC (or a notebook) in real time and on a smart mobile LCD. A sink sensor at the construction site collects the environmental data such as temperature, humidity, intensity of illumination, positioning information of GPS (Global Positioning System), CO2 gas density etc. from wireless node sensors. A microprocessor (LM3S8962) system by the graphic-based C programming processes the obtained data from a sink sensor and then a TinyOS-based PC (or a notebook) executes the compiled data from the microprocessor and displays those data on the screen. The environmental data among a microprocessor system and USN sensors are communicated in serial and by a TCP/IP (Transmission Control Protocol and Internet Protocol) method between a microprocessor system and a PC (or a notebook). The transmitted data to the smart mobile is also available on its LCD (Liquid Crystal Display). The environmental data obtained from the construction site are sent as online data in real time to the head center or the corresponding department. Such data is provided as a basic data to check the present status of constructing spot. Time and efforts are also reduced. Especially such effectiveness is more increased as the construction spot is more distant from the head center or the corresponding office. The obtained environments data from USN sensors in this study showed the appropriately visible information of temperature, humidity, intensity of illumination, positioning information of GPS (Global Positioning System), CO2 gas density etc. at the construction spot. In addition, such data is available to modify the construction design and is able to provide with the related information of the present environmental status or conditions of processing under the construction.

2

Interfaces of a USN

2.1

Structure and Interfaces of a USN

Fig. 1 shows a USN available to the wireless communication and the interfaces of a microprocessor system with a PC (or a notebook).

Fig. 1. Interfaces of a USN

92

Y.-W. Lee

The environmental data related to temperature, humidity etc. are obtained from the wireless node sensors and such data are sent to the microprocessor system through the sink node sensor by the method of a UART (Universal Synchronous Asynchronous Receiver and Transmitter). A microprocessor system processes the graphic-based programming and then shows graphically the results with the processed data on the PC monitor by a TCP/IP communication in real time. 2.2

Interfaces of Microprocessor System

A microprocessor system receives the measured environmental data from wireless node sensors of a USN and processes those data of inputs by using the method of graphic-based programming. After compiling of such programming, a C code is generated and then the results of such TinyOS-based C programming are graphically displayed on a PC or a notebook monitor and the environmental data of the related output are both shown. The temperature output of micro-processing by graphic-based programming is in Fig. 2 as shown.

Fig. 2. The output of micro-processing temperature data

2.3

Interfaces of TinyOS-Based USN System

The environmental data from wireless USN sensors are able to be processed and displayed in a windows-based TinyOS as well as can be processed by the graphicbased programming of a microprocessor system. Fig. 3 shows the results of the output image and data obtained from the USN sensors. The sensors are directly interfaced with the TinyOS-based windows and the measured environments data from those sensors are processed. Such obtained environmental data as the outputs are shown by an image or the output data of appropriate unit.

Study on Micro-processing Implementation of USN Environment Data

93

Fig. 3. The output of a TinyOS-based processing

2.4

The Related Process of a USN Program

The processing procedure of a USN program before downloading to the node sensor is shown as in Fig. 4. The NesC code appropriate to the USN sensors is programmed in the TinyOS.

Fig. 4. Process of a USN program

After the NesC code is compiled into the C code, the code is transformed into the file of machine code (*.hex) by cross-compiling. The machine code is downloaded to the corresponding node sensor through the program downloader of ISP-USB and then wireless sensors can start to measure the environmental data from the corresponding position.

94

Y.-W. Lee

3

Data Acquisition and Processing

3.1

Data Transmission

As the compiled code is downloaded to the measuring wireless sensors and a sink sensor, the related environments data start to be measured. The obtained data are collected from the sink node and are processed and then displayed on a PC screen or a notebook LCD in real time. The output can be displayed on the LCD of a smart mobile. Fig. 5 shows the output of the related image and data on the LCD of a smart mobile.

(a) The output data of a smart mobile (b) The image data of a smart mobile Fig. 5. Outputs of smart mobile

3.2

Data Transmission

The following table 1 shows an example of the CO2 gas density of environments data. The measuring values of CO2 sensor are within the range of 0-3000 ppm and as shown in table 1, CO2 gas density of 378 ppm and 392 ppm are shown respectively in the crowd area of local downtown and on the road of automobile in the vicinity. This value is a little higher than that of 369 ppm in the close-together area of the apartment. It is suggestive that we consider our environmental factors from global warming during construction process. (The Weather Center of Korea reported the value of 392.5 ppm measured by Monitoring Center of Global Atmospheric Research at Anmyeon-Do of Korean island in 2007.) In a USN system, the transmission is done by a wireless method from the node sensors for measuring the environmental data to a sink sensor for collecting those data and then the communication is done in series between a sink node and a microprocessor system. Finally the results of processed data are transmitted by a TCP/IP (Transmission Control Protocol/Internet Protocol) from a microprocessor system to a windows-based PC or a notebook.

Study on Micro-processing Implementation of USN Environment Data

95

Table 1. An example of measured CO2 gas density of environmental data (ppm) Measured Concentrated Crowd Area Road of No. Apartment Area of Downtown Automobile 1 364 383 409 2 367 373 386 3 377 377 380 Average 369 378 392 * Reprinted from Study on Implementation of Environment Data Acquisition and Processing System by Using a USN by Young -Wook Lee and Sang-Ho Yeon.

4

Conclusions

The measuring system of environments data at construction site which is composed of wireless USN (Ubiquitous Sensor Network) sensors is able to obtain and process the related data easily such as temperature, humidity, GPS (Global Positioning System) information, CO2 gas density etc. by making use of the recent USN technique. The environmental data needed for the construction spot are obtained and processed by a USN technique. We can also reduce our efforts and time by monitoring the progressive status of construction and by giving feedback of the related data for the construction process to the necessary head center or the corresponding department in time on demand. In addition, we are able to provide the necessary information considering the environmental factors in case of modifying design for construction process.

References 1. Gay, D., Levis, P., von Behren, R., Welsh, M., Brewer, E., Culler, D.: The NesC Language: A Holistic Approach to Networked Embedded Systems. In: Proceedings of Programming Language Design and Implementation (2003) 2. Hill, J.L.: System Architecture for Wireless Sensor Networks. In: A Dissertation of Ph. D, Univ. of California, Berkeley (2003) 3. 2007 Monitoring Report of the Global Atmospheric Research Program. The Weather Center (2008) 4. Ubiquitous Sensor Network System Using ZigbeX. HANBACK Electronics (2008); ISBN 978-89-90758-12-5 5. HANBACK Newsletter. HANBACK Electronics 119(4) (2008) 6. Control of ARM-Based Microcontroller for Learning by HBE-MCU-LabView. HANBACK Electronics (2009) 7. Lee, Y.-W., Yeon, S.-H.: Study on Implementation of the TinyOS-Based Technique for a Ubiquitous Sensor Network at the Construction Site. In: IWIT 2010 Winter Smart Green IT Convergence Conference, 10th Anniversary, vol. 8(2), pp. 133–134. The Institute of Webcasting Internet and Telecommunication (2010)

Web Based Remote Robot Control for Adjusting Position on Manufacturing System Hwa-Young Jeong1 and Bong-Hwa Hong2,* 1

Humanitas College of Kyunghee University, Hoegi-dong, Seoul, 130-701, Korea [email protected] 2 Dept. of Information and Communication, Kyunghee Cyber University, Hoegi-dong, Seoul, 130-701, Korea [email protected]

Abstract. Manufacturing system, such as factory automation system, was operated by many components. The robot is one of them and very important component to operate and control the machine. This paper has a focus on remote control the robot which is in the factory. The purpose is to support function that manager who isn’t stayed at the factory is able to know the machine status and control the robot position at the distance location. To this process, we construct two sever, operational and management sever in web based management environment. And we also make a matrix to adjust the robot position. Keywords: Robot control, Remote management, Manufacturing system, Remote control.

1

Introduction

Recently, significant advances in Internet and computer technology have made it possible to develop an Internet based control and monitoring for industrial systems [1]. Additionally manufacturing systems have been changing from flexible manufacturing to agile manufacturing in order to meet current market requirements such as the function or design of products. In order to meet current market requirements, it is necessary to shorten the time needed to reflect the production plan in the manufacturing system. Moreover, it is necessary to understand the status of the manufacturing system and reflect it in production planning. To realize the manufacturing system, a remote access system to check the manufacturing system status is required. But FA (Factory Automation) controllers used in manufacturing systems could break down under several conditions. The failure of FA controllers causes reconfiguration of the production plan. Thus, downtime reduction in manufacturing systems is necessary to increase the operating ratio of the manufacturing system and to increase productivity. It is also necessary to provide a way to detect failures, analyze the reasons for the failures, and recover from a failure quickly. Instead of dispatching maintenance staff, a remote access method to the FA *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 96–103, 2011. © Springer-Verlag Berlin Heidelberg 2011

Web Based Remote Robot Control for Adjusting Position on Manufacturing System

97

controllers through a network is required. In existing manufacturing systems, several vendors provide a remote access facility in which remote access software communicates with the FA controllers directly through phone lines. There have been several studies that proposed methods applying internet technology; they were mainly developed for information systems for office work, and for control systems and manufacturing systems [2]. However many research have not enough to supported the remote robot control function using the internet. Therefore, we propose a method that is to support remote control function on the web to user or manager. By the method, user or manager can observe the manufacturing system at a distance location from the FA machine. To provide this function, we make a module structure and insert in main GUI (Graphical User Interface) system of manufacturing system.

2

Remote Access to Controllers for Manufacturing System

2.1

Remote Control for Manufacturing System

The World Wide Web successfully demonstrates how current technology can support information sharing among widely dispersed groups.

Fig. 1. Digital Plant Architecture (Predictive Intelligence Integrated into Automation Architecture)

Manufacturing system construct normally many components that was produced from different company. Fig. 1 shows an example structure of digital plant for manufacturing system. In this contribution it is shown how control engineering laboratory experiments can be implemented and offered such that they can be performed remotely via the Web [3].

98

H.-Y. Jeong and B.-H. Hong

Typically, manufacturing systems are classified in terms of the physical machine arrangement. A production job shop is a manufacturing system in which machines are grouped by function (e.g., turning dept, milling dept., etc). Transfer lines, group technology cells and lean, linked cell manufacturing systems are established so that machines are arranged based on product flow. A manufacturing system is also a subset of an entire manufacturing enterprise. Manufacturing enterprises consist of people, "things", and information. People are deployed to perform various functions such as marketing, design, purchasing, inventory control, inspection, machining, management, safety, service, and security. "Things" range from factories, to machines, materials, transporters, computers, warehouses, vendors of components, and utilities. Information is related to marketing requirements, product design, manufacturing systems and operations, manufacturing processes, human resources, supplier chain systems, and general management. All these elements constitute part of the manufacturing enterprise and thus, the design of manufacturing systems is regarded to be complex [5]. In the field of control systems, the advantages of Internet applications provide new techniques in monitoring and controlling a controlled object remotely. An operator could control a plant from anywhere as long as there is a PC with an Internet or an Intranet network. For industry or laboratory based works or managements, these advantages include: (1) allowing remote monitoring and adjustment of plants, (2) allowing collaboration between skilled plant managers and laboratories situated in geographically diverse locations, (3) allowing the business to relocate the physical location of plant management staff easily in response to business needs [6]. The manufacturing system’s application, remote control system, using the internet has been organized by following a three-tier-based architecture an evolution of the client-server scheme with an additional level, the middleware. Web technology acts as communication mean for the entire architecture. The three levels can be outlined as follows [7]: • • •

Presentation level: it consists of a Web browser and of its extensions which acts as user interface for accessing the clinical database. Middleware: it consists of a Web server and of its extensions which allow access to storage level for information retrieval and formatting for suitable presentation to the client. Storage level: it consist of a standard Electronic Patient Record based database made up of all resources containing information of interest for healthcare applications.

Web Based Remote Robot Control for Adjusting Position on Manufacturing System

2.2

99

Functional Requirements

There are many considerations to design and develop the remote control system for manufacturing system. Not only international standard communication protocols such as TCP/IP but also proprietary communication protocols could be selected for the communication protocol between an application and the FA controller. The selection of the communication protocol depends on the configuration of the manufacturing system. Therefore, proprietary communication protocols usually must be supported in addition to TCP/IP, a communication protocol on the Internet. Moreover, most of the FA controllers are expected to work for over 10 years. Legacy FA controllers and the latest FA controllers may work together on the same production line. Thus, it is necessary to provide extensibility to deal with new communication protocols for highspeed networks. Furthermore, there are many types of FA controllers, such as PLCs and robot controllers, which are used in manufacturing systems. Remote access software must be able to communicate with all types of FA controllers. Each FA controller has its own access method. The FA controllers and computers in the manufacturing systems have restricted computation performance. It is very important that the communication middleware which is running on the manufacturing system be independent of the type of FA controller to be accessed [8]. Kazuhiro, et al [8] depicted implementation method for remote control as shown Fig 2.

Fig. 2. The structure for remote control by Kazuhiro, et al [8]

3

Remote Robot Control

3.1

Robot Control Process

Cheng, et al [9] described a method to robot control. In their research, the robot image with (7 X 7) pixels is considered. Fig. 3 shows the definition of the robot coordinate system, where the center of robot is defined as origin. The coordinate set Sr is as shown in Example 1.

100

H.-Y. Jeong and B.-H. Hong

Example 1. Suppose l = 7, then Sr = {(-3; +3), (-3; +2), (-3; +1), (-3; 0), (-3; -1), (-3; -2), (-3; -3); (-2; +3), (-2; +2), (-2; +1), (-2; 0), (-2; -1), (-2; -2), (-2; -3); (-1; +3), (-1; +2), (-1; +1), (-1; 0), (-1; -1), (-1; -2), (-1; -3); (0; +3), (0; +2), (0; +1), (0; 0), (0; -1), (0; -2), (0; -3); (+1; +3), (+1; +2), (+1; +1), (+1; 0), (+1; -1), (+1; -2), (+1; -3); (+2; +3), (+2; +2), (+2; +1), (+2; 0), (+2; -1), (+2; -2), (+2; -3); (+3; +3), (+3; +2), (+3; +1), (+3; 0), (+3; -1), (+3; -2), (+3; -3)}

Fig. 3. Robot pixel coordinates representation by Cheng, et al [8]

Fig. 4. The playground coordinate system by Cheng, et al [8]

Web Based Remote Robot Control for Adjusting Position on Manufacturing System

101

The coordinate system of the playground has the left-bottom corner functions as the origin of the coordinate system (Fig 4). Since a playground with 150 centimeters in length and 130 centimeters in width is used, the range of the horizontal coordinate of image is defined from 0 to 150, and the vertical coordinate is defined from 0 to 130. 3.2

Proposed Remote Robot Control for Manufacturing System

Fig 5 shows a structure to remote robot control system via the internet.

Operational Server

Operational Data

Internet Management Server

Production Data

Manager

End User / System Operator

Manufacturing System

Robots

Fig. 5. The structure of remote control for manufacturing system

In this environment, end user or system operator always control the operation beside the manufacturing system. However, manager does not stay beside the system in the factory. Therefore manager is able to control the machine component as robot to access the internet. We construct a matrix (8 x 12) to robot control as shown Fig 6. It can control to move from base point (0, 0) to any location, up or down, front or back.

102

H.-Y. Jeong and B.-H. Hong

Fig. 6. The position matrix for robot control

4

Conclusion

In this paper, we have a focus to control the robot in manufacturing system via the internet. The main purpose is to support the robot control function to the manager who is deal in manage the manufacturing system at distance location, and observe and analyze the production information of each the machine which is in the factory. In order to this process, we make a structure that has two server, Operational server and Management server. Especially Operational server deal with get all of operational information, robot data, sensor data, production information, and so on, from the machines. The manager is able to control the robot position by remote access to Operational server through the internet. We also construct a matrix to adjust robot position by (8 x 12).

References 1. Jang, S.D., Lee, W.W., Son, Y.G., Oh, J.S., Cho, M.H.: Web-Based Control System For 150-Mw Pulse Modulator Application. In: Proceedings of LINAC 2002, Gyeongju, Korea (2002) 2. Kusunoki, K., Imai, I., Negi, T., Matsuda, N., Ushijima, K.: Proposal and Evaluation of a Method of Remote Access to FA Controllers via the Internet, Scripta Technica (2002) 3. Schmid, C.: Virtual Control Laboratories and Remote Experimentation in Control Engineering. In: 11th EAEEIE Annual Conference on Innovations in Education for Electrical and Information Engineering (EIE), Ulm, Germany, (April 26-28, 2000) 4. Ellender, D.: Digital Architecture Technology Brings Full-Scale Automation To Remote Oil, Gas Fields, The American Oil & Gas Reporter (2005) 5. Suh, N.P., Cochran, D.S., Lima, P.C.: Manufacturing System Design. Annals of the ClRP 47(2) (1998)

Web Based Remote Robot Control for Adjusting Position on Manufacturing System

103

6. Safavi, A.A.: Web-Based Control and Monitoring Systems: The new challenge. In: Proceeding of 12th Iranian Conference on Electrical Engineering, Mashad, Iran, vol. 1, pp. 119–125 (May 2004) 7. Lamberti, F., Montrucchio, B., Sanna, A., Zunino, C.: A Web-based Architecture Enabling Multichannel Telemedicine Applications. Journal of Systemics, Cybernetics and Informatics 1(1) (2003) 8. Kusunoki, K., Imai, I., Negi, T., Matsuda, N., Ushijima, K.: Proposal and Evaluation of a Method of Remote Access to FA Controllers via the Internet. Electronics and Communications in Japan, Part 1 85(6) (2002) 9. Cheng, C.D., Vadakkepat, P., Ko, C.C., Chen, B.M., Xiang, X.: Robot motion control and image reconstruction over internet. In: Proceedings of the International Conference on Computational Intelligence, Robotics and Autonomous Systems, Singapore, pp. 423–428 (November 2001)

Discrimination of Speech Activity and Impact Noise Using an Accelerometer and a Microphone in a Car Environment Seon Man Kim1, Hong Kook Kim1, Sung Joo Lee2, and Yun Keun Lee2 1 School

of Information and Communications Gwangju Institute of Science and Technology, Gwangju 500-712, Korea {kobem30002,hongkook}@gist.ac.kr 2 Speech/Language Information Research Center Electronics and Telecommunications Research Institute, Daejeon 305-700, Korea {lee1862,yklee}@etri.re.kr

Abstract. In this paper, we propose an algorithm to discriminate speech from vehicle body impact noise in a car. Depending on road conditions such as the presence of large bumps or unpaved stretches, impact noises from the car body may interfere with the detection of voice commands for a speech-enabled service in the car, which results in degraded service performance. The proposed algorithm classifies each analysis frame of the input signal recorded by a microphone into four different categories such as speech, impact noise, background noise, and mixed speech and impact noise. The classification is based on the likelihood ratio test (LRT) using statistical models constructed by combining signals obtained from the microphone with those from an accelerometer. In other words, the different characteristics detected by both acoustical and mechanical sensing enable better discrimination of voice commands from noise emanating from the vehicle body. The performance of the proposed algorithm is evaluated using a corpus of speech recordings in a car moving at an average velocity of 30-50 km/h with impact noise at various signal-to-noise ratios (SNRs) from -3 to 1 dB, where the SNR is defined as the ratio of the power of speech signals to that of impact noise. It is shown from the experiments that the proposed algorithm achieves a discrimination accuracy of 85%. Keywords: Speech enabled service in a car, car impact noise, voice activity detection, accelerometer.

1

Introduction

Interest in speech interfaces for controlling electronic products has grown rapidly because of safety and convenience concerns. In a vehicle, it is particularly essential for the driver to use a speech interface system to control electronic devices, e.g., car navigation or telematics systems. However, the quality of the speech signal in car environments is deteriorated by the numerous noise sources such as the car engine, fan, audio system, wind, road, and conversation among passengers [1-3]. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 104–113, 2011. © Springer-Verlag Berlin Heidelberg 2011

Discrimination of Speech Activity and Impact Noise

105

(a)

(b) Fig. 1. Examples of speech, background b noise, and impact noise recordings: (a) male speeech, (b) noise when driving at 30-50 km/h

Road noise is resulted from f the movement of a vehicle’s tires over the road ssurface, and it is the major sou urce of stationary background noise exposure. In particuular, the tires’ contact with a speeed bump or barrier on a road induces vehicles to vibrrate, which brings about impact noises [4-6]. Furthermore, front and rear tires of vehiccles contribute to two successiv ve impact noises, having durations and waveforms that are similar to speech, as shown n in Fig. 1. Even though impact noises degrade the perfformance of speech interfacee systems such as hand-free communication systems and speech recognition systemss, few studies have been conducted which deal with imppact noises. Therefore, the development of a method of discriminating speech signal frrom impact noise is required fo or the realization of successful speech interface systems in car environments [7-10]. Therefore, this paper prroposes an algorithm for discriminating impact noise and driver’s speech signal in a car environment. This discrimination is accomplishedd by the combination of three decision d rules pertaining to non-background noise, imppact noise, and speech. Recenttly proposed statistical model-based decision rules hhave demonstrated good perform mance by employing the likelihood ratio test (LRT) with the complex Gaussian distributtion [11]. In our experience, however, the statistical moddelbased detectors were incap pable of discriminating impact noises from speech becaause two signals were very simiilar each other, as shown in Fig. 1. Thus, it is necessaryy to obtain information on impaact noise in an alternative way, which should not be signnificantly affected by speech signal. Fortunately, the tires’ contact with a speed bumpp or

106

S.M. Kim et al.

barrier on a road induces impact noise that is further transmitted as vehicle shock vibration over the car body [4]. This transmitted vibration can be easily measured by an accelerometer. Thus, we can propose an impact noise activity detector comprising the signals from an accelerometer instead of those from a microphone. The remainder of this paper is organized as follows. Following this introduction, we review a statistical model-based decision rule in Section 2. In Section 3, we propose a technique for discriminating impact noises from speech. In particular, the approach to utilizing an accelerometer in detecting impact noise activity is proposed. In Section 4, we evaluate the discrimination performance of the proposed method. Finally, we summarize our findings in Section 5.

2

Statistical Model-Based Target Signal Activity Detection

Target signal activity detection can be interpreted as a binary hypothesis test. Let X k (A) be the spectral component of a microphone signal, where k (= 1,2,", K ) is a frequency bin index and A (= 1,2,") is a frame index. Also, let Tk (A) and Dk (A) denote the target and non-target spectral component, respectively. Then, two hypotheses can be described as H T , 0 : X k (A) = Dk (A) and H T ,1 : X k (A) = Dk (A ) + Tk (A) . Assuming that Tk (A) and Dk (A) follow zero-mean complex Gaussian distributions, the likelihood ratio on H T , 0 and H T ,1 under the observation X k (A) , Λ k ( X k (A) ), is given by, Λ k ( X k (A)) =

 ϕ (A)ψ T ,k (A)   exp T ,k  1 + ψ (A )  1 + ψ T , k (A ) T ,k   1

(1)

where ψ T ,k (A) = λT ,k (A ) λD ,k (A) and ϕT ,k (A) = | X k (A) |2 λD ,k (A) . Here, ψ T ,k (A) and ϕT ,k (A) indicate the a priori and a posteriori signal-to-noise ratio (SNR), respectively. In addition, λT ,k (A) and λD ,k (A) indicate target and non-target spectral variance, respectively. Then, a target activity decision rule is established from the average value of the log likelihood ratio for individual frequency bin as, log Λ T (A) =

1 K −1  logΛ k (A) K k =0

≥ η T : H T ,1 < ηT : H T ,0

(2)

where η T is a pre-determined decision threshold. Consequently, the decision rule strongly depends on the reliable estimate of λT ,k (A) / λD ,k (A) and ψ T ,k (A) / ϕT ,k (A). In other words, the statistical model-based target signal activity decision rule requires target and non-target signal spectral variance estimation. Therefore, the decision procedure is rewritten as {H T , 0 or H T ,1} = ℜT (ψ T ,k (A),ϕT ,k (A), ζ D ,ηT ) ∀k

(3)

where ℜ T (⋅) is a statistical model-based decision function and ∀k means that all frequency bins are used to decide the target signal activity.

Discrimination of Speech Activity and Impact Noise

3

107

Proposed Impact Noise and Speech Discrimination Method

Let N k (A) denote the k-th spectral component of the A-th frame of background car noise, distributed over the entire time interval, as shown in Fig. 1(b). In addition, let Vk (A) and S k (A) be impact noise and speech of the k-th frequency bin and the A-th frame, respectively. Then, depending on the presence or absence of Vk (A) and S k (A) , four hypotheses could be constructed as H 0 : X k ( A) = N k ( A)

H 1 : X k (A) = N k (A) + Vk (A)

H 2 : X k (A) = N k (A) + S k (A),

H 3 : X k (A) = N k (A) + S k (A) + Vk (A).

(4)

This paper aims to decide which hypothesis out of four hypotheses, i.e., H 0 , H 1 , H 2 and H 3 , is true. Assuming that N k (A) , Vk (A) and S k (A) follow zero-mean com-

plex Gaussian distributions, 16 likelihood ratios, i.e., Λ ij (i=0,1,2,3; j=0,1,2,3), could be derived directly on H 0 , H 1 , H 2 and H 3 [12]. The approach using 16 likelihood ratios makes it difficult to tune the relevant parameters and optimize its performance. Instead, the 16 hypotheses can be reduced into six ones by properly combining three kinds of target signal activity hypothesis models. The first of them is the hypothesis of non-background noise, H VS , which corresponds to any of impact noise, speech or both. On one hand, H S and H V are defined for only speech and impact noise activity, respectively. They are represented as H VS , 0 : X k (A) = N k (A ),

H VS ,1 : X k (A ) = N k (A) + VS k (A) with VS k (A) = Vk (A) + S k (A) (5a)

H S , 0 : X k (A) = NVk (A), H S ,1 : X k (A) = NVk (A) + S k (A) with NVk (A) = N k (A) + Vk (A) (5b) H V , 0 : X k (A) = N k (A),

H V ,1 : X k (A) = NS k (A) + Vk (A) with NS k (A) = N k (A) + S k (A) .

(5c)

From Eqs. (5a), (5b) and (5c), we can further define four different hypotheses, including only background noise activity, H 0 , speech activity, H 1 , impact noise activity, H 2 , and mixture of speech and impact noise activity, H 3 . That is, H0

←

H VS , 0 & H V , 0 & H S , 0 ,

H1

←

H VS ,1 & H V , 0 & H S ,1

(6a)

H2

←

H VS ,1 & H V ,1 & H S , 0 ,

H3

←

H VS ,1 & H V ,1 & H S ,1 .

(6b)

Fig. 2 shows a block diagram of the proposed approach. First, the power spectral density (PSD) of the acoustic signal, | X k (A) |2 , which is obtained from a microphone, is used to detect non-background noise and speech activity, i.e., H VS , 0 / H VS ,1 . On the other hand, the PSD of the vibration signal, | Yk (A) |2 , which is obtained from an accelerometer, is used to detect impact noise activity, H V , 0 / HV ,1 . In addition, the

108

S.M. Kim et al.

Fig. 2. Block diagram of thee proposed discrimination approach between impact noise and speech

vibration signal-based im mpact noise activity decision, H V , 0 / H V ,1 , and nnonbackground noise and speech activity decision, H VS ,0 / HVS ,1 , are also utilized togetther to estimate the spectral varriance, λV ,k (A) , of impact noise from a microphone, whhich is used to decide speech activity, a i.e., H S , 0 / H S ,1 . Finally, among H 0 , H 1 , H 2 and othesis is true by using the three decision results of H VS H 3 , we decide which hypo V ,0 / HVS ,1 , H V ,0 / H V ,1 , and H S ,0 / H S ,1 .

3.1

Statistical Model-B Based Impact Noise and/or Speech Activity Detection

From the hypothesis on th he presence, HVS ,1 , or absence, H VS ,0 , of non-backgrouund noise in Eq. (5a), we can deetect non-background noise activity using a statistical m model-based decision rule, as deefined in Section 2. That is, {H VS , 0 or H VS ,1 } = ℜVS (ψ VS ,k (A),ϕVS ,k (A),ηVS ) ∀k

(7)

where ℜVS (⋅) denotes a statistical model-based decision function for the nnonbackground noise activity, VS (A) , and ηVS is a threshold. 3.2

Statistical Model-B Based Speech Activity Detection

Contrary to Section 3.1, fro om the hypothesis on the presence, H S ,1 , or absence, H S , 0 , of speech in Eq. (5b), we can also detect speech activity using a statistical moddelbased decision rule, such ass {H S ,0 or H S ,1 } = ℜ S (ψ S ,k (A),ϕ S ,k (A),η S ) ∀k

(8)

Discrimination of Speech Activity and Impact Noise

109

where ℜ S (⋅) denotes a speech activity decision function for S (A) and η S is a threshold. In Eq. (8), ψ S ,k (A) = λS ,k (A) / λ NV ,k (A) and ϕ S ,k (A) =| X k (A) |2 / λ NV ,k (A) . In order to estimate ψ S ,k (A) and ϕ S ,k (A) , the estimate of λ NV ,k (A) , λˆNV ,k (A) , is obtained by λˆNV ,k (A) = λˆN ,k (A) + λˆV ,k (A)

(9)

where λˆN ,k (A) and λˆV ,k (A) are the PSD estimate of background noise and impact noise, respectively. Note that λˆN ,k (A) has been already estimated from the nonbackground noise activity estimation procedure. In addition, we will give a detail explanation on estimating λˆV ,k (A) in the next subsection. In order to estimate ψ S ,k (A) , λˆS ,k (A) should be also estimated before, which is done by a spectral subtraction me-

thod between λˆVS ,k (A) and λˆV ,k (A) by the following equation of λˆS ,k ( A ) = max( λˆVS ,k ( A) − β ⋅ λˆV ,k ( A),0)

(10)

where β is a tuning parameter. 3.3

Impact Noise Activity Detection Using an Accelerometer

As mentioned in Section 1, we utilize an accelerometer for impact noise activity detection. Let Yk (A) be the spectral component of sensing signal by an accelerometer at the k-th frequency bin and A-th frame. Also, let N k′ (A) and Vk′(A ) denote the accelerometer noise (or impact noise) and impact vibration, respectively. Then, two hypotheses can be postulated as H V′ , 0 : Yk (A) = N k′ (A) and H V′ ,1 : Yk (A) = N k′ (A ) + Vk′(A) . Applying a statistical-based decision rule similar to Eq. (3), we can detect the impact vibration activity by {H V′ , 0 or H V′ ,1} = ℜ′V (ψ V′ ,k (A),ϕV′ ,k (A ),ηV′ ) ∀k ,

(11)

where the distributions of N k′ (A) and Vk′ (A) are zero-mean complex Gaussian, ℜ′V (⋅) denotes a decision function of an impact vibration activity, and ηV′ is a threshold. Figs. 3(a) and 3(b) show that the time interval of impact noise is similar to that of impact vibration, which is estimated by a log likelihood-based decision rule depicted in Fig. 3(c). Thus, upon this observation, we alternatively use the impact vibration activity interval instead of the impact sound activity interval in this paper.

110

S.M. Kim et al.

(a)

(b)

(c) Fig. 3. Illustration of the relatiionship between impact noise and impact vibration; (a) the acooustic signal obtained from a micrrophone, (b) vibration signal obtained from an accelerometer and (c) log likelihood estimate of th he impact vibration activity

Although the impact noiise activity detection is approximately accomplished by detecting the impact vibratio on activity, the spectral variance estimate of the imppact noise, λˆV ,k (A) , is required in i Eqs. (9) and (10). In other words, we use a Wiener fiilter in order to estimate λˆV ,k (A) . Also, λˆNS ,k (A) is estimated by a recursive procedure executed only when H V′ , 0 is determined d to be true. That is, we have λˆNS ,k (A) = ζ NS ⋅ λˆNS ,k (A − 1) + (1 − ζ NS ) ⋅ λˆVS ,k (A)

where ζ NS is a smoothing parameter.

((12)

Discrimination of Speech Activity and Impact Noise

111

Fig. 4. Configuration of an acccelerometer sensor and a microphone system in order to colllect impact vibration and speech siignals from a commercial navigation system

(a)

(b)

(c)

Fig. 5. Snapshot of the road conditions c for impact noise; (a) a speed bump, (b) two successsive convex hemispherical surfacess, and (c) a concave surface

4

Performance Ev valuation

Fig. 4 shows a system em mploying a microphone and an accelerometer, which w was constructed in a 2,000 cc class c vehicle in order to collect test data for performaance evaluation. In particular, a commercial navigation platform equipped with a low ccost microphone was used to co ollect speech test data. Here, B&K type 4393 acceleromeeter sensor was used in order to t measure the car impact vibration signal, where a B&K Type 2692 conditional am mplifier was used to adjust the signal power. The desiired speech signal was played from f a speaker mounted below the headrest of the driveer’s seat under a clean condition n. The background noise, impact noise, and vibration siggnal were recorded while drivin ng the car at an average speed between 30 and 50 km m/h subject to three different roaad conditions shown in Fig. 5.

112

S.M. Kim et al.

Table 1. Confusion matrix between the manual decision and the decision by the proposed algorithm

Manual Decision

H0

Case I H1 H2

Decision by the Proposed Algorithm Case II Case III H3 H0 H1 H2 H3 H0 H1 H2

H3

H0

91.4

1.8

6.6

0.1

92.7

2.4

4.7

0.3

95.6

1.6

2.8

0.05

H1

10.3 85.2

3.5

0.9

7.3

89.6

0.3

2.8

9.9

76.2

1.7

12.2

H2

6.1

0

93.9

1.0

3.3

0.9

88.5

7.2

-

-

-

-

H3

-

-

-

-

-

-

-

-

0.3

4.2

10.4 85.2

Located on the dashboard, 10 speech utterances by 5 males and 5 females, 10 impact noise signals, and vibration signals were separately recorded at a sampling rate of 16 kHz. In order to simulate the driving conditions, we mixed speech signals with impact noise in three different ways. For the first case, impact noise occurred at the beginning of the speech interval, which was referred as Case I. For the second one, impact noise appeared at the end of the speech interval (Case II). Lastly, we added impact noise into the speech interval (Case III). Subsequently, 100 mixed utterances were prepared as a test database, where SNRs ranged from -3 dB to 1 dB. The proposed method was applied once every frame whose length was 20 ms long. In order to obtain spectral components, we applied a 320-point discrete Fourier transform (DFT). Throughout the experiment, we set ηVS = 0.5 in Eq. (7), η S = 0.5 in Eq. (8), β = 0.5 in Eq. (10), ηV′ = 0.15 in Eq. (11), and ζ NS = 0.9 in Eq. (12). The performance of the proposed method was measured as an accuracy ratio between the decision outputs from the proposed algorithm and those obtained manually. Table 1 summarizes the results in a confusion matrix form. As shown in the table, the proposed method exhibited more than 85% accuracy in Case I and Case II. The accuracy of Case III was lowered than those of Case I and Case II. This was mainly contributed by the confusion between H 1 and H 3 states.

5

Conclusion

In this paper, we proposed a technique to discriminate between impact noise and speech activity in a car environment. The proposed technique employed three statistical model-based decision rules, for non-background noise, impact noise and speech activities. In particular, an effective impact noise detector using an accelerometer was proposed. Then, each decision rule result was utilized to discriminate between an impact noise and speech activity. From the performance evaluation, the proposed algorithm had a discrimination accuracy of 85 %.

Discrimination of Speech Activity and Impact Noise

113

Acknowledgements. This work was supported in part by the Industrial Strategic technology development program, 10035252, ‘Development of dialog-based spontaneous speech interface technology on mobile platform’ funded by the Ministry of Knowledge Economy, Korea, and by the Basic Science Research Program through the National Research Foundation of Korea funded by the Ministry of Education, Science and Technology (No.2011–0026201).

References 1. Wu, K.G., Chen, P.C.: Efficient speech enhancement using spectral subtraction for car hands-free applications. In: roceedings of International Conference on Consumer Electronics (ICCE), Las Vegas, NV, pp. 220–221 (2007) 2. Ahn, S., Ko, H.: Background noise reduction via dual-channel scheme for speech recognition in vehicular environment. IEEE Transactions on Consumer Electronics 51(1), 22–27 (2005) 3. Kim, S.M., Kim, H.K.: Hybrid probabilistic adaptation model controller for generalized sidelobe canceller-based target-directional speech enhancement. In: Proceedings of ICASSP, Prague, Czech Republic, pp. 2532–2535 (2011) 4. Lee, S.-K., Kim, H.-W., Na, E.-W.: Improvement of impact noise in a passenger car utilizing sound metric based on wavelet transform. Journal of Sound and Vibration 329(17), 3606–3619 (2010) 5. Lee, S.-K., Chae, H.-C.: The application of artificial neural networks to the characterization of interior noise booming in passenger cars. Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering 218(1), 33–42 (2004) 6. Wang, Y.S., Lee, C.-M., Kim, D.-G., Xu, Y.: Sound-quality prediction for nonstationary vehicle interior noise based on wavelet pre-processing neural network model. Journal of Sound and Vibration 299(4-5), 933–947 (2007) 7. Hu, J.S., Cheng, C.C., Liu, W.H., Yang, C.H.: A robust adaptive speech enhancement system for vehicular applications. IEEE Transactions on Consumer Electronics 52(3), 1069–1077 (2006) 8. Kim, S.M., Kim, H.K.: Probabilistic spectral gain modification applied to beamformerbased noise reduction in a car environment. IEEE Transactions on Consumer Electronics 57(2), 866–872 (2011) 9. Park, J.H., Kim, S.M., Yoon, J.S., Kim, H.K., Lee, S.J., Lee, Y.K.: SNR–based mask compensation for computational auditory scene analysis applied to speech recognition in a car environment. In: Proceedings of Interspeech, Makuhari, Japan, pp. 725–728 (2010) 10. Park, J.H., Shin, M.H., Kim, H.K.: Statistical model-based voice activity detection using spatial cues and log energy for dual-channel noisy speech recognition. CCIS, vol. 120, pp. 172–179 (2010) 11. Sohn, J., Kim, N.S., Sung, W.: A statistical model-based voice activity detection. IEEE Signal Processing Letters 6(1), 1–3 (1999) 12. Lee, S.Y., Shin, J.W., Yun, H.S., Kim, N.S.: A statistical model based post-filtering algorithm for residual echo suppression. In: Proceedings of Interspeech, Antwerp, Belgium, pp. 858–861 (2007)

Crosstalk Cancellation for Spatial Sound Reproduction in Portable Devices with Stereo Loudspeakers Sung Dong Jo1, Chan Jun Chun1, Hong Kook Kim1, Sei-Jin Jang2, and Seok-Pil Lee3 1

School of Information and Communications Gwangju Institute of Science and Technology(GIST), Gwangju 500-712, Korea {sdjo,cjchun,hongkook}@gist.ac.kr 2 NSSC Center, Korea Electronics Technology Institute, Goyang, Gyeonggi-do 410-380, Korea [email protected] 3 Digital Media Research Center Korea Electronics Technology Institute, Seoul 137-070, Korea [email protected]

Abstract. To reproduce spatial sound through stereo loudspeakers in a portable device environment, it is important to properly design a crosstalk cancellation algorithm that cancels out acoustical crosstalk signals. In other words, the difference between the direct path of head-related transfer functions (HRTFs) and the crosstalk path of HRTFs is very small at certain frequencies, and this causes an excessive boost of frequencies when designing a crosstalk cancellation filter. To mitigate this problem, we propose a crosstalk cancellation filter design method that allows for the selective attenuation of unwanted peaks in the spectrum by constraining the magnitude of the difference between the direct and crosstalk path. The performance of the proposed method is evaluated by subjective source localization and objective tests. It is shown from the tests that the proposed method can provide improved spatial sound effects with very closely spaced stereo loudspeakers. Keywords: Crosstalk cancellation, inverse filter, HRTFs, fast deconvolution.

1

Introduction

In recent years, numerous portable devices such as mobile phones, laptop computers, MP3 players, portable TV sets, and portable digital imaging devices have become available. Many of these devices are equipped with a pair of small loudspeakers. However, due to the limited size of these devices, the distance between the loudspeakers is very short and this leads to a poor spatial sound effects. Typically, the objective of spatial sound reproduction systems is to synthesize a virtual sound image such that the listener perceives as if the signals reproduced at the listener’s ears would have been produced by a specific source located at an intended position relative to the listener [1]. This type of spatial sound reproduction system can provide an immersive sound environment with variable applications in virtual reality, augmented reality, video games, mobile phones, and home entertainment systems, among others. Typically, there are two main environments for rendering a spatial sound. The first tries to generate a spatial T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 114–123, 2011. © Springer-Verlag Berlin Heidelberg 2011

Crosstalk Cancellation for Spatial Sound Reproduction in Portable Devices

115

sound in a headphone-based environment and the second uses two or more loudspeakers to render a spatial sound. In the case of a headphone-based environment, the spatial sound source can be easily reproduced at the listener`s ears, since the headphone separates binaural sound channels to each ear [2]. In contrast, in a loudspeaker environment, binaural sounds from each loudspeaker are mixed and simultaneously delivered to both of the listener`s ears. That is, each loudspeaker sends sound to the same-side ear, as well as undesired sound to the opposite-side ear. This problem is known as crosstalk and it degrades the spatial sound reproduction [3]. To overcome this problem, various crosstalk cancellation algorithms have been proposed in order to reduce the crosstalk effect by designing appropriate inverse filters of acoustic transfer functions. Note that the concept of crosstalk cancellation was first introduced in the early 1960s [4]. Since then, a number of sophisticated crosstalk cancellation algorithms have been presented which used two or more loudspeakers to render binaural signals [5]. In practice, a crosstalk cancellation algorithm can be implemented by a two-by-two matrix of digital filters. Unfortunately, a severe inversion problem arises when the difference between the direct path head-related transfer function (HRTF) and the cross-talk path HRTF is very small, and so one ends up having to invert an almost singular two-by-two matrix. This undesirable property causes the optimal solution to amplify certain frequencies by a large amount. This problem, usually referred to as illconditioning, is particularly severe at low frequencies when two loudspeakers are positioned close together [6]. In this paper, we propose a crosstalk cancellation method suitable for a portable device environment. In the proposed method, the ill-conditioning at certain frequencies is prevented from constraining the magnitude of difference between the direct and the crosstalk paths of HRTF before designing the inverse filters of the acoustic transfer functions. The organization of this paper is as follows. Following this introduction, a conventional crosstalk cancellation method is briefly reviewed in Section 2. Section 3 describes the fast deconvolution method using frequency-dependent regularization. After that, we propose a crosstalk cancellation algorithm in Section 4. In Section 5, the performance of the proposed method is analyzed. Finally, this paper is concluded in Section 6.

2

Conventional Crosstalk Cancellation

A block diagram of crosstalk cancellation for two (or stereo) loudspeakers is illustrated in Fig. 1, which is positioned symmetrically in front of a single listener. In the figure, z-transform is used to denote the relevant signals and system responses. x1 ( z ) and x 2 ( z ) are the input binaural signals, ν 1 ( z ) and ν 2 ( z ) are the inputs to

the two loudspeakers, and w1 ( z ) and w2 ( z ) are the sound signals generated at the listener`s ears. There are two acoustic transfer functions from stereo loudspeakers to the listener`s ears; the direct path C1 ( z ) and the crosstalk path C 2 ( z ). In addition, H 1 ( z ) and H 2 ( z ) are the crosstalk cancellation functions that transform the binaural

116

S.D. Jo et al.

Fig. 1. Block diagram of a conventional crosstalk cancellation system for stereo loudspeakers

signals, x1 ( z ) and x2 ( z ) , into the loudspeakers input signals, ν 1 ( z ) and ν 2 ( z ) . By inspecting Fig. 1, it is verified that ν 1 ( z )   H 1 ( z ) ν ( z ) =  H ( z )  2   2

H 2 ( z )  x1 ( z )  H 1 ( z )  x 2 ( z )

(1)

 w1 ( z )  C1 ( z )  w ( z )  = C ( z )  2   2

C 2(z)  ν 1 ( z )  . C1 ( z ) ν 2 ( z )

(2)

and

An ideal cross-talk cancellation system reproduces the input binaural signals, x1 ( z ) and x2 ( z ) , at each ear of the listener. Thus, it is straightforward to show that this is achieved when the crosstalk cancellation matrix H(z ) should be the inverse matrix of acoustic transfer functions, C(z ) . That is, H ( z) H( z ) =  1 H 2 ( z) 2.1

H 2 ( z ) 1 =  2 H 1 ( z ) C1 ( z ) − C2 ( z ) 2

C1 ( z ) − C2 ( z )  − C ( z ) C ( z ) . 1  2 

(3)

Inverse Filter Design

The main issue in a crosstalk cancellation system is to invert a matrix C(z ) . In practice, the exact inverse is not possible. It is possible to get a stable and causal inverse system when a system is minimum phase. However, acoustic transfer functions are not likely to be a minimum phase system. Another problem encountered with the exact inversion is in that H 1 ( z ) and H 2 ( z ) become very large when the difference between the direct path C1 ( z ) and the crosstalk path C2 ( z ) is very small. This problem particularly becomes severe at very low frequencies, since the direct path, C1 ( z ) , and the crosstalk path, C2 ( z ) , are almost equal when the wavelength is very long. This undesirable property causes the optimal solution to amplify certain frequencies by a large amount. Thus, the design should carefully consider the inverse

Crosstalk Cancellation for Spatial Sound Reproduction in Portable Devices

117

Fig. 2. Block diagram of the fast deconvolution method

functions in an environment in which stereo loudspeakers are to be positioned close together. Consequently, it is not realistic to expect an exact inverse of the acoustic transfer functions. However, by designing an approximated inverse filter, we can get an optimal solution close to the exact solution. The requirements for designing a crosstalk cancellation filter can be broadly classified into either time domain or frequency domain design method. Time domain design methods guarantee a stable and causal solution by approximating the inverse filter in the time domain, though they have a complex structure and heavy computational complexity. In contrast, frequency domain design methods can be used easily obtain an optimal solution and to theoretically analyze the solution [5]. In this paper, we aim to design a crosstalk cancellation filter for a portable device environment where stereo loudspeakers are positioned close together. Thus, it is very important to mitigate the ill-conditioning problem.

3

Inverse FIR Filter Design Using Fast Deconvolution with Frequency-Dependent Regularization

This section outlines the theory upon which the fast deconvolution algorithm is based. Fast deconvolution is a method based on a fast Fourier transform (FFT) in combination with the least-square approximation method, which is commonly used for designing crosstalk cancellation filters [7]. This method does not try to find the exact solution, but rather the best approximation, which results in minimum errors in a least-squares sense. 3.1

Least-Squares Approximations

Fig. 2 shows a block diagram of the fast deconvolution method. In the figure, it is assumed that the crosstalk cancellation system works in the z-transform domain. Here, x (z ) is a vector of input binaural signals, H(z ) is a matrix composed of the crosstalk cancellation filters, v (z ) is a vector of the loudspeaker input signals, C(z ) is a vector of acoustical transfer functions, w(z ) is a vector of the sound signals reproduced at the listener`s ears, I(z ) is the identity matrix, d(z ) is a vector of the desired signals, and e(z ) is a vector of the performance error signals. In addition, z − m implements a modeling delay of m samples to ensure that the crosstalk

118

S.D. Jo et al.

cancellation system is causal and performs well not only in terms of amplitude but also in terms of phase [8]. The relationships among the signals are denoted as follows v ( z ) = H( z ) x ( z )

(4)

w ( z ) = C( z ) v ( z )

(5)

d( z ) = z − m I ( z ) x ( z )

(6)

e( z ) = d ( z ) − w ( z ) .

(7)

Then, by employing the Tikhonov regularization [9], the filter design is performed by minimizing the cost function of J ( z ) = e H ( z )e ( z ) + βv H ( z ) v( z )

(8)

where the first term e H ( z )e( z ) is the performance error term and β v H ( z ) v( z ) is the effort penalty term. In addition, H represents the Hermitian operator, which transposes and conjugates its argument, and the positive real number β is a regularization parameter that determines the weight given to the effort term. 3.2

Fast Deconvolution Method Based on Frequency-Dependent Regularization

The cost function J (z ) is a minimum in the least-squares sense when H(z ) is given as H( z ) = [C H ( z )C( z ) + βB * ( z ) B( z )I]−1 C H ( z ) z − m

(9)

where * denotes the complex conjugate operator. In this case, the regularization parameter is the product of two components: a gain factor β and a shape factor B(z ) , where B(z ) is the z-transform of a digital filter which amplifies the undesired frequencies boosted by crosstalk cancellation [10]. Thus, we can suppress the signal value boosted at certain frequencies by adjusting the regularization term. In other words, it is important to find the optimal frequency-dependent regularization parameters that compromise crosstalk cancellation in order to minimize the illconditioning problem. To this end, Eq. (9) gives an expression as a continuous function of frequency.

4

Proposed Crosstalk Cancellation

In this section, we propose a crosstalk cancellation design method for providing spatial sound reproduction through a pair of very closely spaced loudspeakers. As mentioned in Section 3, there is a greater potential for the ill-conditioning problem to arise when the two loudspeakers are positioned close together, such as the case in a portable device environment. This is because both the direct and crosstalk paths from

Crosstalk Cancellation for Spatial Sound Reproduction in Portable Devices

119

Fig. 3. Environment of the two close small loudspeakers arrangement

Fig. 4. Magnitude responses of difference between C1 ( k ) and C2 ( k )

the two loudspeakers to the listener’s ears are similar. In order to mitigate this problem, a frequency-dependent regularization method based on fast deconvolution was previously proposed in [7]. However, the purpose of the regularization technique is to impose a subjective constraint on the solution. The constraint may also degrade the performance of crosstalk cancellation because of some losses in the inversion accuracy if the regularization value affects the output signals by suppressing the signal value at frequencies that have no ill-conditioning problem. Thus, we propose a technique to prevent ill-conditioning at certain frequencies by constraining the magnitude difference between the direct and crosstalk paths of HRTF before designing the inverse filters of the acoustic transfer functions. 4.1

Design of Crosstalk Cancellation Filter

Fig. 3 shows the configuration of stereo loudspeakers in this study. Each speaker had 2 cm diameter. It were placed at 30 cm apart from the listener at a listening angle of 10D or - 10D . In this paper, we used the HRTFs data, which were measured on a KEMAR dummy-head [11]. Fig. 4 shows the magnitude responses of the difference between the direct path, C1 ( k ) , and the crosstalk path, C2 ( k ) , according to the configuration of Fig. 3. As shown in Fig. 4, the magnitude responses of the differences between C1 (k ) and C2 ( k ) were very small at low frequencies. In other words, C1 ( k ) and C2 ( k ) were quite similar because stereo loudspeakers were placed very close together. In addition, the magnitudes were quite small at frequencies around 8 kHz since the HRTFs had steep notches at around 8 kHz. This

120

S.D. Jo et al.

(a)

(b)

Fig. 5. Magnitude responses of (a) H 1 ( k ) and (b) H 2 ( k ) calculated using fast deconvolution without any regularization

(a)

(b)

Fig. 6. Magnitude responses of (a) H1 ( k ) and (b) H 2 (k ) calculated using fast deconvolution with a threshold magnitude of TdB = -5 dB

was caused by a pinna reflection and anti-aliasing filters [8]. Therefore, if a crosstalk cancellation filter was designed by inverting the transfer functions, C1 ( k ) and

C2 (k ) , the solution would contain large peaks around the frequencies where C1 ( k ) and C2 ( k ) were similar. This undesirable property could lead to strong boosts at those frequencies. Fig. 5 shows the magnitude responses of the crosstalk cancellation filters, H1 ( k ) and H 2 (k ) , which were calculated using the fast deconvolution method with no regularization. Here, the gain regularization factor β was zero in Eq. (9). As shown in the figure, the magnitude responses of the crosstalk cancellation filters had sharp peaks at the frequencies below around 1 kHz, at around 8 kHz, and over 20 kHz. Importantly, there was a very sharp peak at around 1 kHz, which could cause an excessive boost at the low frequencies. Consequently, a crosstalk cancellation filter should be carefully implemented in order to avoid overloading the loudspeakers. To prevent the sharp peaks at such frequencies, we constrained the magnitude difference to between C1 (k ) and C2 ( k ) before designing the crosstalk cancellation filters. The magnitude difference between C1 ( k ) and C2 ( k ) was defined as m(k ) = 20 log10 ( C1 (k ) − C2 (k ) ) .

(10)

Crosstalk Cancellation for Spatial Sound Reproduction in Portable Devices

121

 Then, the magnitude difference between C1 ( k ) and C 2 (k ) was redefined as m(k ) by using a threshold, TdB , as  m (k ) = max (m(k ), TdB ) .

(11)

Fig. 6 shows the magnitude responses of crosstalk cancellation filters, H1 (k ) and H 2 (k ) , which were calculated using the fast deconvolution method with a threshold of TdB = -5 dB. As shown in the figure, the sharp peaks of magnitude responses disappeared at the frequencies, as compared to Fig. 5.

5

Performance Evaluation

In this section, we evaluated the performance of the proposed method. First, we conducted a subjective localization test in the environment illustrated in Fig. 3. To this end, six subjects participated in each test and the subjects were instructed to sit at a position in front of stereo loudspeakers. The test stimulus was a pink noise and binaural signals were rendered by filtering the pink noise with HRTFs. The directions were on the front horizontal plane from -90° to 90° at a step of 30° on the azimuth. Each stimulus was played 5 times at durations of 25 ms with 50 ms silent intervals. In this test, we measured the perceived source direction for the two kinds of pink noise. One was a pink noise rendered by HRTFs only, and the other was processed via the crosstalk cancellation from the pink noise rendered by HRTFs. The results of the source localization test were shown in terms of target azimuth versus the judged azimuth. Figs. 7(a) and 7(b) show the results of the subjective localization test of the azimuth without and with crosstalk cancellation, respectively. When crosstalk cancellation was not applied, the judged azimuths were measured to within ± 30D of all target azimuths. However, the judged azimuths were closer to the target azimuth after crosstalk cancellation was applied. Next, we measured the channel separation at the listener’s ears, which was defined as the ration between the contralateral magnitude response, C 2 (k ) , and the ipsilateral magnitude response, C1 ( k ) [12]. Fig. 8 shows the channel separation for the proposed crosstalk cancellation method. It was shown from the figure that crosstalk cancellation was indeed effective.

Fig. 7. Results of the subjective localization test of the azimuth (a) without and (b) with crosstalk cancellation

122

S.D. Jo et al.

Fig. 8. Comparison of channel separation: the dashed line is a natural head shadowing channel separation and the solid line shows the channel separation after crosstalk cancellation

6

Conclusion

In this paper, we proposed a crosstalk cancellation design method that provided spatial sound reproduction through a pair of very closely spaced loudspeakers, such as those found in portable devices. The proposed method allowed the selective attenuation of unwanted peaks in the spectrum by constraining the magnitude of the difference between the direct and crosstalk paths. Then, the performance of the proposed method was evaluated by a subjective source localization test and an objective test. It was shown from the tests that the proposed method could provide spatial sound effects using only a pair of very closely spaced loudspeakers.

References 1. Bauck, J.L., Cooper, D.H.: Generalized transaural stereo and applications. Journal of the Audio Engineering Society 44(9), 683–705 (1996) 2. Begault, D.R.: Challenges to the successful implementation of 3-D sound. Journal of the Audio Engineering Society 39(11), 864–870 (1990) 3. Cooper, D.H., Bauck, J.L.: Prospects for transaural recording. Journal of the Audio Engineering Society 37(1/2), 3–19 (1989) 4. Atal, B.S., Schroeder, M.R.: Apparent sound source translator. U.S. Patent (3), 236, 949 (1966) 5. Nelson, P.A.: Active control of acoustic fields and the reproduction of sound. Journal of Sound and Vibration 177(4), 447–477 (1994) 6. Kirkeby, O., Nelson, P.A., Hamada, H.: The “stereo dipole” – a virtual source imaging system using two closely spaced loudspeakers. Journal of the Audio Engineering Society 46(5), 387–395 (1998) 7. Kirkeby, O., Nelson, P.A., Hamada, H., Orduna-Bustamante, F.: Fast deconvolution of multichannel systems using regularization. IEEE Trans. Speech and Audio Processing 6(2), 189–194 (1998) 8. Parodi, Y.L., Rubak, P.: Objective evaluation of the sweet spot size in spatial sound reproduction using elevated loudspeakers. Journal of the Acoustical Society of America 128(3), 1045–1055 (2010)

Crosstalk Cancellation for Spatial Sound Reproduction in Portable Devices

123

9. Tikhonov, A.N.: Solution of incorrectly formulated problems and the regularization method. Soviet Mathematics Doklady 4(2), 1035–1038 (1963) 10. Kirkeby, O., Nelson, P.A.: Digital filter design for inversion problems in sound reproduction. J. Audio Eng. Soc. 47(7/8), 583–595 (1999) 11. Gardner, B., Martin, K.: HRTF Measurements of a KEMAR dummy-head microphone. MIT Media Lab., http://sound.media.mit.edu/KEMAR.html 12. Gardner, W.G.: 3-D audio using loudspeakers. Ph.D. Dissertation. MIT Media Lab, Cambridge (1997)

Perceptual Enhancement of Sound Field Reproduction in a Nearly Monaural Sensing System Chan Jun Chun1, Hong Kook Kim1, Seung Ho Choi2, Sei-Jin Jang3, and Seok-Pil Lee4 1

School of Information and Communications Gwangju Institute of Science and Technology, Gwangju 500-712, Korea {cjchun,hongkook}@gist.ac.kr 2 Department of Electronic and Information Engineering Seoul National University of Science and Technology, Seoul 139-743, Korea [email protected] 3 NSSC Center, Korea Electronics Technology Institute, Goyang 410-380, Korea [email protected] 4 Digital Media Research Center, Korea Electronics Technology Institute, Seoul 137-070, Korea [email protected]

Abstract. In this paper, we propose a method for enhancing the sound field in a nearly monaural sensing system. The proposed method controls the interchannel coherence (ICC) of a stereo sound source by incorporating early reflections into the sound source. The degree of ICC is controlled using both a decorrelator and a principal component analysis (PCA) technique, while a cascaded all-pass filter is used to induce early reflections. To demonstrate the effectiveness of the proposed method on the perception of the sound field, a subjective test is carried out. It is shown from the test that the proposed method can reduce the ICC, thus it improves the perceived sound field. Keywords: Sound field, nearly monaural sensing system, inter-channel coherence (ICC), principal component analysis (PCA), decorrelator, early reflections.

1

Introduction

Spatial quality, which is a subset of sound quality, is an important factor in the audio field [1]. It is attributed to the assessment of an auditory image and indicates whether a listener is satisfied or not with the auditory image. There are several spatial attributes associated with spatial quality. Among them, sound field is mostly related to how a listener perceives the width of an auditory image from the audio source, which is shown in Fig. 1. The sound filed is also called the apparent source width (ASW) [2]. Microphone configuration plays an important role in capturing the spatial quality. There are many techniques [3] for recording the spatial quality. One of the techniques suggests that the distance between auditory sensors should be determined by the T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 124–131, 2011. © Springer-Verlag Berlin Heidelberg 2011

Perceptual Enhancement of Sound Field Reproduction

125

Fig. 1. Concept of the sound field or the apparent source width (ASW)

Fig. 2. Configuration of auditory sensors for capturing the spatial quality

width of the sound source. It is usual that the distance is set from one-third to one-half of the sound source width, as depicted in Fig. 2 [4]. In particular, the distance ranging from 20 to 60 cm is recommended for the appropriate distance between auditory sensors [1]. For many portable video or audio devices, however, it is difficult to arrange auditory sensors according to the recommended configuration. There is a mechanical limitation due to the size of the portable device, which results in a nearly monaural sensing system even if the device has stereo microphones. Thus, the audio information obtained from such devices rarely covers the aspects of spatial quality, i.e., the sound field. In this paper, we propose a sound field enhancement method for a nearly monaural sensing system. First of all, principal component analysis (PCA) is applied [5][6] to decompose the audio signals obtained by stereo auditory sensors into correlated and uncorrelated components. Next, on the basis of the knowledge that inter-channel coherence (ICC) and early reflections affect the perception of the sound field [7][8], the ICC is reduced by using the combination of PCA and a decorrelator, and early reflections are generated by using an all-pass filter.

126

C.J. Chun et al.

Fig. 3. Perception of the sound field depending on ICC

The remainder of this paper is organized as follows. Following the introduction, we describe the basics of spatial hearing with regard to the perception of the sound field depending on the ICC and the intensity of the early reflections in Section 2. In Section 3, we propose a sound field enhancement method that incorporates PCA, a decorrelator, and early reflections in order to control the ICC. Next, we evaluate the performance of the proposed method by calculating the ICC and by measuring the perceived sound field in Section 4. Finally, we conclude our findings in Section 5.

2

Spatial Hearing

In this section, we discuss the basics of spatial hearing with regard to the sound field. Fig. 3 shows the perception of the sound field depending on the ICC between a pair of loudspeakers. The ICC indicates the coherence between two channels and it is expressed as ∞

ICC =

 x L (n) x R (n)

n = −∞ ∞ ∞ 2 2  x L (n )  x R ( n ) n = −∞ n = −∞

(1)

where x L (n ) and x R (n ) are the left and right audio samples at time n, respectively. Fig 3 shows how the ICC affects the auditory perception [7]. When the ICC is almost 1.0, the auditory image is in the middle of two loudspeakers. Thus, the perception of the sound field is actually narrow, as shown in the shadowed area labeled as ‘1’ in Fig. 3. On the other hand, as the ICC is reduced to 0, the perception of the sound field is increased, as illustrated in the shaded area labeled as ‘4’ in Fig. 3. Thus, in order to adjust the ICC, both PCA and a decorrelator are applied in this work.

Perceptual Enhancement of Sound Field Reproduction

127

Fig. 4. Perception of the sound field depending on the intensity of early reflections

Fig. 5. Overall structure of the proposed sound field enhancement method

In addition to the ICC, early reflections are also related to the perception of the sound [8]. Fig. 4 illustrates the relationship between the early reflections and the sound field. It is found that the perception of the sound field increases with the intensity of the early reflections. Thus, in this work, we generate early reflections by using an all-pass filter.

3

Proposed Sound Field Enhancement Method

In this section, we propose a sound field enhancement method in order to enhance the sound field in a nearly monaural sensing system. Fig. 5 shows an overall structure of the proposed method. As shown in the figure, the proposed method incorporates the PCA technique and the generation of early reflections. In addition, it also uses a decorrelator. First of all, the audio signals obtained by stereo auditory sensors are decomposed into correlated and uncorrelated components using the PCA technique. Here, the decomposition into the two signal components is achieved by using a 2×2 covariance matrix, A, which is defined as

cov(x L , x L ) cov(x L , x R )  A= cov(x R , x L ) cov(x R , x R )

(2)

where cov(x p , x q ) is the covariance of x p and x q , and p and q represent the left channel, L, and the right channel, R, respectively. The covariance matrix in Eq. (2)

128

C.J. Chun et al.

Fig. 6. Magnitude and phase response of the decorrelator

has two eigenvectors, which become the basis vectors for a new coordinate system. These eigenvectors are then used as the weight vectors corresponding to the left and right channels to generate the correlated and uncorrelated signal components. Next, the correlated signal component is fed into the input of a decorrelator via the PCA. Fig. 6 illustrates the spectrum of the 640-tapped finite-duration impulse response (FIR) decorrelator. In addition, the correlated component, uncorrelated component, and the decorrelator output are mixed into stereo audio signals, such as

l (n ) = r(n) =

1 3 1 3

(c(n) + u( n) + d (n) )

(3)

(c(n) − u(n) − d (n) )

(4)

where c(n ) and u(n ) are the correlated and uncorrelated components obtained via the PCA, respectively. In addition, d (n ) denotes the decorrelator output. It should be noted here that the stereo audio outputs are divided by 3 for conserving a constant acoustic energy between the input and the processed audio signals. An all-pass filter is used to generate early reflections, as shown in Fig. 7 [9]. The transfer function of the all-pass filter is expressed as

H ( z) =

z − M 1 − g1 z − M 2 − g 2 z −( M 1 + M 2 ) − g1 z − M 2 − g 2 z − M 1 + g1 g 2 ⋅ = 1 − g1 z − M 1 1 − g 2 z − M 2 1 − g1 z − M 1 − g 2 z − M 1 + g1 g 2 z − ( M 1 + M 2 )

(5)

where M 1 and M 2 denote delay lengths of 240 and 840, respectively. The gain coefficients, g1 and g 2 , are set to 0.67 and 0.37, respectively.

Perceptual Enhancement of Sound Field Reproduction

129

Fig. 7. An all-pass filter for generating early reflections

4

Performance Evaluation

1

1

0.5

0.5 ICC

ICC

In this section, we evaluated the performance of the proposed sound field enhancement method. First, we measured the ICC to demonstrate how the PCA and the decorrelator worked together to decompose the audio signals into correlated and uncorrelated components. Fig. 8 shows the ICCs according to frame index, where audio signals are segmented into a sequence of frames whose length is 1024. It was shown from the figure that the ICC was reduced by applying the proposed method.

0

-0.5

-1 0

0

-0.5

200

400 600 Frame index i

(a)

800

1000

-1 0

200

400 600 Frame index i

800

1000

(b)

Fig. 8. ICC according to the frame index; (a) original audio signals and (b) processed audio signals

A subjective test was also conducted to evaluate the degree of the perception of the sound field perception. Six subjects with no auditory diseases participated in the test. We prepared audio files of different genres including four music genres such as ballad, dance, hip-hop, and rock. Fig. 9 illustrates the loudspeaker configuration used in the test. In this test, we measured the perceived sound field. Two types of audio contents were prepared for the test. One was the original audio content, which was obtained in a nearly monaural sensing system. The other was the audio content processed from the original file using the proposed method. For the test, randomly selected audio files were reproduced, and each subject was asked to estimate the length of the perceived sound field by ear.

130

C.J. Chun et al.

Fig. 9. Loudspeaker configuration for the subjective test

Fig. 10. Comparison of the average perceived sound field measured in cm for different genres of audio signals

Fig. 10 illustrates the result of the subjective test. As shown in the figure, the sound filed was improved after applying the proposed method. By averaging the perceived sound field over all the audio signals, it was measured as 30.50 cm for the input audio signals. However, it was increased as 87.63 cm after applying the proposed method.

5

Conclusion

In this paper, we proposed a sound field enhancement method in a nearly monaural sensing system. To this end, the ICC was controlled, and early reflections were incorporated. In other words, to decrease the ICC between the audio signals reproduced by a pair of loudspeakers, a PCA technique and a decorrelator were used. Moreover, an all-pass filter was designed to generate early reflections. The performance of the proposed method was evaluated by measuring the ICC and by performing a subjective test. It was shown from the ICC measurement and the subject test that the proposed method could reduce the ICC and improve the perception of the sound field.

Perceptual Enhancement of Sound Field Reproduction

131

References 1. Rumsey, F.: Spatial quality evaluation for reproduced sound: terminology, meaning, and a scene-based paradigm. Journal of the Audio Engineering Society 50(9), 651–666 (2002) 2. Rumsey, F.: Spatial Audio. Focal Press, Oxford (2001) 3. Gayford, M.: Microphone Engineering Handbook. Focal Press, Oxford (1994) 4. Dooley, W.L., Streicher, R.D.: MS stereo: a powerful technique for working in stereo. Journal of the Audio Engineering Society 30(10), 707–718 (1982) 5. Jolliffe, I.T.: Principal Component Analysis. Springer, Heidelberg (2002) 6. Chun, C.J., Kim, Y.G., Yang, J.Y., Kim, H.K.: Real-time conversion of stereo audio to 5.1 channel audio for providing realistic sounds. International Journal of Signal Processing, Image Processing and Pattern Recognition 2(4), 85–94 (2009) 7. Blauert, J.: Spatial Hearing: The Psychophysics of Human Sound Localization. MIT Press, Cambridge (1997) 8. Barron, M., Marshall, H.: Spatial impression due to early lateral reflections in concert halls: the derivation of a physical measure. Journal of Sound and Vibration 77(2), 211–232 (1981) 9. Schroeder, M.R., Logan, B.F.: Colorless artificial reverberation. Journal of the Audio Engineering Society 9(3), 192–197 (1961)

Quality-Aware Loss-Robust Scalable Speech Streaming Based on Speech Quality Estimation Jin Ah Kang1, Seung Ho Choi2, and Hong Kook Kim1 1

School of Information and Communications Gwangju Institute of Science and Technology (GIST), Gwangju 500-712, Korea {jinari,hongkook}@gist.ac.kr 2 Department of Electronic and Information Engineering Seoul National University of Science and Technology, Seoul 139-743, Korea [email protected]

Abstract. This paper proposes a quality-aware loss-robust scalable speech streaming (QLSSS) method to improve the perceived speech quality (PSQ) of a scalable wideband speech streaming (SWSS) system over IP networks. To this end, the proposed method estimates the PSQ and the packet loss rate (PLR) from the received speech data. Subsequently, it decides the amount of redundant speech data (RSD) that a speech decoder can use to reconstruct lost speech signals for high PLRs. According to this decision, the proposed method optimizes a scalable speech coding mode for current speech data (CSD) and RSD bitstreams in order to prevent speech quality from being degraded under the estimated packet loss condition and maintain the transmission bandwidth. The effectiveness of the proposed method is then demonstrated using the ITU-T Recommendations G.729.1 and P.563 as a scalable wideband speech codec and a PSQ estimator, respectively. It is shown from the experiments that an SWSS system employing the proposed QLSSS method significantly improves speech quality under packet loss conditions. Keywords: Scalable wideband speech streaming, packet loss, perceived speech quality, redundant speech transmission, ITU-T G.729.1, ITU-T P.563.

1

Introduction

Due to the rapid development of Internet protocol (IP) networks over the past few decades, audio and video streaming services are increasingly available via the Internet. Moreover, as these services are extended to wireless networks, the quality of service (QoS) of audio and video streaming becomes even more critical. In particular, speech streaming systems require a minimum level of speech communication quality. Speech quality is largely related to network conditions, such as packet loss or endto-end packet delays [1]. However, when speech streaming is performed over user datagram protocol/IP (UDP/IP) networks, packets may be lost or arrive too late for playback due to inevitable delays. In this case, even though packet losses occur frequently in wireless networks due to bandwidth fluctuations, a typical speech streaming system can only tolerate a few packet losses for real-time services [2][3]. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 132–142, 2011. © Springer-Verlag Berlin Heidelberg 2011

QLSSS Based on Speech Quality Estimation

133

Several packet loss recovery methods have been proposed for speech streaming systems implemented over the Internet and wireless networks. For instance, the techniques proposed in [4] and [5] were sender-based packet loss recovery methods using forward error correction (FEC). With regards to wireless networks, the techniques proposed in [6] and [7] were based on unequal error protection (UEP) methods. In addition, the modified discrete cosine transform (MDCT) coefficients of audio signals have been used as redundant data in order to assist an audio decoder to reconstruct lost audio signals [8]. However, these methods do not take into account time-varying network conditions, i.e., the packet loss rate (PLR) conditions. Thus, in order to recover lost packets based on conventional FEC methods, the redundant data should be transmitted constantly, even if the network conditions claim no packet losses. Therefore, an efficient scalable speech streaming method is needed to improve speech quality that can be degraded due to time-varying PLR. To this end, this paper proposes a quality-aware loss-robust scalable speech streaming (QLSSS) method that transmits redundant speech data (RSD) adaptively according to an estimation of perceived speech quality (PSQ). The PSQ estimation is performed for the received speech data by using a single-ended speech quality assessment. Simultaneously, the PLR of the network is estimated using a moving average method. In addition, a realtime transport protocol (RTP) payload format is suggested as a means of supporting the proposed QLSSS method. In other words, a speech packet combines the bitstreams of the current speech data (CSD) and the RSD when the PSQ is assumed to be low due to an increased PLR. Thus, even if a speech packet is lost, the speech decoder can reconstruct the speech signal corresponding to the lost packet by using the RSD bitstreams from the previous packet. In contrast, when the PSQ is assumed to be high due to a decreased PLR, a speech packet is organized using only the CSD bitstreams that are encoded by a higher bitrate. The effectiveness of the proposed QLSSS method is finally demonstrated using ITU-T Recommendations G.729.1 [9] and P.563 [10] as a scalable speech codec and a PSQ estimator, respectively. The remainder of this paper is organized as follows. Section 2 presents the structure of a scalable wideband speech streaming (SWSS) system based on the proposed QLSSS method with an RTP payload format. Next, Section 3 describes the detailed procedure of the proposed QLSSS method at the sender side as well as the receiver side, and then the performance of the proposed QLSSS method is discussed in Section 4. Finally, Section 5 concludes this paper.

2

Scalable Wideband Speech Streaming (SWSS) System

2.1

Structure of SWSS System Based on the Proposed QLSSS Method

An SWSS system, which was developed to offer better audio quality for integrated services over broadband internet connections [11], improves speech quality from a narrowband packet public switched telephone network (PSTN) quality (300–3400 Hz) to wideband quality (50–7000 Hz) [9]. In the SWSS system, a continuous speech signal is sampled into discontinuous speech frames that are encoded into bitstreams

134

J.A. Kang, S.H. Choi, and H.K. Kim

using a wideband speech codec. Then, it transmits the bitstreams using a real-time streaming protocol after packetizing. Meanwhile, in an opposite side of the SWSS system, the arriving packets are unpacketized into bitstreams that are further decoded to speech frames. Finally, these speech frames are sent to an output device.

Subsystem A Sender Side Input Speech

Scalable Wideband Speech Encoding

RTP Payload Formatting

RTP Packetizing

Output Packet

Receiver Side

Output Speech

Packet Loss Rate Estimation

Perceptual Speech Quality Estimation

Scalable Wideband Speech Decoding

Speech Frame Buffering

RTP Payload Analysis

RTP Unpacketizing

Input Packet

Network Subsystem B Sender Side Output Packet

RTP Packetizing

RTP Payload Formatting

Scalable Wideband Speech Encoding

Input Speech

Receiver Side

Input Packet

RTP Unpacketizing

RTP Payload Analysis

Scalable Wideband Speech Decoding

Speech Frame Buffering

Perceptual Speech Quality Estimation

Packet Loss Rate Estimation

Output Speech

Fig. 1. Packet flow for a scalable wideband speech streaming system employing the proposed QLSSS method, where Subsystems A and B represent the two communication parties

Fig. 1 shows a packet flow for the SWSS system implemented in this paper, where Subsystems A and B represent both parties of the speech stream communication employing the proposed QLSSS method. First of all, the sender side of Subsystem A performs scalable wideband speech encoding for the input speech frame. Next, the sender side generates a packet according to an RTP payload format. Here, the packet includes the CSD bitstreams, including feedback information related to the RSD transmission mode for the opposite SWSS system. Note that the RSD bitstreams are adaptively incorporated in this payload according to the RSD transmission mode received from Subsystem B. Finally, the formatted RTP packet is transmitted.

QLSSS Based on Speech Quality Estimation

135

Meanwhile, as the RTP packet arrives at the receiver side of Subsystem B, the receiver side analyzes the received packet according to the RTP payload format and then extracts the CSD bitstreams and the feedback information. If the RTP payload format includes the RSD bitstreams, the RSD bitstreams are stored to be used for recovering future lost packets. Next, the extracted CSD bitstreams are decoded using the scalable wideband speech decoder and the decoded speech frames are stored in a speech buffer for the PSQ estimation. Finally, the RSD transmission mode for the opposite SWSS system is decided by the estimated PSQ and PLR, and then it is included in the RTP packet that is sent back to Subsystem A. 2.2

RTP Payload Format

As mentioned in Section 2.1, the proposed QLSSS method should have an indicator for a scalable bitrate of wideband speech coding. Moreover, in order to deliver the feedback information from Subsystem A to Subsystem B, or vice versa, there should be a reserved field to accommodate the transmission of the RSD bitstream and feedback information. As a result, we selected the RTP payload format defined in IETF RFC 4749 for the G.729.1 scalable wideband speech codec [12], which is shown in Fig. 2. 0

1

MBS

2

3

4 byte

FT Zero or More Speech Frames at the Same Bitrate

Fig. 2. RTP payload format for a scalable wideband speech codec, ITU-T Recommendation G.729.1, defined in RFC 4749

In the payload format, the ‘MBS|FT’ sequence contains the payload header. The four-bit MBS field is used to tell the opposite SWSS system the maximum bitrate to be received. That is, the speech encoder should not exceed the encoding bitrate indicated by the received MBS, which is assigned a value from 0 to 11, corresponding to an encoding bitrate of 8, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30, and 32 kbit/s, respectively. Note that the speech encoder can change its bitrate at any time, as long as it does not exceed the received MBS. In addition, the FT field, consisting of four bits, indicates the actual encoding bitrate of the contained bitstreams. Thus, this field is also assigned a value from 0 to 11, corresponding to one of the encoding bitrates between 8 and 32 kbit/s, respectively. Note that the number 15 indicates the condition that there is no data to be transmitted, and that the numbers 12 to 14 are reserved for future use. Consequently, in order to realize the proposed QLSSS method with this payload format, we incorporate two new frame indices into the FT field, corresponding to the RSD bitstream and the feedback information. The indices are denoted using the numbers 12 and 13, respectively.

136

J.A. Kang, S.H. Choi, and H.K. Kim

The use of the RTP payload format described above has several advantages. First, the control feature such as the MBS field is retained by using the RTP payload format for the speech codec employed in the implemented SWSS system. Next, the overhead of the control fields for each RSD bitstream only requires four bits of the FT field. Finally, no additional transport protocol for the RSD transmission mode is needed since this feedback is conducted using the RTP packet that delivers the speech bitstream. Consequently, the transmission overhead for the RSD transmission mode is significantly reduced when compared to existing transport protocols designed for feedback, such as the RTP control protocol (RTCP) [13].

3

Proposed Quality-Aware Loss-Robust Scalable Speech Streaming Method

3.1

Adaptive Packet Loss Recovery and RSD Transmission Mode Decision

Fig. 3 shows the procedure for adaptive packet loss recovery at the receiver side of an SWSS system according to the proposed QLSSS method. First, packet loss occurrence is verified by RTP header analysis for a received speech packet, PKT (n ). If there is no packet loss, the CSD bitstreams, CSD ′(n), or the RSD bistreams,

RSD′(n + 1),", RSD′( n + k ), are extracted from the RTP payload analysis. Here, the extracted RSD bistreams are stored in a buffer for future use. In contrast, if there is a packet loss, the lost speech signals are recovered by using the RSD bitstreams, RSD′(n ), or packet loss concealment (PLC) algorithm provided by the speech decoder, depending on the availability of the RSD bitstreams. Finally, the speech frame data, CSD′′(n ), is generated by the speech decoder. The PSQ and PLR are estimated using speech data once the number of speech frames is sufficient to estimate a PSQ score. To this end, as shown in Fig. 4, each

PKT(n)

RTP Header Analysis

Packet Loss True

False RTP Payload Analysis

CSD′(n)

Scalable Decoding

RSD′(n +1) ... RSD′(n + k )

CSD′′(n)

RSD′(n)

Bitstream Buffer

RSD′(n)

exist

no exist Packet Loss Concealment

Fig. 3. Procedure for the adaptive packet loss recovery at the receiver side

QLSSS Based on Speech Quality Estimation

137

decoded or reconstructed frame data, CSD ′′(n ), is stored in a speech buffer when the speech data of N frames are constructed by overlapping with adjacent P frames. ~ In other words, the PSQ, Q ( k ), is estimated after ( N − P ) frames are newly re~ ceived from the opposite SWSS system. In addition, the PLR, L (k ), is estimated from a moving average of the previous PLR, L( k − 1), and the average PLR, L ( k − 1), as ~ L (k ) = (1 − α ) L (k − 1) + α L (k − 1).

CSD′′(n)

CSD′′(k − 2 N − P − 1)

…

Speech Buffer

CSD′′(k + N − P)

(1)

~ Perceptual Q(k ) Speech Quality Estimation

Packet Loss Rate Estimation

RSD Degree ~ L (k ) Decision

k

Fig. 4. Procedure for the estimation of perceived speech quality and packet loss rate

Finally, by comparing the estimated PSQ and PLR with each threshold, the RSD transmission mode for the opposite SWSS system is decided in terms of the degree of RSD, k , according to the following equation of ~ 0, if Q (k ) ≥θ Q1 and L~ (k ) ≤ θ L1  ~ ~ k = 2, if Q (k ) ≤ θ Q 2 and L (k ) ≥ θ L 2 . 1, otherwise 

(2)

~

where θ Q1 and θQ 2 are the thresholds for Q ( k ) , and θ L1 and θ L 2 are the thre~ sholds for L ( k ). 3.2

Scalable Speech Coding and Adaptive RSD Transmission

Fig. 5 shows the procedure for the scalable speech coding and the adaptive RSD transmission at the sender side for the proposed QLSSS method. First, the sender side verifies the degree of RSD, k , which is delivered by the opposite SWSS system. Then, it encodes the CSD and RSD using a scalable speech encoder according to the verification result. In other words, if k is zero, the CSD bitstreams are encoded at the highest bitrate with no RSD bitstreams (see Fig. 6). Otherwise, the bitrate of the CSD is set to a smaller bitrate than the highest bitrate, and the remaining bitrate is assigned to the RSD transmission. Thus, both the CSD and RSD bitstreams are encoded. Finally, after the RTP payload format described in Section 2.2 is configured according to this adaptive RSD transmission, the RTP packets are transmitted to the opposite SWSS system.

138

J.A. Kang, S.H. Choi, and H.K. Kim

CSD(n)

k =0

Highest Bitrate Encoding

CSD(n)

True

CSD′(n) RTP Payload

Formatting

PKT (n)

False CSD(n) Bitrate Adjustment RSD(n + 1)

Scalable Encoding

CSD′(n) RSD′(n + 1)

…

…

RSD′(n + k )

RSD(n + k )

Fig. 5. Procedure for scalable speech coding and the adaptive RSD transmission at the sender side RC 0 k =0

k =1

CSD′(n)

R C1

RR 0

CSD′(n)

RSD′(n + 1)

RC 2 k =2

CSD′(n)

R R1

RR 2

RSD′( n + 1) RSD′( n + 2)

Fig. 6. Bitrate assignment according to the degree of RSD

As described above, the proposed QLSSS method offers several advantages. First, the adaptive operation of the packet loss recovery according to the network conditions is effective since burst packet losses generally occur when the network is congested due to a sudden increase in the amount of data over the network [5]. Second, when compared to a conventional method that transmits the RSD bitstream for each speech packet by using additional network overhead, the proposed QLSSS method generates the RSD bitstream without increasing the transmission bandwidth. Third, in order to estimate the network conditions, the proposed QLSSS method estimates the PSQ. This is done because the PSQ measured as a mean opinion score (MOS) can be considered to be a clearer indicator of the speech quality than other parameters in the SWSS system.

4

Performance Evaluation

4.1

Experimental Setup

In order to demonstrate the effectiveness of the proposed QLSSS method, an SWSS system was first implemented using the ITU-T Recommendations G.729.1 and P.563 as a scalable wideband speech codec and a PSQ estimator, respectively. In this work, the speech signals were sampled at 16 kHz, and then encoded using the G.729.1

QLSSS Based on Speech Quality Estimation

139

speech codec operating at 32 kbit/s. When the degree of the RSD changed, the bitrate of the CSD and RSD also changed as shown in Table 1. By considering the requirements of the ITU-T Recommendation P.563, N was set to 200 frames for the PSQ estimation, which corresponded to 4 s. Moreover, P was set to 150 frames. Therefore, the PSQ was estimated after 50 new frames were received. Table 1. Bitrate assignment according to the degree of RSD Degree of Redundant Speech Data k 0 1 2

Bitrate of Current Speech Data RC0 32 kbit/s 16 kbit/s 16 kbit/s

Bitrate of Redundant Speech Data RR0 RR1 16 kbit/s 8 kbit/s 8 kbit/s

For the PLR estimation, we evaluated the performance of the proposed QLSSS method with different values of α in Eq. (1) and then set α to 0.4. Using a similar process, we set the thresholds for the estimated PSQ and PLR in Eq. (2), θQ1, θ Q 2 , θ L1 , and θ L 2 , to 4.15 MOS, 3.75 MOS, 3%, and 4.5%, respectively. To compare the speech quality within the same transmission bandwidth, we implemented two conventional packet loss recovery methods: a fixed redundant speech transmission (RST) method and a PLC method. The fixed RST method encoded speech signals using the G.729.1 at 16 kbit/s with an RSD transmission of 16 kbit/s. In other words, the fixed RST method always transmitted the RSD bitstream for each speech packet, thus the lost speech signals were recovered using the received RSD bitstream. However, if the RSD bitstream was not received due to burst packet losses, the lost speech signals were then recovered using the PLC algorithm embedded in the G.729.1 decoder. On the other hand, the PLC method encoded speech signals using the G.729.1 encoder at 32 kbit/s without using the RSD transmission. However, the lost speech packets were only recovered using the PLC algorithm embedded in the G.729.1 decoder. For the following experiments, we prepared speech sentences from the NTT-AT speech database [14]. Each speech sentence was about 4 s long and sampled at a rate of 16 kHz. These speech utterances were filtered using a modified intermediate reference system (IRS) filter, followed by the automatic level adjustment [15]. As an evaluation method for the recovered speech quality, we used the perceptual evaluation of speech quality (PESQ) defined in ITU-T Recommendation P.862 [16]. In order to show the effectiveness of the proposed QLSSS method under different PLRs, including burst loss characteristics, we generated eight different PLR patterns of 1%, 3%, 5%, 7%, 9%, 11%, 13%, and 15% by using the Gilbert–Elliot channel model defined in the ITU-T Recommendation G.191 [15]. Here, the burstiness of the packet losses was set to 0.5, and the mean and maximum consecutive packet losses were measured at 1.5 and 4.0 packets, respectively.

140

4.2

J.A. Kang, S.H. Choi, and H.K. Kim

Performance Evaluation for the Proposed QLSSS Method

In order to demonstrate the effectiveness of the proposed QLSSS method, the speech quality of the SWSS system using the proposed QLSSS method was compared to those of the SWSS system using the fixed RST method and the PLC method. Table 2 compares the speech quality measured in MOS using PESQ for the different packet loss recovery methods under different PLRs ranging from 0% to 15%. As expected, if the PLR was 0 (in other words, if there were no packet losses), the speech quality of the proposed QLSSS method was identical to that of the PLC method, but it was better than that of the fixed RST method, as shown in the first column of the table. In addition, it was shown from the table that as the PLR increased, the speech quality obtained by applying the proposed QLSSS method was better than those of other methods. As a result, the proposed QLSSS method yielded an average speech quality of 3.70 MOS, which was higher by as much as 0.25 and 0.04 MOS than the PLC method and the fixed RST method, respectively. Table 2. Speech quality measured in MOS using PESQ for the different packet loss recovery methods under PLRs ranging from 0% to 15% PLR (%) Method PLC method

5

0

1

3

5

7

9

11

13

15

Avg.

4.16 3.93 3.77 3.53 3.36 3.24 3.13 3.03 2.87

3.45

Fixed RST method

3.94 3.92 3.82 3.72 3.66 3.54 3.59 3.44 3.32

3.66

Proposed QLSSS method

4.16 3.90 3.82 3.73 3.67 3.60 3.58 3.49 3.38

3.70

Conclusion

In this paper, we proposed a quality-aware loss-robust scalable speech streaming (QLSSS) method that guaranteed speech quality without increasing transmission bandwidth. To this end, the proposed QLSSS method was designed to transmit redundant speech data (RSD) according to the estimated results of the perceived speech quality (PSQ) and the packet loss rate (PLR). Here, we used a single-ended speech quality assessment and a moving average method to estimate the PSQ and PLR, respectively. The proposed QLSSS method was applied to the receiver and sender sides of a scalable wideband speech streaming (SWSS) system. The receiver side of the SWSS system first decided the degree of RSD based on the estimated PSQ and PLR and then it sent feedback information to the opposite SWSS system about the decision result, via real-time transport protocol (RTP) packets for speech bitstreams. On the other side, the sender side of the SWSS system controlled the degree of RSD according to the received feedback and subsequently optimized the speech coding bitrate in order to maintain the equivalent transmission bandwidth while accommodating the RSD bitstreams. Finally, we evaluated the quality of the speech recovered by the proposed QLSSS method under different PLRs and compared it with that of the

QLSSS Based on Speech Quality Estimation

141

conventional redundant data transmission (RDT) method. Results indicate that the proposed QLSSS method improved the speech quality from 3.66 to 3.70 MOS when compared to the conventional method for PLRs ranging 0% to 15%. Consequently, the proposed QLSSS method could be efficiently applied to SWSS systems in order to improve the speech quality that has degraded due to packet losses. Acknowledgments. This work was supported in part by the “Fusion-Tech Developments for THz Information & Communications” Program of the Gwangju Institute of Science and Technology (GIST) in 2011 and by the MKE (The Ministry of Knowledge Economy), Korea, under the ITRC (Information Technology Research Center) support program supervised by the NIPA (National IT Industry Promotion Agency) (NIPA-2011-C1090-1121-0007).

References 1. Wu, C.-F., Lee, C.-L., Chang, W.-W.: Perceptual-based playout mechanisms for multistream voice over IP networks. In: Proceedings of Annual Conference of the International Speech Communication Association (Interspeech), Antwerp, Belgium, pp. 1673–1676 (2007) 2. Zhang, Q., Wang, G., Xiong, Z., Zhou, J., Zhu, W.: Error robust scalable audio streaming over wireless IP networks. IEEE Transactions on Multimedia 6(6), 897–909 (2004) 3. Park, N.I., Kim, H.K., Jung, M.A., Lee, S.R., Choi, S.H.: A packet loss concealment algorithm robust to burst packet loss using multiple codebooks and comfort noise for CELPtype speech coders. CCIS, vol. 120, pp. 138–147 (2010) 4. Bolot, J.-C., Fosse-Parisis, S., Towsley, D.: Adaptive FEC-based error control for Internet telephony. In: Proceedings of IEEE International Conference on Computer Communications (INFOCOM), New York, NY, pp. 1453–1460 (1999) 5. Jiang, W., Schulzrinne, H.: Comparison and optimization of packet loss repair methods on VoIP perceived quality under bursty loss. In: Proceedings of 12th International Workshop on Network and Operating Systems Support for Digital Audio and Video (NOSSDAV), Miami, FL, pp. 73–81 (2002) 6. Yung, C., Fu, H., Tsui, C., Cheng, R.S., George, D.: Unequal error protection for wireless transmission of MPEG audio. In: Proceedings of IEEE International Symposium on Circuits and Systems (ISCAS), Orlando, FL, vol. 6, pp. 342–345 (1999) 7. Hagenauer, J., Stockhammer, T.: Channel coding and transmission aspects for wireless multimedia. Proceedings of the IEEE 87(10), 1764–1777 (1999) 8. Ito, A., Konno, K., Makino, S.: Packet loss concealment for MDCT-based audio codec using correlation-based side information. International Journal of Innovative Computing, Information and Control 6(3B), 1347–1361 (2010) 9. ITU-T Recommendation G.729.1: An 8-32 kbit/s Scalable Wideband Coder Bitstream Interoperable with G.729 (2006) 10. ITU-T Recommendation P.563: 563: Single-Ended Method for Objective Audio Quality Assessment in Narrow-Band Telephony Applications (2004) 11. Bessette, B., Salami, R., Lefebvre, R., Jelinek, M., Rotola-Pukkila, J., Vainio, J., Mikkola, H., Jarvinen, K.: The adaptive multirate wideband speech codec (AMR-WB). IEEE Transactions on Speech and Audio Processing 10(8), 620–636 (2002)

142 12. 13. 14. 15.

J.A. Kang, S.H. Choi, and H.K. Kim

IETF RFC 4749.: RTP Payload Format for the G.729.1 Audio Codec (2006) IETF RFC 1889.: RTP: A Transport Protocol for Real-Time Applications (1996) NTT-AT.: Multi-Lingual Speech Database for Telephonometry (1994) ITU-T Recommendation G.191: Software Tools for Speech and Audio Coding Standardization (1996) 16. ITU-T Recommendation P.862: Perceptual Evaluation of Speech Quality (PESQ), an Objective Method for End-to-End Speech Quality Assessment of Narrowband Telephone Networks and Speech Codecs (2001)

Artificial Bandwidth Extension of Narrowband Speech Signals for the Improvement of Perceptual Speech Communication Quality Nam In Park, Young Han Lee, and Hong Kook Kim School of Information and Communications Gwangju Institute of Science and Technology (GIST), Gwangju 500-712, Korea {naminpark,cpumaker,hongkook}@gist.ac.kr

Abstract. In this paper, an artificial bandwidth extension (ABE) algorithm from narrowband to wideband is proposed in order to improve the quality of narrowband speech. The proposed ABE algorithm is based on spectral band replication in the modified discrete cosine transform (MDCT) domain with no additional bits. In particular, the patch index search for the replication band is restricted so that the harmonic structure of the wideband speech is maintained after ABE. In the proposed ABE algorithm, we first determine whether the current analysis frame of speech signals is voiced or unvoiced. A harmonic spectral replication or a correlation-based replication approach is then applied for the voiced or unvoiced frame, respectively. The proposed ABE algorithm is finally embedded into the G.729 speech decoder as a post-processor. It is shown from the subjective evaluation using a MUSHRA test that the mean opinion score of the wideband speech signals extended by the proposed ABE method is measured as 75.5, which is higher of around 14% than that of narrowband speech signals. Keywords: Bandwidth extension, artificial bandwidth extension, narrowband speech, wideband speech, perceptual quality improvement.

1

Introduction

Applications using wideband speech are slowly yet inevitably gaining over those using narrowband speech. Such a recent trend requires speech coding technology with an increased quality of decoded signals, instead of concentrating on the absolute compression efficiency [1-3]. In most speech communication systems, the speech bandwidth is limited to a range of 0.3–3.4 kHz. This speech bandwidth represents a good compromise between speech quality and transmission bandwidth for voiced sounds in general, but often a poor one for unvoiced sounds; this fact typically results in muffling on speech quality. In order to mitigate such a problem, wideband speech coding has been proven as an alternative [4]. Indeed, wideband speech, whose bandwidth ranges from 50 Hz to 7 kHz, spans all distinctive speech frequency components, thus it sounds clearer and gives a more natural conversation than that using narrowband speech. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 143–153, 2011. © Springer-Verlag Berlin Heidelberg 2011

144

N.I. Park, Y.H. Lee, and H.K. Kim

However, narrowband speech has been popularly serviced in many applications such as voice communications over a public switched telephone network (PSTN), voice over IP (VoIP), and voice applications in smart phones [5]. Therefore, simply replacing a narrowband speech codec with a wideband one in order to improve the quality of decoded speech is not a right solution. Instead, the extension of speech bandwidth from narrowband to wideband could be an alternative. There are two different kinds of approaches for extending the bandwidth according to whether or not the side information is available. It is usual to realize bandwidth extension by using the side information that is transmitted from the encoder. In other words, the encoder should generate auxiliary information based on the analysis of the high frequency component of the input signal [6-7]. The decoder recovers the high frequency signal from the low frequency signal and then uses the auxiliary information to adjust the generated high frequency signal. For example, the G.729.1 speech coder provides embedded coding with 12 different bit rates between 8 and 32 kbit/s [7]. The baseline coder of G.729.1 is fully compatible with G.729, thereby ensuring narrowband speech quality in 8 kbit/s mode. From the 14 kbit/s mode, whose operation mode is called ‘layer 2,’ a wideband signal can be synthesized using a BWE technique. By allocating additional bits for the BWE, the high band signal can be reconstructed in the decoder. However, this BWE approach requires additional bits and also requires some modification of the encoding process of a speech coder. On the other hand, instead of using the side information, the other type of BWE, which is also called artificial bandwidth extension (ABE), can estimate the high band signal from the low band signal without any side information. The estimation can be done by using a pattern recognition algorithm such as hidden Markov models (HMMs) [8], Gaussian mixture models (GMMs) [9], and so on [10-12]. For example, an ABE method is based on a source filter model of speech production, according to the fact that speech consists of an excitation signal and a vocal tract filter [13]. The vocal track filter is usually modeled by a set of linear prediction (LP) coefficients. Based on statistical recovery, GMM is then applied in order to extend the envelope. Such a method is realized without any additional bits, while it requires a training process. Instead of a model-based approach for predicting the high band from the low band, another conventional method was the spectral band replication (SBR) based ABE. This method copies modified discrete cosine transform (MDCT) spectrum to generate the high-frequency signal, and then adjusts its tonality for improving the subjective quality [14]. However, this method can cause a mismatch of the harmonic component at the boundaries between low and high bands, since it simply copies the low band to construct the high band without considering harmonic structure of speech signal. In this paper, an ABE method based on harmonic spectral replication and correlation-based spectral replication is proposed, where the extension is performed in the modified discrete cosine transform (MDCT) domain. Since the voice signal consists of tones and harmonics, applying a correlation-based method can cause a mismatch of the harmonic characteristics. Hence, we first classify each analysis frame of speech signal into a voiced or an unvoiced frame using a spectral tilt parameter, and then apply the harmonic spectral replication or correlation-based spectral replication

ABE of Narrowband Speech Signals for the Improvement

Narrow band sl (n) Speech (8kHz)

145

Modified Discrete Cosine Transform (MDCT) Sl (k )

Artificial Bandwidth Extension ~ S h (k )

~ S l (k )

Inverse Modified Discrete ~sl (n) Cosine Transform (IMDCT) Inverse Modified Discrete Cosine Transform (IMDCT)

QMF Synthesis

~ sw (n) Wide band Speech (16kHz)

~ s h ( n)

Fig. 1. Block General structure of artificial bandwidth extension from narrowband speech to wideband speech, which is applied in the MDCT domain for the voiced or unvoiced frame, respectively. In other words, the harmonic spectral replication can maintain the harmonic characteristic between low and high band for voiced frames. The remainder of this paper is organized as follows. Following this introduction, Section 2 describes the proposed ABE algorithm, and Section 3 describes how to realize the proposed ABE algorithm on a CELP-type speech decoder. Section 4 then demonstrates the performance of the proposed ABE algorithm. Finally, this paper is concluded in Section 5.

2

Proposed Artificial Bandwidth Extension (ABE) Algorithm

Fig. 1 shows a general structure of ABE applied in the MDCT domain to extend bandwidth from narrowband to wideband. In the figure, the narrowband speech, s (n) , is segmented into a sequence of frames, where frame size is N. Then, each analysis frame is transformed into the frequency domain using a 2N-point MDCT, Sl (k ) . After that, an ABE algorithm is applied in order to obtain high band MDCT

coefficients, Sh (k ) . In this paper, the proposed ABE algorithm based on the combination of harmonic spectral replication and correlation-based spectral replication is applied, which will be explained in Section 2.1. In addition, in order to prevent MDCT coefficients in the boundary between narrowband and high band from being abruptly changed, the MDCT coefficients of narrowband and high band, Sl (k ) ~ ~ and Sh (k ) are modified into Sl (k ) and Sh (k ) . Next, the low band and high band signals, ~ sl (n) and ~ sh (n) , in the time domain are obtained by applying a 2N-point inverse MDCT (IMDCT), respectively. Finally, the bandwidth extended signal, ~ sw ( n ) , is obtained by the quadrature mirror filterbanks (QMF).

146

N.I. Park, Y.H. Lee, and H.K. Kim Low band MDCT Coefficients S l (k )

Energy Calculation in the Sub-band E (b )

Normalization of the Low band Coefficients S l (k )

Yes

T

No St > θSt

Harmonic Spectral Coefficients Replication

Correlation-based Spectral Coefficients Replication S l′( k )

S l′ ( k )

Stretch of Low Spectral Coefficients S h (k ) Spectral Energy Control ~ S h (k )

Wideband Signal Generation Sˆ w ( k )

Wideband MDCT Coefficients

Fig. 2. Block diagram of the proposed ABE algorithm

2.1

U/V-Dependent Spectral Band Replication

Fig. 2 shows a block diagram of the proposed ABE algorithm. The proposed ABE ~ ~ algorithm generates the low and high band MDCT coefficients, S l (k ) and S h (k ) . In this paper, the frame size, N, is set to 512. First, 2N-point MDCT coefficients of the low band, Sl (k ) , are grouped into 16 sub-bands. That is, each sub-band has 32 MDCT coefficients. Then, the energy of the b-th sub-band, E (b) , is defined as

E (b) =

k = 32⋅( b +1) −1



k = 32⋅b

Sl2 (k ) ,

b = 0,1,",15 .

(1)

Next, E (b) is used to normalize each MDCT coefficient belonging to the b-th subband, such as Sl ( k ) =

Sl ( k ) , E ( b)

32b ≤ k < 32(b + 1) and b = 0,1,",15

where Sl (k ) is the k-th normalized low band MDCT coefficient.

(2)

ABE of Narrowband Speech Signals for the Improvement

147

In order to classify each frame into a voice or an unvoiced frame, we define spectral tilt, St , by using the first reflection coefficient as N −1

St =

 s(n ) s( n − 1)

n =1 N −1

(3)

2  s (n )

n =1

where s(n ) is the narrowband speech signal at the n-th time instant as shown in Fig. 1. If St is greater than θ St , this frame is declared a voiced frame; otherwise, it is as an unvoiced frame. In this paper, θ St is set to 0.25 from the preliminary experiment. Next, in order to extract the pitch information for a voiced frame, a normalized autocorrelation function is calculated as N −1

R (τ ) =

 s( n ) s ( n − τ )

n =τ

N −1

.

(4)

 s (n) 2

n =τ

Using Eq. (4), the pitch is obtained by selecting τ at which R(τ ) is maximized. That is, T = arg max Pl ≤τ ≤ Ph R(τ ) , where Pl and Ph are set to 20 and 147, respectively [15]. Specifically, in order to generate a high band signal with the harmonic characteristics, the harmonic period in the MDCT domain is determined as Δv = 2N

T

.

(5)

The k-th harmonic MDCT coefficient, Sl′(k ) , is then expressed as

S l′( k ) = S l (k +

N N − Δ v − mod( N , Δ v ) ) , k = 0,1,", − 1 2 2

(6)

where S l (k ) is the normalized low band signal as described in Eq. (2), and

mod( x, y ) is the modulus operation defined as mod( x, y ) = x % y . Also x  is the largest integer less than or equal to x . On the other hand, in order to patch high band MDCT coefficients from low band MDCT coefficients for an unvoiced frame, the optimal position in Sl (k ) is determined by the equation of Δ uv = argmax [corr ( Sl ( k ), Sl ( k + m))] 0≤ m ≤ N / 4 −1

(7)

148

N.I. Park, Y.H. Lee, and H.K. Kim

where Δ uv is the optimum shift of the patch and corr( Sl ( k ), Sl ( k + m )) is the crosscorrelation between the low band MDCT coefficients, Sl (k ) , and the normalized low band MDCT coefficients, Sl (k ) . In other words, the cross-correlation can be represented as N / 4 −1

corr ( Sl (k ), Sl ( k + m)) =  Sl ( k + k =0

3 N ) Sl ( k + m), 4

m = 0,1, ",

N −1. 4

(8)

Finally, S ′(k ) that is the most correlated to Sl (k ) in a range of 3–4 kHz is expressed as

Sl′( k ) = Sl ( k +

1 N + Δ uv ), 4

k = 0,1,",

N − 1. 2

(9)

The MDCT coefficients obtained from Eqs. (6) or (9), Sl′(k ) , which have a 2 kHz bandwidth, are finally stretched to the extended MDCT coefficients with a 4 kHz bandwidth by the following equation of  S ′ ( k / 2), k = 0,2, " , N − 2 S h (k ) =  l k = 1,3, " , N − 1  0,

(10)

where Sh (k ) is the k-th normalized and extended MDCT coefficient. 2.2

Energy Control

In order to avoid an abrupt change in energy at the high band after patching MDCT coefficients from the low band, the amplitude of each MDCT coefficient for the high band should be adjusted. First of all, the refined energy for the b-th high band, Eh (b) , is defined as

α ⋅ E ( b + 7), if E ( b + 8) > α ⋅ E ( b + 7) Eh (b) =  , b = 0,1,",7 otherwise  E ( b + 8),

(11)

where E (b) is the energy in the b-th low band as defined in Eq. (1) and α is set to 1.1 in this paper. Next, in order to maintain the energy of boundary between the low and high band, the scale factor of the energy, β , is calculated as

β = E (15) E (0) h

(12)

ABE of Narrowband Speech Signals for the Improvement

149

Bitstream

CELP-type Speech Decoder sl (n) T , k1

Modified Discrete Cosine Transform (MDCT) Sl (k )

{S~ (0), S~ (1),", S~ ( N −1)} l

Proposed ABE Algorithm

l

{

~ sl (n)

l

}

~ ~ ~ Sh (0), Sh (1),", Sh ( N −1)

Inverse MDCT Inverse MDCT

~ sh (n)

QMF Synthesis

~ s w (n )

Output Speech (16kHz)

Fig. 3. Application of the proposed ABE algorithm as a post-processor of a CELP-type narrowband speech decoder

where E (15) is the energy of the last sub-band in the low band and Eh (0) is the energy of the first sub-band in the high band. By combining Eqs. (10) and (11), the energy of the b-th sub-band in the high band, Eˆ h (b) , is finally modified as Eˆ (b) = β ⋅ E (b), b = 0,1, " ,7 . (13) h

h

Similarly to Eq. (10), Eˆ h (b) is stretched as  Eˆ (b / 2), b = 0,2,",14 . E h (b ) =  h  E h (b − 1), b = 1,3,",15

(14)

~ Next, the amplitude of S h (k ) is adjusted as

~ Sh ( k ) = Sh ( k ) Eh ( b), b = k / 32 , k = 0,1, " , N − 1 .

(15)

~ Then, the wideband MDCT coefficients, S w (k ) , are constructed by integrating ~ ~ Sl (k ) and S h (k ) , such as ~ ~ ~ ~ ~ ~ ~ S w (k ) = [Sl (0), Sl (1),", Sl ( N − 1), Sh (0), Sh (1),", S h ( N − 1)] .

3

(16)

Application of the Proposed ABE Algorithm as a Post-processor of a CELP-Type Narrowband Speech Coder

The proposed ABE algorithm is applied to the reconstructed speech by the G.729 speech decoder [16] which is one of CELP-type narrowband speech coders. Fig. 3

150

N.I. Park, Y.H. Lee, and H.K. Kim

100 80 60 40 20 0 Reference

3.5 kHz Anchor

7 kHz Anchor

G.729

G.729.1 (Layer 2)

Proposed ABE

Fig. 4. MUSHRA test results

shows how the proposed ABE algorithm works as a post-processor of the decoder. For a given bitstream, the G.729 decoder reconstructs speech signals, sl (n) . Note here that G.729 encodes narrowband speech signals once every 10 msec and compresses them with a bit-rate of 8 kbit/s, thus the reconstructed speech signals of 10 msec long are processed by the proposed ABE algorithm. That is, the frame size, N, is equal to 80, and sl (n ) is transformed into the frequency components using a 160-point MDCT. In order to realize the proposed ABE algorithm, we require the spectral tilt parameter denoted in Eq. (3). Instead of directly computing the spectral tilt parameter from sl (n ) , the first reflection coefficient, k1 , which can be obtained from the spectral envelope parameters during G.729 decoding, is used for St in Eq. (3). In addition, when St is greater than θ St , a pitch period should be computed. Here, the pitch information obtained from the decoding is also used for T in Eq. (5). After applying the proposed ABE algorithm, the low band and high band signals, ~ sl ( n ) and ~ sh (n) , are obtained by a 160-point IMDCT, respectively. Finally, the bandwidth extended signal, ~ sw ( n ) , is synthesized by filtering the 64-QMF [7].

4

Performance Evaluation

In order to demonstrate the effectiveness of the proposed ABE algorithm, a multiple stimuli with hidden reference and anchor (MUSHRA) listening test [17] and a spectrum comparison were carried out. For the MUSHRA test, 6 speech sentences, comprised of the utterances of 3 males and 3 females, were taken from the sound quality assessment material (SQAM) [18]. Especially, since SQAM speech files were recorded with stereo at a sampling rate of 44.1 kHz, the right channel signals of each file were down-sampled from 44.1 kHz to two different versions such as 8 and 16 kHz. In other words, 8 kHz down-sampled

ABE of Narrowband Speech Signals for the Improvement (dB) -20

(dB) -20

-40

-40

-60

-60

-80

-80

-100 0

4

(a)

-100 8 0 (kHz)

(dB) -20

(dB) -20

-40

-40

-60

-60

-80 -100

151

4

(b)

8 (kHz)

-80 0

4

(c)

-100 8 0 (kHz)

4

(d)

8 (kHz)

Fig. 5. Comparison of the spectra obtained from (a) the original speech signals sampled at 16 kHz, (b) the decoded signals by G.729 decoder, (c) the decoded signals by G.729.1(Layer 2), and (d) the speech signal by the proposed ABE algorithm

signals were processed by G.729 and further by the proposed ABE, while 16 kHz down-sampled signals were processed for G.729.1(Layer 2). Next, two anchors with cut-off frequencies of 3.5 and 7 kHz were prepared for the MUSHRA test. Seven people with no auditory disease participated in this test. The listener was presented with several audio stimuli. The first was the reference, which was the original speech down-sampled at 16 kHz. The remainders were processed by G.729, G.729.1(Layer 2) and the proposed ABE method. Each listener gave a score between 0 and 100 depending upon their opinion of the quality. Fig. 4 shows the results of the MUSHRA test. As shown in the figure, the proposed method gave an average score of 75.5, which was higher than G.729 and lower than G.729.1(Layer 2), as we expected. This result indicated that the proposed ABE method enhanced the performance of G.729 without any additional bits. Moreover, compared to G.729, the quality improvement of 43% was achieved by the proposed ABE method. Finally, the spectra of speech signals processed by G.729, G.729.1(Layer 2), and the proposed method were compared, which is shown in Fig. 5. It was shown from the figure that the high band spectrum of the proposed ABE algorithm was quite similar to that of G.729.1(Layer 2).

5

Conclusion

In this paper, we proposed an artificial bandwidth extension (ABE) algorithm from narrowband to wideband to improve the quality of narrowband speech. The proposed ABE algorithm was based on the harmonic spectral replication and correlation-based spectral replication for voice and unvoiced frame, respectively. The proposed ABE

152

N.I. Park, Y.H. Lee, and H.K. Kim

algorithm was embedded into the G.729 speech decoder as a post-processor. It was shown from the subjective evaluation that the proposed ABE method achieved MUSHRA score improvement of 43% compared to G.729. Moreover, when comparing the spectra of speech signal, it was shown that the high band spectrum of speech signals processed by the proposed ABE algorithm was quite similar to that by the layer 2 mode of G.729.1 that was one of standardized wideband speech coders in ITU-T. Acknowledgments. This work was supported in part by the Practical R&D Program of the GIST Technology Initiative (GTI), Gwangju Institute of Science and Technology, Korea.

References 1. Mikko, T., Lasse, L., Anssi, R., Henri, T.: Scalable super-wideband extension for wideband coding. In: Proceedings of ICASSP, pp. 161–164 (2009) 2. Stephen, V.: Listener ratings of speech passbands. In: Proceedings of IEEE Workshop on Speech Coding, pp. 81–82 (1997) 3. Park, N.I., Kim, H.K., Jung, M.A., Lee, S.R., Choi, S.H.: Burst packet loss concealment using multiple Codebooks and comfort noise for CELP-type speech coders in wireless sensor networks. Sensors 11(5), 5323–5336 (2011) 4. ITU-T Recommendation G.830: Subjective Performance Assessment of Telephone-band and Wideband Digital Codec (1996) 5. Goode, B.: Voice over internet protocol (VoIP). Proceedings of the IEEE 90(9), 1495–1517 (2002) 6. Kosuke, T., Kei, K.: Low-complexity bandwidth extension in MDCT domain for lowbitrate speech coding. In: Proceedings of ICASSP, pp. 4145–4148 (2009) 7. Rogot, S., Kovesi, B., Trilling, R., Virette, D., Duc, N., Massaloux, D., Proust, S., Geiser, B., Gartner, M., Schandl, S., Taddei, H., Yang, G., Shlomot, E., Ehara, H., Yoshida, K., Vaillancourt, T., Salami, R., Lee, M.S., Kim, D.Y.: ITU-T G.729.1: an 8-32 kbit/s scalable coder interoperable with G.729 for wideband Telephony and voice over IP. In: Proceedings of ICASSP, pp. 529–532 (2007) 8. Jax, P., Vary, P.: On artificial bandwidth extension of telephone speech. Signal Processing 83, 1707–1719 (2003) 9. Song, G.-B., Martynovich, P.: A study of HMM-based bandwidth extension of speech signals. Signal Processing 89, 2036–2044 (2009) 10. Kornagel, U.: Techniques for artificial bandwidth extension of telephone speech. Signal Processing 86, 1296–1306 (2006) 11. Pulakka, H., Laaksonen, L., Vainio, M., Pohjalainen, J., Alku, P.: Evaluation of an artificial speech bandwidth extension method in three languages. IEEE Transactions on Audio, Speech, and Language Processing 16, 1124–1137 (2008) 12. Kim, K., Lee, M., Kang, H.: Speech bandwidth extension using temporal envelope modeling. IEEE Signal Processing Letters 15, 429–432 (2008) 13. Murali, M.D., Karpur, D.B., Narayan, M., Kishore, J.: Artificial bandwidth extension of narrowband speech using Gaussian mixture model. In: Proceedings of ICASSP, pp. 410–412 (2011)

ABE of Narrowband Speech Signals for the Improvement

153

14. Tsujino, K., Kikuiri, K.: Low-complexity bandwidth extension in MDCT domain for lowbitrate speech coding. In: Proceedings of ICASSP, pp. 4145–4148 (2009) 15. Kondoz, A.M.: Digital Speech: Coding for Low Bit Rate Communication Systems, 2nd edn. Weiley (2004) 16. ITU-T Recommendation G.729: Coding of Speech at 8 kbit/s Using Conjugate-Structure Code-Excited Linear Prediction, CS-ACELP (1996) 17. ITU/ITU-R BS 1534: Method for Subjective Assessment of Intermediate Quality Level of Coding Systems (2001) 18. EBU.: Sound Quality Assessment Material Recording for Subjective Tests (1988)

Improvements in Howling Margin Using Phase Dispersion Jae-Won Lee1 and Seung Ho Choi2 1

Graduate School of NID Fusion Technology Seoul National University of Science and Technology, Seoul 139-743, Korea [email protected] 2 Department of Electronic and Information Engineering Seoul National University of Science and Technology, Seoul 139-743, Korea [email protected]

Abstract. The limitation in the gain of audio system is mainly due to the howling generated by an acoustic feedback circuit. Furthermore, the howling occurs differently depending on the acoustic environments. In this paper, we propose a phase dispersion method to improve the howling margin in the audio amplifier systems. In order to eliminate unexpected potential howling frequencies, the proposed method controls the phase around the howling frequency using allpass filter with phase dispersion. From the experiments, it is shown that the howling margin is improved by around 3 dB. Keywords: Howling margin, phase dispersion, acoustic gain, acoustic feedback circuit.

1

Introduction

The audio amplifier system using microphone can generate howling due to acoustic feedback circuit (AFC) at some specific frequency as shown in Fig. 1 [1-3]. The AFC is unstable and can diverge by positive feedback [4]. Furthermore, the howling may vary depending on the overall environment such as the positions of microphones and loudspeakers, room shape and arrangement, the position and movement of talker, room temperature, etc [1]. Therefore, it is difficult to predict the howling [1]. The detailed explanations for the conditions of howling occurrence and the properties of potential howling frequency (PHF) can be found in [5] and [6], respectively. To increase acoustic gain, the audio amplifier system is generally used. However, the gain is bounded due to the unfavorable sound, howling. The acoustic gain can be partly improved by changing indoor conditions or electric devices. The gain control and frequency shift methods were developed to suppress howling [7]. However, the gain control is unsuitable because of reduced acoustic gain. Also, the frequency shifting can restrict applications. Recently, adaptive notch filters have been employed to suppress the howling [8-11]. However the filters inherently distort the original sound and can generate unexpected PHFs. Essentially, it is preferable to improve T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 154–161, 2011. © Springer-Verlag Berlin Heidelberg 2011

Improvements in Howling Margin Using Phase Dispersion

155

howling margin. This paper presents a novel method to improve the howling margin using phase dispersion without distorting magnitude spectrum. The remainder of this paper is organized as follows. Howling phenomenon and margin are described in Section 2. We present the proposed method for improving the howling margin using phase dispersion in Section 3. The experiments are illustrated in Section 4. Finally, conclusions are given in Section 5.

2

Howling Phenomena and Margin

2.1

Howling Phenomenon

In audio amplifier system as shown in Fig. 1(a) the microphone input signal x(t) is amplified by gain g and then the loudspeaker output signal y t is generated. The y t makes the multiple reflection signals { y t } [1-3]. And then, the { y t } are ∑ α y t τ that is again entered to the summed to the feedback signal y t microphone, where τ is the delay time of y t and α is the attenuation factor of the indoor wall [1, 12].

yd (t ) = y1(t)

x(t)

X(ω)

y0 (t)

Ls

Y(ω)

αe-jωτ

L1

y2 (t) yn (t)

(a)

(b) X(ω)

Ls

Y(ω) R(ω)

Ld

Yd (ω) = R(ω)Y(ω) (c)

(d)

Fig. 1. AFC model and sound level; (a) AFC formation in compliance with the indoor reflection, (b) AFC model in free space, (c) device position and sound level detection, and (d) AFC model in indoor condition

In free space, the direct sound y t becomes only feedback signal. The AFC in free space is shown in Fig. 1(b). The transfer function H ω with the phase and magnitude response is given as follows: H ω

Y ω X ω

g·

α

ωτ α

α ωτ

ωτ α

,

(1)

156

J.-W. Lee and S.H. Choi

|H ω | H ω

α

tan

ωτ α α ωτ α

.

(2)

ωτ

The transfer function changes periodically like a comb filter [1, 6]. The frequency at H ω 2πm becomes a PHF where m is an integer [6]. In other words, a howling is likely to occur at the PHF. In indoor conditions the transfer function can be represented by H ω

Y ω X ω

∑

αe

ωτ

.

(3)

However, it is difficult to determine the parameters since the transfer function is affected by the environment such as the indoor conditions, the positions of microphone and loudspeaker, and the directivity characteristics [1]. Also the howling frequency continuously varies according to the circumferential changes. Thus it is difficult to predict a howling occurrence in advance. 2.2

Acoustic Gain and Margin

In free space, the acoustic gain of audio amplifier system is defined by determined L and L into microphone (see Fig. 1(c)) [5, 12]. In the figure, L is sound pressure level (SPL) from source to microphone and L is SPL from loudspeaker to microphone directly. It can be amplified until L > L [5, 12, 13]. In indoor conditions, the maximum output level of audio amplifier system is lower than free space. The feedback signal Y ω is generated by the room impulse response R ω and the level increases from L to L as shown in Fig. 2. Moreover, a number of PHFs encounter with Y ω . The howling margin (dB) is defined by L L , which is 6 dB in general. Therefore the indoor audio system needs the minimum howling margin of 6 dB.

Fig. 2. Alteration of sound pressure level by the indoor reflection and generation of PHF

Improvements in Howling Margin Using Phase Dispersion

2.3

157

Howling Occurrence

The AFC can easily diverge due to a positive feedback circuit [4]. When a howling occurs, the signal magnitude increases continuously over time at a specific frequency as shown in Fig. 3(a) [12]. An example for the howling phenomenon on a real music signal is given in Fig. 3(b) [1]. y(t)

{

}

n −1

yn (t) = x ( t-nΔt ) * gs ( t ) * gf ( t ) +  ym (t) n

m =0

y1 (t) = x ( t-Δt ) * gs ( t ) * g f ( t ) + y0 (t) y0 (t) = x (t ) * g s (t ) time(s)

(a)

(b)

Fig. 3. Howling occurrence; (a) step response and (b) spectrogram

2.4

Notch Filter and All-Pass Filter

Adaptive notch filters suppress howling by reducing the gain in a specific frequency as shown in Fig. 4 [10]. It requires detecting accurate howling frequency [8-11] and can generate unexpected PHFs as shown in Fig. 4 [10].

0 ° phase

Fig. 4. Response of notch filter

An all-pass filter (APF) passes all frequencies equally, but changes the phase relationship between various frequencies. A simple all-pass filter can be implemented by H z

.

(4)

For example, as shown in Fig. 5, the phase is altered by 90° at the filter frequency without changing the magnitude spectrum. The howling can be suppressed if the phase is properly controlled using APF.

158

J.-W. Lee and S.H. Choi

x(n )

y(n )

z -1

(a)

(b)

Fig. 5. First order all-pass filter; (a) direct form and (b) magnitude and phase response

3

The Proposed Howling Margin Improvement Method

In most cases, the frequency with a peak signal is same as PHF, and a maximum peak signal may firstly generate a howling as shown in Fig. 2. If the phase of the peak signal is changed and the magnitude of the signal is reduced, the howling can be suppressed, which results in howling margin improvement. Considering the phase condition alone, the howling can be predicted by the relationship between distance d and wavelength λ as shown in Fig. 6(a) [14]. If a howling occurs at n = 2 and the APF is used as shown in Fig. 6(b), the howling can be suppressed. However, the howling can generate other howling frequencies due to the phase shift. Fig. 7(a) shows a howling at 1 kHz by 30 dB with phase 0°. The howling is decreased by 6 dB after using APF as shown Fig. 7(b), but the new pair of PFHs are generated around 1 kHz. If the phase response of APF is randomly altered using the phase disperser before the signal level increases, the howling margin can be improved. λn = d / n

λn = d / n

(a)

(b)

Fig. 6. Howling suppression using phase disperser; (a) relation between reflection distance and wavelength and (b) phase dispersion using APF

Improvements in Howling Margin Using Phase Dispersion

(a)

159

(b)

Fig. 7. PHF occurrence examples; (a) PHF occurrence at 1 kHz and (b) new PHFs occurrence

4

Experiments

Fig. 8 shows the configuration for the experiments, which employed the electric reverberator to generate the indoor sound. The signal block provided a music sound or a pink noise. The feedback circuit and system gain were implemented with digital signal processor (DSP). The phase disperser was employed using APF, which were controlled by external controller. The hopping control parameter [Hz] altered pseudo random numbers with the bandwidth of 100 Hz around APF frequency [Hz], and the hopping interval time was 10 msec. At first, the gain g was aligned according to the maximum acoustic margin. And then, the gain g was increased step by step depending on howling. If another howling occurred, another APF could be used.

Fig. 8. Configuration for the experiments

Fig. 9(a) shows the experimental results with the reverberation time of 0.7 sec with the howling margin of 6 dB. The howling sequentially occurred at 3 points, whose gains were increased by around 3 dB as shown in Fig. 9(b). Each howling was suppressed using the phase disperser as shown in Fig. 9(c). Similar results were obtained with the reverberation times of 1.7 sec and 2.7 sec as shown in Fig. 10 and Fig. 11. From the experiments, we found that the howling margin could be improved by around 3 dB.

160

J.-W. Lee and S.H. Choi

(a)

(b)

(c)

Fig. 9. Experimental results for phase dispersion (reverberation time: 0.7 sec); (a) stable state (0 dB), (b) howling occurrence (+3 dB), and (c) howling suppression using phase disperser

(a)

(b)

(c)

Fig. 10. Experimental results for phase dispersion (reverberation time: 1.7 sec); (a) stable state (0 dB), (b) howling occurrence (+3 dB), and (c) howling suppression using phase disperser

(a)

(b)

(c)

Fig. 11. Experimental results for phase dispersion (reverberation time: 2.7 sec); (a) stable state (0 dB), (b) howling occurrence (+3 dB), and (c) howling suppression using phase disperser

5

Conclusion

In this paper, we proposed a phase dispersion method to improve the howling margin in the audio amplification systems. The proposed method controlled the phase around the howling frequency using all-pass filter with phase dispersion in order to eliminate unexpected potential howling frequencies. The experimental results showed that the howling margin could be improved by around 3 dB.

Improvements in Howling Margin Using Phase Dispersion

161

References 1. Lee, J.-W., Choi, S.H.: Detection of howling frequency using temporal variations in power spectrum. Communications in Computer Science 151, 46–51 (2011) 2. Antman, H.S.: Extension to the theory of howlback in reverberant rooms. Journal of the Acoustical Society of America (Letters to the Editor) 39(2), 399 (1996) 3. Waterhouse, R.: Theory of howl-back in reverberant rooms. Journal of the Acoustical Society of America (letters to the Editor) 37, 921 (1965) 4. Nyquist, H.: Regeneration theory. Bell Syst. Tech. J. 11, 126–147 (1932) 5. Lindberg, E.: The Barkhausen criterion (observation). In: Proceedings of the 18th IEEE Workshop on Nonlinear Dynamics of Electronics System (NDES 2010), Dresden, Germany, pp. 15–18 (2010) 6. Troxel, D.: Understanding Acoustic Feedback & Supressors. RainNote, Rain Co (2005) 7. Schroeder, M.R.: Improvement of acoustic feedback stability by frequency shifting. Journal of the Acoustical Society of America 36(9), 1718–1724 (1932) 8. Loetwassana, W., Punchalard, R., Lorsawatsiri, A., Koseeyaporn, J.: Adaptive howling suppressor in audio amplifier system. In: Proceedings of Asia-Pacific Conference on Communications, pp. 445–448 (2007) 9. Gil-Cacho, P.: Regularized adaptive notch filters for acoustic howling suppression. In: Proceedings of EUSIPCO, pp. 2574–2578 (2009) 10. Choi, T.-H., Park, B.-U., Kim, H.-Y.: Quasi-linear phase adaptive notch filter for howling suppression. In: Proceedings of Spring Meeting of the Acoustical Society of Korea, vol. 20(2), pp. 245–248 (2001) 11. Nehorai, A.: A minimal parameter adaptive notch filter with constrained poles and zeros. IEEE Transactions on Acoustics, Speech, and Signal Processing 33, 983–996 (1995) 12. Davis, D., Davis, C.: Sound System Engineering, 2nd edn., pp. 424–426. Focal Press (1997) 13. Boner, C.P., Boner, R.E.: The gain of sound system. Journal of the Audio Engineering Society 17(2), 147–150 (1969) 14. Sogami, A., Kawamura, A., Iiguni, Y.: A distance-based howling canceller with adaptive bandwidth. In: Proceedings of the 23rd ITC-CSCC, pp. 345–348 (2008)

Secure Client-Side Digital Watermarking Using Optimal Key Selection Jing-Jing Jiang and Chi-Man Pun Department of Computer and Information Science, University of Macau Macau SAR, China {ma96525,cmpun}@umac.mo

Abstract. In this paper, we proposed an adaptive secure client-side digital audio watermarking by choosing an optimal embedding key from numerous available embedding keys for the encryption and joint decryption and watermarking procedure of the specific content. In the proposed scheme, the embedding key will determine not only the client-specific watermark content to be embedded, but also the watermarking embedding position where the watermark will be placed. It is obvious that different region of the host signal featured different characteristics, so where to embed the watermark will impact the equilibrium between two basic requirements, known as inaudibility and robustness, of watermarking development. Consequently, how to choose the embedding key for the particular content to be distributed could be of crucial significance. Simulation results show that the watermark in this scheme can survive robustly after being subjected to various common audio attacks, such as filtering, resample and so on. Keywords: Client-Side Embedding, Key Selection, LUT, Digital Watermarking.

1

Introduction

In recent years, we are experiencing a clear trend toward the distribution of massscale multimedia content since the social network becoming more and more popular among the customers. Unlike the movies and music which were only available on physical media in the early 1990s, new electronic distribution channels, such as video-on-demand services and digital music downloads, push the threat of the copyright infringement to a more severe situation, which the record companies and movie studios attempt to struggle against by all means. There a many adopted solutions existing for the content protection which aimed at preventing or deterring copyrights violations from malevolent users. On the prevent side, all kinds of copy control[1] mechanisms are exploited to prevent users from making unauthorized copies and distributing the content without the permission of the content owners. These mechanisms include cryptographic tools and copy control watermarks. Different from the forensic tracking watermarks, the copy control watermarks serve as auxiliary mean that tell the compliant players a set of permitted or prohibited actions, such as “do no copy” or “one copy only”. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 162–168, 2011. © Springer-Verlag Berlin Heidelberg 2011

Secure Client-Side Digital Watermarking Using Optimal Key Selection

163

On the deterring side, the forensic tracking watermarks which link the content to a particular user or device receiving it are among the various means utilized to trace the user of redistributing if unauthorized content is found. Since the content distribution systems are based on server-client architecture traditionally, the most straightforward way to embed forensic tracking watermark to the digital content is to have a trusted server to embed the watermark and then sends the watermarked content to the corresponding client. Such method seems to have securely embedded the watermark into the content without disclosing the unwatermarked copies and information concerning the embedding procedure to the outside world. However, in this scenario, the computation burden regarding the embedding on the server largely relies on the number of users. As the number of the user is expect to be increasingly large, extremely high computation burden could rests on the server side. In addition, since the distribution of individually watermarked copies requires us to resort to a secured communication channels, bandwidth requirements could become prohibitive. A different architecture, namely network-nodes side, proposed by several authors embeds the watermark at intermediate nodes of a network. For instance, in [2] Crowcroft et al. propose exploiting intelligent network nodes that switch between two watermarked streams to encode a unique payload in time. However, this approach is still experiencing the limitation of the bandwidth. In addition, the selection process requires security and cooperation at intermediate network nodes, which may not be practical in today’s widely used internet which integrates all different types of network together. An alternative solution is named client-side watermarking which embed the watermark on the client side rather than on the server side. In this scenario, the server could send a unique copy of the content to all the interested users through broadcasting systems. Since the broadcasting system is utilized to distribute the unique content, the bottlenecks regarding computation burden and bandwidth requirement encountered in the both server-side and network-node side could be effectively eliminated in this case. However, this client-side scheme gives rise to a new problem. It is obvious that the clients are untrusty, so it is possible that the original content or the watermark to be embedded may be exposed to the malevolent clients. Consequently, a new approach, defined as secure watermark embedding, has been proposed to address such a problem. In the secure watermark embedding, a single encrypted version of the original copy is sent to all the interested clients, but client-specific keys allow decryption of the content and implicit embedding of a client-specific watermark at the same time, leading the specific client to get a uniquely watermarked version of the content. The ability of wiping off the bottleneck in the server-side and network nodes side of this approach makes this client-side embedding the prime candidate for forensic tracking in the mass-scale electronic content distribution systems, provided that the security at the untrusty client could be guaranteed. Security of the embedding is based on the design of the particular encryption and decryption methods. In[3], Emmanuel et al. encrypt each video frame by masking it with a noise sequence. Decryption the video frame using a unique decryption key

164

J.-J. Jiang and C.-M. Pun

including a watermark automatically superimposes the corresponding watermark on the content. Another approach, named stream switching[4], divides the content stream in to various segments and prepares multiple versions of each segment by embedding a different watermark and encryption with a different key. In the literature, various approaches for secure watermark embedding have been proposed. Here our attention is focused on the methods that could decrypt a ciphertext to slightly different plaintext given different decryption keys. In such way, the slightly difference between the original and decrypted content can be served as the forensic watermark which could be extracted as evidence when copyright violation occurred in the future. The first scheme of this approach has been proposed in[5], where a special stream cipher named Chameleon is exploited. While encrypting, a sequence of indexes is generated to select four entries from a look-up-table(LUT), called encryption LUT, comprising of random 64-bit words, for each element of the plaintext. These entries are XORed with the plaintext to create the ciphertext. The decryption process is almost the same as the encryption except using a different LUT, named decryption LUT, which is formed by inserting some bit errors in some entries of the encryption LUT. So, a unique watermark can be left on the original digital content after decryption procedure. Recently , Adelsbach [6]et al. and Celic et al.[7] have proposed generalizations of the Chameleon cipher. Especially, the first method applicable to multimedia content is [7], which allow us to use algebraic operations during encryption and decryption procedure. Piva et al. [8] has proposed a new lookup-table-based secure client-side embedding scheme which combine the security of client-side embedding with the robustness of informed embedding methods. In particular, we developed a LUT-based adaptive secure client-side audio watermarking scheme allowing us to exploit embedding key adaptively in the light of different characteristic of the original digital signal to be distributed. In order to better balance the two basic requirements of inaudibility and robustness in the watermarking designing, we choose the optimal embedding key for each digital signal to be distributed in our scheme.

2

Adaptive Secure Client-Side Digital Audio Watermarking

To move one step further from the schemes described in [7], we proposed a Adaptive Secure Client-Side Digital Audio Watermarking by choosing an optimal embedding key from numerous available embedding keys for the encryption and joint decryption and watermarking procedure of the specific content. In this scheme, the embedding key will determine not only the client-specific watermark content to be embedded, but also the watermarking embedding position where the watermark will be placed. It is obvious that different region of the host signal featured different characteristics, so where to embed the watermark will impact the equilibrium between two basic requirements, known as inaudibility and robustness, of watermarking development. Consequently, how to choose the embedding key for the particular content to be distributed could be of crucial significance. The proposed algorithm for encryption is shown in Fig. 1.

Secure Client-Side Digital Watermarking Using Optimal Key Selection

165

Fig. 1. The flowchart of the encryption procedure

The proposed algorithm for joint decryption and watermarking is shown in Fig. 2.

Fig. 2. The flowchart of joint decryption and watermarking procedure

166

J.-J. Jiang and C.-M. Pun

To appraise the inaudibility of the proposed audio watermark scheme, Peak Signal Noise Ratio (PSNR) has been used as the objective measurement of the embedded audio signal quality. Its definition is as follows:

Length * max A2 (i ) PSNR = 10 log10 ( Length −1 ) * 2  [ A(i ) − A (i )] i =0

*

where A , A , and Length represent the original audio signal, the watermarked signal, and the length of the audio signal respectively. If the PSNR value exceeds a certain threshold, the watermarked audio signal can be concluded to be inaudible; otherwise, the watermarked signal can be considered perceptible.

3

Experimental Results

In order to appraise the robustness and inaudibility of our audio watermarking scheme, several music audio signals of different cultural styles, each with a length of 10 seconds, have been used for audio watermarking experiments. All the audio signals used in the experiments are 16 bit signed mono audio signals sampled at 44.1kHz. Three level DWT has been applied to the audio signals; we do not use higher level DWT since a larger level DWT involves a higher computation cost and a smaller level DWT facilitate the proposed watermarking algorithm to yield more robust watermarked audio signals. In view of the embedding strength in this scheme is almost fixed as the variance of LUT is almost constant and the impact of the embedding position on the robustness and the inaudibility, we choose adaptively an optimal embedding key, which will determine where to embed the watermark, based on the characteristics of the audio signals processed by the proposed embedding algorithm. To demonstrate the robustness of the proposed watermarking scheme, many kinds of attacks of the audio signals have been employed, such as low-filter, resampling, and Mp3 compression attacks. In order to estimate the inaudibility of the embedded signals, we first compute the PSNRs for the watermarked audio signals obtained by applying the proposed watermark embedding algorithm to a number of host audio signals. Subsequently, different kinds of attacks are applied to the watermarked signals; we then attempt to retrieve the watermark information from the watermarked audio signals after the attacks. For each watermark retrieved, we calculate its Correlation Coefficients, an objective measurement of the correlation between the detected watermark and the original watermark, with the original watermark.

Secure Client-Side Digital Watermarking Using Optimal Key Selection

167

Table 1. The CC between original watermark and extracted watermark retrieved from audio signal pop.wav using different embedding keys after subjected to different attacks KEY

Attacks

Resample

Requantization

LowFilter

Mp3

Key1

0.5019

0.5029

0.4996

0.5018

Key2

0.6596

0.6592

0.6569

0.6574

Key3

0.7831

0.7808

0.7823

0.7823

Key4

0.5539

0.5534

0.5506

0.5534

Key5

0.6164

0.6169

0.6152

0.6154

Key6

0.6433

0.6434

0.6426

0.6435

Key7

0.6966

0.6965

0.6954

0.6955

Key8

0.5741

0.5722

0.5729

0.5744

Key9

0.4163

0.7154

0.7149

0.7161

Key10

0.7331

0.7327

0.7322

0.7315

Table 2. The CC between original watermark and extracted watermark retrieved from host audio signals using optimal keys after subjected to different attacks, and their PSNR values Audio Signals

Jazz

Classic

Rock

46.76

47.35

43.86

49.24

0.031

0.019

0.039

0.046

0.7831

0.6648

0.5018

0.6166

0.7808 0.7823 0.7823 0.7701

0.6652 0.6625 0.6615 0.6608

0.5016 0.5005 0.5016 0.4994

0.6147 0.6158 0.6149 0.6107

Attacks

Requantize LowFilter Mp3-96 Mp3-64

Correlation Coefficients

PSNR No Watermark Resample

4

Pop

Discussions and Conclusion

In this paper, an adaptive secure client-side audio watermarking method has been proposed. Many techniques have been applied in this scheme. First of all, a client-side watermarking was adopted in this scheme due to its efficiency and simplicity. In clientside watermarking, the watermark is embedded on the client side, and the server could distribute a unique copy through broadcasting system. It is no doubt that this method could effectively eliminate the bottlenecks regarding computation burden and bandwidth requirement encountered in the both server-side. Then, in order to guarantee the robustness against all kinds of audio attacks that may be probably applied to the audio signals, we chose an optimal embedding key for the particular audio signal so that better robustness ability can be achieved given an acceptable level of inaudibility.In addition, to rule out the human factors that may impact the appraisement

168

J.-J. Jiang and C.-M. Pun

of the inaudibility and robustness of the propose watermark scheme, we calculate the PSNR and CC in our algorithm. What’s more, to tradeoff the calculation expense and robustness, we embed the watermark information using level-3 Discrete Wavelet Transform. Finally, taking the advantage of the inaudible feature of the low frequency coefficients after DWT, we choose the low frequency components to place the watermark information. The Correlation Coefficients between retrieved watermark from different audio signals which have been subjected to various attacks and the original watermark are shown in Table2. When using different embedding key for a single host audio signals, the experiment results show that there will be an optimal key available in the distribution system which brings better robustness, measured by CC, than the others, after watermarked audio signals having been subjected to different attacks, (Table1).

References 1. Maes, M., et al.: Digital watermarking for DVD video copy protection. IEEE Signal Process Mag. 17(5), 47–57 (2000) 2. Crowcroft, J., Perkins, C., Brown, I.: A method and apparatus for generating multiple watermarked copies of an information signal. WOPtent/56059 (2000) 3. Emmanuel, S., Kankanhalli, M.: Copyright protection for MPEG-2 compressed broadcast video. In: ICME 2001, IEEE Int. Conf. on Multimedia and Expo., pp. 203–209 (2001) 4. Jin, H., Lotspiech, J.: Attacks and forensic analysis for multimedia content protection. In: Proc. of ICME. IEEE Int. Conf. on Multimedia and Expo. (2005) 5. Anderson, R.J., Manifavas, C.: Chameleon-A new kind of stream cipher. In: Proc. 4th Int. Workshop on Fast Software Encryption, London, U.K, pp. 107–113 (1997) 6. Adelsbach, A., Huber, U., Sadeghi, A.-R.: Fingercasting-Joint fingerprinting and decryption of broadcast messages. In: 11th Australasian Conf. Information Security and Privacy, pp. 136–147 (2006) 7. Celik, M., et al.: Secure embedding of spread-spectrum watermarks using look-up tables. In: Proc. Int. Conf. Acoustics, Speech and Signal Processing, Honolulu, HI, pp. II-153–II-156 (2007) 8. Piva, A., Bianchi, T., Rosa, A.D.: Secure Client-Side ST-DM Watermark Embedding. IEEE Trans. on Information Forensics and Security 5(1), 13–26 (2010)

Effective Electronic Advertisement Auction System Tokuro Matsuo1 and Satoshi Takahashi2 1

2

Graduate School of Science and Engineering, Yamagata University 4-3-16, Jonan, Yonezawa, Yamagata, 992-8510, Japan [email protected] http://www.tokuro.net Graduate School of System and Information Engineering, University of Tsukuba 1-1-1 Tennoudai, Tsukuba Ibaraki, Japan [email protected] http://infoshako.sk.tsukuba.ac.jp/˜stakahashi/

Abstract. A structure of the Internet advertisement is that the service providers decide order of placement of many advertisements and advertising fees by auctions when advertisers offer their promotions. It is known that Generalized Second Price Auction (GSP) mechanism is most efficient auction mechanism of the advertisement auction. Some searching sites employ GSP mechanism basically. There are a lot of researches on GSP in order to analyze and clarify its feature and advantages. However, these researches assume that traded advertisements are mutually independent. That is means that each advertisement does not influence other advertisements. Also these researches do not consider a value of advertisement, which means some criterions of a name value of a company, an effectiveness and an importance, that is dependently each other. This paper proposes a new advertisement auction mechanism based on GSP with considering the value of advertisement. We analyze the auctioneer’s profit in comparison between normal GSP, normal VCG (Vickrey-Clarke-Groves Mechanism) and our proposed mechanism. The contribution of our research includes to clarify the features and advantages of advertisement auctions and effects to search service sites’ profit rate.

1 Introduction Advertisements on the webpage provide good opportunity to get new customers. In recent years, a lot of webpages providing a search service have advertisements, which are related with searched word by user. Some searching company has a advertising space on his/her webpages and allocates it for some advertisers based on an advertising fee. As same as items trading in the Internet auctions, a displayed advertisement on web page is also based on the auction, called the Internet advertisement auction. Internet advertisement auction is one of important income source for some search engines such as Yahoo! and Google[1][2]. When users search for some words on the search engine, an advertisement related with the searched keywords is displayed with result of search[4]. The order of advertisements to be displayed is determined based on bid value in an auction. Advertisers can set up the interval and period to display the advertisement as a time slot. The advertising fee is determined based on the Generalized Second Price T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 169–178, 2011. c Springer-Verlag Berlin Heidelberg 2011 

170

T. Matsuo and S. Takahashi

Auction, which is known higher revenues than the Generalized Vickrey Auction[3]. A winner in the auction gets a space to display their advertisement and the web page owner allocates time and position in the web page to show the advertisement. There are a lot of contributions about GSP(Generalized Second Price Auction) researches in electronic commerce research. In this auction, bidding and winner determination are conducted multiple time. Advertiser advertiser can change his/her bid value because the auction is continued with repetition. When advertisers try to bid in an auction, they bid on their strategy. However, GSP has an envy free equilibrium and webpage owner providing advertisement space can get larger benefit compared with VCG (VickreyClark-Groves) Mechanism. In previous research, the value of advertisement is assumed as independent with each other. Otherwise, some of their researches do not refer the value of the advertisement. However, each advertisement has a certain value for users. It means some criterions of a name value of a company, an effectiveness, an importance and an attribution, that is dependently each other. When same or similar item is soled in two e-commerce sites, the price on the advertisement is different from another one. If a buyer considers the price is important attribute to choose item, the advertisement selling items at low price has more value for the buyer. For example, a shop A gives an advertisement to sell an item for $100. When a shop B gives the advertisement to sell the same item for a shop $80, its value of the advertisement is higher than shop A’s value if the condition of item and other situations between shop A and B. In this paper, we focus on such situation and simulate the revenue of advertisers. Also, we analyze a result of simulation of Internet advertisement auction with relationship between value of each advertisement. After the simulation, we reformulate our proposed model and mechanism based on the preliminary simulation. Concretely, we discuss about dynamical environment. It is more realistic situation of the advertisement market on the Internet.

2 Preliminaries Suppose that there are n advertisers and k slots. A slot is a place of advertisement on a webpage. Let ci be a click-through-count (CTC) of the advertisement placed on the slot i. CTC is the number of clicks of the advertisement per an unit time. We assume following rule for each ci : ci−1 ≥ ci , for 2 ≤ i ≤ k. This rule means that CTC of the slot i is lower than the slot i − 1 for 2 ≤ i ≤ k. When an advertiser j bids a pair of a keyword and value per click to use a slot as {“keyword”, bj }, a payment of the advertiser, who was allocated a slot i, is defined by bj · ci . Figure 1 shows the generally advertisement auction model. In this model, there exists an advertisement auction system which decides some winners of the auction and management of slots. First of all, each advertiser bids some pairs of a keyword and value per click for advertisement slot to the auction system. After that, the auction system decides some winners of the auction and allocates the advertisements to the slots based on the bidding values. Also the auction system announces CTC to the winners, and the winner pays decided payment to the auction system. The auction system employs some auction mechanisms for a winner determination. The auction mechanism is a rule of allocation and decision of payment. Generally, the

Effective Electronic Advertisement Auction System

171

Fig. 1.

auction mechanism describes the auction system. We introduce some typical auction mechanisms for the Internet advertisement auction. We assume every following auction  satisfies Nash equilibrium. The Nash equilibrium shows that a strategy SA is a best strategy for agent A if every agent without agent A chooses an optimal strategy S  . Vickrey-Clark-Groves (VCG) mechanism. VCG mechanism is generalized from Vickrey auction, which has dominant strategy as truthful bidding[6]. Each advertiser j bids own value per click for auctioneer. The auction system allocates a slot for the advertiser by descending order of bids. Suppose A˜ is a set of winners of the auction and A˜−j is a set of winners of the auction which eliminates the advertiser j, we define a payment per click pj of the winner as follows,   bk − bk − bj . pj = ˜−j k∈A

˜ k∈A

VCG mechanism satisfies incentive compatibility and Pareto efficiency. The Incentive compatibility (Strategyproofness) means that each advertiser choice an optimal strategy without influence of other advertisers. The Pareto efficiency means a total utilities of each advertiser and the auction system[3]. We show an example, suppose that there are two slots and three advertisers. Advertiser 1, 2 and 3 bids $300, $200 and $100 per click, respectively. In this case, the auction system allocates slot 1 and 2 to advertiser 1 and 2, and advertiser 1 and 2 pays $100 and $100 per click, respectively. Also, the auction system’s gain is ($100 + $100) = $200. GSP (Generalized Second Price Auction) protocol. GSP protocol is an auction protocol which is natural extended form second price auction[7]. The auction system sorts all bided values by descending order, and allocates slot i to i-th highest valued advertiser for all slots. The advertiser who is allocated slot i pays bi+1 per click for the auction system.

172

T. Matsuo and S. Takahashi

It is known that GSP protocol does not satisfies incentive compatibility[8]. Therefore, the truthful bidding is not dominant strategy in GSP. On the other hands, GSP converges on Locally Envy Free equilibrium[4]. The auction is Locally Envy Free equilibrium, if an advertiser who gets a slot i does not increase own utility neither getting a slot i − 1 nor getting a slot i + 1[9]. Hence, the slot i is an optimal position which maximizes the advertisers’ utility. We consider the same example in VCG. Suppose that there are three advertisers and two slots, and advertiser 1, 2 and 3 bids $300, $200 and $100 per click. In this case, the advertiser 1 and 2 gets the slot 1 and 2, and pays $200 and $100 per click, respectively. The gain of auction system is $200 + $100 = $300. Therefore, the GSP protocol is better than VCG mechanism in the advertisement auction, since the auction systme gains $200 on the VCG mechanism. Note that if there is one slot, then the result of auction is the same on both GSP protocol and VCG mechanism. Their mechanisms are employed not only the advertisement auction, but also the general Internet auction. Next we discuss an auction system of the Google Adwords. 2.1 Google Adwords Google Adwords is an auction protocol similar to GSP protocol. Google Adwords employs CTR (Click-Through-Rate) and QS (Quality-Score). CTR is a ratio of click deClick-through-count of an advertisement noted by CTR = Number . Quality score is decided by Google of page view of an advertisement from CTR and relationship between text of the advertisement and searching keyword. Also, Google sets a minimum bidding value. An allocation of slots are based on descending order of multiplying the value by quality score, called evaluation score. It means that if high quality score has a possible to get a good position of slot by cheap payment. Google requires all advertisements positioned on upper slots must have a certain quality score level. Let q be a quality score of an advertiser allocated on a slot i, and bi (b1 > b2 > · · · > bi > · · · > bk ) be a evaluation score. A payment p per click is denoted by p = bi+1 + 1. q It is known that CTR is proportional to order of slots. N. Brooks[5] say that there is a strong correlation between CTR and order of slots. The report also shows the ratios of CTR when a first ordered CTR is 100%. in this result, a second ordered is 77.4%, and third is 66.6%. However, Google suggests there is an exception. For example, some famous companies positioned lower slots has larger CTR than some upper positioned companies, since the famous companies get many click-through-counts even lower position. On the other hand, Sponsored search which is derived by Yahoo! Search Marketing has technique similar to Google Adwords, but, there is a difference that order of slots is descending order of only bidding value. We consider the same example in VCG. Suppose that there are three advertisers and two slots, and advertiser 1, 2 and 3 bids $300, $200 and $100 per click. Also advertiser 1, 2 and 3’s quality score is 2, 1.5 and 1, respectively. In this case, the evaluate scores are 600, 300 and 100, respectively. The advertiser 1 and 2 gets the slot 1 and 2, and pays $300/2 = $150 and $100/1.5 = $66 per click. The gain of auctioneer is $150 + $66 = $216.

Effective Electronic Advertisement Auction System

173

2.2 Our Advertisement Auction Model Above mechanisms do not consider some value aside from advertising fee. It means a criterion of name value, effectiveness and importance for users. This criterion is not independent with each other, since users compare two or more advertisements for increasing them utilities. Thus each advertisement has a co-dependent value and it is expressed by linear to be evaluated. Let Ad (|Ad| ≤ k) be a set of advertisements which are now placed on the site, Co(j) ⊆ {1, ..., n} ∩ Ad be a set of co-dependent advertisements with advertisement j. When a value of company j’s advertisement changes Ajaf ter from Ajbef ore ,  co-dependent value of other advertisement  with company j is shown Bbef ore and    Baf . B is changed a value effected by all advertisement in Co() to Baf ter bef ore ter . That is   Baf ter = Bbef ore + α



(Ajaf ter − Ajbef ore )

(1)

j∈Co()

The condition of the above equation is given as 0 ≤ α ≤ 1. α = 0 shows independent with each advertisement. Quality score in the GSP auction protocol used in Google Adwords is placed a value in which we have defined above definition in the simulation. Figure 2 is an example of the model of our proposed mechanism. When a user clicks the link of the advertisement, its value is increased. Relatively, other advertisement’s value becomes going down. When low-ranked advertisement is clicked by users, the advertisement is regarded as valuable comparing with high-ranked advertisement. In the Figure 2, we assume all of advertisers participate to bid for the first time. After bidding, the winners are determined be the auction. Then, each advertisement is displayed at the website as (A). Users click the advantages and the value of each advertisement changes based on number of click as (B). After one period passes, advertisers bid at second round auction to keep their advertisement in the website. We also assume all advertiser bid same price comparing with first round auction. The order of advertisement is changed based on both bid price and advertisement’s value. In this case, although advertisement 1’s value decreased in (B), position of advertisement 1 is kept at top because bid price is very high as (C). Because advertisement 4’s value is quite high in (B), the rank of advertisement in (C) becomes second although bid price is the lowest in other three advertisers. Next, we show some preliminary experiments for evaluation of our proposed model, and for finding new conditions or characteristics.

3 Preliminary Experiment 3.1 Condition We set 3-10 slots to be put advertisements and 10-50 advertisers (companies to join in the auction) who bid to get a space for their advertisement. The lowest bid price in the auction is set $10 and advertiser’s bid value is defined a uniform distribution between

174

T. Matsuo and S. Takahashi

Fig. 2. Concept of the proposed mechanism

$10 and $100. Initial value of each advertisement is defined on a uniform distribution between 0.2 and 2.2. Number of clicking by end-user is assumed on a uniform distribution between 1 and 100 in a time slot. 3.2 Procedure of Trade We now simulate our proposing mechanism by using following procedure. The procedure gives in above condition and we run this procedure at 100 thousands times. (1)Decide a number of slots for advertisements. (2)Decide a number of clicks for each slot in a period. (3)Decide each advertiser’s bid value and advertisement value. (4)Allocate each slot in descending order according to a valuation that is multiplied by a bid price and a value of advertisement. (5)Calculate each advertiser’s payment and benefit. (6)Change a value of advertisement of a certain advertiser. (7)Compute a new value of advertisement based on equation (1). (8)Conduct step 4) and 5) based on the new value of advertisement and bid price. 3.3 Results Table 1 shows result of simulation in which value of advertisement is changed. There are 20 advertiser advertisers and value of a certain advertisement is reduced and it effects other values of advertisement. When number of slot is changed from 4 to 10 and a value of one advertisement is reduced, 54,000 auctions make whole profit in the market increase in 100,000 trial. Average of the increased profit is $21.38. We discuss result of simulation from Table 1.

Effective Electronic Advertisement Auction System

175

Table 1. A value of one advertisement is reduced Number of slot Increase (%) Decrease (%) Average of increased/decreased profit 4 54.1 45.9 $22.38 6 55.3 44.7 $25.92 8 55.7 44.3 $27.56 10 56.3 43.7 $30.22

1. Averages of profit is normally increased and the profit increases when number of slots increases. 2. Possibility of profit increase is increased when number of slot increases. This feature is apparent because the curve in Table 1 is monotonic increase. As same as the above, table 2 shows the case where 20 advertisers join in the auction and value of one advertiser’s advertisement is increased. The number of slot is changed from 4 to 10 in each trial. We discuss result of simulation from Table 2. Table 2. A value of one advertisement is increased Number of slot Increase (%) Decrease (%) Average of increased/decreased profit 4 6 8 10

43.8 43.0 42.2 42.0

56.2 57.0 57.8 58.0

-$32.85 -$36.63 -$38.81 -$40.31

1. Averages of profit is normally decreased and the profit decreases when number of slots increases. 2. Possibility of profit increase is decreased when number of slot increases. This feature is also apparent because the curve in Table 1 is monotonic decrease. Table 3 is a result where number of slot is fixed as 5 and one advertiser changes value of his/her advertisement. The number of advertiser is changed 10 to 50 in each trial. Rate of increase/decrease of value of advertisement is assumed by uniform distribution. The result shows a comparison of profits between non-affective and affective. 1. Averages of profit is normally decreased and the profit decreases when number of slots increases. 2. Possibility of profit increase is decreased when number of slot increases. Average of profit is negative because possibility that the profit decreases is large. From above simulation and analysis, we find out the following features. First, total profit of webpage owner reduces when each advertisement has co-dependence between its value. Second, when the size of auction becomes large, average of profit is decreased.

176

T. Matsuo and S. Takahashi Table 3. Number of slot is fixed as 5

Number of advertisers Increase (%) Decrease (%) Average of increased/decreased profit 10 49.8 50.2 -$2.81 20 49.1 50.9 -$4.87 30 48.7 51.3 -$5.17 40 48.3 51.7 -$6.27 50 48.1 51.9 -$6.85

3.4 Comparison to VCG Table 4 shows the result of simulation when the number of advertisers is 20 and number of slots are changed from 4 to 10 in each trial. When number of advertiser increases, our proposed GSP mechanism makes large profit comparing with general VCG mechanism. Table 4. A value of one advertisement is increased Number of slot Increase (%) Decrease (%) 4 80.3 19.7 6 83.6 16.4 8 83.7 16.3 10 83.8 16.2

Table 5 shows the result of simulation when the number of slot is fixed as 5 in comparison between our proposed GSP and general VCG mechanism. Our proposed GSP makes larger profit compared with normal VCG with monotonic increase when the number of advertisers increases. When number of advertisers is not many, the increase rate is high. After number of advertisers is 30, increase rate becomes less and it seems to become convergence. Table 5. Number of slot is fixed as 5 Number of advertisers Increase (%) Decrease (%) 5 10 20 30 40 50

51.6 62.3 82.9 89.5 92.8 94.5

48.4 37.7 17.1 10.5 7.2 5.5

4 Discussion The preliminary experiment is static environment, however, the real is dynamic environment. The co-dependent evaluation, which is in our proposed model, should be decided dynamically. Figure 3 shows an image of dynamic changing of co-dependent values.

Effective Electronic Advertisement Auction System

177

In this figure, there are two advertisements, they influence each other. In dynamic environment, there is a deadline of co-dependent value decision. This deadline shows a time in which a final co-dependent value is decided. White circle and gray circle shows two advertisements’ co-dependent value. In this figure, their initial states are fixed on time 0. When the white circle is going down, then this phenomenon influence the gray circle. Each co-dependent value iterates this phenomena until deadline. This image also shows in left part of figure 3. In the dynamic environment, there are many advertisements which influence each other. Hence we should reformulate our proposed model and create efficient mechanism for a dynamic environment.

Fig. 3. Image of dynamically decision of co-dependent value

4.1 Reformulation In this section, we reformulate our proposed model and an auction mechanism used by the model. Now we redefine some terms and formulas. Suppose that there are n advertisers and k slots in the advertisement auction. Let Ad (|Ad| ≤ k) be a set of advertisements which are now placed on the site and Co(j) ⊆ Ad be a set of co-dependent advertisements with advertisement j. Also let Cjt be a codependent value of j at time t. Cjt is computed by the following:  Cjt = Cjt−1 + βkt (Ckt − Ckt−1 ), (2) k∈Co(j)

where β t is a condition parameter. Also Cjt ’s range is (0, 1]. We define β t as a function on a co-dependent vector C t−1 . Hence, βjt := βjt−1 + f (C t−1 ). The auction system is able to compute the formulation 2 by using a simultaneous equation. Next we reformulate an auction mechanism for the dynamic auction model. The auction system allocates each slot by descending order of a function of pair of the

178

T. Matsuo and S. Takahashi

co-dependent value and advertiser’s bids gj (Cjt , bj ). Note that since an advertiser j’s bidding value bj does not change among the auction, the advertisement’s position of slots is decided by only co-dependent value. If a ordered sequence of bid value is b1 > bj+1 b2 > · · · > bj > bj+1 > · · · bn , a payment per click ptj is denoted by ptj = C t + 1. j

5 Conclusion We proposed codependent value-based GSP mechanism in the Internet advertisement auctions, and ran a preliminary simulation based on multi-agents. Our analysis showed that total profit changes in different auctions mechanism GSP, VCG, and our proposed mechanism. Particularly, from the analysis, auctioneer changes the auction protocol based on his/her estimate profit. However, our auction protocol has an advantage where the website provides more useful advertisement for users, because the order of allocation is based on both price and value. Also we reformulated our proposed model and mechanism for dynamic environment. The co-dependent value is changing dynamically among a few times. Our model showed this phenomena.

References 1. http://www.yahoo.com 2. http://www.google.com 3. Vickrey, W.: Couterspeculation, Auctions, and Competitive Sealed Tenders. Journal of Finance 16(1), 8–37 (1961) 4. Edelman, B., Ostrovsky, M., Schwarz, M.: Internet Advertising and the Generalized Second Price Auction: Selling Billions of Dollars Worth of Keywords. American Economic Review 9(1), 242–259 (2007) 5. Brooks, N.: The Atlas Rank Report: How Search Engine Rank Impact Traffic, Insights, Atlas Institute Digital Marketing (2004) 6. Milgrom, P.: Putting Auction Theory to Work. Cambridge University Press (2004) 7. Varian, H.R.: Position auctions. International Journal of Industrial Organization 25, 1163–1178 (2007) 8. Abrams, Z., Schwarz, M.: Ad Auction Design and User Experience. In: Deng, X., Graham, F.C. (eds.) WINE 2007. LNCS, vol. 4858, pp. 529–534. Springer, Heidelberg (2007) 9. Guruswami, V., Hartline, J.D., Karlin, A.R., Kempe, D., Kenyon, C., McSherry, F.: On profitmaximizing envy-free pricing. In: Proceedings of the Sixteenth Annual ACM-SIAM Symposium on Discrete Algorithms, pp. 1164–1173 (2005)

Energy-Efficient Fire Monitoring Protocol for Ubiquitous Sensor Networks Heemin Kim, Ae-cheoun Eun, Sunyoung Han, and Young-guk Ha* Department of Computer Science and Engineering, Konkuk University 1 Hwayang, Gwangjin, Seoul 143-701, Korea {procan,syhan}@cclab.konkuk.ac.kr, [email protected], [email protected]

Abstract. Many countries are looking for ways to fight the forest fire at early stage using sensor network by integrating IT technologies. Studies are conducted in fire-related sensor network field in line with those changes, and the studies are broadly divided into efficient processing of fire data between sensor nodes and sensor network energy efficiency in case of fire. Thus, this study forms multi-layer cluster hierarchy dynamically according to the direction of fire spreading on cluster-based network hierarchy appropriate for FireMonitoring, and intends to propose Energy-efficient Fire Monitoring Protocol that can reduce energy consumption by the entire sensor network by performing efficient transmission of fire data. Keywords: Fire-monitoring, forest fire detection, Fire-detection ,Energyefficient routing, USN, Cluster-based sensor network.

1

Introduction

Sensor network is evaluated as one of the best systems that can be applied at current environment. However, sensor network has many problems in fire-Monitoring environment due to limited battery problems. Flat based routing-based Fire-Monitoring like SPIN consumes energy very quickly since all nodes detecting fire start communicating individually. The energy use is inefficient for layer routing like LEACH regarded as appropriate for Fire-Monitoring since cluster individually transmits fire data like flat-based routing.[1] The use of energy is inefficient in these problems since the number of data transmission by nodes increases as the region gets wider and the number of nodes increases.[2][3][4]Therefore, this study proposes Energy-efficient Fire Monitoring Protocol(EFMP) that increases energy efficiency by reducing the number of transmissions of temperature data measured by sensors detecting fire from Fire-Monitoring environment to Sink Node by forming multi-layer cluster hierarchy dynamically according to the direction of fire spreading from cluster-base. *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 179–188, 2011. © Springer-Verlag Berlin Heidelberg 2011

180

H. Kim et al.

2

Design of Energy-Efficient Fire Monitoring Protocol

2.1

Structure of Energy-efficient Fire Monitoring Protocol

The features of EFMP are electing a Master Head that collects and manages information from Slave Nodes managing Cluster. The location of Master Head changes according to information on fire (direction of fire). The passage defines structural design of EFMP and algorithm and packet type for deciding a Master Head. The changes in the election of a Master Head according to Fire Monitoring refers to the forming of optimal dynamic hierarchical cluster by electing an optimal Master Head according to the changes in case of fire. It results in energy efficiency by reducing the number of transmission paths at the time of transmitting information on fire unlike existing sensor routing. Fig. 1. is EFMP Protocol Stack which designed by this paper and broadly composed of sections responsible for hardware and sections forming protocol.

Fig. 1. Energy-efficient Fire Monitoring Protocol stack

2.2

Type of Packets

The features of EFMP are electing a Master Head that collects and manages information from Slave Nodes managing Cluster. The location of Master Head changes are according to the information on fire (direction of fire). The passage defines structural design of EFMP and algorithm and packet type for deciding a Master Head. 1) SIG_FIRE : Refers to a packet for detecting fire and sending detected information to Slave Node or Master Head and the conditions of sensor detection are decided by the following (1). tn −1 TEMP(tn) -    TEMP t / n − 1  t =t1

()(

) > ΔTEMPMAX 

Condition for detecting fire by SIG_FIRE Packet from EFMP

(1)

Energy-Efficient Fire Monitoring Protocol for Ubiquitous Sensor Networks

TEMP(tn)

tween

is currently measured temperature.

tn −1  t = TEMP t / n − 1 t =t 1

()(

)

181

is the sum be-

and TEMP(tn - 1) . It means average temperature prior to current temtn−1 perature when  t = TEMP t is divided by (n-1). From here, it means differences in t =t1 TEMP(tn)

()

average temperature of current temperature if

TEMP(tn) -

(

tn−1  TEMP t / n − 1 t =t 1

()(

)

)

> ΔTEMPMAX

and the differences can be regarded as temperature increase. Also, if the value ΔTEMPMAX is max, then it can be regarded as the highest temperature increase among increases in temperature. Therefore, it recognizes fire by detecting fire when it is higher than ΔTEMPMAX and sends information to Slave Node or Master Head. The composition of SIG_FIRE Packet is shown in Fig. 2.

Fig. 2. SIG_FIRE Message Format

2) SIG_DATA : Sensor monitors Fire and sends its information to Slave Node by recognizing fire. Sensor node receiving information relays data on fire (SIG_FIRE) it collected to a Master Head and SIG_DATA is used when sending information from Slave Node to the Master Head.

Fig. 3. SIG_DATA Message Format

3) SIG_INFORM : SIG_INFORM packet is used when a new Master Head provides its own information to Slave Nodes. It is used when the first Slave Node detecting fire within Cluster is changed to the Master Head.

Fig. 4. SIG_INFORM Message Format

4) SIG_QUERY, SIG_RESP : Slave Node within Cluster transforms to the Master Head from Slave Node once the sensor detects fire within its own Cluster. Slave Node becomes a Master Head candidate and the previous Master Head questions whether it is appropriate as a Master Head to nearby Master Head candidates. The message format sending this information is called SIG_QUERY and response to this is called SIG_RESP.

182

H. Kim et al.

Fig. 5. SIG_QUERY Message Format & SIG_RESP Message Format 5) SIG_LISTEN : Address information on new Master Head should be sent to nearby Slave Nodes once a new Master Head is elected. SIG_LISTEN Packet provides address information of newly elected Master Head.

Fig. 6. SIG_LISTEN Message Format

6) SIG_TRANS & SIG_RESET : SIG_TRANS is used for registering information on new Master Head by receiving SIG_LISTEN packet from the new Master Head and is also used when Slave Node registers its own address information. SIG_RESET packet is used when the previous Master Head returns back to Slave Node by handing over its authority as a Master Head to the newly elected Master Head and when reverted previous Master Head is registering its information with the new Master Head. 2.3

Role of Cluster Head According to the Changes in Case of Fire Monitoring

Cluster Head decides Master Head according to the changes in case of FireMonitoring. Fig. 7 shows the changes in roles of nodes according to changes in case of fire and all nodes are initialized while in Watch Mode.

Fig. 7. Role of Cluster Head according to the changes in case of fire Monitoring

Energy-Efficient Fire Monitoring Protocol for Ubiquitous Sensor Networks

183

Watch Mode refers to initial state where sensors don't detect fire. Cluster Head that detects fire for the first time changes from Watch Mode to Master Mode. It transforms nearby Cluster Heads in Watch Mode into Slave Mode since it is in Master Mode itself under Master Mode. That is, EFMP System displays much better energy efficiency than existing cluster network since the system forms hierarchical networks dynamically by changing Cluster Head into Cluster Head Watch Mode, Slave Mode, or Master Mode according to the conditions of fire. 2.4

EFMP Detailed Operation Procedures According to Spreading of Fire

Cluster Head of EFMP operates in Watch Mode before detecting fire, monitors fire from own zone and collects sensor information. Sensor of Cluster Head in Watch Mode as shown in Fig. 8 sends fire information to own Cluster Head once fire is detected within own Cluster. Cluster Head receiving information of fire from sensor recognizes fire within own Cluster for the first time and sends information to Sink Node by converting to Master Head from Watch Mode.

Fig. 8. Method of electing Master Head initially from Watch mode

First elected Master Head sends information on fire within own Cluster to Sink Node. It, then, sends Master Head information to nearby Cluster Head in Slave Mode through SIG_LISTEN message that is Master HEAD and updates. Cluster Head where fire broke out is elected as the first Master Head. A new Master Head according to the progress of fire sends collected data to Sink Node as shown in Fig. 9. It acts as the Master Head continuously until Master Head candidate appropriated for Master Head is elected again.

Fig. 9. Master Head Information Update through SIG_LISTEN Message

184

H. Kim et al.

The header detecting fire for the first time is the most appropriate for a Master Head optimally since fire is not broken out from other areas for cases of Fig. 8, 9. However, reelection of optimized Master Head is necessary according to the range and information of fire if the range of fire extends to B, C Clusters as well as A Cluster as shown in Fig. 10. An optimal Master Head (Cluster Head detecting fire within own Cluster) elects a Master Head that has the least number of Hop counts between the number of Hops from Master Head candidate to Sink Node. The number of Hop on the sum of distance from other Cluster Head to Master Head candidate according to the range of fire. The least number of Hops refers to shorter distance for sending and using the least amount of energy. In other words, sensor information is collected by Cluster Head and collected information is not sent to individually to Sink Node by Cluster Head but by Master Head in batch. Master Head collecting information. Master Head can achieve efficiency in distance and energy used for sending information to Sink Node when selecting the least amount of Hop Count. thus the location of Master Head can be regarded as having direct effect on the efficiency of sensor network. The most optimized Master Head among Master Head candidates is the least number of Hop-Counts among candidate masters and the number of Cluster Heads increases detecting fire within own Cluster according to the spread of fire. Cluster Head detecting fire becomes a Master Head candidate that can become the Master and makes inquires to current Master Head on the conditions for becoming optimized master through SIG_INFORM message.

Fig. 10. Procedure of electing a new Master Head according to the range of fire

Fig. 11. Procedure of electing a new Master Head

Energy-Efficient Fiire Monitoring Protocol for Ubiquitous Sensor Networks

185

The number of Cluster Heads detecting fire within own Cluster increases as fire spreads widely in Fig. 11. Cluster Head detecting fire can become a Master H Head candidate can be master an nd inquires to current Master Head on the presence of cconditions for becoming optim mized master through SIG_INFORM message. Master H Head receiving SIG_INFORM message m sends SIG_QUERY message to send the tootal number of Hops to each Maaster Head candidate and Master Head candidates calcuulate the total number for Hops to Sink and send the information to the master throuugh SIG_RESP message.

Fig. 12. Procedure of updating Slave Node by the changes in case of fire

When Master Head recceiving the information, Master Head compares the tootal number of Hops on distancce for sending messages to Sink and the number for H Hops from each Master Head. Then, T it informs a candidate with smaller number of H Hops than itself if available as the t next Master Head, while it sends message that it w will maintain the role of a curreent master if the total number of Hops from Master H Head candidates are bigger. Masster Head moved to Cluster Head B from Cluster Headd A since the fire spreads wideely than in Fig. 13. Cluster Head B sends SIG_LISTEN N, a packet requesting for updaating new Master Head information to each Master H Head candidates since it becamee the Master Head and Master Head candidates receivving messages update Master Heead information.

Fig. 13. Re-eleection of a third Master Head by the spread of fire

186

H. Kim et al.

The previous Master Head updates information on the next Master Head by itself and does not update after receiving SIG_LISTEN since it knows which one is the next Master Head and sends SIG_TRANS message indicating completion of update thus elects a new Master Head. Elected Master Head relays information to Sink Node in batch by receiving information from each Cluster Head as the previous Master Head has been doing. The fire spreads wider than Fig. 13 to Cluster D. Cluster Head D realizes that it detected fire and notifies Master Head that it is now a Master Head candidate. Master Head receiving information from Cluster Head D elects the next Master Head by comparing Hop-Count between own Hop-Count and Hop-Count of Master Head candidates. It is elected as a Master Head since Hop-Count from Cluster Head D is smaller than Cluster Head C. Each Cluster Head sends information collected from own sensors to Sink Node in previous Cluster Head method but there are problems of different energy efficiency in transmission by Cluster according to the location of Cluster Head, thus, the location of Cluster Head needs to be regularly changed to solve this problem. Otherwise, the entire energy efficiency on sensor network worsens according to several wrong locations of Cluster Head.

3

Performance Evaluation

OMNET++ 4.1 was used to test performance of EFMP in this study. MiXiM module was used to implement sensor network. Fig. 14 is a graph showing average power consumption required for forming cluster without including data between node and cluster for 1 hour (3600 seconds). The yaxis of the graph shows average power consumption per sensor node as shown in the following formula. C(Si,t) is power consumed for time t by sensor Si and N is the total number of sensor nodes from (3).

( )

N  C Si , t i =1 N

Formula for calculating average power consumption according to time

Fig. 14. Average power consumption per sensor by time

(2)

Energy-Efficient Fire Monitoring Protocol for Ubiquitous Sensor Networks

187

We know that existing clusters consume large amount of power to maintain connections from each cluster head to sink node when forming cluster head from FireMonitoring environment. However, other cluster heads send information only to Master Head and the Master head maintains connection to Sink Node once it is decided in EFMP under Fire-Monitoring environment. It is energy efficient since the number of transmissions from the entire sensor nodes are reduced by the Master Head. The results of testing show that EFMP is better in forming clusters and power for maintaining connections to Sink Nodes than existing clusters.

Fig. 15. Total accumulated power consumption according to time

Fig. 15 is a graph comparing accumulated power consumption according to the total time for sensor node participating under Fire-Monitoring environment. tc indicates current time and N indicates the total number of nodes participating in fireMonitoring from (3). indicates the total power consumption by sensors according to time under Fire-Monitoring environment. The result shows the total accumulated power consumption by sensor nodes according to time as shown below. tc  N    C( Si , t )   t = 0  i =1 

(3)

Total accumulated power consumption by sensor nodes according to time

The results of test show that General Cluster Protocol used 4,346,675mA for 1 hour while EFMP used 3,933,713mA. EFMP reduced energy by about 9.5% according to the testing.

4

Conclusion

Sensor network is being evaluated as one of the most applicable system in firemonitoring environment. However, sensor network has problems in fire-Monitoring

188

H. Kim et al.

environment due to limited battery capacity. The energy efficiency drops when general cluster sensor network is used when multiple sensors monitor fire over a wide area due to characteristics of Fire-Monitoring.[5][6]Therefore, this study proposed EFMP that reduces energy consumption from the entire sensor network by dynamically forming multi-layer cluster hierarchy according to the spread of fire and efficiently transmitting data in case of fire to cluster-based network hierarchy appropriate for the characteristics of Fire-Monitoring. The system reduced energy by about 9.5% than general cluster system through testing. EFMP showed better energy efficiency as the number of nodes increases in case of fire-Monitoring environment based on the regression analysis. The future study will be conducted to examine the size of fire by the spread of fire by expanding EFMP in fire-Monitoring environment. It is hoped that more studies on protocols will result in protecting lives and properties much quickly from fire.

References 1. Akyildiz, I.F.: Wireless Sensor Networks: A Survey. Elsevier Sci. B. V. Comp. Networks 38(4), 393–422 (2002) 2. Muruganathan, S.D., Ma, D.C.F., Bhasin, R.I., Fapojuwo, A.O.: A centralized energyefficient routing protocol for wireless sensor networks. IEEE Communications Magazine 43(3), S8–S13 (2005) 3. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: Energy-efficient communication protocol for wireless microsensor networks. In: Proceedings of the Conference on 33rd Annual Hawaii International System Sciences 2000, p. 10 (2000) 4. Yu, L., Wang, N., Meng, X.: Real-time forest fire detection with wireless sensor networks. In: 2005 International Conference on Wireless Communications, Networking and Mobile Computing, September 23-26, vol. 2, pp. 1214–1217 (2005) 5. Zhang, J., Li, W., Han, N., Kan, J.: Forest Fire Detection System Based on a ZigBee Wireless Sensor Network, pp. 369–374. Higher Education Press (2008); co-published with Springer-Verlag GmbH 6. Fok, C.-L., Roman, G.-C., Lu, C.: Efficient Network Exploration and Fire Detection using Mobile Agents in a Wireless Sensor Network. In: ONR-MURI Review, Baltimore, MD (2004) 7. Gui, C., Mohapatra, P.: Power conservation and quality of surveillance in target tracking sensor networks. In: Proceedings of the ACM MobiCom (2004) 8. He, T., Krishnamurthy, S., Stankovic, J., Abdelzaher, T., Luo, L., Storelu, R., Yan, T., Gu, L., Hui, J., Krogh, B.: Energy-efficient surveillance system using wireless sensor networks. In: Proceedings of the 2nd International Conference on Mobile Systems (2004)

Design of Optimal Combination for New and Renewable Hybrid Generation System Kun Hyun Park, Chul Uoong Kang, Gi Min Lee, and Joung Hwan Lim Department of Mechatronics, School of Engineering, Jeju national University, 102 Jejudaehakno, Jejusi, South Korea {3313park,cukang,jhlim}@jejunu.ac.kr, [email protected]

Abstract. Hybrid generation system is basically merging systems of two or more different types of generation systems. Hybrid Generation System is more effective than utilization of single renewable energy resource. The combination of different types of generation systems are vary important because the total cost for generation is dependent on the combination. This study presents a method to design the optimal combination for new and renewable hybrid system. The method aims at finding the configuration, among sets of system components, that meets the desired system requirements, with the lowest value of the energy cost. Keywords: Optimal combination of Hybrid Generation System, new and renewable energy.

1

Introduction

Recently, the hybrid generation system has became significant because of the complementary characteristics among the new and renewable energy resources. To use the energy resources of hybrid system more efficiently, the combination of different types of generation systems that constitute the hybrid system is very important because the energy cost depends on the kinds of new and renewable energy resources. However, the combination of the hybrid system is performed on the basis of experience and intuition, which is not attained the optimum efficiency. Since new and renewable energy resources have stochastic behaviour, the major aspects in the design of the hybrid system are the reliable power supply of the consumer under varying atmospheric conditions and the cost of energy. In order to use the new and renewable energy resources more efficiently and economically, the design of optimal combination of hybrid system with battery plays an important role in this respect. Various optimization techniques of hybrid system have been reported in the literature. Kellogg [1] and Chedid [2] reported the linear programming technique. On the other hand, Karaki [3] and Bagul [4] developed the probabilistic approach, and iterative technique was developed by Kellogg [1]. Musgrove [5] presented dynamic programming, and Yokoyama et al. [6] developed multi-objective method. Yang et al. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 189–198, 2011. © Springer-Verlag Berlin Heidelberg 2011

190

K.H. Park et al.

[7] and Beyer et al. [8] have obtained the set of different configurations which meet the load using the autonomy level of the system. Protogeropoulos et al. [9] presented general methodology by considering design factor such as autonomy for sizing and optimization. Recently, Diaf at al.[10] suggest very accurate mathematical approach for characterizing PV module, wind generator and battery. However, these techniques are only for the optimal sizing of the hybrid system given the combination of energy resources. This paper presents a methodology for the design of optimal combination of hybrid generation system with storage batteries. The methodology adopted LPSP concept which was presented by Diaf [10]. By modifying the LPSP, we present a methodology to perform the optimal combination of a new and renewable hybrid generation system. The methodology aims at finding the combination, among sets of system components, that meets the desired system requirements, with the lowest value of the energy cost. The availability of the method is demonstrated with the results produced through sets of simulations.

2

Mathematical Modeling of Hybrid System Components

There are many kinds of new and renewable energy resources. In this paper, we consider the typical new and renewable energy resources such as wind, PV, tidal energy resources. Any other new and renewable energy resources can also be considered to be the component that constitutes the hybrid system as far as it can be modeled mathematically. 2.1

Modeling of Wind Generator System

There are many types of wind generators that have different power output performance curves, so that the model used to describe the performance of wind generators is expected to be different. Some authors assume that the turbine power curve has a linear, quadratic or cubic form. Other authors approximate the power curve with a piecewise linear function with a few nodes. In this study, we use the original mathematical model of output power for wind generation system. This may be somewhat different from the actual power curves. The model, however, can be applied any types of wind generation system. The mathematical model of wind turbine output can be defined as: (1) where, ρ is air density, A is diameter of rotor, v is wind speed. For a specific wind generator, a mathematical model can be developed according to its power output performance curve that is given by the manufacture. The power output is modeled through interpolation of the values of the data provided by the manufacturers. Since the power curve can be assumed to be smooth, they can be

Design of Optimal Combination for New and Renewable Hybrid Generation System

191

approximated using a line interpolation. The fitting equation of the output characteristic of wind generator can be expressed as:

(2)

Where, PW (v) is the output power of wind generator at wind speed v, Pr is the rated power, v is the wind speed, vs, vr and vc are the cut-in, rated and cutout wind speeds, respectively. n is the number of cubic spline interpolation functions corresponding to n +1 values couples (speed, power) of data provided by the manufacturers and a,b are the coefficients of the line interpolation functions which depend on the wind turbine generator type. 2.2

Modeling of PV System

If the PPV solar radiation on the tilted surface, the ambient temperature and the manufacturers data for the PV modules are available, the power output of the PV generator, PPV, can be calculated according to the following equations. (3) where, η0 is the instantaneous PV generator efficiency, N is number of modules, Am is the area of a single module used in a system andGt is the global irradiance incident on the titled plane. The instantaneous PV generator efficiency η0 is represented by the following equation. (4) where, ηr is the PV generator reference efficiency, η0t is the efficiency of power tracking equipment which is equal to 1 if a perfect maximum power point tracker is used, TC is the temperature of PV cell (°C), Tr is the PV cell reference temperature and βt is the temperature coefficient of efficiency, ranging from 0.004 to 0.006 °C per for silicon cells. However, to simplify the model, we use the general PV generator efficiency that is used many practical approach. 2.3

Modeling of Tide Generator System

There are many types of tidal generation system. Among them the most popular one is a horizontal axis blade type of tidal generation system which is basically no different from the horizontal axis wind turbine system. In this study, the horizontal axis blade type of tidal generation system is assumed. It, therefore, has the same mathematical model as that of wind generation system stated in Eq.(1). That is, (5)

192

K.H. Park et al.

where, ρsec is sea water density, A is diameter of rotor, vcur is the speed of current. Since ρsec is about 1052.2 kg/m3 the tidal energy is much greater than wind energy when the current speed is equal to wind speed. 2.4

Modeling of Battery System

Since the state of battery is related to the previous state of charge and to the energy production and consumption situation of the system during the time from t-1 to t, it should be modeled differently according to the generation and load conditions. When the total power from the hybrid generation system is greater than the load required, the battery is in charging state and modeled as follows; (6) Where, C(t) is battery bank capacity, EG(t) is total power of the hybrid system, EL(t) is the power needed by the load at time t, σ is self discharge rate of the battery, ηbat is the battery efficiency, and ηinv is the inverter efficiency. During discharging process, the battery discharging efficiency was set equal to 1, and during charging, the efficiency is 0.65 to 0.85 depending on the charging current. On the other hand, when total power is less than the load demand, the battery is in discharging state and modeled as follow; (7)

3

Design Model of Optimal Combination

3.1

The RLP Model

To determine the optimal combination of the new and renewable energy sources and to achieve the optimal configurations of the hybrid system in term of technical analysis, the RLP model is developed, which is modified from the method of LPSP. The total power, Ptot, generated by the hybrid generation system at time t is calculated as follow: (8) where, Pi is the output power of the i-th individual generation system that constitutes the hybrid system. n is the total number of the generation systems in the hybrid system. Then, the inverter input power, Pinv (t), is calculated using the corresponding load power requirements. (9) where, Pload (t)is the power required by the load at time t, is the inverter efficiency.

Design of Optimal Combination for New and Renewable Hybrid Generation System

193

The following two different situations may appear during the operation of the hybrid generation system.

① The total power generated by the hybrid system is greater than the power ②

needed by the load Pinv. In this case, the energy surplus is stored in the batteries and the new battery capacity is calculated using Eq.(5) until the full capacity is obtained, the remainder of the available power is not used. The total power generated by the hybrid system is less than the power needed by the load, Pinv, the batteries supply the energy deficit, and a new battery capacity is calculated using Eq.(7).

In case when the total power generated by the hybrid generators is equal to the power needed by the load Pinv, the batteries remains unchanged, and this case can be considered as special case of . If the power generated by the hybrid system is less than the load demand, The batteries should supply the energy deficit. However, if the battery capacity reaches to the minimum capacity state, Cmin, in which the batteries cannot discharge anymore, the hybrid system can no more supply energy deficit. In this case the power deficit must be supplied from the external energy system. The power deficit in this case is called as ' Lack of power', PLP, and can be defined as:

①

(10) Where, PG(t) and Pload(t) are total power and load power requirement. Pload(t) ∆t represents total load demand power, and the last term represents the power consumed by the load. In Eq. (10), it is assumed that power generated by the hybrid system during ∆t is unchanged. The amount of power discharged until the battery capacity reaches Cmin , Cout, is written as, (11) The ratio of lack of power, RLP, for a period T, can be defined as the ratio of total lack of power over the total load required during that period.

(12)

Using Eq. (12), the design of optimal combination of the hybrid system components is performed. 3.2

Economical Model

For hybrid generation systems the most important concern is to achieve the lowest energy costs, and the economical approach can be the best benchmark of cost

194

K.H. Park et al.

analysis. Several methods are used to get different options for energy system; the levelised cost of energy is often the preferred indicator [11]. However, the method is not easy to apply in a practical application because it is very complicated. In this study, a simple economical model is developed. Let Ti be the output power of the i-th individual generation system that constitutes the hybrid system per Wh and n be the total number of the generation systems in the hybrid system. Also let Ei be the generation capacity of the i-th individual generation system. The total cost per Wh, Ttot can be expressed as follows. (13) Where, Tbat is the cost of battery per Wh, and C is battery capacity

4

Simulation Results

Fig.1 shows the algorithm of the design of optimal combination. First, it assumes combination of the size of each component of the hybrid system. Using the given data, it calculates the total power generated by the hybrid system. The power is then compared with the power required by the load. During the process PLP is calculated and summed for total period T. Finally, RLP is calculated. If the resulting RLP satisfies the required RLP, the assumed combination of the size is a candidate for optimal combination. Among the many candidates, it finds optimal combination by applying the economical model. The major new and renewable energy sources, such as wind, PV and tide energy, are selected for the simulation. Fig.2 through Fig. 4 shows wind data, solar irradiance data, and tidal current data respectively. The data were acquired for 3 days at somewhere in Jeju island. Design condition is shown in table 1. The load is assumed as 20 Wh and operates for 24 hours a day. The required RLP is 0, which meet the stand-alone system that need no external supply of energy. For given design conditions the algorithm finds optimal combination of the hybrid system that meet the lowest cost for the generation of the power required by the load. Table 2 shows the results of optimal sizing for various combinations of hybrid system. It also shows the total generation cost for each combination. The cost of the wind-tide combination is higher most than those of the other combinations. On the other hand, the optimal combination that satisfies the lowest cost is the wind-PV-tide hybrid system. Combination is the lowest cost of hybrid system. The algorithm yields only one combination for the optimum solution, where the cost of Wh energy is a minimum. These results can be different when the cost of Wh for each component is changed. However, the algorithm can still find the optimal solution with consistency.

Design of Optimal Combination for New and Renewable Hybrid Generation System

195

1. Input wind, solar, current data 2. for j=1 to N ► N is the assumed number of sets for valid combinations 3. Assume combination (Ew,EPV,Et) 4. v_n 0 ► v_n : the total number of valid combinations 5. for i=1 to total period ( total number of data set) 6. Ptot Eq.(8) 7. if Ptot == Pload then 8. PLP 0 9. end 10 if Ptot > Pload then 11.C(t) Eq.(6) 0 12. PLP 13. end 14. if Ptot < Pload then 11. C(t) Eq. (7) Eq.(10) 12. PLP 13. end Fig. 1. Optimal design algorithm

10 9 8 7 6 5 4 3 2 1 0 1

6

11

16

21

26

31

36

Fig. 2. Wind data

41

46

51

56

61

66

196

K.H. Park et al.

Fig. 3. Current data

1200 (j/M2) 1000 800 600 400 200 0 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 (t) Fig. 4. Irradiance data Table 1. Design Conditions

Cost Type

Unit size

Total size (Won/kWh)

Wind

5W

To be designed

105

PV

30 W

//

640

Tide

5W

//

201

Battery

5 Wh

//

320

Wind

5W

20 Wh

-

Design of Optimal Combination for New and Renewable Hybrid Generation System

197

Table 2. Results of optimal combinations Size

cost

remarks

combination wind (W) PV(W)

battery(Wh)

(won)

-

145

95.3

-

tide(W)

Wind-PV

100

60

Wind-Tide

5

-

270

135

100.43

-

PV-Tide

-

30

105

95

71.65

-

Wind-PV-Tide

40

30

45

80

58.45

optimal

5

Conclusion

In this paper, method for design of optimal combination for hybrid renewable generation system has been studied. A simple mathematical modeling for typical new and renewable energy resources such as wind, PV, tidal energy resources were developed. Using the models, the method of designing the optimal combination of the hybrid system was developed. The method is based on RLP (Ratio of Lack of Power) and economical model. This aims at finding the configuration, among a set of system components that meets the desired system requirements, with the lowest value of the energy cost. The method was applied to hybrid wind-PV-tide generation system. The availability of the methodology was successfully demonstrated with the field data acquired from sets of experiments. Acknowledgement. Following are results of a study on the "Human Resource Development Center for Economic Region Leading Industry" Project, supported by the Ministry of Education, Science & Tehnology(MEST) and the National Research Foundation of Korea(NRF).

References 1. Kellogg, W., Nehrir, M.H., Venkataramanan, G., Gerez, V.: Optimal unit sizing for a hybrid PV/wind generating system. Electric Power System Research 39, 35–38 (1996) 2. Chedid, R., Saliba, Y.: Optimization and control of autonomous renewable energy systems. Int. J. Energy Res. 20, 609–624 (1996) 3. Karaki, S.H., Chedid, R.B., Ramadan, R.: Probabilistic performance assessment of autonomous solar-wind energy conversion systems. IEEE Trans. Energy Conv. 14, 766–772 (1999) 4. Bagul, A.D., Salameh, Z.M., Borowy, B.: Sizing of stand-alone hybrid PV/wind system using a three-event probabilistic density approximation. Solar Energy 56, 323–335 (1996) 5. Musgrove, A.R.D.: The optimization of hybrid energy conversion system using the dynamic programming model – RAPSODY. Int. J. Energy Res. 12, 447–457 (1988)

198

K.H. Park et al.

6. Yokoyama, R., Ito, K., Yuasa, Y.: Multi-objective optimal unit sizing of hybrid power generation systems utilizing PV and wind energy. J. Solar Energy Eng. 116, 167–173 (1994) 7. Yang, H.X., Burnett, J., Lu, L.: Weather data and probability analysis of hybrid photovoltaic wind power generation systems in Hong Kong. Renewable Energy 28, 1813–1824 (2003) 8. Beyer, H.G., Langer, C.: A method for the identification of configurations of PV/wind hybrid systems for the reliable supply of small loads. Solar Energy 57, 381–391 (1996) 9. Protogeropoulos, C., Brinkworth, B.J., Marshall, R.H.: Sizing and techno-economical optimization for hybrid solar PV/wind power systems with battery storage. Int. J. Energy Res. 21, 465–479 (1997) 10. Diaf, S., Diaf, D., Belhamel, M., Haddadi, M., Louche, A.: A methodlogy for optimal sizing of autonomous hybrid PV/wind system. Energy Policy 35, 5708–5718 (2007) 11. Nelson, D.B., Nehrir, M.H., Wang, C.: Unit Sizing of Stand Alone Hybrid Wind /PV/Fuel Cell Power Generation Systems. IEEE Power Engineering Society General Meeting 3, 2116–2122 (2005)

Parameter Optimization of UWB Short Range Radar Detector for Velocity Measurement in Automobile Applications Purushothaman Surendran1, Chul-Ung Kang1, and Seok Jun Ko2,* 1

The authors are with Jeju National University 2 Jeju National University Jeju, Korea [email protected]

Abstract. For designing UWB-SRR detectors we must understand the required design parameters and specific algorithm for target detection. In this paper, we optimize the parameters such as the number of coherent integration, pulse repetition interval, Doppler frequency resolution and FFT measurement time of UWB Short Range Radars for measuring the relative velocity of the target in automobile applications. The proposed detector with optimized parameters can measure the target minimum relative velocity of about 6.99 km/hr in very short measurement time. The minimum number of FFT points required to process Doppler Frequency by Fast Fourier Transform (FFT) is 32 points. The detection is based on one transmitted pulse against a background of white Gaussian noise (WGN). The performance of the proposed detector with optimized parameters is analyzed and simulation has been done in order to verify. Keywords: FFT, Velocity resolution, Coherent integration, UWB-SRR Detector, Doppler Frequency, Pulse Repetition Interval.

1

Introduction

In automobile industries, the vehicle safety has improved in the last decades with the increase in new safety technologies. The 24 GHz Short Range Radars are mainly used to give information such as range and velocity of the target to the driver and in some cases they are connected to a computer that performs some guiding actions to reduce collision and minimize the injuries. The source for target detection is the radar signals reflected by the target that is a mixture of noise and varied signals. The designed system must provide the optimal technique to obtain the desired target detections and to measure the relative velocity between the radar and target. The preferred detection can be determined by using specific algorithm for measuring the energy of the received signals. For this reason radar systems in 24 GHz have good performance in range and velocity measurement therefore they can be applied in different automobile applications like parking-aid, pre-crash detection [1]. *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266 , pp. 199–207, 2011. © Springer-Verlag Berlin Heidelberg 2011

200

P. Surendran, C.-U. Kang, and S.J. Ko

Previous works [4]-[5] has been focused on influence of increasing range resolution on detection ability of targets and also designed the algorithm with Fast Fourier Transform (FFT) method mainly. In this paper, we present the parameters optimization such as the number of coherent integration, the number of FFT points and the velocity resolution by using very narrow pulse width such as UWB signal. Our paper is made on the assumption of single moving target and shows the results by using the Monte Carlo simulation. The organization of this paper is as follows. In Section 2, the system model is described. In Section 3, we propose the detector for velocity measure. In Section 4, the parameter optimization method is expressed. In Section 5, simulation result is shown. Finally, conclusion is presented in Section 6.

2

System Description

The block diagram of a UWB radar system as shown in figure 1 is split in to two parts, the transmitter part and the receiver part.

Fig. 1. Block diagram of UWB radar system

In the transmitter part, the pulses are initiated by the Pulse Repetition Frequency (PRF) generator which triggers the pulse generator which in turn generates Gaussian pulses with sub-nano second duration as shown in figure 2. The Pulse Repetition Interval (PRI) is controlled by the maximum range of the radar. The maximum range for unambiguous range depends on the pulse repetition frequency and can be written as follows Rmax =

c 2 ⋅ f PRF

(1)

Parameter Optimization of UWB Short Range Radar Detector

201

where fPRF is the pulse repetition frequency and c is the velocity of light. And the range resolution can be written as ΔR =

c ⋅ TP 2

(2)

where Tp is the pulse width and c is the velocity of light. Then the transmitted signal can be written as follows s (t ) = AT ⋅ cos( 2πf ct + ϕ 0 ) ⋅

+∞

 p (t − n ⋅ T

PRI

)

(3)

n = −∞

where p(t) is the Gaussian pulse as follows   t  p (t ) = exp − 2π   τ    p 

2

   

(4)

And the parameters employed in this UWB radar system are described as follows; AT is the amplitude of single transmit pulse, φ0 is the phase of the transmit signal, fc is the carrier frequency, and TPRI is the pulse repetition time.

Fig. 2. Transmitted signal and received baseband signal

Since the range resolution of the UWB radar system is much less than the extent of the target, the echo signal is the summation of the time-spaced echoes from the individual scattering centers that constitute the target [3]. Therefore, in this paper, we can assume that the target has L independent reflecting cells. The target model is written as L −1

h(t ) =  α l ⋅ δ (t − τ l )

(5)

l =0

where the number of scatters L, the amplitude of the scatters αl, and the time delays of the scatters τl are all unknown. If the target is moving with relative velocity ν [km/hr]

202

P. Surendran, C.-U. Kang, and S.J. Ko

between the radar and the target, then the baseband complex received signal r(t) is written as r (t ) = AT

+∞ L −1

 α

l

⋅ e jθl p (t − nTPRI − τ l ) + n(t )

(6)

n = −∞ l =0

where n(t) is the complex additive white Gaussian noise (AWGN) with two-sided power spectral density N0/2 and θl is the arbitrary phase of l-th scatter that can be written as θl = -2πfcτl+ϕ0. The sampling rate is same to the pulse width and the wavelength λ is c/fc and c is the velocity of light, the Doppler shift is denoted as ωd = ±4πν /λ = ±4πνfc /c. In the Doppler shift, the positive sign (+) indicates the closing target and the negative sign (-) means the receding target.

3

Proposed Detector

In this section, we propose the algorithm for measuring the relative velocity of the target. First, as shown in Figure 3, we can detect the relative velocity of the target by performing DFT. Typically DFTs are performed with Fast Fourier Transform (FFT). The proposed detector consists of coherent integration, Discrete Fourier Transform (DFT) algorithm and square law detector. The sampling frequency is set depending on the pulse width, the baseband received signal is sampled in an in phase (I) and quadrature (Q) channel at every Tp, and the sampling rate Tp is same to the pulse width of 2ns and the range resolution can be 30cm from (2) and also it is assumed that the maximum target range can be 20m by using (1). From the above mentioned range resolution and maximum target range, the range gates of atleast 67 are required to detect the target range and velocity. Therefore the range gates are equal to the number of memory in the coherent integration. The sampled valued at every TP is applied to switch I and Q of the coherent integration. The switch I is shifted at every Tp sample, i.e., the range gate. It takes N·TPRI to coherently integrate and dump for all range gates. The coherent integration for the i-th range gate in I branch can be expressed as follows X I (i ) =

1 Nc

Nc

 Re {r n =1

n

( iT P ) }

(7)

where L−1

rn (iTP ) = AT αl ⋅ e jθl p((nTPRI + iTP ) − nTPRI −τ l ) + n' (nTPRI + iTP )

(8)

l =0

The DFT formed for a range cell provides a direct measure of the Doppler frequency. The result for one complete measurement cycle is a matrix of range gates and Doppler frequencies. Detection statistics (such as square law detection statistic) are performed for DFT outputs according to the prescribed detection strategy.

Parameter Optimization of UWB Short Range Radar Detector

203

Fig. 3. Block Diagram of the proposed detector

The DFT algorithm performed by FFT operates at every N·TPRI and the 32 point output of the DFT algorithm is a complex output which is squared and added can be represented as Y (i ) =

4

1 N

 {X N

n =1

I

( i ) e − jΩ n

} + {X 2

Q

(i ) e − j Ω n

}

2

(9)

Parameter Optimization Method

The detector of UWB radar must determine that a signal of interest is present or absent, and then the UWB radar processes it for some useful purpose such as range determination and velocity measure. In this paper, we propose the importance of optimizing various design parameters such as the number of coherent integration, sampling time, velocity resolution, Doppler frequency resolution, FFT measurement time and the required FFT length. The most important part of radar receiver is to understand how the ultra-short pulses are sampled in resolution (range-Doppler) cell. The required sampling frequency depends on the Doppler frequency range which corresponds to the relative velocity between the detector and the target. The minimum Doppler frequency resolution that can be achieved is given by

204

P. Surendran, C.-U. Kang, and S.J. Ko

Δf D = (1 / TFFT ) / N

(10)

where TFFT is the sampling time of the FFT input which is the product of number of coherent integration (Nc) and pulse repetition interval (TPRI). And N is the number of FFT points TFFT = N c × TPRI

(11)

The relation between Doppler frequency and relative velocity is fD = −

2vf c c

(12)

where v is the relative velocity and fc is the transmitter frequency. Thus the velocity resolution that can be achieved from the minimum Doppler frequency is Δv = −

Δf D ⋅ c 2 ⋅ fc

(13)

The required FFT measurement time also depends on the velocity resolution as shown in the figure below, For a FFT length of 32 points the FFT measurement time is given by TFFT measure = N × TFFT

(14)

If the number of FFT points increases then the measurement time also increases correspondingly.

5

Computer Simulation Result

The purpose of the simulation is to find the relative velocity between the UWB detector and the target. In the simulation we assume that the total energy reflected from the target is 1 and the system parameters are used as mentioned in table I. The signal-to-noise ratio (SNR) is defined as Ē/N0, where Ē is the total average energy reflected from the target. Fig. 6 shows the simulation result of a single moving target in frequency domain against a background of additive white Gaussian noise (AWGN). Fig. 7 shows the variance (dB) versus the Signal-to-noise ratio (Ē/N0) in time domain and frequency domain. A large enough number of trials are performed to obtain the variance at various Ē/N0. The number of trials is about 1000000 times. We can predict that the variance (which is proportional to noise power of AWGN) in frequency domain is less than the variance in the time domain, therefore the performance of the detector will be superior in frequency domain than compared to time domain.

Parameter Optimization of UWB Short Range Radar Detector 650

FREQUENCY RESOLUTION (Hz)

600

COHERENT INTEGRATION = 200

550 500 450 400 350 300 250 200 150 100 50 0 0

16

32

48

64 FFT POINTS

80

96

112

128

14

16

Fig. 4. FFT points Vs Doppler frequency resolution

MEASUREMENT TIME (ms)

250

200

150

100

50

0 0

2

4

6 8 10 12 VELOCITY RESOLUTION (Km/Hr)

Fig. 5. Velocity resolution Vs measurement time Table 1. System Parameters Parameters

Notation

Value

Pulse Repetition Interval

TPRI

500ns

Pulse Width

Tp

2ns

Maximum Target Range

Rmax

20m

Range Resolution

ΔR

30cm

Number of Coherent Integration

Nc

200

Velocity Resolution

Δv

6.9948 km/hr

Number of FFT points

N

32

205

206

P. Surendran, C.-U. Kang, and S.J. Ko

Fig. 6. 3D Graph of Detector output

1 VARIANCE_TIME_DOMAIN

0.1

VARIANCE_FREQUENCY_DOMAIN

VARIANCE (dB)

0.01

0.001

0.0001

0.00001

0.000001 -15

-12

-9

-6

-3

0

3

Ē/N0

Fig. 7. Ē/N0 Vs Variance

6

9

12

15

Parameter Optimization of UWB Short Range Radar Detector

6

207

Conclusion

In this paper, we have optimized the design parameters such as coherent integration, pulse repetition interval, target velocity resolution and the FFT measurement time of Ultra Wide Band Short Range Radar (UWB-SRR) detectors in Automobile applications. Thus the optimized parameters enhance the performance of the proposed detector, so that the target relative velocity of up to 6.9948 km/hr can be detected in very short FFT measurement time of about 3.2ms by using minimum FFT points of 32 efficiently. Finally, we show that the noise power is reduced in frequency domain when compared with the time domain thus increases the detection probability of the detector. Therefore the proposed detector is memory and time efficient detector for automobile applications which is the critical problem in conventional FFT method. Acknowledgments. This work was supported by the Korean Ministry of education & Science Technology, 331-2007-1-D00265.

References [1] Strohm, K.M., et al.: Development of Future Short Range Radar Technology. In: Radar Conference (2005) [2] Taylor, J.D., et al.: Ultra-Wideband Radars Technology: Main Features Ultra-Wideband (UWB) Radars and differences from common Narrowband Radars. CRC Press (2001) [3] Taylor, J.D., et al.: Introduction to Ultra-Wideband (UWB) Radars systems, Boca Raton, FL (1995) [4] Klitz, M.: An Automotive Short Range High Resolution Pulse Radar Network. Ph. D. Dissertation (January 2002) [5] Minkler, G., Minkler, J.: The Principles of Automatic Radar Detection In Clutter: CFAR. Magellan Book Company (1990) [6] Proakis, J.G.: Digital Communications. McGraw-Hill (2001) [7] Surendran, P., Ko, S.J., Kim, S.-D., Lee, J.-H.: A Novel Detection Algorithm for Ultra Wide Band Short Range Radar in Automobile Application. In: IEEE VTC 2010. Springer, Heidelberg (2010) [8] Skolnik, M.I.: Introduction to radar systems, 3rd edn. McGraw-Hill (2001) [9] Oppermann, I., Hamalainen, M., Iinatti, J.: UWB Theory and Applications. John Wiley & Sons Ltd (2004)

Data Signature-Based Time Series Traffic Analysis on Coarse-Grained NLEX Density Data Set Reynaldo G. Maravilla Jr., Elise A. Tabanda, Jasmine A. Malinao, and Henry N. Adorna Department of Computer Science (Algorithms and Complexity Lab) University of the Philippines, Diliman, Quezon City 1101, Philippines {jamalinao,hnadorna}@dcs.upd.edu.ph

Abstract. In this study, we characterize traffic density modeled from coarse data by using data signatures to effectively and efficiently represent traffic flow behavior. Using the 2006 North Luzon Expressway Balintawak-North Bound (NLEX Blk-NB) hourly traffic volume and time mean speed data sets provided by the National Center for Transportation Studies (NCTS), we generate hourly traffic density data set. Each point in the data was represented by a 4D data signature where cluster models and 2D visualizations were formulated and varying traffic density behaviors were identified, i.e. high and low traffic congestions, outliers, etc. Best-fit curves, confidence bands and ellipses were generated in the visualizations for additional cluster information. We ascertain probable causes of the behaviors to provide insights for better traffic management in the expressway. Finally, from a finer-grained 6-minute interval NLEX Blk-NB density data set, the coarser-grained hourly density data set were validated for consistency and correctness of results. Keywords: Data Signatures, Traffic Density Analysis, North Luzon Expressway, Non-Metric Multidimensional Scaling.

1

Introduction

Previous traffic behavior studies dealt only with volume analysis[2]. If we are to consider congestion, density is an accurate indicator. Density considers the occupied space of the road and the speed of the vehicles and it can give a better estimate of the real behavior of the traffic flow. Expressways, most of the time, should exhibit very low densities and high speeds, but spikes in the density graphs of 2006 North Luzon Expressway Balintawak - North Bound (NLEX BlkNB) segment data show otherwise. Domain experts identified inconsistencies and pointed out that the outliers determined are unrealistic for expressways. The study aims to show that the proposed density model is effective in estimating the traffic behavior of NLEX, with emphasis on u being the space mean speed and not time mean speed. The data set recorded and provided by National Center for Transportation Studies (NCTS) is in time mean speed. The data set, T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 208–219, 2011. c Springer-Verlag Berlin Heidelberg 2011 

Data Signature-Based Time Series Traffic Analysis

209

therefore, will be preprocessed to produce and represent realistic characterizations of traffic flow in NLEX. With the densities produced by the model from a study recently conducted[1], we will analyze the traffic flow by building a model for hourly traffic space mean speed and volume in NLEX. Data signatures will then be produced to represent the hourly traffic density data points. These data signatures and the time-domain data set cluster model will be visualized using the Non-Metric Multidimensional Scaling[11] and Data Images, respectively. Then, the intercluster and intracluster relationships of these data points will be examined. Data set outliers and potential outliers will be identified and analyzed using the methods in [2,8]. We validate our results in the hourly density data with the 6-minute data set. Section 1.1 discusses the definitions and basic notations used in this paper. Section 2 shows the the step by step building of the density-based model for the NLEX Blk-NB segment. This section also shows how the density data set is represented as a data signature and further visualized using nMDS, and Data Image. The results of these steps are discussed in Section 3. Finally, the conclusions and recommendations for further studies are in Section 4. 1.1

Definitions

The Data Sets. The data set provided by NCTS in this study on the NLEX Blk-NB segment in the year 2006 is periodic. The first data set contains hourly time mean speed and mean volume of the said segment. The second data set contains 6-minute time mean speed and mean volume of the said segment. The data set is preprocessed in a previous study in which average time mean speeds must meet the minimum speed requirement of 40 kph. Eleven weeks are eliminated for the first data set, leaving us with 41 weeks at 168 hours each. To maintain consistency, eleven weeks are also eliminated for the second data set, leaving us with 41 weeks at 1680 hours each. Traffic Flow 1. Volume q. Volume is the hourly mean of the number of vehicles per lane. 2. Time Mean Speed ut . Time mean speed is the mean of the speeds ui of the n vehicles  npassing through a specific point within a given interval of ui time, i.e. ut = i=1 . n 3. Space Mean Speed us . Space mean speed is the speed based on the average travel time of n vehicles in the stream within a given section of road, i.e. us =  nn ui . i=1 4. Density k. Density k is the number of vehicles over a certain length of a road. It is estimated as in k = uqs . Space mean speed is used in estimating the density because it considers the space between the vehicles.

210

R.G. Maravilla Jr. et al.

5. Estimation of Space Mean Speed from Time Mean Speed. Since the data set provided contains only the time mean speed and space mean speed is required in determining density, we estimate the space mean speed from the time mean speed using Rakha-Wang equation[4] to get us , where σ2 us ≈ ut − utt There will be a 0 to 1 percent margin of error in the estimation. Data Signature. A data signature, as defined in [6] is a mathematical data vector that captures the essence of a large data set in a small fraction of its original size. These signatures allow us to conduct analysis in a higher level of abstraction and yet still reflect the intended results as if we are using the original data. Various Power Spectrum-based data signatures[7,8] had been employed to generate cluster and visualization models to represent periodic time series data. Fourier descriptors such as Power Spectrums rely on the fact that any signal can be decomposed into a series of frequency components via Fourier Transforms. By treating each nD weekly partitions in the NLEX BLK-NB time-series traffic volume data set[8] as discrete signals, we can obtain their Power Spectrums through the Discrete Fourier Transform(DFT) decomposition. Power Spectrum is the distribution of power values as a function of frequency. For every frequency component, power can be measured by summing the squares of the coefficients ak and bk of the corresponding sine-cosine pair of the decomposition and then getting its square root, where k = 0, 1, . . . , n − 1. The Power Spectrum Ak of the signal, k = 0, 1, . . . , n − 1 is given by, Ak = ak 2 + bk 2 . Studies have shown that the set {A0 , A7 , A14 , A21 } is an optimal data signature for both visualization[7] and clustering [9]. Methods in [9] validate the optimality of the 4D data signature by showing an improved Dunn-like index. The 4D data signature used for clustering achieved statistical competence among all other data signatures. The study achieved ≈ 97.6% original data reduction for production of an optimal cluster model for Dunn-like variables. Data Visualization. In this study, we incorporate two methods, namely NonMetric Multidimensional Scaling and Data Images to analyze the data set. The first method projects the 4D signatures into a simpler 2D visualizations for traffic analysis. The second one is used to present the time-domain traffic density data. 1. Non-Metric Multidimensional Scaling. Non-Metric Multidimensional Scaling (nMDS)[11] is another visualization technique that maps multidimensional data set onto a 2D space. It computes the dissimilarity of the data points using Euclidean distance, Correlation, Mahalanobis, and other distance measures discussed in the literature[11]. nMDS includes a minimization of the stress or loss function to determine an optimal projection of the points in the Euclidean space given the known relationships in the higher dimension.

Data Signature-Based Time Series Traffic Analysis

211

2. Data Image. Data image is a graphical representation that transforms the given multidimensional data set into a color range image. Observations are made through the colors’ given characteristics and respective magnitudes. In our given data set, weeks are represented by the y-axis arranged by their cluster membership and days by the x-axis (with 1 as Sunday, 2 as Monday, and so on). The weeks are arranged according to their clusters. Clusters are determined by using the X-Means Clustering algorithm[5] that takes the 4D data signatures of the weeks in the data set as its input. 3. Confidence Intervals. Confidence Interval is a statistical analysis on given data sets to determine the reliability of the estimated population parameter. – Best Fit Curve. Linear, quadratic, cubic, quartic, and quintic curves are generated and fitted to their respective clusters. Root Mean Squared Deviation (RMSD) formula is applied to the curves to determine the best fit curve. RMSD gets the difference of the actual nMDS y values observed and the predicted y values (ˆ y ) of the curve model. The one with the lowest value will then be the best fit curve. – Confidence Bands. From the constructed best fit curve, the confidence √  SS band is extended above and below the curve by c DF tα (DF ), where c = G|x × Σ × G |x, G|x is the gradient vector of the parameters at a particular value of x, G |x is the transposed gradient vector, Σ is the variance-covariance matrix, SS is the sum of squares for the fit, DF is the degrees of freedom, and tα (DF ) is the student’s t critical value based on the confidence level α and the degrees of freedom DF . – Confidence Ellipse. A confidence ellipse, as defined in [12], uses intervals for both X and Y values of the scatterplot. The interval is projected horizontally and vertically respectively. The confidence ellipse is formed using the equation Z ± R × I, where Z is the mean of either X or Y , R is the range of either X or Y , and I is the confidence level 1 − α. 4. Potential Outliers. Potential outliers, as previously defined in [8,12] are points projected “near”or at the periphery of a region occupied by its cluster in the 2D visualization. – Absolute Potential Outliers. An absolute potential outlier is a data point that lies outside the confidence band and ellipse of its respective cluster. This point is not represented by its cluster’s best fit curve. – Valid Potential Outliers. A valid potential outlier is a data point that lies outside the confidence ellipse, but lies within the confidence band of its cluster and is still represented by the best fit curve. – Ambiguous Potential Outliers. An ambiguous potential outlier is a data point that is bounded by either two confidence bands or two confidence ellipses of different clusters or is inside a confidence ellipse, but outside of its cluster’s confidence band.

212

2 2.1

R.G. Maravilla Jr. et al.

Methodology Building an Effective Model from Sparse Data Points

1. From the preprocessed data set, we extract the mean volume and time mean speed per hour. 2. We estimate the space mean speed from the time mean speed by first getting the variances among the time mean speeds of the four lanes. We apply the Rakha-Wang equation to get the space mean speed per hour of the segment. To maintain consistency, the computed space mean speeds undergo preprocessing to eliminate values that are below 40 kph. 3. We estimate the density k using the given mean volume and space mean speed per hour. 4. To validate the produced hourly density data, steps 1-3 for modeling hourly traffic density are also conducted for modeling 6-minute traffic density. 2.2

Data Signature-Based Cluster and Visualization Models of the NLEX Traffic Density Model

1. Given the hourly density data generated from the previous section, the values of each week is transformed from its time domain to its frequency domain representation through the Discrete Fourier Transform and generate its 168D (and 1680D) Power Spectrum values for the hourly (and 6-minute interval) data set. Then, a 4D data signature is constructed from the Power Spectrum values of each week consisting of the components A0 , A7 , A14 , A21 . 2. Using all the data signatures of the weeks in the density data set as input to the X-Means clustering algorithm[5], we build the data set’s cluster model to identify groups of weeks that may have high, regular, and low traffic density (i.e. congestion) and pinpoint outliers and potential outliers in the model. We also pinpoint time frames where these various traffic behavior are identified. The traffic density analysis will be presented to the domain experts for their assessment. With the resulting assessment, we will provide suggestions to traffic control management for business-related decisions. 3. Produce the Non-Metric Multidimensional Scaling 2D visualization using the 4D data signature representations of each week in the data set, incorporating the results of X-Means clustering algorithms by coloring the 2D projections of the data signatures with respect to the assigned color information of their cluster. 4. Generate confidence bands, confidence ellipse, and best fit curve at 90% confidence interval per cluster to determine its set of potential outliers. 5. Visualize the traffic density values of the time domain data set where rows represent the values of each week, structured contiguously based on the clustering result, and each pixel is colored based on the actual values of the density in a time slot. Darkened lines separate the clusters and outliers from one another.

Data Signature-Based Time Series Traffic Analysis

213

6. We also perform the data signature visualization of the 6-minute density data set, producing this data set’s own set of nMDS scatterplot and Data Image. We then compare its clusters’ data points and Data Image with the hourly density data set’s. Using this data set’s nMDS scatterplot, we also construct confidence bands, confidence ellipse, and best fit curve per cluster to determine its own set of potential outliers. We compare the produced outliers of the 6-minute density data set with those of the hourly density data set. After validating the hourly density data set with the 6-minute data set, we perform analysis on this data set’s potential outliers to find their enclosing cluster’s and its elements’ behavior.

3 3.1

Results and Discussions Density Graphs of the Data Sets

From the preprocessed data, we computed the variances of the hourly time mean speed. The variances are consistent, but there are relatively high values of variance are found. It is because other lanes are congested during a specific hour, therefore, time mean speed variation is evident. From the computed variances, the hourly space mean speeds of the segment are produced. Spikes from processed space mean speeds are still observed, but they are relatively shorter than the spikes produced in the raw space mean speed graph. With the new set of space mean speeds (from 6888 hours, we now have 6880 hours), consistency in the density graph is expected. The calculated hourly and 6-minute densities from the mean volume and processed space mean speeds are shown in Figure 1(a) and 1(b), respectively. The graphs shows consistent values of density except on spikes where traffic incidents could have happened. The graphs show similar behavior of traffic density. Consistency in this matter shows us that the produced hourly density model is precise. To further validate the hourly density model, we then perform data visualization techniques in comparing the two data sets.

Fig. 1. Hourly Densities of the Segment

214

3.2

R.G. Maravilla Jr. et al.

Data Signature-Based Cluster Models

Hourly Data Set Visualization. The nMDS visualizations of the data signatures from the hourly traffic density data set is shown in Figure 2, respectively. The points which are reflected with the same symbols and coloring belong to the same cluster. The shown cluster model was generated using the data signatures of the hourly traffic density data set. Using the nMDS visualization, curves are fitted to each cluster’s data points. The best fit curves determined by RMSD are found to be linear. No curve was fitted for Cluster 3 due to the lack of data necessary in constructing the curve. The best fit curves are used to determine the confidence bands and confidence ellipse of each cluster. With the resulting best fit curves, confidence ellipses, and confidence bands, outliers and potential outliers are examined. The resulting visualizations with each cluster’s best fit curves, confidence bands and ellipse is shown in Figure 2.

Fig. 2. Data Signature-based nMDS Visualization of the Hourly Traffic Density Data

Cluster 4’s ellipse is not considered because it only has a few points and it covers all points of Cluster 2, making all these points ambiguous. Cluster 4 of the hourly density data set contains no potential outliers because it is relatively far from the points of other cluster, preventing them from being covered by other clusters’ ellipses. Cluster 3 is an outlier of the hourly density data set because it is the only has one data point. All potential outliers are found to be ambiguous since all points are covered by their own confidence ellipse. The details and evidences of the points being potential outliers in their clusters is found in the appendix.

Data Signature-Based Time Series Traffic Analysis

215

Table 1. Potential Outliers of Hourly Density Data Set’s Clusters Cluster Ambiguous Potential Outliers 0 Wk1, Wk8, Wk28, Wk29, Wk31, Wk35 1 Wk2, Wk3, Wk5, Wk7, Wk11, Wk33, Wk37 2 Wk13, Wk16, Wk18, Wk20, Wk23, Wk38, Wk45, Wk46, Wk47

The Data Image of the hourly density data set is also analyzed to determine the time frames of regular and irregular densities. The weeks of the hourly data set’s Data Image are arranged according to their clusters. Figure 3 shows the Data Image of the hourly traffic density data set.

Fig. 3. Data Image of the Segment’s Hourly Traffic Density Data

The densities that are consistent in their value with respect to the same day throughout the year are the regular densities. Irregular density values are inconsistent with respect to the regular values of their day. As seen in Figure3, week 18’s Day 2 (Monday) has irregular density because it has relatively lower density than the other Monday’s of the year. Week 15’s Day 4 (Wednesday) has higher density than the other Wednesday’s of the year, making it irregular. The weeks of Cluster 2 should be the time frames of regular density since Cluster 2 has the highest number of weeks among the clusters. But since incidents are inevitable, irregular densities can be observed in Cluster 2. The details describing these discovered irregularities are discussed in the appendix. 6-minute Interval Data Set Visualization. Using the DSIV tool, the data signature for the 6-minute density data set is produced and the clusters of the weeks are generated. The weeks of the hourly and 6-minute density data sets

216

R.G. Maravilla Jr. et al.

have the same cluster model except for two variations. Cluster 2 in the hourly model is split into two different clusters (Clusters 2 and 4). The weeks of Cluster 4 of the first model were separated as outliers in the second one. Thus, Clusters 3, 5, 6, 7, and 8 in the 6-minute density data set are all outliers. These divisions occurred because of the weeks’ high intracluster distance. 6-minute density data set has more points to consider than the hourly density data set. The Data Image of the 6-minute density data set is also generated. The weeks of the 6-minute data set’s Data Image are arranged in such a way that they follow the order of the weeks of the hourly data set’s Data Image. This is done for a more convenient comparison between the two Data Images.

Fig. 4. Data Image of the Segment’s 6-minute Traffic Density Data

Figures 3 and 4 exhibit similar behaviors. This further validates the accuracy of the hourly density model. With the validation, the hourly data set is sufficient in representing the whole traffic data set. From the data signatures of the 6-minute density data set, we produce its cluster model, nMDS visualizations, and the best fit curves, confidence bands and ellipses. The 6-minute density data set has more potential outliers than the hourly density data set. It can be ascertained that most of the potential outliers in the hourly model are also discovered as such in the 6-minute data, validating our initial result.

4

Conclusions and Recommendations

We have shown in this paper that data signature-based density analysis can provide an efficient and effective representation of traffic behavior. Using the space mean speed instead of time mean speed produces realistic results because it considers the rate of movement of vehicles within a given section. Density analysis, together with thorough preprocessing of the data set, produces an effective congestion indicator.

Data Signature-Based Time Series Traffic Analysis

217

With the data signature representation of the hourly density output data, analysis on traffic outliers can be conducted efficiently. With the same preprocessing and procedures done on the 6-minute density data set, comparison with the hourly density data set yielded similar results. Thus, the hourly density data set is validated and is accurate enough to be used in traffic congestion analysis. With the validation of a larger scaled data set, there are less data points to process, providing efficiency without compromising its accuracy. With the outliers and potential outliers determined by our study, expressway management can have an efficient analysis of traffic behavior that can be used in anticipating traffic flow patterns. Traffic incidents can be addressed more efficiently to reduce accidents and other traffic obstructions. Additionally, to come up with a more generalized behavior of the whole expressway, it is recommended that other NLEX segments be also analyzed. Acknowledgements. The researchers like to thank Dr. Ma. Sheilah GaabucayanNapalang and Dr. Jose Regin Regidor for validating our results and providing the data sets. This work is partially supported by a grant from DOST-PCIEERD through an ERDT project entitled Information Visualization via Data Signatures.

References 1. Maravilla, R., Tabanda, E., Malinao, J., Adorna, H.: Traffic Density Modeling on NLEX Time Series Data Segment. In: Proceedings of the National Conference for Information Technology Education (2011) 2. Malinao, J., Juayong, R.A., Corpuz, F.J., Yap, J.M., Adorna, H.: Data Signatures for Traffic Data Analysis. In: 7th National Conference on IT Education (2009) 3. Sigua, R.G.: Fundamentals of Traffic Engineering, 42–66 (2008) 4. Rakha, H., Wang, Z.: Estimating Traffic Stream Space-Mean Speed and Reliability from Dual and Single Loop Detectors (2005) 5. Pelleg, D., Moore, A.: X-means: Extending K-means with efficient Estimation of the Number of Clusters. In: Proceedings of the 17th International Conf. on Machine Learning (2000) 6. Wong, P., Foote, H., Leung, R., Adams, D., Thomas, J.: Data Signatures and Visualization of Scientific Data Sets. In: Pacific Northwest National Laboratory. IEEE, USA (2000) 7. Malinao, J., Juayong, R.A., Oquendo, E., Tadlas, R., Lee, J., Clemente, J., Gabucayan-Napalang, M.S., Regidor, J.R., Adorna, J.: Gabucayan-Napalang, Ma.S., Regidor, J.R., Adorna, J.: A Quantitative Analysis-based Algorithm for Optimal Data Signature Construction of Traffic Data Sets. In: Proceedings of the 1st AICS/GNU International Conference on Computers, Networks, Systems, and Industrial Engineering, CNSI 2011 (2011) 8. Malinao, J., Juayong, R.A., Becerral, J., Cabreros, K.R., Remaneses, K.M., Khaw, J., Wuysang, D., Corpuz, F.J., Hernandez, N.H., Yap, J.M., Adorna, A.: Patterns and Outlier Analysis of Traffic Flow using Data Signatures via BC Method and Vector Fusion Visualization. In: Proc. of the 3rd International Conference on Human-centric Computing, HumanCom-2010 (2010)

218

R.G. Maravilla Jr. et al.

9. Malinao, J., Tadlas, R.M., Juayong, R.A., Oquendo, E.R., Adorna, H.: An Index for Optimal Data Signature-based Cluster Models of Coarse- and Fine-grained Time Series Traffic Data Sets. In: Proceedings of the National Conference for Information Technology Education (2011) 10. Johnson, R.: Visualization of Multidimensional Data with Vector-fusion. IEEE Trans., 298–302 (2000) 11. Cox, T., Cox, M.: Multidimensional Scaling, 42–69 (1994) 12. Oquendo, E.R., Clemente, J., Malinao, J., Adorna, H.: Characterizing Classes of Potential Outliers through Traffic Data Set Data Signature 2D nMDS Projection. Philippine Information Technology Journal 4(1) (2011)

Appendix For the hourly traffic density data set, the we were able to find evidences of some of the points in the cluster model that we have produced as being potential outliers in Table 1, as shown below. We also state here the identified in traffic flow using Figure 3. We look for the events that triggered the irregular densities in the Data Image by focusing on the time frames of relatively high, relatively low, and extremely high density values. – Certain days of Cluster 0’s weeks 1, 8, 29, and 35 have relatively lower density than the common density for that particular day in the cluster. For example, week 1’s Sunday has lower density than other Sundays of the same cluster. On the other hand, certain days of this cluster’s weeks 28, and 31 have relatively higher density than the common density for that particular day in the cluster. – Some of Cluster 1’s weeks have similar behavior with Cluster 0’s weeks. This is because the confidence ellipses of Clusters 0 and 2 cover these weeks, making them ambiguous. However, these weeks belong to Cluster 1 because of distinct behaviors exhibited only by the weeks in Cluster 1. We can see this in week 2, where its density is relatively lower on Friday, but relatively higher on Saturday which is a characteristic of Cluster 1. – Cluster 2’s potential outliers have days that have a different density value than the common one. Week 13, for instance, has relatively lower densities on Friday and Saturday, as compared to its co-weeks in the cluster. – Extremely low densities are not classified specifically because this is a regular occurrence in expressways during midnight until early morning. – Extremely high density values are observed during important holidays in the country. This is reflected by the time frames with a dark red range. The weeks are weeks 15 and 44. The start of Holy Week happens on week 15 (Day 4) and the All Saints’ Day holiday happens on week 44 (Day 3). These are the days when people visit their provinces located at the north of Luzon. – Sudden increase of density values during days with consistent density values usually observed are irregular densities that are classified as relatively high density values. Reasons behind this sudden increase include accidents on weeks 12 (Day 6), 20 (Day 7), and 28 (Day 3) and departure for holidays on weeks 50 and 52.

Data Signature-Based Time Series Traffic Analysis

219

– Relatively low densities are observed on sudden decrease in densities the same way relatively high densities are observed on sudden increase in densities. We observe that the days after the highest density value occurred have relatively low density values. This is due to the large number of people departing at the same time at the official start of the vacation for the Holy Week. Majority of the people planning on a vacation have already left, leaving a few to depart on the following days (week 15’s days 5, 6, and 7 and week 16’s day 1). – During weeks 39 and 44, typhoons Milenyo and Paeng, respectively, struck the country. Travel advisory from weather domain experts prevented the people from traveling which is why there is a low density turnout on the said time frames. There is also a low density turnout during some holidays. Christmas Eve (week 32’s day 1) and New Year (week 1’s day 1) are observed in the country with people staying inside their houses to celebrate. Day 2 of week 18 also has relatively low density. This is due to the Labor Day holiday. Most of the professional drivers who pass by NLEX are on holiday. – The Data Image also reflects relatively low densities during Day 1 of weeks 4, 27, and 47. This is due to the many people watching the fights of the boxer Manny Pacquiao. Pacquiao-Morales 2 happened during week 4 (January 22, Philippine time). Pacquiao-Larios happened during week 27 (July 2, Philippine time). Pacquiao-Morales 3 happened during week 47 (November 19, Philippine time). For some of the potential outliers in the 6-minute density data set’s model which are not found in the hourly density data set’s, the following observations were derived. Cluster 0’s week 27 has relatively lower density on Sunday than the usual density value of Sundays in the same cluster. Week 43 of Cluster 2 and week 42 of Cluster 4 are both similar to each other’s cluster. The same goes for the majority of the potential outliers in both clusters. This is due to their clusters’ confidence ellipses enclosing each others points. This is further supported by the hourly data set’s clustering wherein both clusters’ points belong to one cluster (Cluster 2 of hourly data set).

Automated Video Surveillance for Monitoring Intrusions Using Intelligent Middleware Based on Neural Network Ana Rhea Pangapalan1, Bobby D. Gerardo1, Yung-Cheol Byun2,*, Joel T. De Castro1, and Francisca D. Osorio1 1

Institute of ICT, West Visayas State University Luna St., Lapaz, Iloilo City, Philippines [email protected] 2 Dept. of Computer Engineering, Jeju National University Jeju City, Korea [email protected]

Abstract. Automated Video Surveillance Using Intelligent Middleware presented a Java based system that detects human activities in a security sensitive area and provides alarm for illegal activities identified. The system was developed using Netbeans IDE 6.8 [7] as the working environment, while Java as the programming language [8]. This study enhanced and strengthened existing security, therefore minimizing possibility of missed events which might be a threat to an area. The system composes three major processes: Motion Detection, Subject Identification and Behavior Classification. Motion Detection captures image of any movement detected. Subject Identification screens every captured image by classifying whether the motion is made by human and eliminating those which are caused by wind, animals and other non-human entity. Behavior Classification categorizes the image passed as to what action and outputs alarm if it is considered as illegal. In order to carry out these complex functionalities, a middleware was utilized to maintain continuous data flow from capturing to image processing and to reduce bulk of inputs that are being processed. Neural network [9, 10] was employed as the information processing paradigm for human or non-human and behavior classification. The result shows that the system processed video continuously as it classified behavior automatically. Keywords: Intrusion detection, intelligent middleware, neural network, security.

1

Introduction

Security of human lives and property has always been a major concern for civilization for several centuries. Video surveillance systems have been widely used to solve issues arising from illicit human activities. As more and more surveillance cameras are deployed in a facility or area, the demand for automatic methods for video processing is increasing [6], since in a conventional system one important event could be missed by a simple distraction from the one in charge of monitoring and watching the video streams. *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 220–228, 2011. © Springer-Verlag Berlin Heidelberg 2011

Automated Video Surveillance for Monitoring Intrusions

221

Rapid developments from the past years to fully automate these systems have occurred since motion analysis in video has attracted many researchers. Smart Surveillance in itself is one of the most challenging problems in computer vision [1] due to its promising applications in many areas. With these developments, issues regarding the functional aspect of the automated systems arise. Common processes of these systems are motion detection, extraction of human body structure from images, tracking across frames, and behavior classification. In motion detection itself, different changes in input like fluctuating illumination and shadows can create movement therefore providing useless data for processing. Different algorithms have been developed to support complex processes and inconsistency from inputs due to changes in real world scenarios. From time to time, new algorithms are created and new methods are applied to overcome limitations of previous developed systems. These issues motivated the researchers in developing the proposed system which covered motion detection, subject identification and behavior classification. This paper documents all the information and procedures necessary to achieve the goals of the system.

2

Related Studies

In the study of Oh et al. [10] about View-Point Insensitive Human Pose Recognition Using Neural Network, they proposed a view-point insensitive human pose recognition using neural network. Their system consists of silhouette image capturing module, data driven database, and neural network. First, they capture 3D human model from different camera angle to generate a 2D silhouette image. Different poses from different camera are captured. Captures are stored in a database and will be used for neural network training. They use 2/3 features of all and 1/3 features are utilized as test data. The researchers used trained and non-trained data for testing. It was observed that the average precision is 75.3% at non-training data, and 81.24% at training data. The paper on ACE (Annotated Critical Evidence) Surveillance [5] presents an automated video surveillance technology that is developed by National Research Council of Canada (NRC) in 2008 which states that, in real-time monitoring mode, the problem is that an event may easily pass unnoticed due to false or simultaneous alarms and lack of time needed to rewind and analyze all potentially useful video streams. The system was developed for the purpose of enabling more efficient use of surveillance systems and one of its concerns is data storage and manageability for post-incident investigation. Developed by the Video Recognition Systems team of the National Research Council of Canada’s Institute of Information Technology (NRC), it was implemented as a software that runs on an ordinary desktop computer, performs real-time analysis of captured video streams for the purpose of automatically extracting and annotating Critical Evidence Snapshots, which are used to automatically alarm the system and which allow efficient summarization and browsing of captured video data.

222

A.R. Pangapalan et al.

On the other hand, a paper on Video-Based Human Motion Estimation System [11] came-up with the system designed to estimate body silhouette representation from sequences of images. Milanova and Bocchi in the same paper seeks to explore the hypothesis that the two building blocks of the accuracy of human motion estimation, the measured data and the prior model are critical components, using extremely high accuracy measured data and shape of body motion priors, so that the objective function is more precise and less noisy, resulting in an easier solution. Their main goal is to develop a new module for extracting accuracy measured data from video imagery. In a paper on Middleware for Distributed Video Surveillance [4], the researchers used middleware for the system and suggested the use of intelligent middleware for addressing the real-life challenges that include managing a large scale of cameras which will be difficult only when utilizing computer vision algorithms required in detecting and interpreting activity in video. Also, video surveillance systems manned by security personnel will be ineffective because even trained operators lose concentration and miss a large percentage of significant events. They utilized a middleware for support of network of cameras where the input is coming from. In the paper about Robust Techniques for Background Subtraction in Urban Traffic Video[2], explored the different background subtraction techniques in urban traffic video sequences have been compared and tested. For experiment, Cheung and Kamath [4] have selected four publicly-available urban traffic video sequences. Two algorithms that produce good performance are Mixture of Gaussian and Median Filtering. The MoG (Mixture of Gaussian) appears to have the best precision and recall and MF (Median Filtering) is a very close second. MoG still, has its own drawbacks. It is computationally intensive and its parameters require careful tuning and are very sensitive to global illumination. The researchers concluded that MoG produces best results while adaptive median filtering offers a simple alternative but still having a competitive performance. In [3, 12], the researchers build a motion decomposition approach to analyze the periodicity in human actions and also study about motion detection. In this paper [12], the researcher proposed a novel video compression idea to compress these extracted periodic activities from the videos. The method exploits the correlation between the frames over longer length of time, of the order of the period of the activity, as compared to the traditional video compression algorithms which use correlation between a few neighboring frames for motion prediction and compensation. He also considers the problem of silhouette normalization for activity analysis.

3

System Architecture

One of the goals of this study was to develop middleware that supported the simultaneous data flow from capturing to video processing. During the simulation and testing, the researchers utilized only one camera to minimize the size of input data that would be handled by the system for real time processing. The system was limited only to single, stationary cameras and video input; thus, pan, tilt and zoom features were not included. Another concern was that, from different area, varying behaviors were perceived to be illegal. For the scope to be visible, a list of different behaviors

Automated Video Surveillance for Monitoring Intrusions

223

was given and activities which were illegal from the area covered were selected, if not, the default settings would be followed where illegal activities were pre-defined. Due to limited knowledge in the language and the paradigm used, more time had been spent on researching causing time allocation to be affected therefore, creating some constraints on coding, debugging and polishing. 3.1

Framework

Figure 1 shows the foundation of the proposed system. It starts from a video stream that serves as the main input. From there motion detection initializes which is followed by the activation of neural networks. Take notice that middleware at that stage then subject detection and behavior classification until the alarm is generated takes place.

Alarm

Behavior Classification Subject Detection Middleware

Activation of Neural Network Motion Detection

Video Stream

Fig. 1. Framework of the proposed system

3.2

Architectural Design

Figure 2 shows the Architectural Design. The hardware included in the proposed system is a camera and a workstation/computer unit. The data being passed by a middleware includes the raw and the classified images. Those were being processed until the system outputs an alarm.

224

A.R. Pangapalan et al.

Fig. 2. Architectural Design of the Proposed System

3.3

Procedural Design

When the system is initiated, it will ask for a user name and a password for user authentication and it will give an option to start the surveillance. The user will then be prompted to choose a human behavior to be identified as illegal by the system. Through this, networks which correspond to the chosen behaviors will be activated. However, if the user fails to choose anything the network for all default illegal behaviors will be used by the system. The procedural design of the proposed system is presented in figure 3. Start

Input user

Yes

Start

Activate Networks

Surveillance

No Start

Fig. 3. Procedural Design of the Proposed System

Automated Video Surveillance for Monitoring Intrusions

4

225

Simulation and Results

As the system starts-up the user is required to input the password for authentication to start the surveillance. After that, user can define behaviors which are considered to be illegal in the area chosen for the system to deploy. A list is given and if the user fails to select, pre-defined/default human behaviors will be perceived as illegal and all neural networks will be activated. Output comes in the form of notification and if an instance of intrusion occurs. The log-in window and the action selector is shown in figure 4 and 5, respectively. Figure 5 also shows action that could be selected for activation of Neural Network. While launching of motion detection software is shown in figure 6.

Fig. 4. Log in window of the system

Fig. 5. Motion Selector of the system

226

A.R. Pangapalan et al.

Fig. 6. Launching of Motion Detection Software

Fig. 7. Sample Output for Identified Illegal Action

Whenever a human is detected, the middleware saved the image and passed it for behavior classification. Lastly, when an illegal behavior is detected, an alarm is generated and prompted as illustrated in figure 7.

5

Conclusions and Recommendations

The purpose of the proposed system was to offer security over an area without much need of human judgment and monitoring. It was planned, developed, tested and evaluated and was functional and working according to the processes. In one way or another, though limited, the system still achieved its goals and objectives since the researchers were able to identify the key processes which supported the complexity of an automated video surveillance system, determined the right algorithm that ensured efficiency, carried out real-time processing through the middleware, found ways to reduce the bulk of data needed and utilized only the inputs necessary to run the system.

Automated Video Surveillance for Monitoring Intrusions

227

Furthermore, it can also be concluded that the system served a good start and a foundation for a more sophisticated and polished security application. It shows that the limitations in terms of the system’s capacity can be solved since the idea and the basic concepts, processes and functions were achieved and made. For the improvement of the proposed system the researchers recommend that the future system should be capable of deleting previous captured and stored images to maximize space, place activation of networks, capturing and system alert in the same window for easier navigation and show detected image together with a sound alert. Acknowledgments. This research was financially supported by the Ministry of Education, Science Technology (MEST) and Korea Industrial Technology Foundation (KOTEF) through the Human Resource Training Project for Regional Innovation. This research is not possible without the help of the research team composed of Yssel Tarra Modina, Alvin Jorge Tambalo, Christy Villano, Joy Vista and some Faculty of the Institute of ICT at West Visayas State University.

References 1. Kosmopoulos, D.I., Antonakaki, P., Valasoulis, K., Katsoulas, D.: Monitoring human behavior in an assistive environment using multiple views. In: Proceedings of PETRA (2008) 2. Cheung, S.-C.S., Kamath, C.: Robust techniques for background subtraction in urban traffic video. In: Proceeding of SPIE, vol. 5308, p. 881 (2004) 3. Danjou, N.: Motion Detection (2006), http://noeld.com/programs.asp?cat=video#MDetect (retrieved March 1, 2010) 4. Detmold, H., van den Hengel, A., Dick, A., Falkner, K., Munro, D., Morrison, R.: Middleware for Distributed Video Surveillance. IEEE Distributed Systems Online 9(2) (2008) 5. Gorodnichy, D.O.: ACE Surveillance: The Next Generation Surveillance for Long-Term Monitoring and Activity Summarization. First International Workshop on Video Processing for Security (VP4S-2006), June 7-9, Quebec City, Canada 6. Kotikalapudi, U.K.: Abnormal event detection in video, Master’s Thesis, Supercomputer Education Research Center, Indian Institute of Science (2010), http://www.serc.iisc.ernet.in/graduation-theses/Uday.htm (retrieved March 1, 2010) 7. Java (2010), http://www.java.com/en/download/whatis_java.jsp (retrieved March 1, 2010) 8. NetBeans, http://netbeans.org/community/releases/68/relnotes.html (retrieved March 1, 2010) 9. Neural Network, http://www.merriam-webster.com/dictionary/neural%20network (retrieved March 1, 2010)

228

A.R. Pangapalan et al.

10. Oh, S., Lee, Y., Hong, K., Kim, K., Jung, K.:: View-point insensitive human pose recognition using neural network (2009), http://www.waset.org/journals/waset/v44/v44-50.pdf (retrieved 1 March, 2009) 11. Milanova, M.G., Bocchi, L.: Video-Based Human Motion Estimation System. In: HCI, vol. (11), pp. 132–139 (2009) 12. Varsha Chandrashekhar, H.: Human Activity Representation, Analysis and Recognition. (2006), http://www.security.iitk.ac.in/contents/publications/more/ human/activity/representation.pdf

SMS-Based Automatic Billing System of Household Power Consumption Based on Active Experts Messaging Mark Dominic Cabioc1, Bobby D. Gerardo1, Yung-Cheol Byun2, 1

2

*

Institute of ICT, West Visayas State University Luna St., Lapaz, Iloilo City, Philippines [email protected]

Dept. of Computer Engineering, Jeju National University Jeju City, Korea [email protected]

Abstract. The study about SMS-based Automatic Billing System of Power Consumption aimed to change the conventional way the power utility provider gathers and handles billing data. The system is composed of two basic parts such as the remote site and the base station. The former calculates and sends power consumption while the latter retrieves meter readings, calculates billing charges and processes payment of the customers. Microsoft Visual Studio 2008 was used to develop the Main Server Software with Visual Basic 2008 as the Integrated Development Environment (IDE) or programming tool and Visual Basic as the programming language. The database was created in Microsoft SQL Server 2005 and ActiveXperts Messaging Server 4.1 was use as an SMS framework that allows the system to send, receive and process SMS. The use of this system will give greater benefit to the electric company and its customers because of the ease and less impediment in gathering meter readings in remote locations and an instant delivery of billing statements to the customer’s cellular phones through SMS technology. Keywords: SMS messaging, automatic billing, Active Expert, power management.

1

Introduction

Conventional electricity billing system in the country has been lagging in terms of technology in gathering and processing power consumption data for billing purpose. It is done by an assigned person who visits each meter location periodically and reads the meter manually. Data collected are then processed either manually or via specialized software into the customers‟ bills which in turn are distributed manually to each respective customer. As such, conventional meter reading poses several problems wherein adverse effects extend to the customer himself. Misreading of the power meter for instance can *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 229–238, 2011. © Springer-Verlag Berlin Heidelberg 2011

230

M.D. Cabioc et al.

inadvertently increase or decrease the customer’s bill. The recent technology of automatic meter reading (AMR) [1] has solved the problems of conventional meter reading. AMR has promised fast and accurate collection of meter readings, however, incorporating this with manual computation of electric bills does not fully utilized its functionality. With the development of the text messaging in the country and the dependence of the people on it, the researchers took advantage of the GSM network and SMS technologies available to solve these inefficiencies. The convergence of SMS and GSM network allow more mobile and wireless applications to be implemented such as automatic meter reading and billing. Although many related projects have been presented in some other countries, the development and implementation of this study in the Philippines is not yet totally accomplished. The proposed SMS-based Automatic Billing System of Power Consumption will elaborate further on some existing technologies and studies that focus on Automatic Meter Reading and Remote Meter Reading since it will give more effort and attention in developing the billing system interface of power consumption. The development of the system will not only focus on the communication network but on the application side wherein a secure and reliable Automatic Billing System will be developed. The use of this system will give greater benefit to the electric company and its customers because of the ease and less impediment in gathering meter readings in remote locations and an instant delivery of billing statements to the customer's cellular phones.

2

Related Studies

The study on Fixed-Network Automatic Meter Reading (AMR) System [2] was a graduation project of Awad and Gosh presented to the Department of Electrical Engineering, University of Jordan. The project discusses about utilizing already available fixed communication networks (e.g., the cellular network) for exchanging data to minimize cost and human effort. The purpose of their project was to introduce a Fixed-Network AMR design that manages the reading of the electricity meters at the consumers' side. This design was intended to replace the existing manual methods of gathering data. The approach to the solution for this problem was made with the use of the GSM network and a custom RF solution. Different hardware modules were introduced to help exchange data between a central office and any node in the system (i.e., customer's side). The GSM and RF communication media were fully utilized by introducing their own GSM and RF protocols. A paper on Networked Remote Meter-Reading System Based on Wireless Communication Technology [4] studied meter-reading system based on Bluetooth wireless communication technology and GSM. The remote meter-reading system employs distributed structure, which consists of measure meters, sensors, intelligent terminals, management centre and wireless communication network. The intelligent terminal which designed based on embedded system and Bluetooth technology was used to realize acquisition information submitted from meters and sensors control the energy-consuming devices moreover in residence.

SMS-Based Automatic Billing System of Household Power Consumption

231

While in a paper on Real-time Energy Management over Power-lines and Internet [17], explore the creation of an infrastructure for energy management that should enable the enhancement of existing applications like automatic meter reading, distribution grid management, and remote control. The system allows for direct communication with equipment at the customers’ premises via a two stage hierarchical power line communication system and an IP (Internet Protocol)-based private network. A paper on Novel Approach for Remote Energy Meter Reading Using Mobile Agents [15] incorporated power or energy meter systems with embedded controllers such as micro web-servers with Ethernet port to transmit the reader data over the Internet. Such data can be then fed and integrated into existing energy management systems located at power companies and organizations. Mobile agents are executing programs that migrate during execution and present methods for maintaining and using distributed systems. However, the problem of efficiently collecting data from a large number of distributed embedded Web-servers in the energy meters is still a challenging problem. On the other hand, in [1] developed hardware structure consisting of a digital energy meter module, another digital meter for water and a telephone module, all lined with a single chip microcontroller were equipped with credit card reading capability to automatically read and charge the consumption on site. Also, all service metering modules were facilitated with an automatic service connection and disconnection based on the available credit. The software structure commands the whole process via the microcontroller input/output ports. Furthermore, the study of Jabundo et al. [9] about prepaid kilowatt-hour meter aimed at reducing certain problems such as pilferage and excessive use of electric energy in establishments. This system was composed of a customized digital power meter that was prepared for prepaid use. Occupants of the establishment will purchase energy load–equal to its corresponding energy based on the current energy rate–from the establishments’ owner. When the occupant totally consumes the energy load, the current will be cut off by the magnetic switch. Based on the reviewed literatures, the researchers are motivated to study an automatic billing system of power consumption that automatically acquires meter readings from a remote station, calculates billings and sends SMS bills to the customers. The proposed system used a digital power meter designed to send meter readings directly to the system through GSM network, an SMS Messaging Server as its middleware and develop a multi-level system that handles data collection, billing computation and payment processing.

3

System Architecture

Theoretically, the proposed system will help ease the collection of billing data, thus saving the utility provider and customer time, money, effort and inconvenience. It also eliminated the cost incurred by the traditional meter data collection process, thus lessening the burden cast upon the customers. The proposed automatic billing system is composed of a remote site and base station as shown in Figure 1. The digital meter that is located in the remote site (homes and buildings) will send the meter readings in the base station’s billing system via SMS using the stable GSM network for cellular phones.

232

M.D. Cabioc et al.

Fig. 1. The proposed System Architecture

3.1

Remote Site

The remote site is the digital power meter customized to receive queries and send its own meter reading to the base station via SMS. The device consists of a microcontroller unit, GSM modem, LCD display, current and voltage sensor attached to the main power supply.

Fig. 2. Architectural design of the remote site

As shown in Figure 2, the remote site’s current and voltage sensors receive energy that passes through the meter reader from the power line to the customer’s workload — appliances or other AC devices at a maximum of 2 amperes. The two sensors measure current and voltage respectively, which can be viewed from the meter reader’s LCD display.

SMS-Based Automatic Billing System of Household Power Consumption

233

Microcontroller unit calculates the input from the current and voltage sensors, converting the results into kilowatt per second, representing the customer’s power consumption. The resulting data are then sent to the GSM modem in the event the remote site receives an SMS request from the base station with this format:’check data’. The data are then processed into an SMS message in this format: “Meter ID, Reading”. The message is then sent to the server. 3.2

Base Station

The base station acts as the central server where all the commands and processing of data are executed. It is composed of another GSM modem, an SMS Messaging Server and the main server software.

Fig. 3. Architectural Design of the Base Station

As shown in Figure 3, GSM modem is used by the base station to communicate with the remote site by sending an SMS request to the customized digital meter, which sends back the current kWh reading in return. The GSM modem is also used to send billing information to the customer via SMS.

4

Simulation and Results

The SMS Messaging Server manages SMS messages sent and received by the GSM modem. Running in the background, it acts as the system’s middleware. Meanwhile, the Main Server Software is multi-level software that handles processes such as retrieval of meter readings directly from the remote station, computation of bills and processing of payments for the said bills. The software is composed of a database and data collection, computation and analysis, SMS billing statements generation and payment processing modules. The Main Server Software’s Database is connected to all the modules of the software. This is where data in the system is stored and retrieved. The Data Collection Module communicates data with the SMS Messaging Server. From there it

234

M.D. Cabioc et al.

retrieves relevant SMS messages such as meter readings to be saved in the database. The computation and analysis, module retrieves data from the database, calculates and analyzes the overall consumption and relays the information to the billing statements generation module.

Fig. 4. Input Window for Customer’s Details for Registration

Fig. 5. Assigning meter details for each customer

Fig. 6. Updating charge rates

SMS-Based Automatic Billing System of Household Power Consumption

235

The SMS billing statements generation module generates detailed reports of customers monthly bill consisting of the current and previous kWh readings, total kWh used, charge rates, amount due and due date. Billing statements are sent to the consumer’s cellular phone in SMS format. This is presented in Figure 7.

Fig. 7. Processing Customers’ Payment

On the other hand, the payment processing module is responsible for processing the payments of the customer. It also includes the handling of unpaid due amounts and other functions related to payment system.

Fig. 8. Billing Information (SMS) Received by the Customer

236

M.D. Cabioc et al.

The GSM network, when interfaced with an SMS gateway like ActiveXperts can be utilized as an effective data communications system. The ActiveXperts SMS gateway on the other hand, is good third-party software for this matter. It efficiently manages received data into databases and is capable of sending data via SMS. Figure 8 shows the SMS sent to the customer for its billing information.

Fig. 9. Customer’s Monthly Consumptions Report

On the other hand, Figure 9 shows the monthly power consumption of the customer. The system could also provide master list of the customers and as well as daily report of consumption and payment reports.

5

Conclusions and Recommendations

In this study, a user-friendly software application for the base station was created. We demonstrated the automation of the retrieval, storage and sending of data. In addition, the system’s software was able to compute the data acquired from the remote site and generate billing statements for the customer. A prototype digital meter that measures power consumption and sends data via SMS was successfully adopted and integrated into the system. Moreover, effective communication between the remote site and the base station was successfully established using the GSM network gateway. To further improve the efficiency of the software system, the researchers recommend that functions such as bulk sending of SMS billing information to the customers, one click computation of all customers’ monthly bills and receiving of inquiries from customers via SMS shall be included in the future study. Also the researchers recommend that energy consumers should monitor their power consumption by inquiring their own meter readings directly from the customized digital meter at any given time.

SMS-Based Automatic Billing System of Household Power Consumption

237

Acknowledgments. This research was financially supported by the Ministry of Education, Science Technology (MEST) and Korea Industrial Technology Foundation (KOTEF) through the Human Resource Training Project for Regional Innovation. We would like to express our gratitude to the efforts made by the research team composed of Elera Marie O. Joaquin, Denmark S. Padua, Jey Mark P. Palma, John Niño C. Requina, , Christian Roy S. Somcio, and some Faculty members of the Institute of ICT of West Visayas State University.

References 1. Al-Qatari, S.A., Al-Ali, A.R.: Microcontroller-based Automated Billing System (1995) 2. Awad, A.J., Abu Ghos, R.T.: Fixed-Network Automatic Meter Reading (AMR) System. Department of Electrical Engineering, University of Jordan. FixedNetwork-AutomaticMeter-Reading-AMR-System, http://www.scribd.com/doc/950863/ (retrieved January 2010) 3. Base station. In: TechTerms.com, The Tech Terms Computer Dictionary. from http://www.techterms.com/definition/basestation (Retrieved 2005) 4. Cao, L.: Networked Remote Meter-Reading System Based on Wireless Communication Technology. In: IEEE International Conference on Information Acquisition (2006) 5. Database: Dictionary of Computer and Internet Terms (10th ed.). Barron’s Educational Series, Inc., New York (2009) 6. Downing, D., Covington, M.: Dictionary of Computer and Internet Terms (10th ed.) Barron‟s Educational Series, Inc.,New York (2009) 7. GSM. In: VoIP Dictionary (2003), http://www.voipdictionary.com (retrieved January 2010) 8. GSM Modem: In Now SMS (2002). from http://www.nowsms.com/gsm%20modems.htm (retrieved January 2010) 9. Jabundo, D.J.R., et al.: Prepaid Kilowatt-hour Meter. Prepaid Kilowatt-hour Meter. College of Engineering, Central Philippine University, Iloilo City, Philippines (2010) 10. Microsoft, SQL. Tutorials, http://www.sqlcourse.com (retrieved from 2011) 11. Microsoft Visual Basic (2008), http://www.microsoft.com (retrieved 2008) 12. Ofrane, A., Harte, L.: Introduction to Wireless Billing. Usage Recording, Charge Processing, System Setup, and Real Time Billing 2006, http://www.billingdictionary.com/billing_dictionary_billing_ system_definition.html (retrieved January 2010) 13. Power Consumption. In: Dictionary Babylon (1997), http://www.dictionary.babylon.com/ (retrieved January 2010) 14. SMS: In Tech Terms, The Tech Terms Computer Dictionary (2005). SMS Messaging Server, http://www.techterms.com/definition/sms (retrieved January 2010) 15. Tahboub, R., et al.: Novel Approach for Remote Energy Meter Reading Using Mobile Agents. In: Third International Conference on Information Technology: New Generations, ITNG (2006) 16. Thiele, T.: Power Meter - How an Electrical System Works (2008), http://electrical.about.com/od/panelsdistribution/ss/elecsys works_3.htm (retrieved January 2010)

238

M.D. Cabioc et al.

17. Treyl, A., et al.: Real-Time Energy Management Over Power Lines and Internet. In: Proceeding of the 2003 Joint Conference of the 4th International Conference on Information and Communications and Signal Processing, 2003 Pacific Rim Conference on Multimedia, pp. 144–147 (2003) 18. Visual Basic Tutorials, http://www.msdn.microsoft.com (retrieved from 2011)

Hierarchical Clustering and Association Rule Discovery Process for Efficient Decision Support System Bobby D. Gerardo1, Yung-Cheol Byun2,*, and Bartolome Tanguilig III3 1

Institute of ICT, West Visayas State University Luna St., Lapaz, Iloilo City, Philippines [email protected] 2 Dept. of Computer Engineering, Jeju National University Jeju City, Korea [email protected] 3 Technological Institute of the Philippines, Cubao, Quezon City [email protected]

Abstract. This paper proposed a model based on hierarchical Clustering and Association Rule, which is intended for decision support system. The proposed system is intended to address the shortcomings of other data mining tools on the processing time and efficiency when generating association rules. This study will determine the data structures by implementing the cluster analysis which is integrated in the proposed architecture for data mining process and calculate for associations based on clustered data. The results were obtained using the proposed system as integrated approach and were rendered on the synthetic data. Although, our implementation uses heuristic approach, the experiment shows that the proposed system generated good and understandable association rules, which could be practically explained and use for the decision support purposes. Keywords: Data mining, decision support system, clustering, association rules.

1

Introduction

Often, there are many attributes or dimensions that are contained in the database, and it is possible that subsets of such dimensions are highly associated with each other. The dimensionality of a model is determined according to the number of input variables used. Clustering can be used to group data into clusters so that the degree of association is strong between members of the same cluster and weak between members of different clusters [1], [9]. Thus, each cluster describes the class to which its members belong. For that reason, cluster analysis can reveal similarities in data which may have been otherwise impossible to find. Data cubes allow information to be modeled and viewed in multiple dimensions and such cubes are then defined by the dimensions and facts [1]. They defined *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 239–247, 2011. © Springer-Verlag Berlin Heidelberg 2011

240

B.D. Gerardo, Y.-C. Byun, and B. Tanguilig III

dimensions as entities with respect to which an organization wants to keep records of. Data cubes may be used in theory to answer query quickly, however, in practice they have proven exceedingly difficult to compute and store because of their inherently exponential nature [7]. Moreover, issues that other researchers observed in the data mining tasks were computing speed, reliability of the approach for computation, heterogeneity of database, and vast amount of data to compute [1], [2], [7]. This paper explore the formulation of the cluster analysis technique as integrated component of the proposed model to partition the original data prior to implementation of other data mining tools. The model that we proposed uses the hierarchical nearest neighbor clustering method and apriori algorithm for association mining implemented on transactional databases.

2

Related Studies

Association rule mining tasks includes finding frequent patterns, associations, or causal structures among sets of items or objects in transactional databases and relational databases. Data mining uses various data analysis tools such as from simple to complex and advanced mathematical algorithms in order to discover patterns and relationships in dataset that can be used to establish association rules and make effective predictions. 2.1

Cluster Analysis

The goal of cluster analysis is categorization of attributes like consumer products, objects or events into clusters or groups, so that the degree of correlation is strong between members of the same cluster and weak between members of different clusters. Each group describes the class in terms of the data collected to which its members belong. It may show structure and associations in data, although not previously evident, but are sensible and useful once discovered. The results of cluster analysis [9] may contribute to the definition of a formal classification scheme, such as in taxonomy for related animals, insects or plants; suggest statistical models with which to describe populations; indicate rules for assigning new cases to classes for identification and diagnostic purposes; provide measures of definition, size and change in what previously were only broad concepts. 2.2

Apriori Algorithm

There are varieties of data mining algorithms that have been recently developed to facilitate the processing and interpretation of large databases. One example is the association rule algorithm, which discovers correlations between items in transactional databases. The Apriori algorithm is used to find candidate patterns and those candidates that receive sufficient support from the database are considered for transformation into a rule. This type of algorithm works well for complete data with

Hierarchical Clustering and Association Rule Discovery Process

241

discrete values. Some limitations of association rule algorithms, such as the Apriori is that only database entries that exactly match the candidate patterns may contribute to the support of that candidate pattern. In the past years, there were lots of studies on faster, scalable, efficient and cost-effective way of mining a huge database in a heterogeneous environment. Most studies have shown modified approaches in data mining tasks which eventually made significant contributions in this field. However, there are limitations on generated rules, like producing enormous, unclear and sometimes irrelevant rules.

3

System Architecture

The proposed architecture for the data mining system is shown in Figure 1. Its refinement is presented in the subsequent sections. Phase 1

Phase 2

Phase 3

Hierarchical Clustering

Association Rule Mining

Distributed DB Discovered Knowledge Aggregated Data

Clustered data

Fig. 1. The general view of the proposed model

Figure 1 shows the proposed three phase architecture, where the first phase is the data preprocessing stage that performs data extraction, transformation, loading and refreshing. This will result to an aggregated data cubes as shown in the same figure. Phase 2 shows the implementation of the hierarchical nearest neighbor clustering, while Phase 3 is the implementation of Apriori algorithm to generate rules. Phase 2

Hierarchical clustering using nearest neighbor method

Randomly Select k objects

Calculate Mean op for each cluster value

Iterate until criterion function converges

Aggregated Data

Clustered Data

Fig. 2. Refinements of the model at Phase 2

242

B.D. Gerardo, Y.-C. Byun, and B. Tanguilig III

Figure 2 shows the implementation of the cluster analysis using the hierarchical nearest neighbor clustering algorithm while Figure 3 is the implementation of association rule discovery method. Figure 3 shows the refined view of Phase 3. In this illustration, association rule algorithm is used as part of the data mining process. The successions of transforms for association rule algorithm which are represented by bubbles are shown in the shaded rectangle. Phase 3

Association Rules Generation

Calculate frequent sets

Clustered Dataset

Compute association rules

Generate rules based on constraints

Discovered Association Rules

Fig. 3. Refinements of the model at Phase 3

Phase 3 is the final stage in which the association rule algorithm will be implemented to generate the association rules. This calculates for the frequent itemsets and then compute for the association rules using the threshold for support and confidence. The output is given by the last rectangle showing the discovered rules. In this study, the discovered rules are provided in the table showing the support count and the strength of its confidence which are presented in section 5. The process allows the data to be modeled and viewed in multiple dimensions. Cluster analysis will generate partitions of the dataset, and then the association rule discovery process will be employed. The data cubes will reveal the frequent dimensions, thus, could generate rules from it. The final stage is utilization of the result for decision support. The proposed architecture will implement the association rule generation on a clustered database and would expect better data mining results.

4

Cluster Analyses and the Proposed Model

Among the most popular hierarchical clustering methods are Nearest-Neighbor, Farthest-Neighbor, and Minimal Spanning Tree while for non- hierarchical methods are K-Means, Fuzzy K-Means, and Sequential K-Means. This study put more emphasis on the use of hierarchical method as shown in the experimental results. 4.1

Types of Cluster Analysis

Cluster analysis is a method used for partitioning a sample into homogeneous classes to create an operational classification. Such classification may help formulate hypotheses concerning the origin of the sample, describe a sample in terms of a typology, predict the future behavior of population types, optimize functional processes for business site locations or product design, assist in identification as used

Hierarchical Clustering and Association Rule Discovery Process

243

in the medical sciences, and measure the different effects of treatments on classes within the population [9]. Nearest-Neighbor clustering is one of the simplest agglomerative hierarchical clustering methods, which is also known as the nearest neighbor technique. The defining feature of the method is that distance between groups is defined as the distance between the closest pair of objects, where only pairs consisting of one object from each group are considered [10]. An agglomerative hierarchical clustering procedure produces a series of partitions of the data, Pn, Pn-1 until P1. The first Pn consists of n single object clusters, while the last P1 consists of single group containing all n cases. At each particular stage the method joins together the two clusters which are closest together or are most similar. Figure 4 shows the algorithm for the nearest neighbor clustering. Given: A set X of objects {x1,...,xn}, A distance function dis(c1,c2) 1. for i = 1 to n ci = {xi} end for 2. C = {c1,...,cb} 3. l = n+1 4. while C.size > 1 do a) (cmin1,cmin2) = minimum dis(ci,cj) for all ci,cj in C b) remove cmin1 and cmin2 from C c) add {cmin1,cmin2} to C d) l = l + 1 end while

Fig. 4. The nearest neighbor clustering algorithm

Although there are several other hierarchical clustering methods, in this study, the nearest neighbor had been utilized as part of our proposed model for clustering the data. 4.2

Implementation of the Proposed Model

The models in section 3 as reflected in Figures 1, 2 and 3, respectively, will be implemented in a heuristic process. In our experiment, we will calculate for the clustered data based on the proposed model. And then the outputs will be processed by implementing the Apriori for association mining. Tables showing the comparison of the results on original dataset, the proposed model and the discovered rules are presented in section 5.

5

Simulation and Results

The simulation was done on the database containing 30 attributes comprising of six (6) major dimensions and a total of 1,000 tuples of e-commerce and transactional

244

B.D. Gerardo, Y.-C. Byun, and B. Tanguilig III

types of data. The evaluation platforms utilized in the study were IBM compatible computer, Windows OS, C++, and Python. For the purposes of illustrating the database used in the experiment, we present the Dataset showing partially the data as revealed in Table 1. The abbreviated notations for the attributes stand as follows: An= books and its corresponding subcategories, Bn = Electronics, Cn = Entertainment, Dn= Gifts, En = Foods, and Fn = Health. Furthermore, An Book attribute is consist of subcategories like A1= Science, A2=social, A3=math, A4=computer, A5=technology, A6=religion, and A7=children books. Other dimensions are written with notations similar to that of An. The discrete values indicated by each record are corresponding to the presence or absence of the attribute in the given tuples. Supposed that we consider the problem of determining how often consumers buy products and the probability of purchasing some items online. The results which will be presented in the subsequent sections will answer this problem. Most literatures assumed that the hierarchical clustering procedure is suitable for binary or counts data types [1], [7], [8], [10]. The method that we considered for cluster analysis, which is integrated in proposed model is just suitable for the dataset that we assumed. Consumers respond to questions by giving their agreement or disagreement on buying some products online. 5.1

Hierarchical Clustering Results

The simulation will identify relatively homogeneous groups of cases based on selected characteristics. It is observed that a total of 4 clusters had been created and the group membership of each case is shown in Table 1. In the clustering result, the minimum distance of each case indicates its membership to the cluster. In summary, cluster 1 has a total of 433 cases (43.3%), cluster 2 has 235 cases (23.5%), cluster 3 has 165 (16.5%) and cluster 4 has 167 cases (16.7%). 5.2

Comparison of Data Mining Result after the Implementation of the Model

The data mining results using the two approaches are shown on Table 1 which also shows their corresponding values. The same table presents the number of cases that belong to the respective clusters. After implementing the clustering, we then employed the association rule algorithm (Apriori property). The results is shown in Table 2. The use of such algorithm is for discovering association rules that can be divided into two steps: (1) find all itemsets (sets of items appearing together in a transaction) whose support is greater than the specified threshold. Itemsets that meet the minimum support threshold are called frequent itemsets, and (2) generate association rules from the frequent itemsets. All rules that meet the confidence threshold are reported as discoveries of the algorithm.

Hierarchical Clustering and Association Rule Discovery Process

245

Table 1. Clusters and the Discovered Rules, Support >0.90

Original Dataset

Clustered (Cases, 433, 235, 165, 176)

Clusters 

Number of Rules

All

1

2

3

4

(1,000)

(433)

(235)

(165)

(176)

1,758

1,154

708

650

548

Table 2. Comparison of the Discovered Rules Models

Discovered Rules (showing first 5 rules generated)

Support

Confidence

Original (1,758 rules)

A6=Buy -> A2=Buy F4=Buy A6=Buy -> A2=Buy A3=Buy A6=Buy -> A2=Buy C4=Buy F4=Buy A6=Buy -> A2=Buy D2=Buy F4=Buy A6=Buy -> A2=Buy A3=Buy C4=Buy

0.935 0.927 0.916 0.915 0.910

0.942 0.934 0.922 0.921 0.916

Cluster Analysis 1(1154rules)

A6=Buy -> A2=Buy F4=Buy A6=Buy -> A2=Buy A3=Buy A6=Buy -> A2=Buy D2=Buy F4=Buy A6=Buy -> A2=Buy C4=Buy F4=Buy A6=Buy -> A3=Buy F4=Buy

0.924 0.919 0.905 0.903 0.901

0.939 0.934 0.920 0.918 0.915

2 (708 rules)

A6=Buy -> A2=Buy F4=Buy A6=Buy -> A2=Buy A3=Buy A6=Buy -> A2=Buy A6=Buy -> A2=Buy C4=Buy A6=Buy -> A2=Buy D2=Buy

0.912 0.910 0.963 0.942 0.940

0.923 0.921 0.974 0.953 0.951

3 (650 rules)

D2=Buy -> F3=Buy D2=Buy -> F2=Buy D2=Buy -> C4=Buy D2=Buy -> A2=Buy D2=Buy -> A3=Buy

0.921 0.909 0.958 0.958 0.945

0.938 0.926 0.975 0.975 0.963

4 (548 rules)

D2=Buy -> F3=Buy D2=Buy -> F2=Buy D2=Buy -> A2=Buy D2=Buy -> C4=Buy D2=Buy -> A6=Buy

0.922 0.910 0.958 0.952 0.946

0.939 0.927 0.976 0.970 0.963

246

B.D. Gerardo, Y.-C. Byun, and B. Tanguilig III

The result only shows the first five rules generated for each of the cluster. The support threshold that we set prior to the experiment was 0.90. In the original dataset, those who buy A6 (books on religion) will most likely buy A2 (books on social science) and F4 (Health supplement) with support of 0.935 and confidence of 0.942 (94.2% probability of buying). The same fashion of explanation and analysis could be done to other rules. In cluster 1, those who buy A6 (books on religion) will most likely buy A2 (books on social science) and F4 (Health supplement) with support of 0.924 and confidence of 0.939 (93.9%). Similar approach of analysis could be made for other rules in this cluster. And a similar fashion of explanation could also be done for other rules discovered such as in clusters 2, 3 and 4, respectively. In principle, there would be an improvement in processing time since the computation of rules is based on chunks of data, i.e. clustered data. Shorter processing time had been observed to compute for smaller clusters attributes implying faster and ideal processing period than processing the entire dataset. 5.3

Further Analysis and Implications

The blending of cluster analysis and association rule generation in the proposed model specifically isolate groups of correlated cases using the hierarchical nearest neighbor clustering and then using of the extended data mining steps like the algorithm for association rule generation. The model identify relatively homogeneous groups of cases based on selected characteristics and then employed the Apriori algorithm to calculate for association rules. This resulted to some partitions where we could conveniently analyze specific associations among clusters of attributes. This further explains that the generated rules were discovered on clusters indicating highly correlated cases which will eventually implies simplification of analysis of the result, thus beneficial to be used for decision support purposes.

6

Conclusions and Recommendations

The model reveals clusters that have high correlation according to predetermined characteristics and generated isolated but imperative association rules based on clustered data which in return could be practically explained for decision support purposes. The rules generated based on clustered attributes indicates simple rules, thus it could be efficiently used for decision support system such as in policy making or top level decision making. For future works, upgrade of the model based on extended clustering methods like divisive and non-hierarchical clustering may be needed to check if it performs well with other mechanisms. Acknowledgments. This research was financially supported by the Ministry of Education, Science Technology (MEST) and Korea Industrial Technology Foundation (KOTEF) through the Human Resource Training Project for Regional Innovation.

Hierarchical Clustering and Association Rule Discovery Process

247

References [1] Han, J., Kamber, M.: Data Mining Concepts & Techniques. Morgan Kaufmann, USA (2001) [2] Pressman, R.: Software Engineering: a practitioner’s approach, 6th edn. McGraw-Hill, USA (2005) [3] Hellerstein, J.L., Ma, S., Perng, C.S.: Discovering actionable patterns in event data. IBM Systems Journal 41(3) (2002) [4] Multi-Dimensional Constrained Gradient Mining, ftp://fas.sfu.ca/pub/cs/theses/2001/JoyceManWingLamMSc.pdf [5] Chen, B., Haas, P., Scheuermann, P.: A new two-phase sampling based algorithm for discovering association rules. In: Proceedings of ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2002) [6] Margaritis, D., Faloutsos, C., Thrun, S.: NetCube: A Scalable Tool for Fast Data Mining and Compression. In: 27th Conference on Very Large Databases (VLDB), Roma, Italy (September 2001) [7] Han, E.H., Karypis, G., Kumar, V., Mobasher, B.: Clustering in a high-dimensional space using hypergraph models (1998), http://www.informatik.uni-siegen.de/~galeas/papers/general/ Clustering_in_a_High-Dimensional_Space_Using_Hypergraphs_ Models_%28Han1997b%29.pdf [8] Cluster Analysis defined, http://www.clustan.com/what_is_cluster_analysis.html [9] Determining the Number of Clusters, http://cgm.cs.mcgill.ca/soss/cs644/projects/siourbas/ cluster.html#kmeans [10] Using Hierarchical Clustering in XLMiner, http://www.resample.com/xlminer/help/HClst/HClst_intro.htm [11] Ertz, L., Steinbach, M., Kumar, V.: Finding Topics in Collections of Documents: A Shared Nearest Neighbor Approach. In: Text Mine 2001, Workshop on Text Mining, First SIAM International Conference on Data Mining, Chicago, IL (2001) [12] Hruschka, E.R., Hruschka Jr., E.R., Ebecken, N.F.F.: A Nearest-Neighbor Method as a Data Preparation Tool for a Clustering Genetic Algorithm. In: SBBD, pp. 319–327 (2003)

Implementation of Energy Efficient LDPC Code for Wireless Sensor Node Sang-Min Choi1 and Byung-Hyun Moon2 1

D&D Division Taihan Electric Wire Co., Ltd. 785, Gwangyang-2dong, Dong-gu, Anyang, Gyeonggi, 431-810, Korea [email protected] 2 Dept. of Computer and Communication Engineering, Daegu Univ. 15, Neari, Jinryang, Gyeongsan, Gyeongbuk, 712-714, Korea [email protected]

Abstract. The energy efficiency of error control scheme is very important because of the strict energy constraints of wireless sensor networks. Wireless sensor node requires simple error control schemes because of the low complexity request of sensor nodes. Automatic repeat request(ARQ) and forward error correction(FEC) are the key error control strategies in wireless sensor networks. In this paper, we implemented the efficient QC-LDPC encoder which does not require matrix inversion to improve the complexity of the encoder. It is shown that the efficient QC-LDPC code obtained 17.9% and 36% gain respectively in the mean number of transmission for the transmission power of -19.2dBm and -25dBm. Keywords: QC-LDPC code, Wireless sensor networks.

1

Introduction

In recent years, the idea of wireless sensor networks has produced lots of research, because of wireless sensor networks can be applied widely in many fields. In wireless sensor networks, erroneous transmission can happen by wireless channel noise. Automatic repeat request (ARQ) detects errors using cyclic redundancy check (CRC) and retransmission of data is used as error control scheme for wireless sensor networks. Sensor node requires long lifetime with limited battery, but retransmission of data is the primary source of energy consumption and reduces the lifetime of sensor node. Therefore, error control scheme of forward error correction (FEC) for wireless sensor networks is necessary [1]. Low-density parity-check (LDPC) codes were first introduced by Gallager in 1962 and rediscovered by Mackay and Neal in 1996 and come into the spotlight for next generation communication system[2][3]. The LDPC codes construction can be categorized into Mackay's random construction and sub-block based structured construction. The Mackay's random construction LDPC codes show very good performance. However, this construction is computationally intensive implementation because of large memory requirement. The sub-block based structured construction T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 248–257, 2011. © Springer-Verlag Berlin Heidelberg 2011

Implementation of Energy Efficient LDPC Code for Wireless Sensor Node

249

scheme can be implemented with less complexity than the Mackay’s random construction [4]. In this paper, the efficient quasi-cyclic(QC) LDPC code on ATmega128 based sensor node using sensor network OS platform(SenWeaver OS) is implemented. The efficient QC-LDPC code for wireless sensor networks has small size parity check matrix and can be implemented easily. The encoding scheme of efficient QC-LDPC code is simplified version of Richardson’s encoder and does not require matrix inversion so that this encoding scheme is suitable for sensor node which has limited computation ability. This paper is organized as follows. In section 2, the efficient QCLDPC encode is given. In section 3, the specifications of the sensor node and the OS platform (SenWeaver OS) are introduced. In section 4, the experiment results for the efficient QC-LDPC code are given. Finally, the conclusion is made in section 5.

2

The Efficient QC-LDPC Code

2.1

QC-LDPC Code

For wireless sensor node applications, we consider a subclass of LDPC code, QCLDPC, whose parity-check matrix consists of circulant permutation matrices or the zero matrix. Ii is the Ns × Ns permutation matrix which shifts the identity matrix I to the right by i-times for any integer i, 0 ≤ i ≤ Ns. Let I1 be the Ns × Ns permutation matrix given by 0 0  I1 =  #  0 1 

1 0 " 0 0 1 " 0 # # % #  0 0 " 1 0 0 " 0

(1)

Using this notation parity check matrix H can be defined by

 I s0 , 0  I  s1, 0 H =  I s2 , 0   # I s  ( m −1), 0

I s 0 ,1 I s1,1

I s0 , 2 I s1, 2

I s 2 ,1

I s2 , 2

# I s( m −1),1

# I s( m −1), 2

I s0 ,( n −1)  I s1,( n −1)  " I s2 ,( n −1)   % #  " I s( m −1), ( n −1)   " "

(2)

where si, j is the shift value corresponding position (i, j ) sub-block. This value is one of the {0, 1, 2, …, Ns-1} for nonzero sub-block and si , j = − 1 for zero matrix. The size of H is mNs × nNs. [4].

250

S.-M. Choi and B.-H H. Moon

2.2

Encoding of LDPC Codes

Block type LDPC encodin ng process has Richardson’s encoding algorithm if M × N Parity check matrix is divid ded into the form

A B T  H =  C D E 

(3)

where A is (M-l) × (N-M), B is (M-l) × l, T is (M-l) × (M-l), C is l × (N-M), D is l × l, E is l × (M-l). All these matrices are sparse and T is a lower triangular w with one along the diagonal [5]. Fig. 1 shows the parity check matrix for the Richardsoon’s encoder.

Fig. 1. Parity check matrix for Richardson’s encoder

Let the codeword c={u, p1, p2} where u denotes the systematic part, p1 and p2 dennote the parity parts, p1 has leng gth l, and p2 has length (M-l). The codeword c satisfies the following equations.

 uT  A B T   T  T HcT =    p1  = 0 C D E   T   p2  Au T + Bp B 1T + Tp2T = 0

(− ET

−1

)

(

)

A + C u T + − ET −1 B + D p1T = 0

(4)

(5)

Implementation of Energy Efficient LDPC Code for Wireless Sensor Node

251

Let Φ = − ET −1 B + D and d assume that Φ is nonsingular. Then, we can obttain parity bits as the following equations

( (Au

)

p1T = − Φ −1 − ET −1 A + C u T p =−T T 2

2.3

−1

T

+ Bp

T 1

)

(6)

The Efficient QC-L LDPC Code

The encoding scheme of efficient e QC-LDPC code is to simplify the Richardsoon’s encoder. The goal is to elim minate the matrix inversion of Φ −1 and T −1 in equationn (6). If Φ and T are identity y matrixes, the encoder can be simplified as shownn in equation (7). Also, the sim mplified encoder does not require matrix inversion and encoding can be done without sub-matrix D and T. This will reduce amountt of computations required to en ncode at wireless sensor nodes. Fig. 2 shows the simpliffied Richardson’s encoder.

p1T = − (− EA + C )u T

(

p 2T = − Au T + Bp1T

)

(7)

Fig. 2. The simplified Richardson’s encoder

The parity check matrix forr the simplified Richardson’s encoder is constructed byy 3 × 6 sub-matrices as shown n in equation (8).

 A1,1  H =  A2 ,1  A3,1 

A1, 2

A1, 3

A1, 4

A1, 5

A2, 2

A2,3

A2, 4

A2, 5

A3, 2

A3, 3

A3, 4

A3,5

A1, 6   A2, 6  A3, 6 

(8)

In equation (8), to mak ke T matrix as identity matrix, let A1,5 = A2,6 = I and

A1, 6 = A2,5 = I −1 = 0 . To makke Φ as an identity matrix from equation (9), equattion (10) has to be satisfied.

252

S.-M. Choi and B.-H H. Moon

Φ = − ET −1 B − D  A1, 4  A3, 6   + A3, 4  A2 , 4  = A3,5 A1, 4 + A3, 6 A2 , 4 + A3, 4 = I

[

]

= A3,5

(9)

Also, if we let A3, 4 = I wh hich corresponds to matrix D as identity matrix, equattion (10) is obtained.

A3,5 A1, 4 + A3,6 A2, 4 = 0

((10)

Assume A3,5 A1, 4 = A3,6 A2, 4 , then Φ becomes identity matrix. In order to satiisfy equation (10), we let A1, 4 in B matrix and A3, 6 in E matrix as an identity mattrix. And, let A3,5 and A2, 4 are a equivalent matrixes. The parity check matrix for the efficient QC-LDPC code is shown in Fig. 3.

Fig. 3. Parity y check matrix for the efficient QC-LDPC code

In this paper, we propose 96 × 192 parity check matrix that is made by shifting thee 32 × 32 identity matrix as shown in the following equation (11). I H =  I1  I13

I

I

I

I

I3 I11

I5 I9

I7 I

I −1 I7

I −1  I  I 

((11)

Implementation of Energy Efficient LDPC Code for Wireless Sensor Node

253

k matrix of 96 × 192 for the efficient QC-LDPC code Fig. 4. Parity check

Code rate 1/2 Mackay's ran ndom constructed LDPC code with column weight 3 and code length 192 has six cycle-4. c However the efficient QC-LDPC code has zzero cycle-4 and girth of 6. The parity check matrix for the efficient 96 × 192 QC-LDPC C is shown in Fig. 4

3

Specifications off Sensor Node and OS Flatform

We have implemented an energy e efficient QC-LDPC code on an ATmega 128 baased sensor node as shown in Fig. F 5. The detailed specifications of the sensor node are shown in Table 1. The exteended 32Kbyte SRAM by Chiplus is used to overcome the shortage of the memory during the encoding and decoding of QC-LDPC. T The CC2420 RF module is useed to transmit the LDPC coded message. The multi-laayer chip antenna for the freq quency range between 2400-2455MHz with 100 M MHz bandwidth is used. The sensor node uses SeenWeaver OS that is developed by the UTRC(Ubiquittous Technology Research Centter) in Daegu University [6]. The features of SenWeaaver OS are as following    

Priority based schedulling/multi-threading Vertical and horizontaal layered architecture Provide ANSI C based d API(Application Program Interface) Semantic modular architecture (energy efficiency by dynamic softw ware reprogramming) mming like PC that operate a number of API withhout  Provide multiprogram alteration  Provide hierarchical reconstruction of API a automatic code generation of variable sensor nnode  Provide abstraction and hardware. The software architecture for f sensor nodes designed such that a layer modular iis a separate block. This allowss easy replacement a block according to hardware withhout changing upper layers. The SenWeaver OS architecture overview is shown in Fig. 66.

254

S.-M. Choi and B.-H H. Moon

Fig. 5. Sensor nodee used for the efficient QC-LDPC code implementation

Fig. 6. SenWeaver OS architecture overview Table 1. Detailed specifications of a sensor node

MCU

Model ATmegaa128L

Ext. SRAM RF module

CS18LV V02563 CC2420 0

Antenna

SWBBL L1

Specification 8bit processor, 128Kbytes flash, 4Kbyte SRAM 32Kbytes SRAM 2.4GHz, IEEE 802.15.4/ZigBee-ready RF transceiver Multilayer chip antenna 2400~2485MHz, 100MHz BW

Implementation of Energy Efficient LDPC Code for Wireless Sensor Node

4

255

Experiment Results

As shown in Fig. 6, the transmitter and the receiver are located on the ceiling to ensure the line of sight communication. By monitoring the 2.43GHz frequency band, no other source of interference is observed. In order to measure performance of the efficient QC-LDPC code, 15,000 12byte long MAC frames as shown in Fig. 7 are transmitted.

Fig. 7. Experiment environment of performance measurement

Fig. 8. LDPC coded simple MAC frame

The experiment results for the uncoded and the efficient QC-LDPC coded cases are summarized in Table 2. The transmitting power is varied from -7 dBm to -25 dBm.. Table 2. Measured BER performance of efficient QC-LDPC code TX power -7 dBm -10 dBm -15 dBm -19.2 dBm -25 dBm

Uncoded case 3.4787 × 10-6 8.5301 × 10-5 8.3478 × 10-4 7.0767 × 10-3 1.6713 × 10-2

Efficient QC-LDPC coded case 0 5.7418 × 10-5 4.6648 × 10-4 5.0378 × 10-3 1.2139 × 10-2

256

S.-M. Choi and B.-H. Moon

Given the packet length and BER performance, we calculate the PER (Packet Error Rate) to be

Pep = 1 − (1 − BER ) L

(12)

where Pep is PER, L is packet length[7]. The calculated PER using equation (12) is shown in table 3. Table 3. PER performance analysis of efficient QC-LDPC code TX power -7 dBm -10 dBm -15 dBm -19.2 dBm -25 dBm

Uncoded case -4 3.3390 × 10 -3 8.1558 × 10 7.7043 × 10-2 4.9428 × 10-1 8.0171 × 10-1

Efficient QC-LDPC coded case 0 5.4971 × 10-3 4.3804 × 10-2 3.8421 × 10-1 6.9040 × 10-1

The mean number of transmission, R, required for success can be calculated as ∞

R = (1−Pep )  (k + 1) Pepk = k =0

1 1 − Pep

(13)

Also, the gain of mean number of transmission Rgain is given by

Rgain =

Rcoded Runcoded

(14)

Where Runcoded is mean number of transmission for uncoded case, Rcoded is that of the efficient QC-LDPC coded case. Table 4 shows the mean number of transmission using equations (13) and (14). Table 4. Mean number of transmission for success TX power

Uncoded case

-7 dBm -10 dBm -15 dBm -19.2 dBm -25 dBm

1.0003 1.0082 1.0835 1.9774 5.0430

Efficient QC-LDPC coded case 1.0000 1.0055 1.0458 1.6239 3.2300

Rgain 0.0003 0.0027 0.0348 0.1787 0.3595

It is shown that the implemented QC-LDPC code showed limited gain for the TX powers between -7dBm and -10dBm. When the TX powers are between -19.2dBm

Implementation of Energy Efficient LDPC Code for Wireless Sensor Node

257

and -25dBm, the implemented QC-LDPC code obtained about17.9% and 36% gain over uncoded case, respectively.

5

Conclusion

In this paper, we propose FEC using efficient QC-LDPC code for the error control in the wireless sensor networks. The efficient QC-LDPC code has small size parity check matrix and easy implementation scheme. The encoding of the efficient QCLDPC code that constructs sub-matrix T and Φ of Richardson’s encoder into identity matrix and requires less computation compare to the Mackay’s random constructed LDPC code. We implemented the efficient QC-LDPC code on ATmega128 based sensor node using sensor network OS platform (SenWeaver OS) and measured the BER performance at TX powers between -7dBm to -25dBm. Also, we calculated the packet error rate and the mean number of transmission required for successful transmission. It is shown that the implemented QC-LDPC code obtained performance gain in the mean number of transmission about 17.9% and, 36% gain over the uncoded case for the transmission powers of -19.dBm and -25dBm, respectively. Acknowledgments. This research was financially supported by Daegu University Research Grant, 2011.

References 1. Jeong, J., Tien Ee, C.: Forward error correction in sensor Networks, U.C.Berkeley Technical Report (May 2003) 2. Gallager, R.G.: Low-Density parity-Check Codes. IRF Trans. on Info. Theory 8, 21–28 (1962) 3. MacKay, D.J.C., Neal, R.M.: Near Shannon limit performance of low density parity check codes. Electron, Lett. 32(18), 1645–1646 (1996) 4. Fossorier, M.P.C.: Quasi-Cyclic Low Density Parity Check Codes from Circular Permutation Matrices. IEEE Trans. Information Theory 50, 1788–1794 (2004) 5. Richardson, T.J., Urbanke, R.L.: Efficient Encoding of Low-Density Parity-Check Codes. IEEE Trans. IT 47, 638–656 (2001) 6. Kim, T.-H., Kim, H.-C.: Software Architecture for Highly Reconfigurable Sensor Operating System. Journal of IEMEK 2(4), 242–250 (2007) 7. Lettieri, P., Fragouli, C., Srivastava, M.: Low power error control for wireless links. In: Proceedings of 3rd annual ACM/IEEE Intl. conference on Mobile computing and networking (MOBICOM) (1997)

A Multi-layered Routing Protocol for UWSNs Using Super Nodes Abdul Wahid, Dongkyun Kim*, and Kyungshik Lim Kyungpook National University, Daegu, Korea [email protected], {dongkyun,kslim}@knu.ac.kr

Abstract. Underwater Wireless Sensor Networks (UWSNs) have peculiar characteristics such as high propagation delay, limited bandwidth and high error rates. Therefore, communication protocols for UWSNs are highly required to cope with them. Specifically, the design of an efficient routing protocol for UWSNs is one of the most important challenges. The routing protocols can take advantage of the localization of sensor nodes. However, the localization itself is a crucial issue in UWSNs, which remains to be solved yet. Hence, the design of a non-localization based routing protocol is a preferable alternative. In this paper, we therefore propose a non-localization based routing protocol named MRP (Multi-layered Routing Protocol) for UWSNs. In MRP, the concept of super nodes is employed. The super nodes are the nodes having high capacity such as high energy and transmission power. Our proposed protocol, MRP, consists of two phases: layering and data forwarding phases. During the layering phase, multiple layers are formed by the super nodes, whereas, in the data forwarding phase, data packets are forwarded based on these layers. Based on the simulations using NS2, we observe that our proposed routing protocol, MRP, contributes to the performance improvements. Keywords: Underwater wireless sensor networks, routing, super node.

1

Introduction

Underwater Wireless Sensor Networks (UWSNs) have attracted much research attention both from academia and research community due to the need of a number of applications e.g. harbor inspection, ecological monitoring, oil/gas spills monitoring, homeland security etc. Typically, in UWSNs, acoustic signals are employed as a physical media because of the poor performance of radio and optical signals in water. The radio signals propagate long distance at extra low frequencies (30 – 300 Hz), which requires large antennas and high transmission power. The optical signals do not suffer from such high attenuation, but are affected by scattering. Due to the employment of the acoustic signals, UWSNs have some unique challenges which are different from the terrestrial wireless sensor networks. The acoustic signals have large propagation delay (i.e.1500 m/sec), limited bandwidth (i.e. <100 KHz) and high error rates. Therefore, the protocols proposed for terrestrial sensor networks are not applicable in UWSNs and the protocols designed specifically for UWSNs are highly *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 258–267, 2011. © Springer-Verlag Berlin Heidelberg 2011

A Multi-layered Routing Protocol for UWSNs Using Super Nodes

259

demanded. The demand of the protocols specific to UWSNs triggered the design of a variety of protocols for each layer of communications i.e. MAC, routing, and transport layers. Specifically, the routing protocols require the utmost attention since data transmission from a source towards a destination is based on the routing protocol. Therefore, various routing protocols have been devised by taking into account the characteristics of UWSNs. Generally, the existing routing protocols can be classified into two classes, namely localization based and non-localization based routing protocols. Localization based routing protocols require exact location information of underwater sensor nodes, which itself is a vital issue in UWSNs. In contrast, since non-localization based protocols do not require the location information of sensor nodes, the non-localization based routing protocols are highly recommended by research community. In this article, we therefore propose a non-localization based routing protocol for UWSNs. In our proposed routing protocol MRP, two different types of underwater sensor nodes named as super nodes and ordinary sensor nodes are utilized. The super nodes are the sensor nodes with high capacity such as extensive energy and high transmission power. On the other hand, the ordinary sensor nodes have typical energy and transmission power. It is assumed that a limited number of super nodes are deployed underwater at different water columns at certain depths. The ordinary sensor nodes are deployed at the higher depths at bottom. The ordinary sensor nodes sense data and send it to the nearest super node. These super nodes then transmit the data towards the sink node by relaying it through other super nodes installed at lower depths. For data transmissions from the ordinary sensor nodes towards the super nodes, a routing metric based on a layering structure is used. Different layers are formed by the super nodes and a cost (i.e. layer number) is assigned to each ordinary sensor node based on its existence in the corresponding layer. The rest of the paper is structured as follows. In Section 2, we review related routing protocols and their problems. In Section 3, the proposed routing protocol is described in detail. Section 4 presents the performance evaluation of the proposed protocol. Finally, conclusions are drawn in Section 5.

2

Related Work

In this section, we present the related routing protocols available in the literature. We take into account the well known routing protocols and the protocols/schemes which utilize the concept of the super nodes for UWSNs. We divide the related work into three sub-sections: localization based routing protocols, non-localization based routing protocols and the protocols which utilize the concept of the super nodes. 2.1

Localization Based Routing Protocols

In this sub-section, we present the routing protocols which are based on the localization of the sensor nodes. In [1], a vector based routing protocol called VBF (Vector Based Forwarding) was proposed. In VBF, it is assumed that each node knows its geographic position/location. The data forwarding process is as follows. A vector is computed from the source node towards the sink node. Then, the packet is

A Multi-layered Routing Protocol for UWSNs Using Super Nodes

260

flooded towards the sink node. The nodes residing in a predefined radius (i.e. called routing pipe), around the computed vector, participate in forwarding. The restriction of the predefined radius for forwarding affects data transmissions, since, in case of sparse density, sensor nodes might not be available in the predefined radius (i.e. the routing pipe). In [2], a routing protocol called HHVBF (Hop by Hop Vector Based Forwarding) was proposed. HHVBF inherits the concept of VBF such that HHVBF is also based on vector based routing. However, HHVBF computes the routing vector at each hop from the sender towards the sink node. Since the routing vector is not restricted to a single routing pipe, the HHVBF is less affected by the unavailability of the sensor nodes. However, HHVBF has the inherited assumption of the localization of the sensor nodes. In [3] a protocol called FBR (Focused Beam Routing) was proposed. In FBR, various transmission power levels (i.e. P1 to PN) are employed. The various power levels are used to select a relay node during the data packets’ forwarding. The source node broadcasts an RTS packet with a transmission power P1. If a reply (i.e. CTS packet) is received from a relay node, the data packet is transmitted to the relay node. Otherwise, the transmission power is increased to the next level. The process is repeated until all the transmission power levels are attained. The limitation of FBR is the increased end-to-end delay due to the RTS/CTS messages. In [4], a routing protocol called DFR (Directional Flooding-based Routing) was proposed. In DFR, the data packets are flooded in a limited flooding zone. The scope of the flooding zone is based on the quality of the links among the sensor nodes. In case of better qualities links, the scope is limited to a few forwarding nodes. In contrast, more nodes are allowed to participate for bad qualities links. DFR is also based on the localization of sensor nodes, which limits its applicability. In [7], SBRDLP (Sector Based Routing with Destination Location Prediction) was proposed. In the SBR-DLP, it is assumed that a mobile sink is available in the network and that each node is aware of the movement schedule of the mobile sink. The data forwarding process is as follows. A source node broadcasts a chk_ngb packet. The neighboring nodes reply with a chk_ngb_reply packet. The chk_ngb_reply packet also contains the sector number of the neighboring nodes. The sectors are computed based on the distance from the vector (i.e. the vector between the source and the sink node). Upon receiving chk_ngb_reply packets from its neighbors, the source node assigns priorities to the neighboring nodes based on the sector number and the neighboring node closest to the mobile sink is selected as a forwarder. 2.2

Non-localization Based Routing Protocols

In this sub-section, the non-localization based routing protocols are presented. In [5], a routing protocol called DBR (Depth Based Routing) was proposed. In DBR, all the nodes are assumed to be equipped with a depth sensor. The data packet’s forwarding decision is taken on the basis of the depths of sensor nodes. Upon receiving a data packet, each node compares its own depth to the depth of the sender of the packet. Among receiving nodes, the node having a minimum depth is allowed to forward the packet. Each node holds the packet for a certain time before forwarding, based on the depth difference to the sender of the packet. DBR has some advantage in that no localization of the sensor nodes is required. However, DBR suffers from the drawback of redundant packets’ transmissions. As the network density increases, the

A Multi-layered Routing Protocol for UWSNs Using Super Nodes

261

number of redundant packets’ transmissions also increases. Since the forwarding nodes hold the packet based on their depth difference from their previous hops, the nodes having similar depths also have similar holding time. Hence, these nodes forward the packet at the same time, resulting in excessive energy consumption and many collisions among the packets. The unbalanced energy consumption is its another drawback. Since only the depth of sensor nodes is taken into account during forwarding, the nodes having smaller depths forward packets most of the time. Therefore, the energy of such nodes is exhausted soon, which creates the routing holes (i.e. void regions) in the network. In [6], H2-DAB (Hop by Hop Dynamic Addressing Based) protocol was proposed. In H2-DAB, each node is assigned a unique ID called HopID. The HopIDs are assigned through Hello packets. The Hello packets are originated by sink nodes. A HopID consists of two digits, the first digit denotes the number of hops towards a sink (e.g. sink 1). Similarly, the second digit denotes the number of hops towards another sink (e.g. sink 2). During the forwarding of data packets, these HopIDs are used to select a next forwarding node. The process is as follows. The source node broadcasts an inquiry packet including its own HopID. Upon receiving the inquiry packet, neighboring nodes send reply packets including their HopIDs. The source node then selects the neighboring node having a minimum HopID. The limitations are the long delays due to the inquiry and reply packets and the unbalanced energy consumption. In [8], Winston et al. proposed a virtual sink architecture where multiple sink are assumed to be connected to each other. In the proposed scheme, each sink broadcasts a Hello packet (called hop count update packet). Upon receiving the Hello packet, each sensor node is assigned a hop count value. During the forwarding of data packets from a source towards the sink node, these hop count values are used for selecting a next forwarding node. Its limitations are redundant transmissions (i.e. the transmission of the same packet towards multiple sinks) and the assumption of the connectivity among the sink nodes. 2.3

Protocols Utilizing the Concept of Super Nodes

In this sub-section, we present the schemes proposed for UWSNs which utilize the concept of super nodes. In [9], a synchronization based protocol called MobiSync was proposed. In MobiSync, it is assumed that super nodes are well synchronized with surface buoys. These super nodes are used as a reference clock for the nonsynchronized nodes. In the proposed scheme, the super nodes are utilized to assist the non-synchronized nodes during the synchronization process. In [10], two energy balancing strategies were introduced in order to maximize the network life-time of the UWSNs. In the proposed strategies, one of the strategies utilizes the concept of super nodes for energy balancing. The process is such that two basic nodes and one super node are grouped together to form a cluster. Basic nodes send sensed data directly to the super node, and the super node then transmits the data to the nearest super node towards the sink node. In the proposed scheme, multiple underwater sink nodes are considered available in the network. In [11], UW-MAC (Underwater Sensor Networks MAC protocol) was proposed. In UW-MAC, the super nodes are used as the cluster-head nodes.

A Multi-layered Routing Protocol for UWSNs Using Super Nodes

262

In contrast to the above mentioned protocols, MRP utilizes the concept of super nodes for routing. In MRP, a limited number of super nodes are deployed at various depths and these super nodes are anchored with the bottom/surface of the ocean. The super nodes at the highest depths collect data packets from ordinary sensor nodes. These super nodes residing at the highest depths then forward data packets to the super nodes at the lower depths. The process is repeated until the data packets are delivered to the sink node residing at the water surface.

3

Proposed Scheme: MRP (Multi-layered Routing Protocol)

In this section, MRP is described in detail. We divide this section into three subsections: proposed network architecture, layering phase and data forwarding phase. 3.1

Proposed Network Architecture

Figure 1 shows the typical architecture of an UWSN. In this type of architecture, sink nodes are deployed at the water surface and sensor nodes are deployed underwater from the top to the bottom of the deployment region. The sensor nodes collect data and send it towards the sink node. The data transmissions occur from a sensor node towards the sink node in a multi-hop fashion. The sensor nodes send the data by relaying it through neighboring sensor nodes available at the lower depths. In this way, the data packets travel from the source nodes, residing at the bottom of the ocean, towards the sink node at the water surface with the help of sensor nodes deployed at various depths. In this type of architecture, the availability of sensor nodes at various depths is very crucial for data transmissions. Since sensor nodes move with water currents, there are great chances of void problem where there is no node available for relaying the data packets. Figure 2 shows our proposed architecture. In the proposed architecture, we utilize two different types of underwater sensor nodes i.e. super nodes and ordinary sensor nodes. Super nodes are the sensor nodes having high capacity such as high energy and transmission power. The super nodes are static in nature such that they are anchored with the bottom/surface of the ocean. The difference from the typical architecture is that, instead of utilizing sensor nodes as relays for data transmission from the source

Fig. 1. Typical architecture of an underwater wireless sensor network

A Multi-layered Routing Protocol for UWSNs Using Super Nodes

263

Fig. 2. Proposed architecture of underwater wireless sensor network

towards the destination/sink, we employ the super nodes to act as relays. The advantage of utilizing the fixed super nodes overcomes the problem of void regions. In addition, in UWSNs, the sensor nodes contain limited resources such as limited battery and transmission power. The sensor nodes having limited resources are exhausted soon which affects the performance of the UWSNs especially in terms of network lifetime. In our proposed architecture, we handle this problem by utilizing the super nodes having high capacity for data transmissions. Since these super nodes have high energy and transmission power, these nodes can be used for long time. 3.2

Layering Phase

During this phase, a set of concentric shells (i.e. layers) are formed around the super nodes. Each concentric shell is assigned an ID called layer ID. During the forwarding of data packets, these layer IDs are used as a routing metric by sensor nodes to send data packets to the super nodes. The process of layer formation is as follows. The super nodes broadcast a probe packet with a transmission power p1. All the sensor nodes that receive the probe packet are assigned as layer 1 nodes. Then, the super node rebroadcasts the probe packet with a higher transmission power p2. The nodes receiving the probe packet with a transmission power p2 are assigned as layer 2

Fig. 3. Diagram illustrating different layers around the super node

A Multi-layered Routing Protocol for UWSNs Using Super Nodes

264

nodes. This process is repeated until the maximum transmission power of the super nodes is attained. The different transmission powers of the super nodes can be adjusted/decided according to the area covered and the positioning of the neighboring super nodes. It is assumed that the super nodes are positioned in such a way that all the sensor nodes in the network are covered. It is possible that a sensor node receive probe packets from more than one super nodes. In such a case, the sensor nodes select the nearest super node as its destination based on the layer ID (i.e. lower ID means closer to the super node). Figure 3 shows an example of the network layout after the formation of the multiple layers. 3.3

Data Forwarding Phase

In this phase, data packets are forwarded from source nodes towards the sink nodes. A source node sends a data packet to the super nodes, and the super nodes then forward it to the sink node by relaying it through the super nodes deployed at lower depths. The process is such that, during the forwarding of the data packet, the source node includes its own layer ID in the data packet. The receiving nodes compare their layer IDs and only the nodes which have lower layer ID number will forward the packet. The lower layer ID number shows that this layer exists closer to the super node. Hence, the packet is transmitted from higher layers towards the lower layers, which ensures that the packet is transmitted towards the super node. For instance, if a sensor node at layer N has a data packet to send, it includes its layer ID (i.e. layer N) in the data packet. Upon receiving the data packet, the nodes residing at layer N – 1 become eligible for forwarding the data packet. Hence, the nodes at layer N – 1 further forward the packet including their own layer ID. This process is repeated until the packet reaches the super node. Upon receiving the data packet, the super node transmits the data packet towards the sink. The super node utilizes other super nodes (i.e. available at lower depths) as relay nodes for the data packet’s transmission towards the sink node. The super node transmits the data packet with its maximum transmission power when it transmits the data packet towards the relay super node. Since the maximum transmission power is used, all the sensor nodes residing in the super node’s region, towards the bottom, overhear the data packet. This overhearing acts as an ACK for the sensor nodes and the source node. Hence, there is no need for an extra ACK packet which is typically required by the source of the packet.

4

Performance Evaluation

4.1

Simulation Settings

We have performed the simulations using a commonly used network simulator called NS-2. Simulations were performed with a different number of sensor nodes (i.e. 100, 150, 200 and 250). Two different types of topologies were implemented i.e. grid and random topologies. In each topology, a source node was selected from the bottom of the deployment region. The source node generated a data packet of a size of 64 bytes

A Multi-layered Routing Protocol for UWSNs Using Super Nodes

265

every 15 seconds. The 802.11 protocol was used as an underlying MAC protocol. For all topologies, the results were averaged from 5 runs. Since our proposed routing protocol, MRP, is a non-localization based routing protocol, therefore, we compared our routing protocol with a representative nonlocalization based routing protocol in UWSNs called DBR [5]. 4.2

Performance Metrics

We used the following metrics for evaluating the performance of our proposed routing protocol MRP. End-to-end delay

The end-to-end delay is the time taken by a packet from a source to the destination/sink node. Delivery ratio

The ratio of the number of packets successfully received at the sink node to the number of packets transmitted from the source node. 4.3

Simulation Results and Analysis

The end-to-end delay of both the schemes (i.e. DBR and MRP) in grid topology is compared as shown in Figure 4. MRP has lower delay than DBR, since DBR requires sensor nodes to hold the packet for a certain time, while MRP does not require such holding time. In MRP, packets are forwarded as soon as they are received by the nodes having lower layer IDs than the sender of the packet. As shown in Figure 4, the increase in the number of nodes does not affect the end-to-end delay. Hence, MRP is considered a highly scalable routing protocol. In contrast, in DBR, the end-to-end delay increases with the increase in the number of nodes, since more nodes are involved in holding the packet. The end-to-end delay of both the schemes in random topology is also shown in Figure 5. Similar to the grid topology, MRP shows much lower delay than DBR in random topology. According to the results, the change in topology (i.e. from the grid to random topology) has no effect on the performance of MRP. The reason is that the layers are formed in a similar fashion in both the topologies and the forwarding decision is made based on the layer ID. Therefore, as soon as there are sensor nodes having lower IDs, data packets are delivered to the destination. Figure 6 shows the delivery ratio of DBR and MRP. We believe that the suppressions of some packets cause the delivery ratio of DBR to be lower than the proposed MRP protocol. In DBR, during the holding of a data packet by a sensor node, when this sensor node receives the same packet from another node, it discards its packet. Figure 7 shows the delivery ratio in random topology. Similar to grid topology, MRP enables all the packets to be successfully received in random topology.

A Multi-layered Routing Protocol for UWSNs Using Super Nodes

Fig. 4. Comparison of end-to-end delay in grid topology

Fig. 6. Comparison of delivery ratio in grid topology

5

266

Fig. 5. Comparison of end-to-end delay in random topology

Fig. 7. Comparison of delivery ratio in random topology

Conclusions

Routing is one of the challenging issues in UWSNs due to the intrinsic properties of UWSNs. In this paper, a novel routing protocol named MRP (Multi-layered Routing Protocol) was proposed. MRP is a non-localization based routing protocol which does not require any localization technique of sensor nodes. MRP employed two types of sensor nodes for routing i.e. super nodes and ordinary sensor nodes. The super nodes are the nodes having high transmission power and high energy, while the ordinary

A Multi-layered Routing Protocol for UWSNs Using Super Nodes

267

sensor nodes have typical transmission power and energy. MRP has two phases: layering and forwarding phases. During the layering phase, multiple layers are formed by the super nodes using different transmission powers. Each of ordinary sensor nodes is assigned a unique ID called Layer ID. During the forwarding phase, data packets are forwarded from the ordinary sensor nodes towards the super nodes based on the assigned Layer IDs. Through simulations, MRP was compared with a well known non-localization based protocol called DBR. Through simulation results, it was proved that MRP has improved performance over DBR. Acknowledgment. This work was supported by Defense Acquisition Program Administration and Agency for Defense Development under the contract UD100002KD.

References 1. Xie, P., Cui, J.-H., Lao, L.: VBF: Vector-based forwarding protocol for underwater sensor networks. In: IFIP Networking 2006, Portugal, pp. 1216–1221 (2006) 2. Nicolaou, N., See, A., Xie, P.: Improving the robustness of location-based routing for Underwater Sensor networks. In: OCEANS 2007, Europe, pp. 1–6 (2007) 3. Jornet, M., Stojanovic, M., Zorzi, M.: Focused beam routing protocol for underwater acoustic networks. In: Third ACM Workshop on Underwater Networks (WUWNet 2008), San Francisco, pp. 75–82 (2008) 4. Hwang, D., Kim, D.: DFR: Directional flooding-based routing protocol for underwater sensor networks. In: OCEANS 2008, Quebec City, Canada, pp. 1–7 (2008) 5. Yan, H., Shi, Z., Cui, J.: DBR: Depth-based Routing for Underwater Sensor Networks. In: IFIP Networking 2008, Singapore, pp. 16–1221 (2008) 6. Ayaz, M., Abdullah, A.: Hop-by-Hop Hop Dynamic Addressing Based (H2-DAB) Routing Protocol for Underwater Wireless Sensor Networks. In: IEEE International Conference on Information and Multimedia Technology (ICIMT 2009), Jeju Island, Korea, pp. 436–441 (2009) 7. Chirdchoo, N., Soh, W.-S., Chua, K.-C.: Sector-Based Routing with Destination Location Prediction for Underwater Mobile Networks. In: IEEE International Conference on Advanced Information Networking and Applications (WAINA 2009), pp. 1148–1153 (2009) 8. Winston, K.G., Tan, H.-X.: Multipath Virtual Sink Architecture for Underwater Sensor Networks. In: IEEE OCEANS, Singapore, pp. 1–6 (2006) 9. Liu, J., Zhou, Z., Peng, Z., Cui, J.-H.: Mobi-Sync: Efficient Time Synchronization for Mobile Underwater Sensor Networks. In: IEEE Global Telecommunications Conference (GLOBECOM), Miami, FL, pp. 1–5 (2010) 10. Luo, H., Guo, Z., Wu, K., Hong, F., Feng, Y.: Energy Balanced Strategies for Maximizing the Lifetime of Sparesely Deployed Underwater Acoustic Sensor Networks. Sensors, 6626–6651 (2009) 11. Watfa, M.K., Selman, S., Denkilkian, H.: UW-MAC: An underwater sensor network MAC protocol. Int. J. Communication Systems, 485–506 (2010)

Experimental Measurement for EVM Performance Enhancement of Wireless Repeater System Daesik Ko1 and Hwase Park2 1

Mokwon University, Daelim University, Korea

Abstract. This paper suggests the adaptive multi-feedback filter system which, among a number of adaptive feedback filters in the wireless repeater system, applies the most appropriate feedback system according to the channel environment to enhance the performance of removing interference. The experiment result shows that the ITU 3GPP (3rd Generation Partnership Project) recommended specifications are met on both down-link and up-link. The cancellation was 25dBm, the window size was 500ns/1us, and the max. power was +10dBm. By using algorithms selectively according to the change of environment, the system reduced unnecessary use of hardware resources, and enhanced the convergence rate. Keywords: Wireless Repeater System, Adaptive Feedback, Normal Least Mean Square, Error Vector Magnitude.

1

Introduction

A repeater is used to enhance quality of signal for wireless communication in a shadow area or in an area with low signal strength. There are different types of repeater in use. Among those, wired repeaters are most widely used. Wired repeaters can meet the situation or environment under which installation can become complicated or difficult. Installing this type of repeater requires wiring, and it takes much cost for installation and maintenance. Wireless repeaters are easier and cheaper to install than other types of repeater. This type of repeater, however, has a disadvantage that part of the output signal from the Tx antenna feeds back to the Rx antenna. [1,2,3] This paper suggests the adaptive multi-feedback filter system which, among a number of adaptive feedback filters in the wireless repeater system, applies the most appropriate feedback system according to the channel environment to enhance the performance of removing interference. The suggested system can enhance the convergence rate and the hardware structure of the wireless repeater system as it selects the adaptive feedback filter according to the channel environment.[4,5] In this study, the adaptive multi-feedback filter is designed, the multi-feedback filter is implemented with FPGA and DSP, and the EVM and interference cancellation performance was measured on the home wireless repeater system of less than 10mW. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 268–273, 2011. © Springer-Verlag Berlin Heidelberg 2011

Experimental Measurement for EVM Performance Enhancement

2

269

Hardware Structure of Multi-feedback Filter for Cancellation of Interface of Wireless Repeater

There are various feedback filters with various structures. The most widely used feedback filter model is the adaptive digital feedback filter. The performance of the adaptive digital feedback filter is measured with the algorithm that finds the optimum coefficient to be applied to the filter. The typically used algorithm to find the optimum coefficient is the LMS algorithm with superior stability and simple structure [6][7]. The suggested multi-feedback system is composed of a feedback signal generator which sums up power from numbers of adaptive sub-filters and generates the feedback signal; an original signal detector which deduces the feedback signal from the input signal and detects the service signal only; an input channel filter which removes unnecessary signal from the input signal; a feedback signal channel filter which generates only the signal in the same band with the input channel filter; and an output signal control which controls the delay of feedback signal caused due to device and environment coefficients. The input channel filter is located at the leading edge of the original signal detector, and the adaptive filter calculates the optimum weight for the selected band. Because the same channel filter is used at the output of the feedback signal, the system suppresses the signal of the band which is generated unnecessarily during the weight update process. Channel filters are applied to the output unit for the system using the existing channel filter, while they are applied to the input unit and the feedback signal for the proposed. This structure removes unnecessary signal from the input signal used in the adaptive filter algorithm, and corrects errors which may occur due to out-of-band signal. Figure 1 illustrates the hardware structure of the suggested system. In order to efficiently handle the entered signal, the A/D converter must be selected in consideration of the frequency band, strength of signal and noise to be handled when converting the received analog signal into the digital signal. In this research, Altera's EP3C55F484 was used as FPGA. Most of the adaptive feedback filter algorithm was processed by FPGA with VHDL [8]. For controlling, monitoring and signal processing, 32-bit DSC (Digital Signal Controller) is used. In this research TI's TMS320F2812 was used [9]. A D/A converter is used to output the result calculated with DSC and FPGA algorithms. The D/A convert used in this research was Analog Device's AD9775. In this research, the under sampling method was used for AD sampling. In this research, sampling was made with 64.11MHz, and for the sampling image area, the system is designed to cover 15.36MHz-7.5MHz and 15.36MHz+7.5MHz as the in-band where no aliasing occurs. In order to remove the quantizing noise, a digital filter is installed in FPGA. The signal received from RF and entered into FPGA through the down converter is divided into In-Phase and Quad-Phase signal. NCO (Numerically Controlled Oscillator) is installed in FPGA to take the frequency of this signal to the baseband by FA. To handle the received signal, a decimation filter using a channel filter and a halfband filter is installed in FPGA. Using the decimation filter, FPGA extracts the

270

D. Ko and H. Park

Fig. 1. The proposed system digital system block

feedback signal, and restores the original signal. The restored signal is sent to DAC through the interpolation block. The entire digital system is synchronized with FPGA, and ADC and DAC are sampled. The system generates the feedback signal for each assigned sub-block by multiplying the filter coefficient updated by the adaptive filter sub-block update unit by the reference signal from the output signal delay unit. The system adds up the feedback signals of each sub-block, generates the final feedback signal, deducts the output signal of A/D converter, and detects the original signal.

3

Experiment and Result

To verify the LMS algorithm to be applied to the interference cancellation device of the wireless repeater system using multi-feedback filter, simulation was performed on the convergence time of weight according to isolation. Figure 2 shows the change of weight by isolation under the NLMS algorithm when the received power is -35dBm. The figure shows that, in the most cases, the weight becomes stabilized after 1800-sample (32us) time is passed (1 sample = 20ns). In this figure, I-40 indicates that the isolation is 40dB, the values on the y-axis are weights, and those on the x-axis are times. To verify the performance of ICR made based on the suggested algorithm, a 10mW miniature ICR hardware was developed. FPGA and DSP programs were developed to implement the functions of interference cancellation device of ICR. Test environment was deployed and the equipment performance was verified to assess the hardware structure and the system performance. The operating frequency of ICR was Down-Link 3FA (2112.5~2122.5) and Up-Link 3FA (1922.5~1932.5) in the WCDMA frequency domain, and has 5MHz per FA. In order to verify performance of the wireless repeater using multi-feedback filter, EVM was measured.

Experimental Measurement for EVM Performance Enhancement

271

Fig. 2. Weight Value in Isolation to the Case of Received Power of -35dBm

Figure 3 shows the experimental system used to measure EVM performance of the wireless repeater. Feedback characteristics were experimented for differential isolation by measuring EVM performance with variable center frequency for each FA.

Fig. 3. Experimental System

To verify stability of the system, EVM characteristics were compared on the two conditions: when the feedback signal is entered into ICR interference cancellation device, and when it is not. The interference cancellation device of the produced wireless repeater was equipped with the suggested multi-feedback filter. Figure 4 and 5 show the EVM characteristics on both down-link and up-link with or without feedback signal.

272

D. Ko and H. Park

Fig. 4. EVM Characteristic by uplink FA without Feedback Signal

Fig. 5. EVM Characteristic by uplink FA wit Feedback Interference Signal

The improved LMS algorithm was added to each sub feedback filter algorithm. Under a normal condition when no feedback signal is entered, EVA in 1FA was 7.2%. Under the center frequency 1922.5MHz, input power -65dBm, isolation 60dB and system gain 70dB, EVM of FA1 on the Up-Link channel was 8.2%. The above tables show that in order to meet the specifications, the EVM on the entire FA on both down-link and up-link must be lower than 12%, which is the value recommended by ITU GPP (3rd Generation Partnership Project) [10][11]. Any feedback signal entered was removed in a stable manner. To determine the performance of the wireless repeater, the cancellation performance, the window size and the max. power were measured. For the suggested ICR, the cancellation was 25dBm, the window size was 500ns/1us, and the max. power was +10dBm.

Experimental Measurement for EVM Performance Enhancement

4

273

Conclusion

This paper suggests the adaptive multi-feedback filter system which, among a number of adaptive feedback filters in the wireless repeater system, applies the most appropriate feedback system according to the channel environment to enhance the performance of removing interference. In proposed system, the size of isolation and the environmental change rate are added as the coefficients that determine the size of . In this way, the adaptive LMS algorithm is changed to enhance performance of the system by enhancing the adaptation rate according to the change of the environment. This paper also suggests the enhanced LMS algorithm having multi-feedback filters by applying n sub-blocks. The experiment result shows that the ITU 3GPP (3rd Generation Partnership Project) recommended specifications are met on both down-link and up-link. The cancellation was 25dBm, the window size was 500ns/1us, and the max. power was +10dBm. By using algorithms selectively according to the change of environment, the system reduced unnecessary use of hardware resources, and enhanced the convergence rate.

References 1. Lattice Semiconductor Corporation, LMS Adaptive Filter, Reference Design RD1031 (December 2006) 2. Kim, I.: The Improved 8-PSK Space-Time Trellis Codes on Fast Fading Channels for High Data Rate Transmission. Korea Institute of Information Technology 7(6), 93–98 (2010) 3. Song, Y.J., Kang, I.: New Rapid Synchronization Method for Indoor Wireless Channel Environment. Korea Institute of Information Technology 9(4), 105–112 (2011) 4. Dasgupta, S., et al.: Sign-Sign LMS convergence with Independent Stochastic Inputs. IEEE Trans. on Information Theory 36(1), 197–201 (1990) 5. Kwong, R.H., Johnston, E.W.: A variable step size LMS algorithm. IEEE Trans. Signal Processing 40(7), 1633–1642 (1992) 6. Rani, S., Subbaiah, P.V., Chennakesava Reddy, K.: LMS and RLS Algorithms for Antennas in a CDMA Mobile Communication Environment. International Journal of The Computer, the Internet and Management 16(3), 12–21 (2008) 7. http://www.altera.com 8. http://www.ti.com 9. Lee, M., Keum, B., Son, Y., Joo-Wan, K., Lee, H.S.: A New Low-Complex Interference Cancellation Scheme for WCDMA Indoor Repeater. In: IEEE Region 8 International Conference on Computational Technologies in Electrical and Electronics Engineering, SIBIRCON, pp. 457–462 (July 2008) 10. Kwong, R.H., Johnston, E.W.: A variable step size LMS algorithm. IEEE Trans. Signal Processing 40(7), 1633–1642 (1992) 11. Lan, T., Zhang, J.: FPGA Implementation of an Adaptive Noise Canceller. In: International Symposiums on Information Processing, pp. 553–558 (2008)

Power Model and Analysis of Wireless Transceiver System Jae-Hoon Choi and Heung-Gyoon Ryu Department of Electronic Engineering, Chungbuk National University, Cheongju, Korea [email protected], [email protected]

Abstract. Wireless Communication and mobile computing devices basically work by battery power. It is very important to calculate and apply the power consumption and furthermore to develop nice feasible techniques for reducing the power consumption of wireless communication devices. RF and analog parts in the mm-wave and EHF frequency domain typically consume more energy compared to the digital parts. So, to design the wireless battery-driven system more power efficiently, we have to investigate the system level energy model for the RF front-end of a wireless transceiver. Also, the effects of the signal bandwidth, PAR, date rate, modulation level, distance, specific attenuation of frequency band, and the signal center frequency on the RF front-end energy consumption and system capacity are considered. Eventually, we analyze the relationship between energy per bit and the data rate with the variation of the system bandwidth so that we can find the minimum energy per bit in the several Gbps data rate. Keywords: Power model, Energy consumption, Power consumption, Energy per bit.

1

Introduction

Wireless communication and mobile computing device are widely used in everyday life such as cellular phone, smart phone, PDA, tablet PC, RF ID tags, and so on. All of these devices are powered by batteries with a limited lifetime. Therefore, capacity of battery and power consumption of device are very important at these devices. Since the advance in battery technology have failed to keep up with increasing current consumption wireless communication and mobile computing device, efficient techniques to reduce the power consumption devices have to developed. The design of technique for low power wireless communication systems constantly attracts a great deal of researchers' attention. Different approaches of low power wireless communication have been addressed in recent years. These change of the modulation [1]-[2], multi-hop [3], scheduling method [4]. These approaches are focused on power consumption at digital parts. However, the wireless communication system consumes more power at RF parts. For example, about 75% of the power is consumed by RF front-end in an IEEE 802.11-b T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 274–282, 2011. © Springer-Verlag Berlin Heidelberg 2011

Power Model and Analysis of Wireless Transceiver System

275

wireless LAN card based on Intersil’s PRISM II chipset. Therefore, power model of RF parts is required to design of wireless system. Recently, the RF transceiver power model has been provided [5]. This paper was modeling of transmitter and receiver each device part. Also, the analyses the system quality apply to RF power model. But, this paper is only consider free space loss and that has not been analyzed the system performance according to frequency band. In this paper, we are considered the effect of signal bandwidth, PAR, symbol rate, modulation level, transmission distance, specific attenuation of frequency band and the signal center frequency on the RF front-end energy consumption and system capacity. Also, we analyze the relationship between energy per bit and the system bandwidth.

2

Power Consumption Model

2.1

Transceiver Block

The wireless transmitter and receiver structure that we have used is described in Fig, 1 and Fig.2 [5]. The analog device blocks of transmitter are DAC, reconstruction filter, mixer, RF synthesizer, power amplifier and RF filter. Also, the analog blocks of receiver are band select filter, LNA, mixer, RF synthesizer, baseband amplifier, baseband filter and ADC. Also, we assume that the transmitter and receiver works in three

Pulse shaping filter

Antenna

Reconstruction filter Mixer RF filter

DAC RF synthesizer

Fig. 1. Basic block diagram of the transmitter

Antenna Band Select Filter

LNA

Mixer

Baseband Amplifier

ADC Baseband filter

RF synthesizer Fig. 2. Basic block diagram of the receiver

276

J.-H. Choi and H.-G. Ryu

states : (1) active state when the signal is transmitted, (2) sleep state when there is no signal transmission, and (3) transient state when the transmitter switches from sleep state to active state or active state to sleep state. Therefore the total energy consumption is given by [5]

Etotal = PactiveTactive + PsleepTsleep + PtransientTtransient

(1)

In this paper, we only consider active state power consumption because the power consumption of active mode is dominant. Since transmission energy is delivered by PA [6]. Eactive = ( PPA + 2 Pmix + 2 PFS + PLNA + Pfilter + PBA + PDAC + PADC )Tactive

(2)

where PPA , Pmix , PFS , PLNA , Pfilter and PBA are the power consumption of the PA, mixer, frequency synthesizer, low noise amplifier, filter and baseband amplifier, respectively. 2.2

Power Model

In this section, we present the power models for each of the components in the analog signal chain of a transmitter and receiver using the existing RF power model [5]. The power model of DAC can be express as a function of PAR (Peak-to-average ratio), SQNR(Signal-to-quantization-noise ratio) signal bandwidth B and resolution. PDAC = Vdd ⋅ I 0 ⋅ ( 2SQNR ( dB ) + PAR ( dB )−4.77 dB /6.02 − 1) SQNR(dB) + PAR(dB) − 4.77 dB + 0.5 ⋅ ⋅ C p ⋅ OSR ⋅ B ⋅ Vdd 2 6.02

(3)

The power consumption of baseband active analog filter can estimate as follow [7],

Pfilter = n ⋅ kT ⋅ Q ⋅ f 0 ⋅ SNR 2

(4)

where n is a proportionality constant depending on the filter topology. The power consumption of integer-N PLL frequency synthesizer with multiplication ratio of N can be estimated as follow [5],

Ppll = b1 ⋅ C1 ⋅ Vdd 2 ⋅ FLO + b2 ⋅ C2 ⋅ Vdd 2 ⋅ Fref

(5)

where C1 and C2 represent the total parasitic capacitance loading of the RF circuits,

Fref is the reference frequency and Vdd is the supply voltage, which is also

assumed to be equivalent to the LO voltage swing. The power consumption of VCO is given by [5], PVCO = R ⋅ I pk 2 = C

R 2 R V pk = RC 2ωc 2V pk 2 = 2 2 V pk 2 L L ωc

(6)

Power Model and Analysis of Wireless Transceiver System

where

277

Vpk and I pk is the peak voltage and current amplitude inside the tank cir-

cuit. The power model of the mixer is a function of the noise figure K [6].

P

mixer

NF and the gain

= kmixer ⋅ K / NF

(7)

LNA amplifies the received signals with low input referred noise. LNA determines the noise figure of receiver. The power model of LNA is a function of the noise figure NF and the gain A [6].

P The efficiency [7]

LNA

= k LNA ⋅ A / NF

(8)

η of Class A PA is proportional to the rms value of the output power η=

Prms P K = out ⋅ K = PPA Pout _ max PAR

(9)

The power model of PA is thus given by [6] 2

−1  1  16 ⋅ π 2 ⋅ d 2 ⋅ L b 1  PPA = (2 − 1) ⋅ N ⋅  Q −1  1 − b / 2  SER   PAR. 2 4 2    3Gr Gt λ ⋅ K   

where Gr and Gt are the transmitter and receiver antenna gain,

(10)

d is free space

propagation at distance(meter). L is the system loss factor not related to propagation, and λ is the carrier wavelength. This equation only consider MQAM. Therefore, for other modulation scheme, the PA model is similar but Q function is different. Table 1. RF power consumption[5]

3

PA Mixer F.S

Power model function P (PAR, d, b, SER) P (K, NF) P ( ω c , FLO , Fref )

PAR = 10dB 246 mW 30.3 mW

LNA ADC DAC Filter BA

P (A, NF) P (PAR, SNR, f) P (PAR, SNR) P (SNR, f) P (B, α BA )

20 mW 5.85 mW 2.43 mW 5 mW 5 mW

67.5 mW

System Capacity

In this section, we study on the effect of channel propagation on the system capacity. We use channel propagation model (ITU-R 676-1) that is dry air and water vapour

278

J.-H. Choi and H.-G. Ryu

model. Also, we simulate the system capacity according to frequency band using Shannon capacity formula. 3.1

Specific Attenuation

The specific attenuation due to dry air at ground level (pressure = 1, 013 hPa) and at a D

temperature of 15 C is given by the following equation [8], γ dry = {7.19 × 10 −3 +

6.09 4.81 } f 2 10 −3 dB/km + f 2 + 0.227 ( f − 57) 2 + 1.50

for f < 57GHz

Also, the specific attenuation due to water vapour at ground level following equation, γ w = {0.050 + 0.0021ρ + +

where, f

(11)

is given by the

3.6 10.6 + ( f − 22.2) 2 + 8.5 ( f − 183.3) 2 + 9.0

( 12)

8.9 } f 2 ρ 10−4 dB/km for f <350GHz ( f − 325.4) 2 + 26.3

is frequency expressed in GHz, and

is the water vapour density ex-

pressed g / m3 . 2

10

Wator vapour Dry air Wator vapour+Dry air

1

Specific attenuation (dB/Km)

10

0

10

-1

10

-2

10

-3

10

-4

10

-5

10

5

10

15

20 25 30 35 40 Frequency band (1-60GHz)

45

50

55

60

Fig. 3. Specific attenuation - dry air and water vapour condition (1-60GHz frequency band)

3.2

System Capacity

The system SNR of the receiver is defined as the receiving power over the system noise figure, F. So, the SNR at the input of the demodulator can be written by [9]

Power Model and Analysis of Wireless Transceiver System

SNRdem =

SNRr P ⋅G ⋅G = out r t F PL ⋅ F ⋅ KTB

279

(13)

where, PL is free space loss and F is noise factor. This equation only consider free space loss. So, we add the specific attenuation. Therefore, SNRdem is given by

SNRdem = where,

Pout ⋅ Gr ⋅ Gt SNRr = F PL ⋅ F ⋅ KTB ⋅ Lsp

(14)

Lsp is specific attenuation that is dry air and water vapour.

The Shannon capacity formula is

C = B log 2 ( SNR + 1)

(15)

Therefore, system capacity is given by

C = B log 2 ( SNRdem + 1)

(16)

8

10

Free space (d=1Km) Dry+Water vapour (d=1Km)

7

Capacity (bps)

10

6

10

5

10

4

10

3

10

0

10

20 30 40 frquency band (1 - 60GHz)

50

60

Fig. 4. System capacity using Shannon capacity formula (d=1km, Pout =10 dBm, Noise figure = 6dB, bandwidth = 1.5GHz, G r =1, G t =1 )

Fig.4 shows the system capacity according to frequency band. In this figure, we can see that the system capacity has been falling sharply in 53 GHz - 60GHz because specific attenuation by dry air. Also, the system capacity is decrease at the high frequency band. Because that the free space loss of high frequency band is higher than low frequency band.

280

4

J.-H. Choi and H.-G. Ryu

Power Consumption

In this section, we simulate the total power consumption according frequency band. We use the RF power model from section 2. Table 2 summarizes the related parameters for the power consumption of PA and ADC. From the power models described in Section 2, we can see that center frequency directly affects the power consumption of ADC, PA and filter. Therefore, we simulate the ADC, PA and filter according the center frequency. Table 2. PA and ADC simulation parameter.

Bandwidth Frequency band Distance Modulation

Parameter 20MHz 1-60GHz 1km QPSK, 16QAM

SER G_r G_t Loss

Noise power Roll-off factor

-101 dBm 0.2

V_dd L_min

Parameter 10^-4 1 1 0.8 3V 0.4 um

From the equation (15), we can define the SNR at the input of the demodulator. So, equation (16) can be written as ε ⋅R

SNRdem = 2 B − 1 =

Pout ⋅ Gr ⋅ Gt KTB ⋅ PL ⋅ F ⋅ Lsp

(18)

where ε ≥ 1 and is a pure number. So, the out power of the transmitter becomes Pout =

KTB ⋅ PL ⋅ F ⋅ Lsp  εB⋅ R  ⋅  2 − 1 Gr ⋅ Gt  

(19)

The total energy is given by

P  Etot =  out + Pmixer + PFS + Pfilter + PDAC  ⋅Tt + Pr ⋅ Tr η 

(20)

Pr is receiver power consumption in LNA, BA, mixer, frequency synthesizer, filter, base-band amplifier and ADC. So, we find that Pr is 427.73 mW at 60GHz.

Pr = Pmixer + PFS + PLNA + Pfilter + PBA + PADC

(21)

Also, the total power consumption is given by

Ptot = Etot ⋅

1 Tt + Tr

(22)

Power Model and Analysis of Wireless Transceiver System

281

Next, the energy per bit is becomes

Eb = Ptot ⋅

1 Rb

(23)

Therefore, the energy consumption per bit, Eb , is found as  KTB ⋅ PL ⋅ F ⋅ Lsp Eb =  Gr ⋅ Gt ⋅η 

 1  ε ⋅R  1 ⋅  2 B − 1 ⋅ Tt + Pr ⋅ Tr  ⋅ ⋅ T T R +  t r b   

(24)

-10

x 10

Bw=1GHz Bw=2GHz

14

Energy per bit (J/bit)

12

10

8

6

4

2

X: 9.805e+009 Y: 2.865e-011 8

10

9

10

10

10 R (Data Rate)

Fig. 5. Energy per bit (center frequency = 60GHz, d = 20m,

11

10

ε =1. η =5%, F=12.6, Tt =

Tr =1 ). Fig.5 shows the effect of bandwidth and R on the energy per bit. At R<5Gbps, the energy per bit is equal bandwidth 1GHz and bandwidth 2GHz. But, energy per bit of bandwidth 1GHz is higher than energy per bit of bandwidth 2GHz at R > 5Gbps. Because this energy per bit is increase the according to 2 ^ ( R / B ) . Therefore, the energy per bit has an affinity with the bandwidth.

5

Conclusion

In this paper, we analyze the system power consumption using the RF power model. Also, we analyze the relations between energy per bit and system bandwidth. The system power consumption is increase according to frequency band. Also, the power consumption of PA is more dominate the RF parts. And fixed the distance and compare the system capacity, we can see that higher frequency band, the lower the system

282

J.-H. Choi and H.-G. Ryu

capacity. But, high frequency band can be easily securing the wide system bandwidth. Therefore, the system frequency band and bandwidth is more important for low power consumption. Acknowledgment. This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology(No. 2010-0007567).

References 1. Kim, D.K., Lee, H.S.: Phase-Silence-Shift-Keying for Power-Efficient Modulator. IEICE Trans. Communication E-92B(6) (June 2009) 2. Choi, J.H., Ryu, H.G.: A QAPM(Quadrature Amplitude Position Modulation) for Low Power Consumption Communication. In: Proc. ISWPC (February 2011) 3. Ramanathan, R., Rosales-Hain, R.: Topology control of multihop wireless networks using transmit power adjustment. In: Proc. IEEE INFOCOM 2000 (March 2000) 4. Huang, W., Letaief, K.B.: Cross-layer scheduling and power control combined with adaptive modulation for wireless ad hoc networks. In: Proc. GLOBECOM 2005 (December 2005) 5. Li, Y., Bakkaloglu, B.: A System Level Energy Model and Energy-Quality Evaluation for Integrated Transceiver Front-Ends. IEEE Trans. Very Large Scale Integration Systems 15(1), 90–103 (2007) 6. Wambacq, P., Vandersteen, G., Donnay, S., et al.: Higher-level simulation and power modeling of mixed-signal front-ends for digital communications. In: Proc. IEEE ICECS, pp. 525–528 (1999) 7. Chen, P.F., Larson, L.E., et al.: High-efficiency power amplifier using dynamic powersupply voltage for CDMA applications. IEEE Trans. Microw. Theory Tech. 47(8), 1471–1476 (1999) 8. ITU-Rec. 676-1, Attenuation by Atmospheric Gases in the Frequency Range 1-350GHz (1992) 9. Li, X., Baltus, P., et al.: Wireless Wire-the 60GHz Ultra-Low Power Radio System. In: Proc. Radio and Wireless Symposium, RWS 2009 (January 2009)

Feedback Scheduling for Realtime Task on Xen Virtual Machine Byung Ki Kim , Kyung Woo Hur, Jae Hyuck Jang, and Young Woong Ko Dept. of Computer Engineering, Hallym University, Chuncheon, Korea {bkkim,firehur,jaehyuck,yuko}@hallym.ac.kr

Abstract. In virtual machine environments, it is difficult to allocate CPU resource to virtual machine efficiently because virtual machine lacks knowledge of each domains workload. Especially, realtime tasks in guest domain have to finish before their deadline, however, virtual machine scheduler is not aware of guest-level tasks and how much resources guest domain requires. In this paper, we present a virtual machine scheduling framework based on feedback mechanism. The proposed mechanism exploits various scheduling information from each domain. Xen scheduler controls the CPU allocation by increasing or decreasing CPU slices. We evaluate our prototype in terms of realtime task performance over diverse workload. Our experiment result shows that feedback mechanism effectively allocates CPU resources for guest domain in varying workloads. Keywords: Xen, realtime, scheduler, QoS, resource monitor.

1

Introduction

To support realtime tasks on a virtual machine, virtual machine monitor (VMM) should predict guest domains CPU requirement exactly. For example, guest domain for multimedia streaming service must be executed in a time-sensitive manner, or it will fail to support multimedia task. Therefore, to satisfy quality of service (QoS) of multimedia, guest domain must receive appropriate timeliness guarantees from virtual machine monitor. However, multimedia streaming server has diverse workloads and it is difficult to predict exact amount of requirement of workloads. Since this lack of knowledge about future workloads makes VMM difficult to allocate CPU resources efficiently. VMM could not track which domain is busy and need more CPU because VMs are consolidated. It will degrade realtime guest domains performance and responsiveness. This paper present a realtime-aware virtual machine scheduling mechanism based on feedback mechanism. Our goal is to improve the QoS of realtime task under the condition of 

This research was financially supported by the Ministry of Education, Science Technology (MEST) and National Research Foundation of Korea(NRF) through the Human Resource Training Project for Regional Innovation and Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology(2011-0006040).

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 283–291, 2011. c Springer-Verlag Berlin Heidelberg 2011 

284

B.K. Kim et al.

varying workloads. We designed and implemented VMM monitoring tools and measured the workload of each domain. By predicting CPU usage of each domain, we can dynamically increase and decrease CPU resource of each domain. The reminder of this paper is organized as follows: Section 2 describes the design and implementation of the proposed scheduling mechanism. Section 3 demonstrates and discuses experimental results and Section 4 explains related works. Finally, in Section 5, we concludes our works and discuss future works.

2

Feedback Scheduling Framework

To allocate CPU efficiently, VMM should provide admission control mechanism in real time. However, it is difficult because VMM lacks knowledge of each domains workloads as we mentioned earlier. Figure 1 shows the overall architecture of the proposed system. The proposed system is composed of three parts; first, QoS monitor module in guest operating system checks whether realtime tasks miss their deadline. If there is deadline miss in a certain level, QoS monitor

Domain-0

Guest Domain N

Guest Domain 1

CPU allocation (period, Slice)

RT RTRT Task Task

Task

RT RT NRT Task Task Task

RT RTRT Task Task

Task

RT RT NRT Task Task Task

CPU Utilization CPU Utilization Monitoring

QoS Monitor

Monitoring CPU utilization and allocated CPU slice

QoS Monitor

RT task missed deadline

*[RGTXKUQT Return unused CPU time Hypercall to reduce slices Re-allocate less Slice : slice - utilization

Request more CPU time

Hypercall interface

Hypercall to increase slices Allocate more Slice : slice++

SEDF on the Xen

Fig. 1. The general architecture of the proposed system

Feedback Scheduling for Realtime Task on Xen Virtual Machine

285

requests more CPU slice to VMM scheduler. In this work, we made additional hypercall interface to adjust slices. Second, Scheduling tool in domain0 manages system resource and controls scheduling policy. User can change scheduling information using scheduling tool by changing period and slice of each domain. This tool periodically checks available CPU resource and reallocates CPU resource dynamically. Third, in hypervisor, we implemented feedback scheduling module which receives hypercall from guest operating system and adjusts slice of each domain. When VMM receives the hypercall from guest operating system then SEDF will increase its slice for 1 millisecond. SEDF stop increasing slice when there is no more slice request. In this work, if there is no enough CPU resource to allocate for a VM requesting slice, we have to prohibit excessive resource allocation. 2.1

Requesting CPU Resource

Credit scheduler is a default scheduler in Xen hypervisor and it automatically allocates the same weight to the domains. Default weight value is specified as 256. In credit scheduler, user can adjust weight value differently, however its difficult to allocate fixed amount of CPU resource. SEDF is a second scheduler supported by Xen. We can use SEDF algorithm if we specify SEDF as a default scheduler during boot time. SEDF can execute domain as a realtime manner, however it has drawback for supporting global scheduling which means SEDF cannot distribute workloads between processors. In our work, we used SEDF as a main scheduler, so, each guest domain was initially allocated period and slice. We implemented QoS monitoring tool with library routine. Therefore, all the realtime tasks running on a guest domain calls this library routine. Library routine requests more CPU allocation to SEDF scheduler by calling hypercall routine. Therefore, whenever a realtime task misses deadline, library routine hypercall to VMM. 2.2

Returning CPU Resource

To prevent excessive CPU request by a domain, we have to isolate each domain in a certain level. In this work, scheduling tools running on a domain0 periodically evaluates CPU usage of each domain. For every second, VMM tracks domains CPU utilization. When CPU utilization of current a domain is lower than assigned amount of CPU time, VMM will reduce its CPU slices by one slice for every second. Its heuristic approach, but our prototype focuses on allocation CPU slice real time in diverse workloads. For example, if a domain is assigned 20% of CPU resource (period: 10 ms, slice: 2 ms) then the domain have to consume 20% of CPU. When CPU utilization of a domain has decreased to 15%, the slice must be decreased to 5%. In VMM scheduling modules, there is a runnable queue where each VCPU is sorted by ascending deadlines. SEDF schedules domains from head of PCPU on run-queue. Once a VCPU has consumed its slice for current period, this VCPU removed from run-queue and moved to waitqueue to receive a fresh slice and period. The wait queue is sorted in ascending

286

B.K. Kim et al.

order of the start time of next period. Only User (Administrator) can configure these scheduling parameters. Before the administrator adjusts these parameters, VM scheduled as best-effort mode. SEDF has two more queues, utilization and penalty queue called extra queue. VMs scheduled under best-effort mode, every VCPU is on extra queue and scheduled every 500 micro seconds by round robin manner. When VCPUs increase on a PCPU its scheduling delay increases proportionally.[8] proposed to improve the worst case scheduling response time. They remove its short unblocking situation. But there is still scheduling delay exist. To reduce the effect of context-switch overhead, when the domain is blocked before the end of the current period, SEDF removes current VCPU to extra queue.

3

Experiment and Evaluation

In this experiment, we focus on deadline miss ratio for measuring realtime task. We made a various experiments to draw all the aspects of the proposed system capabilities. Our experimental platform consists of the following components. As shown in table 1, our hardware platform has dual core processor. The software platform is based on CentOS Linux kernel that is widely used in Xen virtualization. We installed 2 domains on Xen hypervisor. Table 1. Experiment Environment

H/W

S/W

CPU Cache Memory Network Disk VMM Dom0 DomU

Intel Core i7 920 2.66Ghz, Dual L2: 256KB * 4, L3 : 8MB DDR3 PC-10600 2G * 3 Gigabit Ethernet Seagate 1TB 7200 RPM Xen 3.4.3 CentOS 5 2.6.18.8-xen0 CentOS 5 2.6.18-164.el5xen

In this experiment, we made two realtime domains, two non-realtime domains and domain0. Figure 2 shows VCPUs on each PCPU. Domain0 is pined on PCPU0 and other VCPUs are pined on PCPU1. Only VMs can share 100 percent of PCPU1 resources. We set these VCPUs on a PCPU because VMs are not affected by Domain0. To evaluate our proposed mechanism, we implemented time-driven periodic realtime task that periodically produce hash work. These types of workloads are periodically executed on real-time domain. We made a periodic task using MD5 hash function, which executes E time units during P period. We can control execution time E by varying hash block size. The period of realtime task is set 33, so every second the task can produce 30 hash data. If we try to change workloads then we simply increase/decrease hash data size. To eliminate additional overhead incurred by I/O system, we generate data which consumed by MD5 hash function.

Feedback Scheduling for Realtime Task on Xen Virtual Machine

RT Guest

Domain0

287

NRT Guest

Hypervisor PCPU

PCPU

Dom0 VCPU-0

Guest1 VCPU-0 Guest2 VCPU-0 Guest3 VCPU-0 Guest4 VCPU-0

Realtime

NonRealtime

Fig. 2. VCPU configuration for experiment

3.1

Workload Setup

Figure 3 shows how realtime domains work when the workloads of each realtime domain changed. X-axis is processing time and Y-axis is hash data size. To make varying workloads, we changed hash data every 60 seconds. The total hash size is between 0.8 MByte to 1 MByte. In this experiment, two realtime tasks are executed concurrently on each realtime domains. At first, we setup SEDF parameter based on start time workloads. As time goes, the workloads are fluctuating whenever the hash data size is changed. If a domain increases hash data size, it will need more CPU resource. Accordingly, the system starts to increase slice, then the domain may prevent deadline miss for a realtime task. After period of time, the domain may have enough CPU resource and not used for realtime task. At this point, available CPU resource should be returned to help other realtime domain. 3.2

Deadline Miss Result

Figure 4 presents how many times each domain missed deadline. X-axis means hash data size and black line means deadline miss. As we can see in this figure, when hash data size is increased, there are lots of deadline misses. However, if hash data size is decreased then there is no deadline miss. Figure 5 shows the proposed system result. As mentioned earlier, if realtime task misses its deadline, QoS monitor requests more CPU slice with hypercall interface. As shown in figure 5, there exists instant deadline misses when hash data size is increased,

288

B.K. Kim et al.

RT−Guest 1 RT−Guest2 Total

1.0

Workloads (MB)

0.8

0.6

0.4

0.2

0

60

120

180 Time (sec)

240

300

360

Fig. 3. Workloads for each domain and total workloads

however with slice increasing for the domain, deadline miss is eliminated. Figure 6 shows how much slices were allocated to VMs at each workload. All of guest domain receives 100 ms period as default. Default SEDF schedules each VM by round-robin fashion. Every VM has 500 microseconds. If 4 domains were booted and work with same workload, their utilizations are 25% for each. It means Xen hypervisor guarantees the fairness, but we are focused on performance enhancement of realtime domain. Although our mechanism does not guarantee the fairness, we guarantee performance of realtime domain which has higher priority.

4

Related Work

Performance analysis and Resource allocation have been well conducted on the Xen VMM[1]. Many researches are roughly focused on improving I/O performance, network response, CPU allocation, resource monitoring and real time guarantee. Three scheduling mechanisms are well analyzed in terms of I/O performance and CPU utilization according to different parameters in various workloads [2][3]. Analyzing I/O bound tasks is complicate because isolated driver domain (IDD) processes I/O processing on behalf of VMs[9]. To improve I/O processing like disk I/O and communication via NICs, SEDF-DC introduces a performance monitoring and profiling tool that reports VPU usage of VMs and VM scheduler with feedback [4]. Govindan et al. introduced communication-aware VM scheduling mechanism for high throughput of network intensive workloads.

Feedback Scheduling for Realtime Task on Xen Virtual Machine

0

0.4

0.5

0

0.5

0.3

0.6

0.4

0.4 Different workloads (MB)

0.3

0.2

0.6

Fig. 4. Deadline miss rate in different workloads with default SEDF

0

0

0.4

0.5

0.5

0.3

0.6

0.4

0.4 Different Workloads (MB)

0.3

0.2

0.6

Fig. 5. Deadline miss rate in different workloads on proposed system

289

290

B.K. Kim et al.

Fig. 6. Slice allocation result on different workloads

Their scheduling mechanism is picking a domain that is likely to experience the most overall reduction in scheduling delay. The domain that has received the most number of packets has the highest priority [5]. Lee et al. suggests soft realtime scheduler for Xen hypervisor by modifying credit scheduler to calculate scheduling priority dynamically. They defined laxity value that provides an estimate of when the task needs to be scheduled next. When a VCPU of a real-time domain is ready, it is inserted where its deadline is expected to be met. This approach deals with low-latency tasks to be executed timely manner. However, it has difficult to guarantee real-time workloads because there is no admission control mechanism. Therefore, if there increase workloads, it cannot meet realtime tasks deadline[6]. RT-Xen introduces a hierarchical real-time scheduling framework in Xen. The key idea is twofold, first, RT-Xen provides a sporadic server root scheduler for Xen that is compatible with eh RM scheduling. Second, they use 1 ms scheduling resolution while incurring moderate overhead. However, RT-Xen cannot support an admission control mechanism for real-time domain, therefore in an excessive workloads, its difficult to guarantee real-time task [7].

5

Conclusion

In this paper, we show a practical scheduling mechanism for realtime guest domain based on task feedback. When a domain needs more CPU then VMM catch this event immediately and adjust CPU resource to guarantee QoS of each domain. In case of varying workloads, VMM has difficulty to predict how much CPU allocation to each domain. On default SEDF scheduler every VM works as

Feedback Scheduling for Realtime Task on Xen Virtual Machine

291

best-effort mode, therefore it is difficult to satisfy varying workloads of realtime domain. In this work, we propose a mechanism that VMM scheduler let notify scheduling event to VMM. Experiment result shows that our prototype system excessively minimizes deadline miss ratio compared to default SEDF. For future work, we consider self-adaptable scheduler that supports varying I/O workloads.

References 1. Barham, P., Dragovic, B., Fraser, K., Hand, S., Harris, T., Ho, A., Neugebauer, R., Pratt, I., Warfield, A.: Xen and the art of virtualization. In: Proceedings of the Nineteenth ACM Symposium on Operating Systems Principles, pp. 164–177. ACM, USA (2003) 2. Cherkasova, L., Gupta, D., Vahdat, A.: Comparison of the three CPU schedulers in Xen. SIGMETRICS Perform. Eval. Rev. 35, 42–51 (2007) 3. Cherkasova, L., Gupta, D., Vahdat, A.: When virtual is harder than real: Resource allocation challenges in virtual machine based it environments. Tech. Rep. HPL2007-25 (2007) 4. Gupta, D., Cherkasova, L., Gardner, R., Vahdat, A.: Enforcing performance isolation across virtual machines in Xen. In: Proceedings of the ACM/IFIP/USENIX 2006 International Conference on Middleware, pp. 342–362. Springer, New York (2006) 5. Govindan, S., Nath, A.R., Das, A., Urgaonkar, B., Sivasubramaniam, A.: Xen and co.: communication-aware CPU scheduling for consolidated xen-based hosting platforms. In: Proceedings of the 3rd International Conference on Virtual Execution Environments, pp. 126–136. ACM, USA (2007) 6. Lee, M., Krishnakumar, A.S., Krishnan, P., Singh, N., Yajnik, S.: Supporting soft real-time tasks in the xen hypervisor. In: Proceedings of the 6th ACM SIGPLAN/SIGOPS International Conference on Virtual Execution Environments, pp. 97–108. ACM, Pennsylvania (2010) 7. Xi, S., Wilson, J., Lu, C., Gill, C.: RT-Xen: Real-time virtualization based on hierarchical scheduling. Washington University Technical Report WUCSE-2010-38 (2010) 8. Masrur, A., Drossler, S., Pfeuffer, T., Chakraborty, S.: VM-Based Real-Time Services for Automotive Control Applications. In: Proceedings of the 2010 IEEE 16th International Conference on Embedded and Real-Time Computing Systems and Applications, pp. 218–223. IEEE Computer Society (2010) 9. Kim, H., Lim, H., Jeong, J., Jo, H., Lee, J.: Task-aware virtual machine scheduling for I/O performance. In: Proceedings of the 2009 ACM SIGPLAN/SIGOPS International Conference on Virtual Execution Environments, pp. 101–110. ACM, Washington, DC, USA (2009)

DTAR: Deduplication TAR Scheme for Data Backup System Sung Woon Kang1, , Ho Min Jung1 , Jung Geun Lee1 , Jin Haeng Cho2 , and Young Woong Ko1 1

Dept. of Computer Engineering, Hallym University, Chuncheon, Korea {upersbird,chorogyi,jeonggun.lee,yuko}@hallym.ac.kr 2 Boston Design Center AMD, MA, USA [email protected]

Abstract. Tar archive format does not support file deduplication scheme, therefore it has drawbacks for utilizing disk storage system efficiently. In this paper, we propose an extended TAR file format called DTAR (Deduplication Tape Archive) which provides file-level deduplication. Key idea of our work is to provide block aligned compressed file that can speed up file insertion and deletion for archiving. Furthermore, DTAR shows performance enhancement using file similarity scheme for distributing files into several storage nodes. Experiment results show that the proposed system can reduce data storage space efficiently and diminish network data traffic compared to general file transfer system. Keywords: Deduplication, TAR, DTAR, Distributed File System.

1

Introduction

Recently, there is digital data explosion caused by high speed internet and multimedia data, which requires an efficient disk management. Therefore we have to consider an efficient data management when we design a data storage system. Data deduplication scheme is a strong technique for eliminating coarse grained redundant data management and often adapted to various storage systems including backup system [1], FTP mirror, virtualization system. In this paper, we propose a simple and powerful data management scheme by extending TAR file format, which is developed for tape backup and commonly used to collect many files into one larger file for distribution or archiving purpose. Our key idea is to adapt data deduplication scheme when handling files in TAR format. Generally, we can find many duplicated files in file sets (group of files handled by archiving system) including Linux distribution, backup system and P2P. Therefore, 

This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology(2011-0002439) and by the Ministry of Education, Science Technology (MEST) and National Research Foundation of Korea(NRF) through the Human Resource Training Project for Regional Innovation.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 292–300, 2011. c Springer-Verlag Berlin Heidelberg 2011 

DTAR: Deduplication TAR Scheme for Data Backup System

293

if we eliminate duplicated files in file set, we can reduce the size of TAR file. To accomplish this goal, first, we adapted file-level deduplication for archiving system that minimizes the capacity of storage system. Second, we provide block aligned compressed file in a DTAR system, which can speed up file insertion and deletion in the DTAR system. Furthermore, our approach exploits file similarity based load balancing mechanism when distributing files to storage nodes. The rest of this paper is organized as follows. In Section 2, we describe related works about deduplication system. In Section 3, we explain the design principle of proposed system and implementation details for DTAR system. In Section 4, we show performance evaluation result of DTAR scheme and we conclude and discuss future research plan.

2

Related Work

Venti [2] is a block-level network storage system which is similar to the proposed system. Venti identifies data blocks by a hash of their contents, because of using a collision-resistant hash function (SHA1) with a sufficiently large output, the data block can be used as the address for read and write operations. The Low-Bandwidth File System [3] makes use of Rabin fingerprinting [4] to identify common blocks that are stored by a file system client and server, to reduce the amount of data that must be transferred over a low bandwidth link between the two when the client fetches or updates a file. Data domain [5] is among the earliest research in the inline storage deduplication area. They present two techniques that aim to reduce lookups on the disk-based chunk index. First, a bloom filter [6] is used to track the chunks seem by the system so that disk lookups are not made for non-existing chunks. Second, upon a chunk lookup miss in RAM, portions of the disk-based chunk index are prefetched to RAM. Lillibridge et al. [7] use the technique of sparse indexing to reduce the in-memory index size for chunks in the system at the cost of sacrificing deduplication quality. The system chunks the data into multiple megabyte segments, which are then lightly sampled (at random based on the chunk SHA-1 hash matching a pattern), and the samples are used to find a few segments seen in the recent past that share many chunks. Obtaining good deduplication quality depends on the chunk locality property of the dataset whether duplicate chunks tend to appear again together with the same chunks. DEDE [8] is a decentralized deduplication system designed for SAN clustered file systems that supports a virtualization environment via a shared storage substrate. Each host maintains a write-log that contains the hashes of the blocks it has written. Periodically, each host queries and updates a shared index for the hashes in its own write-log to identify and reclaim storage for duplicate blocks. Unlike inline deduplication systems, the deduplication process is done out-of-band so as to minimize its impact on file system performance. HYDRAstor [9] discusses architecture and implementation of a commercial secondary storage system, which is content addressable and

294

S.W. Kang et al.

implements a global data deduplication policy. Recently, a new file system, called HydraFS [10], has been designed for HYDRAstor. In order to reduce the disk accesses, HYDRAstor uses bloom filter in RAM.

3 3.1

Design and Implementation of DTAR System DTAR Architecture

TAR file format was developed to write directly to sequential I/O devices for tape backup purposes, it is now commonly used to collect many files into one large file for distribution or archiving, while preserving file system information such as user and group permissions, dates, and directory structures. The TAR file aggregates one or more files, and each file is preceded by a 512 byte header record. The file data is written unaltered except that its length is rounded up to a multiple of 512 bytes and the extra space is zero filled. The end of an archive is marked by at least two consecutive zero-filled records. Basically TAR file format use 512 byte block but block size can be increased because of tape and system properties. In this work, we target DTAR system for backup purpose. In TAR-based backup system, full back up stores every file on a file system, whether file has changed or not. TAR doesnt support file version information and data deduplication, therefore, TAR files are not efficient for backup system using versions. In this work, we propose DTAR file format. DTAR supports data deduplication, compression and version management. DTAR format consists of single header field and file block field. Header has information such as file size, offset, file path, file properties. In addition, DTAR can check whether there exists duplicated files using SHA1 file hash data. We made file hash data for distinguishing duplicated file in DTAR format. Generally, backup data has similar files compared with previous backup version. Therefore, if we compare hash data of a file with old version of DTAR file format, then we can get rid of duplicated files. Furthermore, each file in DTAR format is compressed for diminish storage capacity. A TAR file is simply an archive, no compression techniques are used to reduce the size of the file. If the archive needs to be compressed then additional compress tools must be used. Therefore a TAR file is combined with other tools such as GZIP or BZIP2. This creates a file such as .tar.gz, or .tgz. In this work, we first compress each file and then archive it. When we append each compressed file to DTAR format, we align the file according to the system block size. The reason why we adapt ”first compress and later archive strategy” is to support efficient file add and remove. Figure 1 shows the architecture of DTAR system. DTAR processes file-level deduplication for a file set then it only stores non-duplicated files on DTAR format. DTAR compresses all files and makes one file image by inserting dummy data at the end of each file to accomplish block align between compressed files. DTAR processing steps are as follows: (1) Open the files in the selected directory sequentially and DTM (De-duplication TAR Management) send it to the server with hash data of files to be backed up. The server checks for duplicated files in the hash list and then send non-duplicated hash list to the client. The client makes deduplication work for the directory and file list by hash list received

DTAR: Deduplication TAR Scheme for Data Backup System

295

Fig. 1. Theoverview of DTAR system

from the server. Also DTAR system calculates space of header and block field in advance. (2) Files are compressed to reduce the disk size. We used zlib API for file compression. First, open in the gzopen function. Second, read the contents of the file placed in the buffer and buffer compress in the gzwrite function. At the same time calculate file hash. (3) After compress step, SHA1 hash and file information input the header file. Header file is aligned with 512 byte in advance. (4) Combine header file and file blocks. 3.2

Load Distribution Based on Files Similarity

If there are lots of DTAR files on a server, it’s difficult to find out duplicated DTAR files. Therefore, we have to search all of the DTAR files in server nodes. In this work, we applied file similarity scheme for distributing DTAR files on several nodes. The main idea is to send a DTAR file to a node that has similar DTAR files. This means that high similarity DTAR file may have duplicated compressed files. In this work, we can compute a file similarity by extracting and comparing hash data set of each file. First, DTAR computes each chunk using a sliding window and a rolling hash function. We used the Rabin hash[4] as a rolling hash function. After hashing work, we sort all the hash data and takes 10 high hash data for each DTAR file. If we compare the 10 data between files then we can compare if the file has duplicated blocks in a file. Figure 2 presents how file similarity mechanism works. For example, if a file has two same hash data then we expect that there will be 20% of duplicated blocks between the files. The higher the value means high probability of similar file.

296

S.W. Kang et al.

Fig. 2. Measuring files similarity

3.3

DTM Architecture

DTM (De-duplication Tar Management) is a system configuration utility that works in the server and client as a daemon, which supports DTAR backup, restore, duplication and resource management.

Fig. 3. DTM system architecture

Figure 3 shows the DTM architecture. File deduplication process is to compare SHA1 hash of file between the client and server DTAR header. TEMP space is used as a temporary buffer for cache memory. RB Tree used as a data structure for hash comparison, which is an efficient algorithm for frequent insertion and comparison of hash data. Figure 4 illustrates the backup process using file similarity. The client requests backup for collections of files then DTM check if

DTAR: Deduplication TAR Scheme for Data Backup System

297

Fig. 4. DTM backup process

the file is located on the server by comparing hash data. DTM determine the location of file through count and link of DTAR header.

4

Performance Evaluation

To evaluate the proposed system, we implemented the client and deduplication server. The server and client platform consist of 3GHz Pentium 4 Processor, WD-1600JS hard disk and 100Mbps network. In our experiment, we used Linux kernel source distributed by kernel.org that has 170MByte size and roughly 17,000 files. Table 1 shows experiment data for back-up which is several version of Linux kernel data. Table 1. Experiment data set : Linux kernel source

Data size Number of files

4.1

Set 1

Set 2

Set 3

512 MB 47,991

1002 MB 96,880

1629 MB 147,101

Computation Time and Storage Reduction

First, we conducted performance test for DTAR system. We compared DTAR with tar.gz format. As we mentioned previous section, DTAR provides archiving and compression, however Tar format only supports archiving. Therefore we made tar.gz format by adding compress option using Tar command. Figure 5 shows performance result by varying duplicate rate. In this experiment, we made experiment data by adding duplicated files continuously form 0% to 100%, 0%

298

S.W. Kang et al.

(a) Storage reduction

(b) Computation time

Fig. 5. Experiment result 1 : varying duplicate rate

means there is no any duplicated files and 100% means all files are same. If file set contains more duplicated files then DTAR can minimize storage capacity by using file-level deduplication. Computation time also dramatically decreased when there are more duplicated files in a file set. As we can see figure 5, tar.gz format minimize storage space compared to TAR approach because it compress files. However, Tar and compress Tar cannot eliminate duplicated files in file set. Therefore, file size is always same regardless of duplication ratio. DTAR shows similar result when the duplication ratio is small, however as duplication ratio is increased, file size is dramatically reduced. In terms of computation time, DTAR can save computation time for processing file back up. The main contribution of minimizing computation time is file transfer time. DTAR only sends nonduplicated files to the server, therefore as duplication ration is increased, total bytes transferred to the server is minimized.

(a) Computation time

(b) Storage capacity

Fig. 6. Experiment result 2 : Linux kernel version

DTAR: Deduplication TAR Scheme for Data Backup System

299

In figure 6, we conducted performance test using Linux kernel source with several version in Table 1. DTAR format minimizes computation time and can reduce storage capacity efficiently. In terms of computation time and storage reduction size, DTAR minimizes roughly 30% compared to tar.gz format. We believe that if we conduct version based back up data then there will be more performance gains than Linux kernel source experiment. 4.2

File Similarity Test

In this experiment, we conducted file similarity test for several files that have duplicated blocks from 20% to 90%. File similarity module produces 10 hashes for each file and compares it with actual duplicated blocks in the file. In Table 2, we can find almost all files can detect duplicate rate with high probability. For example, Set 2 contains 40% duplicated blocks and the file similarity result shows 4, which means file similarity module can predict exact duplicate rate in a file. Table 2. File similarity test result

Duplicate Rate (%) Number of Same Hash error

5

Set 1

Set 2

Set 3

Set 4

Set 5

Set 6

20% 2 0

40% 4 0

60% 7 1

70% 6 1

80% 9 1

90% 10 1

Conclusion

In this paper, we propose a novel archiving system, called DTAR that can provide file-level deduplication. DTAR provides block aligned compressed file that can speed up file insertion and deletion for archiving. It also exploits file similarity for distributing DTAR files into several storage nodes. The key idea of DTAR is to minimize the capacity of storage system using file-level deduplication. Additionally, DTAR can speed up file insertion and deletion of compressed file using block alignment scheme. In our work, we can boost up performance by exploiting file similarity based load balance mechanism when building storage node. In experiment results, we show that the proposed system can reduce data storage space efficiently. For future work, we will study how to optimize compressed files in DTAR system considering compress algorithms. Moreover, we will design the optimized deduplication module by continuously monitoring deduplication server.

References 1. Cox, L.P., Murray, C.D., Noble, B.D.: Pastiche: Making backup cheap and easy. ACM SIGOPS Operating Systems Review 36, 285-298 (2002) 2. Quinlan, S., Dorward, S.: Venti: a new approach to archival storage. Proceedings of the 1st USENIX conference on File and storage technologies, pp. 7-7. USENIX Association, Monterey, CA (2002)

300

S.W. Kang et al.

3. Muthitacharoen, A., Chen, B., Mazieres, D.: A low-bandwidth network file system. ACM SIGOPS Operating Systems Review 35, 174-187 (2001) 4. Rabin, M.O.: Fingerprinting by random polynomials. Center for Research in Computing Techn., Aiken Computation Laboratory, Univ. (1981) 5. Zhu, B., Li, K., Patterson, H.: Avoiding the disk bottleneck in the data domain deduplication file system. Proceedings of the 6th USENIX Conference on File and Storage Technologies, pp. 1-14. USENIX Association, San Jose, California (2008) 6. Broder, A., Mitzenmacher, M.: Network applications of bloom filters: A survey. In: Internet Mathematics, pp. 1-14. AK Peters, (2004) 7. Lillibridge, M., Eshghi, K., Bhagwat, D., Deolalikar, V., Trezise, G., Camble, P.: Sparse indexing: large scale, inline deduplication using sampling and locality. Proccedings of the 7th conference on File and storage technologies, pp. 111-123. USENIX Association, San Francisco, California (2009) 8. Clements, A.T., Ahmad, I., Vilayannur, M., Li, J.: Decentralized deduplication in SAN cluster file systems. Proceedings of the 2009 conference on USENIX Annual technical conference, pp. 8-8. USENIX Association, San Diego, California (2009) 9. Dubnicki, C., Gryz, L., Heldt, L., Kaczmarczyk, M., Kilian, W., Strzelczak, P., Szczepkowski, J., Ungureanu, C., Welnicki, M.: HYDRAstor: a Scalable Secondary Storage. Proccedings of the 7th conference on File and storage technologies, pp. 197-210. USENIX Association, San Francisco, California (2009) 10. Ungureanu, C., Atkin, B., Aranya, A., Gokhale, S., Rago, S., Calkowski, G., Dubnicki, C., Bohra, A.: HydraFS: a high-throughput file system for the HYDRAstor content-addressable storage system. Proceedings of the 8th USENIX conference on File and storage technologies, pp. 17-17. USENIX Association, San Jose, California (2010)

Effect of Maximum Node Velocity on GA-Based QOS Routing Protocol (QOSRGA) for Mobile Ad Hoc Network Jiwa Abdullah Dept of Communication Engineering, Faculty of Electrical and Electronic Engineering, Universiti Tun Hussein Onn Malaysia, Johor, Malaysia [email protected]

Abstract. Multiobjective formulations are realistic models for many complex engineering optimization problems such as QoS routing protocol for mobile ad hoc network. The paper presents QoS routing protocol for MANET with specialized encoding, initialization, crossovers, mutations, fitness selections and route search using genetic algorithm with multiple objectives. The aim is to find the best QoS route in order to optimize the design of MANET routing protocols. This NP-hard problem is often highly constrained such that random initialization and standard genetic operators usually generate infeasible networks. The effect of maximum node velocity on the protocol performances is done conclusively shows that QOSRGA had a potential to be the protocol for MANET. Keywords: QoS Routing, Mobile ad-hoc networks, genetic algorithm, fitness function, performances, maximum node velocity.

1 Introduction Next generation of wireless communication would see a seamless connectivity and integration of various technology platform. It is anticipated that wireless mobile ad hoc network (MANET) would be an additional component in the upcoming LTE (Long Term Evolution) deployment. Ultimately, it would be carrying a diverse multimedia applications such as voice, video and data coupled with high security feature. For enhanced quality delivery of delay sensitive applications it would be imperative that MANET[1] provides QoS Routing support, in which it could manages bandwidth-delay [2] constraints and node-connectivity issues. Various mechanism of routing protocols are already available [5][8] at the research level. However, studies on those protocols[3][4][6] showed that some are more susceptible to performance degradation than others. Most on-demand protocols, performed better than the table oriented protocols. Among the on-demand QoS routing protocols proposed in [7][8][9], a CDMA/TDMA MAC layer is commonly used to eliminate the interference between different transmissions. A very promising approach is to establish multiple paths between source and destination. Hence, it would be wise to T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 301–311, 2011. © Springer-Verlag Berlin Heidelberg 2011

302

J. Abdullah

design the protocols which take advantages of multiple paths to improve the overall performance. In the design of computer network, Kumar et al. [8] uses genetic algorithm (GA), as the optimization technique. The authors considered diameter, average distance, and computer network reliability as the optimization parameters. Coley et al. [10] outlines fields of engineering where GA had been applied, such as VLSI routing and communication networks. M. Gen et al. [11] produced detailed study of various GA-based industrial engineering applications such as that applied to scheduling, transportation and reliability. R. Elbaum et al. [12] used GA in designing LAN with an objective to minimise the network delay. S. Mao et al. used GA to optimize the routing problem for multiple description video transmission in MANET[13]. Researchers have applied GA to the shortest path routing problem [14], dynamic channel allocation problem [15] and routing problem[16]. Munetomo[17] proposed GA with variable-length chromosomes, whilst Inagaki [18] proposed GA employing fixed length chromosomes for networking problems. In Section 3, we dwelled on the implementation details of the GA-based The rest of the paper is organized as follows. Section 2 outlined strategies facilitating QoS route selection, introduced MANET model and multiple objective formulations QoS route algorithm. Lastly, Section 4 concludes the paper.

2 QoS Route Selection and Optimization 2.1

Route Selection Algorithm

MANET QoS routing algorithm is a NP-completer it considers two additive or multiplicative metric, or one additive and one multiplicative metrics19]. The QoS route selection algorithm should be efficient and scalable. Typically the heuristics for the solution to this problem would be solved by the following techniques; (1) the ordering of QoS metrics[19]; (2) sequential filtering[20]; (3) scheduling discipline of QoS metrics[4]; (4) admission control techniques[4][21]; (5) control theory approach[9]. In this paper it proposed multiobjective QoS routing algorithm using genetic algorithm. MANET is modeled as a graph G = {E , Q(nci, BAVA , DE 2 E , DMaC )} where E is a set of mobile nodes in the network; and Q is a set of QoS routing constraints which set the limits on the performance of the network. Each mobile node i ∈ E has a unique identity and moves arbitrarily. A circular plane, radius R defines a coverage area within which each node could communicate directly to each other. Neighbours of node i are defined as a set of nodes which are within radius R and directly reachable. Every pair of neighbours can communicate with each other in both directions. Hence, there exists a connectivity between neighbours i and j with an index of nci[25]. If the pairs are moving towards each other or away from each other, the node pair connectivity index, nci should be a positive value which describes the quality of connectivity between any two adjacent nodes. The least nci value indicates good quality connectivity, in which the node pair connectivity time is larger compared to high nci value. The node connectivity index, nci is defined as,

Effect of Maximum Node Velocity on GA-Based QOS Routing Protocol

  10 5 ⋅ b  a −  10 5.c − npem ;     10 5 ⋅ b nci =  ; 5  10 .c + npcm   0; 

303

for P2 < P1 (1)

for P2 < P1 for P2 = P1

where ,

npem= (1/(t2 − t1 )) ((1/ P1 ) − (1/ P2 )) and npcm = (1 /( t 2 − t1 ))(( 1 /

P2 ) − (1 /

P1 )) .

The variable npcm and npem are positive quantities; npcm is due to the node moving toward another neighbour node; npem is due to the node moving away from that neighbour node. The values of npcm is high positive values and npem is low positive values. These two quantities must be integrated forming a single metric which indicate the quality of connectivity between the two adjacent mobile nodes. A node with npcm , indicated that its node pair connectivity time is longer than the node with npem. P1 and P2 is the power measured away from the each other’s node. During operation, a route, R is created from source, s to destination, t as a sequence of intermediate nodes, such that R(s, t) = {s,… i, j, k, l, …t} without loop. The node pair connectivity index, nci(I,j) associated with a node pair is specified by the matrix, C=[ nci(I,j) ], defined as follows,

 nci0 ,0 " ncio,k −1    C= # % #  ncik −1, 0 " ncik −1, k −1 

(2)

The connectivity matrix is built at the source, upon receiving the route reply, RREP packets from the destination after a time lapse due to route request packet, RREQ. The value of nci changes continually as the topology changes. A connectivity indicator Li,j , provides the information on whether the link from node i to node j is included in the routing path. It is defined as follows, 1 Li , j =  0

if there exist connectivity (i, j ) . if otherwise.

(3)

The diagonal elements of L must always be zero. Another formulation in describing the MANET topology is node sequence of the routes, such that,

1, Nk =  0,

if node N k ∈ route if otherwise.

(4)

Using the above definitions, QoS routing can be formulated as a combinatorial optimization problem minimising the objective functions. The sum of nci of the selected route should be minimum, since this would be the most preferred route due to

304

J. Abdullah

the higher probability of being connected longer with next hop neighbours. Then, the formulation statement is to minimise the sum of node connectivity index of the route,

C

k . {SUM

( S ,T ) }

=

T



C

j= S

k , j

⋅ Lk,j

(5)

The sum of nci of the route R(s, t) constitutes the cost of the packet transmission process. In this approach, longer connectivity lifetime, indicate lower the cost of the route. 2.2

Multiple Objectives Optimization Formulation

In many real-life problems, objectives under consideration conflict with each other. Therefore, a perfect multiple objective solution that simultaneously optimizes each objective function is almost impossible. The operation of GA will minimise the sum of node connectivity index of the route, Csum(S,T) , subject to the following constraints. 2.2.1 There Must Be No Looping This constraint ensures that the computed result is indeed an existing path and without loop between a source, S and a designated destination, T such that,

if i=S 1  Li , j −  L j ,i = − 1 if i =T  j=S j =S  j ≠i j ≠i  0 otherwise. T

T

(6)

2.2.2 Available Node Bandwidth Must Be Greater Than the Requested Bandwidth This constraint ensures that the node bandwidth can manage the request bandwidth such that, B AVA,i ≥ BREQ , and for the whole route, BREQ ≤ min (BS, …, Bi,Bj, ...BT ) , where BREQ is the bandwidth of the transmitted message. The node bandwidth must be greater than the demand bandwidth. Generally, for QoS operation to be effective, the bandwidth available for the node in question must be considered. Since the shared medium is being dealt with, CSMA/CA, as the link layer of the mobile ad hoc network, the problem of medium contention among the nodes within the transmission range must be taken into account. Hence, it is necessary to estimate the instantaneous bandwidth available, BAVA,i and bandwidth consumed, BCON,i for the node concerned. Part of the cooperative protocols that are developed, is the Node State Monitoring protocol (NSM), where a method of monitoring bandwidth available and bandwidth consumed is established. 2.2.3 Total Delay Is a Minimum The constraints in terms of link delay and node delay must be considered. m

Dw ≥ {  i =1

S →T

 Di, j ⋅ Li, j + j =1 j ≠1

S →T

 D .N j

i =1

j

}

(7)

Effect of Maximum Node Velocity on GA-Based QOS Routing Protocol

305

If several routes exist, then the total delay for a route to be selected is the one that is the least.

3 The QOSRGA Implementation 3.1

Encoding and Limited Population Initialization

The chromosome consists of sequences of positive integers, which represent the identity of nodes through which a route passes. Each locus of the chromosome represents an order or position of a node in a route. The gene of the first and the last locus is always reserved for the source node, S and destination node, T respectively. The length of the chromosome is variable, but it should not exceed the maximum length which is equal to the total number of nodes in the network. The information can be obtained and managed in real-time by the Node State Monitoring(NSM) protocol and the non-disjoint multiple routes discovery protocols(NDMRD)[22]. The the initial population was obtained by extracting the existing potential solutions from the result of NDMRD protocol [22]. 3.2

Fitness of the Multiple Objectives Function

Fitness value of each route is based on various QoS parameters: bandwidth, node delay, end to end delay and the node connectivity index, nci. According to M. Gen et al. [11] , each objective function is assigned a weight. These weighted objectives are combined into a single objective function. Fitness function operates to minimise the weighted-sum F, F = α .F1 + β .F2 + γ .F3 where F1, F2 and F3 are the objective functions which describe nci, delay and bandwidth respectively. F1, F2 and F3 are defined as follows, F1 =

s →t

C

ij

⋅ Lij ,

| s →t |   | S →t | F2 =   Dij ⋅Lij +  d j ⋅ N j  ,   j =1

1 / Bi if F3 =  1000 if

Bi − BQOS > 0 . Bi − BQOS ≤ 0

(8)

(9) (10)

The weights α , β and γ are interpreted as the relative emphasis of one objective as compared to the others. The values of α , β and γ are chosen to increase the selection pressure on any of the three objective functions. The fitness function F measures the quality and the performance of a specific node state. Having described these parameters, which are the bandwidth, nci, medium access delay and end to end delay, the next issue is how importance each parameter on QoS Routing protocol as a whole. The significance of each parameter is defined by setting appropriate weighting

306

J. Abdullah

coefficients to α , β and γ in the fitness function that will be minimised by the GA operations. The values of these weighting coefficients were determined based on their equal importance towards the overall QoS Routing performance, hence α , β and γ are set to 10-3, 10-4 and 10-3 respectively. The function which involved bandwidth, we need to find the minimum bandwidth among the nodes and compare this with the demand bandwidth, BQOS . If the minimum bandwidth is less than the BQOS , the fitness is set to a high value so that in the selection process it will be eliminated. 3.3

Mobile Nodes Crossover

In our scheme, the two chromosomes chosen for crossover should have at least one common gene, except for source and destination nodes. A set of pair of nodes which are commonly included in the two chosen chromosomes but without positional consistency is first determined, the potential crosspoint. Then, one pair is randomly chosen and the locus of each node becomes a crossing point of each chromosome. It may be possible that loops are formed during crossover. A simple restoration procedure was designed to eliminate the infeasible chromosomes. The procedure for crossover operation is follows: Step 1) Input a matrix which consists of a number of routes, as Eqn 11. Step 2) Test for crossover rate. If the generated crossover rate is more than the given crossover rate, then skip the step. If not proceed. Initialise the random number generator and the new route array. The population size must be positive and even. Step 3) Consider a pair of variable length chromosomes denoted as parents, V1 and V2 , starting from the last chromosome within the population. Step 4) Locate the potential pair of crossing sites. Step 5) If more than one pair of crossing sites exist, apply a random number to establish only one particular pair of crossing sites. Step 6) Do the crossover of V1 and V2. Two offsprings, V1’ and V2’ are produced.  n0,0  n  1, 0 ROUTE _ ARRAY =  n2,0   " nm−1,0 

3.4

n0,1

n0, 2

⋅"

"

"

"

"

"

"

"

"

"

"

"

"

n0,k −1  "  "   "  nm−1,k −1 

(11)

Route Mutation

Mutation is used to change randomly the value of a number of the genes within the candidate chromosomes. It generates an alternative chromosome from a selected chromosome. The procedure for the mutation process is outlined below: Step 1) Input two matrices, population matrix(Eqn. 12) and connectivity matrix (Eqn. 13). Step 2) Select randomly a parent chromosome V, from the POP_MATRIX. It is selected with the probability Pm.

Effect of Maximum Node Velocity on GA-Based QOS Routing Protocol

307

Step 3) Randomly select a mutation node i from V, . Step 4) Generate the first subroute r1 from source node, S to node i by deleting a set of nodes in the upline nodes after the mutation node. POP _ MATRIX =  n0, 0  n  1, 0  n 2, 0   " n m−1, 0 

CONNECTIVI TY MATRIX ,

n 0,1 "

n0, 2 "

⋅"

" " "

" " "

" " "

"

n 0, k −1  "  "   "  n m−1, k −1 

(12) Li , j

 l1,1 l1, 2 l1,3 ⋅ " l  2,1 " " " = l3,1 " " "  " " " " l n ,1 " " " 

l1,n  "  "  " ln ,n 

(13)

Step 5) Generate a second subroute r2 from i to the destination node T. It is done as follows. Step 5-1) Determine node degrees of I , deg(i), neighbours of i. If deg(i)=1 and { deg(i) } = T , then terminate the search, since the second subroute consist of T . If deg(i) =1 and { deg(i) } # T , then terminate the mutation process. If deg(i) > 1 go to Step 5-2. Step 5-2) Select node {1, 2, 3, … deg(i) }. If deg(1)=1 and {deg(1)}=T then second subroute is generated. Proceed with 2 and so on. If deg(1)=1 and {deg(1)}#T, proceed with 2 and so on. If deg(1)>1 go to Step 5-3. Step 5-3) Select node { 1, 2, 3, … deg(1) }. If deg(1)=1 and {deg(1)}=T then second subroute is generated. Proceed with 2 and so on. If deg(1)=1 and {deg(1)}#T, proceed with 2 and so on. If deg(1)>1 terminate. We search for the second subroute up to two stages so that the effort will not take much processing time. Step 5-4) If the number of second subroute generated is more than one, then choose the least hop. Step 6) Combine the first subroute and second subroute forming a new route. Add to the POP_MATRIX. Step 7) If any duplication of nodes exists between r1 and r2, discard the routes and do not perform mutation. Otherwise, connect the routes to make up a mutated chromosome. 3.5

Selection Schemes, GA Parametric Evaluations and Preferences

Choosing genetic algorithm parameters such as selection schemes, population size, mutation rate and crossover rate is a very difficult task. Each combination of parameters may produce a variety of outcomes. Haupt et al [23] outlined a general procedure for evaluating these parameters. In our case, four selection methods were considered, namely the roulette wheel selection (RWS), tournament selection (TS), stochastic universal selection (SUS) and elitism technique. Next, the parameters Pc, Pm and population size are considered.

308

J. Abdullah

3.5.1 Determination of Population Size by Finding Average Minimum Cost, CAMC The effect of population was investigated by fixing the mutation rate ( Pm = 0.01 ) and changing the population size and recording the average minimum cost. The simulation was run for 2000 generations. The results reinforced the view that population size below 100 is appropriate for both the Elitism and Tournament selection. In fact, a population of 20 could be used and still produce good fitness. Hence, tournament selection was chosen. 3.5.2 Crossover and Mutation Probability Another very important parameters for GA implementation are the crossover probability Pc and the mutation probability Pm. These probabilities determine how many times crossovers and mutations occurred within a transmission period. The occurrence of crossover and mutation increases the convergence rate. De Jong [10], tested various combinations of GA parameters and concluded that mutation was necessary to restore lost genes, but should be kept at low rate, avoiding random search phenomenon. Further study by Schaffer et al. [24], suggested that the parameters should have these recommended ranges: population size of 20 ~ 30, mutation rate of 0.005 ~ 0.1 and crossover rate of 0.75 ~ 0.95. Another study by Haupt et al [23] concluded that the best mutation rate lies between 5% and 20% while the population size should be less than 16. In this paper, where GA operation is done in real time, the value of Pc and Pm is taken to be between 0.4 and 0.9 and between 0.05 and 0.2 respectively. The population size is limited up to the number of routes discovered. The limit is also imposed on the number of generations, that is the maximum number of generations to 20. Simulation experiments were run by setting MANET scenario running the protocol, with 20 nodes placed within an area of 1000 meter x 1000 meter. Each node has a radio propagation range of 250 meters and channel capacity of 2 Mbps. Up to 10 sources were initiated transmitting CBR with a data payload of 512 bytes. The investigation concluded that the crossover probability and mutation probability could be taken as 0.7 and 0.1 respectively.

4 Effect of Maximum Velocity on QOSRGA Performances Node mobility generally influences the overall performance of the network. The influenced of velocity on the APDR, AETED and ARLR was investigated. The number of CBR sources was set to five. The simulation was ran by varying uniform velocity distribution with mean outcome of Vmax as 1, 2, 5, 10, 15, 20 and 25 m/s. For stationary nodes, the RWP parameter setting was removed altogether. The source data rates used were 40 kbps and 200 kbps. The aimed of the simulation experiment was to relate the mobility of nodes and its effect on the overall performance of QOSRGA. The results were shown in Figure 1 to Figure 4. Figure 1 shows the APDR against maximum velocity for 40 kbps and 200 kbps. With the increase in velocity, the 40 kbps remained constant at approximately 82%. The performance of the 200 kbps traffic shows a decreasing trend as the velocity

Effect of Maximum Node Velocity on GA-Based QOS Routing Protocol

309

Fig. 1. Average Packet Delivery Ratio as a Function of Max Velocity

Fig. 2. Average Packet Delay as a Function of Max Velocity

Fig. 3. Average Routing Load as a Function of Max Velocity

Fig. 4. Average Throughput as a Function of Max Velocity

increased. It dropped substantially from 78% to 42% at 5 m/s, then to 19% at 20 m/s and improved a little to 22 % at 25 m/s. QOSRGA performed effectively with low source data rate throughout the range of velocities but not with high source data rate. With higher data rate and faster node movement, more packets are dropped due to short node pair connectivity time, that is, high nci value. Figure 2 shows the average delay against maximum velocity. With 40 kbps source, the average delay was much less than that with a source of 200 kbps. The 200 kbps source generated more packets and resulted in higher congestion, and hence produced a greater delay. Nevertheless, it was still below the 100 ms, which was the maximum delay allowed for most multimedia services. The reading also shows a constant trend and not an increasing trend. It was due to the delay reading, which was based on all the packets that actually arrived at the destinations, thus not considering the dropped packets. Figure 3 shows the variation of average routing load ratio (ARLR) against maximum velocity. The routing traffic for 200kbps is 100% more than that of 40kbps. Also, as the velocity increases, the ARLR for the 200kbps source shows an incremental trend. The 40kbps source on the other hand remains constant at approximately 0.06%. As the velocity increased, the higher source needs more overhead packet for every successful packet reception at the destination. Figure 4 shows the variations of throughput. Once again the CBR 40 kbps source produced a constant

310

J. Abdullah

throughput for all velocities. The throughput for CBR 200 kbps showed an immediate decline, after which the decline was gradual. At higher velocity, the node pair connectivity time was less, the packet dropped tend to increase, and hence the bit rate transfer to the destination was reduced considerably.

5 Conclusions A scheme has been presented for multiple objectives QoS routing protocol for MANET based on Genetic Algorithm. In the proposed scheme of QoS routing, selection of a route was based on node bandwidth availability, short end to end delay and the longest node pair connectivity time indicated by node connectivity index (nci ). The route selection algorithm was outlined and implemented. The variable length chromosomes represented the routes and genes represented the nodes. The algorithmic process was initialised by introducing a limited population, accumulated during the route discovery by the Node non-Disjoint Multiple Route Discovery (NDMRD) protocol. The fitness calculation was done using the weighted sum approached, combining the entire objective functions into a single objective. The performance study was done to study the effect of maximum node velocity on the average throughput, average packet delivery ratio, average delay and average routing load. The performances indicated that the protocol is feasible for a reasonable node velocity.

References [1] Perkins, C.E., Bhagwat, P.: Highly Dynamic Destination-Sequenced Distance-Vector Routing (DSDV) for Mobile Computers. Computer Communications Review, 234–244 (1994) [2] Sobrinho, J.L., Krishnakumar, A.S.: Quality-of-Service in Ad Hoc Carrier Sense Multiple Access Wireless Networks. JSAC 17(8), 1353–1368 (1999) [3] Mohapatra, P., Li, J., Gui, C.: QoS In Mobile Ad Hoc Networks. IEEE Wireless Communications 20, 44–52 (2003) [4] Lee, S.B., Gahng-Seop, A., Zhang, X., Campbell, A.T.: INSIGNIA: An IP-based quality of service framework for mobile ad hoc networks. Journal PADC 60, 374–406 (2000) [5] Lin, C.R., Gerla, M.: Asynchronous multimedia multihop wireless networks. In: Proc. IEEE INFOCOM, pp. 118–125 (1997) [6] Johnson, D.B., Maltz, D.A., Hu, Y.C.: The Dynamic Source Routing Protocol for Mobile Ad Hoc Networks (DSR). In: IETF MANET Working Group, INTERNET-DRAFT (2007), http://www.ietf.org/rfc/rfc4728.txt (last accessed: May 30, 2007) [7] Perkins, C.E., Royer, E.M.: Ad-hoc On-Demand Distance Vector Routing. In: Proc. IEEE Mobile Computer Systems and Applications, pp. 90–100 (1999) [8] Kumar, R., Parida, P.P., Gupta, M.: Topological design of communication networks using multiobjective genetic optimization. In: Proceedings of the 2002 Congress on Evolutionary Computation, CEC 2002, May 12-17, vol. 1, pp. 425–430 (2002) [9] Li, B., Nahrstedt, K.: A control theoretical model for quality of service adaptations. In: Proceedings of Sixth IEEE International Workshop on Quality of Service, pp. 145–153 (1998)

Effect of Maximum Node Velocity on GA-Based QOS Routing Protocol

311

[10] Coley, D.A.: An Introduction to Genetic Algorithms for Scientist and Engineers. World Scientific Publishing, Singapore (1999) [11] Gen, M., Cheng, R.: Genetic Algorithms and Engineering Optimization. WileyIntersciences Publication, Canada (2000) [12] Elbaum, R., Sidi, M.: Topological Design of Local Area Networks Using GA. IEEE/ACM Transactions on Networking 4, 766–778 (1996) [13] Mao, S., Hou, Y.T., Cheng, X., Sherali, H.D., Midkiff, S.F.: Multipath routing for multiple description video in wireless ad hoc network. In: IEEE INFOCOM (2005) [14] Leung, Y., Li, G., Xu, Z.B.: A genetic algorithm for the multiple destination routing problems. IEEE Transactions on Evolutionary Computation 2, 150–161 (1998) [15] Wong, S.H., Wassell, J.: Dynamic channel allocation using a genetic algorithm for a TDD broadband fixed wireless access network. In: Proc. IASTED International Conference in Wireless and Optical Communications, Banff, Alberta, Canada, July 17-19, pp. 521–526 (2002) [16] Shimamoto, N., Hiramatus, A., Yamasaki, K.: A dynamic routing control based on a genetic algorithm. In: Proceedings of the IEEE International Conference on Neural Networks, pp. 1123–1128 [17] Munetomo, M., Takai, Y., Sato, Y.: A migration scheme for the genetic adaptive routing algorithm. In: Proceeding of the IEEE International Conference on Systems, Man, and Cybernetics, pp. 2774–2779 (1998) [18] Inagaki, J., Haseyama, M., Kigajima, H.: A genetic algorithm for determining multiple routes and its applications. In: Proceedings of the IEEE International Symposium on Circuits and Systems, pp. 137–140 (1999) [19] Wang, J.C.Z.: Quality of Service Routing For Supporting Multimedia Applications. In: IEEE JSAC, vol. 14, pp. 1228–1234 (1996) [20] Ma, Q., Steenkiste, P.: Quality-of-service routing for traffic with performance guarantees. In: Proceedings of IFIP Fifth International Workshop on Quality of Service (1997) [21] Ahn, G.S., Campbell, A.T., Veres, A., Sun, L.H.: Supporting Service Differentiation for Real-Time and Best-Effort Traffic in Stateless Wireless Ad Hoc Networks (SWAN). In: IEEE TMC, vol. 1, pp. 192–207 (2002) [22] Abdullah, J.: The Design of QOSRGA Protocol Employing Non-Disjoint Multiple Routes in MobileAd Hoc Networks. In: Proc. Of The MMU International Symposium on Information and Communication Technologies (M2USIC 2007), PJ Hilton, Petaling Jaya, Malaysia, November 19-20, pp. 983–43160 (2007) [23] Haupt, R.L.: Optimum population size and mutation rate for a simple real genetic lgorithm that optimizes array factors. In: Antennas and Propagation Society International Symposium, 2000, vol. 2, pp. 1034–1037. IEEE (2000) [24] Schaffer, J.D., Caruana, R.A., Eshelman, L.J., Das, R.: A study of control parameters affecting online performance of genetic algorithms for function optimization. In: Proceedings of the Third International Conference on Genetic Algorithms, pp. 51–60. George Mason University, Morgan Kaufmann Publishers Inc. (1989) [25] Abdullah, J., Parish, D.J.: Node connectivity index as mobility metric for GA based QoS routing in MANET. In: Proceedings of the 4th International Conference on Mobile Technology, Applications, and Systems, Singapore, September 10 -12, pp. 104–111. ACM, New York (2007)

Application of Wireless Accelerometer System for Evaluating Osteoarthritis Dong Rak Kwon1 and Ho-Cheol Lee2 1

Department of Rehabilitation Medicine, Arthritis & Autoimmunity Research Center, Catholic University of Daegu, School of Medicine, 25, Jangdeungsan 1-gil, Nam-gu Daegu 705-718, Korea 2 School of Mechanical and Automotive Engineering, Catholic University of Daegu, 330 Hayang-up, Gyeongsan-si, Gyeongbuk, 712-702, Korea [email protected]

Abstract. The purpose of this research paper is on the quantification of activity using a wireless acceleration sensor. Two different promising algorithms were presented here of which results can be used as a tool for the activity of a rabbit and discriminate the differences between before and after the triggering of osteoarthritis. It was shown that two important signal processing methods such as high-pass filtering and interpolation should be applied to the raw signals in order to get error-free results. All the algorithms and signal preprocessing methods were verified with experiments for rabbits. Keywords: Osteoarthritis, Wireless Acceleration Sensor, Activity Quantification, High-pass Filtering, Interpolation.

1 1.1

Introduction Research Backgrounds and Trends

Apart from the smart phone, huge changes in our every daily life are being made by the technological development in wireless communication area in recent days. Revolutionary accomplishments in semi-conductor technology have lead to miniaturization in semiconductor chip size and ultra-low consumption in power requirement which were considered as impossible in the near future. These changes forced many researchers to concentrate on the sensors based on this wireless technology. There are so many industrial fields where the wireless sensor technology has proven its potential strength and even succeeded commercially. But among others, the research topic of this paper is the sensor which can quantify the activity of a person or an animal. The most representative sensors for measuring activity are accelerometers and gyroscopes. The state-of-the-art based on MEMS (micro-electromechanical system) has made it possible to fabricate these kinds of sensors in a small form factor, with low power consumption and low price. Under these conditions, the combination of wireless technology with MEMS sensors has made the wireless activity measuring sensors applicable to various industrial and research areas. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 312–319, 2011. © Springer-Verlag Berlin Heidelberg 2011

Application of Wireless Accelerometer System for Evaluating Osteoarthritis

313

As mentioned above, though the application fields of the wireless activity measuring sensors are very diverse, some of examples are useful for understanding the recent trends and explaining the purpose of this paper. First of all, Benocci and et al applied this combinational sensor to measuring the motion of older people in order to predict the fall action. They chose ZigBee as a wireless communication protocol and acceleration sensor as an activity estimating tool. According to their results, their sensors showed good performances so that 65 kinds of various falling events could be forecast with just 2 cases of malfunctioning. [1] Especially, this technology has been successfully applied to personal anti-fracture airbag system by some research groups. [2-3] This research used Bluetooth as a wireless communication protocol. The game industry, one of the most promising industrial areas, is not an exception. In order to adapt this wireless activity measuring technology to playing interactive game, some attempts have been made. [4] Analysis of normality of human motions is also tried in recent research. [5] Although the application area of the wireless activity sensor has been expanding rapidly, the related research activities do not appear to be prominent in medical area. Of course, some attempts have been made to utilize these sensors for the sake of estimating human motions in rehabilitation field [6] and this research is a part of those attempts as well. 1.2

Research Goal and Contents

Osteoarthritis is a kind of disease by which cartilage tissues are damaged and inflammation is progressing. This disease has been drawing much attention as many advanced countries became an aging society. In order to prevent and cure this disease, the methods to find the cause of the disease or to estimate the treatments become indispensible. Until now, the most prevailing methods of investigating the origin of osteoarthritis and evaluating how much new medicines are effective can be summarized as follows. First of all, after a drug capable of inducing osteoarthritis, such as collagenase, is inject to knee part of an animal like rabbits, the change in activity is investigated by any method. After this treatment, curing medicine is injected again into the same rabbit and the degree of improvement in its activity is investigated again. The activities during these 3 phases are compared one another. [7] However, only the qualitative estimation methods have been used in this research field so that there is no way to quantify the activity of target rabbits: how much its activity deteriorates after causing osteoarthritis or how much its activity improves after curing treatments. This situation seems to originate from the fact that the classical wired measuring methods have much influence on the activity of the target animal like rabbits. [8] In this research, the limitation of the classical wired method for activity measurement will be overcome by means of adoption of wireless communication device. Not only utilization of wireless communication method (ZigBee) but also the adoption of MEMS acceleration sensors makes it possible to accomplish the small size enough to have little effect on the movement of rabbits. This combination definitely gives more reliability to estimation of osteoarthritis proposed in this paper.

314

D.R. Kwon and H.-C. Lee

Two basic evaluating algorithms are present in this paper. One is physically equivalent to the momentum and the other is very intuitively devised. Since raw data has proved to be unsuitable for the above-mentioned algorithms in some viewpoint, two kinds of preprocessing methods should be added before the raw data are processed by these algorithms. One is for eliminating unwanted DC offset components and the other for resolving the uneven sampling problem. The experiments on 3 rabbits verified that two proposed algorithms are meaningful enough to distinguish the activity between before and after of osteoarthritis on the condition of two preprocessing treatments.

2

Experimental Methods and Results

2.1

Experimental Methods

Fig. 1 shows the sensor, its attaching device (upper figure) and a target rabbit after the sensor is installed. The sensor utilized in this research is the upgraded version of the one used in precedent research [4] and it is called as ZStar3 by Freescale Inc. This sensor is based on the ZigBee protocol family and gives information on acceleration values in 3 different orthogonal directions. (Fig. 1(b) shows axis of coordinates after installation of this sensor) The maximum sampling frequency can be adjusted to 3

Fig. 1. (a) Wireless sensor and (b) attachment of the device to a target rabbit; figure (b) shows the axes of coordinates after installation of the sensor as well

Application of Wireless Accelerometer System for Evaluating Osteoarthritis

315

different levels such as 30Hz, 60Hz and 120Hz. As shown in Fig. 1(b) the rubber bands with elasticity help the stable fixation of the sensor to the waist portion of the rabbit. The maximum frequency is set to 30Hz which are assumed to be sufficient for measuring movement of rabbits during all experiments. The rabbits are released to move freely in a room of 3m x 3m during approximately 3 minutes. The duration of measurements is not of consequence because all the activity algorithm methods include time-averaging method. In advance, the acceleration signals of 3 healthy rabbits are acquired for comparison later. The 1mg of collagenase is injected to knee joint of the same 3 rabbits respectively. The acceleration signals of the rabbits are acquired once more after a few days.

Fig. 2. Measured acceleration signals from the wireless sensor. Each graph from the top indicates accelerations in X, Y and Z direction respectively. The coordinate axes are shown in Fig. 1(b).

Fig. 2 shows the example signals of acceleration acquired under the situation described above. Each signal corresponds to 3 different directions, X, Y and Z as indicated in the titles of the graphs. Many peak signals are observed over all the range. These peaks reflect the abrupt acceleration components induced by the hopping motion of a rabbit. Apparently there seems to be no problems in these raw acceleration signals. But a little more careful inspection reveals the following 2 problems and these 2 problems will cause a serious error during the algorithm calculation process. One is DC offset problem which is depicted in Fig. 3(a) and the other is automatic variation of sampling frequency which is shown in Fig. 3(b).

316

D.R. Kwon and H.-C. Lee

Fig. 3. (a) Raw acceleration signal from the wireless sensor (blue solid line) which includes many DC offset components in places and its high-pass filtered version (black dotted line) which does not contain any DC offset components; (b) The upper graph shows variably sampled raw signal and the lower one shows its interpolated version which has uniform sampling period

First of all, as shown in Fig. 3(a) many DC offset components are observed here and there of the acquired raw data. These temporary DC offsets seems not to be related to any real actions of rabbits. It is supposed that these components originate from the capacitive characteristics of ZStar3 MEMS acceleration sensor. Since the ZStar3 operates based on the capacitance variation, the charging and discharging process can induce temporary DC electrical potential. These DC offset components should be removed before the activity estimation algorithms are applied because they will definitely result in overestimation of the activity. In order to attack this problem, high-pass filtering is simple and relevant. Reflecting on the speed of normal rabbits, high-pass filters with cut-off frequency less than 1Hz are applied. Butterworth filter is chosen as filter design method. The cut-off frequency is used as a parameter and proved to have a significant effect on the correctness of activity algorithms. In Fig. 3(a), the signals before and after high-pass filtering are drawn together. Even if there is some information loss such as peak amplitude decrease, information about the hopping action still remains almost the same after the filtering. On the contrary, all DC components are eliminated over the whole range.

Application of Wireless Accelerometer System for Evaluating Osteoarthritis

317

Another problem is on the variation of sampling frequency. ZStar3 has the function to automatically change its sampling frequency according to the variation speed of acquired signal. In other words, if the acceleration is changing very fast the ZStar3 makes the sampling frequency as high as possible within its maximum frequency value. But it the acceleration is changing relatively slowly, the sampling frequency is adjusted to be as low as possible. This is helpful from power consumption point of view. However, as a result of this adjustment, the sampling period is not uniform as shown in the upper graph of Fig. 3(b). If this unequal sampling rate is kept and the raw signal is not modified properly before it is applied to any activity measuring algorithm, it also giver rise to a serious error as much as the DC offset components do. In this research, all signals go through interpolation before they are passed to any activity calculation algorithm. The sampling frequency after interpolation is fixed to 30Hz which is the lowest maximum frequency of ZStar3. The lower graph of Fig. 3(b) indicate that this uneven sampling problem is gotten rid of after applying the interpolation process. Two different algorithms based on acceleration signals are presented with a view to estimating activity of rabbits. One method evaluates the overall activity of a rabbit. After the absolute values are taken from the original acceleration signal, they are integrated with respect to time and divided by the whole time span in order to eliminating the effect of measuring time length, i.e. for averaging. Since the weight of the rabbit is kept almost the same before and after induction of osteoarthritis, the calculated index value corresponds to mechanical momentum in a physical meaning. The equation used in calculation is given in Eq. (1), where T is the whole time span, ai is acceleration signal value and t indicates the sampling period after interpolation. The acquired signals are all discrete so that the integration is substituted by summation over the same time span. This index is denoted as Aint.

Aint =

t  ai T

.

(1)

The second index for activity measurement is to count the number of hopping action of rabbits. This index is devised from the fact that the most active motion of rabbits is hopping and the hopping makes relatively high peak in acceleration signal. In other words, if the peak magnitude of any hopping action is greater than the chosen threshold acceleration value, then the hopping is counted as real hopping. Since the hopping motion is accumulated as time goes on, the same averaging process as the first index Aint should be applied. Moreover, since there is no reasonable way to define criteria for hopping threshold, only the effect of the threshold are investigated. This index is denoted as Ahop. 2.2

Analysis Results

The acceleration signals are acquired and two proposed preprocessing methods – high-pass filtering and interpolation- are applied to these raw data. Using the signals which have gone through two preprocessing, two activity calculation algorithms are calculated. The algorithm using integration, Aint is shown in Fig. 4.

318

D.R. Kwon and H.-C. Lee

Fig. 4. The graphs are the results of calculated activity indices Aint. The numbers 1, 2 and 3 in X axis indicate the identification number of rabbits. Y axis means calculated activity indices Aint. Each graph from the left corresponds to 0.1Hz, 0.3Hz and 0.5Hz cut-off frequency respectively.

From Fig. 4, it can be concluded that the first proposed activity index Aint shows meaningful difference between the cases of normal rabbits and osteoarthritis-induced rabbits. The three graphs correspond to the experiments where the different cut-off frequencies are applied. From the left 0.1Hz, 0.3Hz and 0.5Hz cut-off frequencies are used respectively. Though it is apparent that 0.3Hz is the most effective cut-off frequency to distinguish the occurrence of osteoarthritis, it cannot be definitely concluded that 0.3Hz is always the optimal frequency. Nevertheless the capabilities of discrimination of the cases 0.1 and 0.5Hz are worse than that of 0.3Hz and it naturally means that there is some optimal point in cut-off frequency.

Fig. 5. The graphs are the results of calculated activity indices Ahop. The numbers 1, 2 and 3 in X axis indicate the identification number of rabbits. Y axis means calculated activity indices Ahop. Each graph from the left corresponds to 0.3g, 0.6g and 0.9g threshold level respectively.

Fig. 5 shows the second calculated activity index Ahop is also a good criteria for evaluating the movements of rabbits. The three graphs use three different threshold acceleration levels for determining whether the hopping is accepted or not. From the left threshold values are 0.3g, 0.6g and 0.9g respectively, where g means gravitation acceleration 9.8m/s2. Similar to the case of integration activity index, it can be concluded that there must be some optimal threshold value. However the threshold should be more carefully investigated in this hopping index from the fact that the index brings out the erroneous result. (See the number 2 and 3 rabbit of the most left

Application of Wireless Accelerometer System for Evaluating Osteoarthritis

319

graph.) For both indices, the parameter optimization remains as a task to be solved. These two indices can be assembled with weighting factor in order to give better results.

3

Conclusion

This paper presents the methods to evaluate the activity using miniaturized and lowpower consuming wireless acceleration sensor. Two different activity measuring indices are proposed. It is shown that high-pass filtering and interpolation should be applied before the indices are calculated. From the experiments on rabbits, these two methods are proven to be effective and meaningful way of evaluating activity. These methods are being applied to determine whether any medicine has an effect on osteoarthritis or not. Acknowledgements. This work was supported by research grants from the Catholic University of Daegu in 2010.

References 1. Benocci, M., Tacconi, C., Farella, E., Benini, L., Chiari, L., Vanzago, L.: Accelerometerbased Fall Detection Using Optimized ZigBee Data Streaming. Microelec. J. 41, 703–710 (2010) 2. Shin, G., Chan, C.S., Li, W.J., Leung, K.S., Zou, Y., Jin, Y.: Mobile Human Airbag System for Fall Protection Using MEMS Sensors and Embeded SVM Classifier. IEEE Sensor J. 9, 495–503 (2009) 3. Tamura, T., Yoshimura, T., Sekine, M., Uchida, M., Tanaka, O.: A Wearable Airbag to Prevent Injuries. IEEE Trans. On Inform. Tech. in Biomed. 13, 910–914 (2009) 4. Jung, Y., Cha, B.R.: Gesture Recognition Based on Motion Interial Sensors for Ubiquitou Interactive Game Contents. IETE Tech. Rev. 27, 158–166 (2011) 5. Tien, I., Glaser, S.D., Bajcsy, R., Goodin, D.S., Aminoff, M.J.: Results of Using a Wireless Inertial Measuring System to Quantify Gait Motions in Control Subjects. IEEE Trans. on Inform. Tech. in Biomed. 14, 904–915 (2010) 6. Ishigaki, N., Kimura, T., Usui, Y., Aoki, K., Narita, N., Shimizu, M., Hara, K., Ogihara, N., Nakamura, K., Kato, H., Ohira, M., Yokokawa, Y., Miyoshi, K., Murakami, N., Okada, S., Nakamura, T., Saito, N.: Analysis of Pelvic Movement in the Elderly during Walking using a Posture Monitoring System Equipped with a Triaxial Acceleromter and a Gyroscope. J. Biomech. 44, 1788–1792 (2011) 7. Kim, S.B., Kwon, D.R., Kwak, H., Shin, Y.B., Han, H., Lee, J.H., Choi, S.H.: Additive Effects of Intra-articular Injection of Growth Hormone and Hyaluronic Acid in Rabit Model of Collagenase-induced Osteoarthritis. J. Korean Med. Sci. 25, 776–780 (2010) 8. Gushue, D.L., Houck, J., Lerner, A.L.: Rabbit Knee Joint Biomechanics: Motion Analysis and Modeling of Forces during Hopping. J. of Orthopaedic Res. 23, 735–742 (2005)

A Performance Evaluation of a Novel Clustering Scheme Considering Local Node Density over WSN Jeong-Sam Kim1 and Tae-Young Byun2,* 1

Division of Computer Technology, Yeungnam College of Science and Technology, Deagu, Rep. of Korea [email protected] 2 School of Computer and Information Communications Engineering, Catholic University of Daegu, Gyeongsan, Gyeongbuk, Rep. of Korea [email protected]

Abstract. This paper is an extension of previously proposed a novel clustering scheme, called D-LEACH, that is based on the calculation of local node density to increase the WSN lifetime by efficiently reducing energy consumption at the sensor node. In this paper, we present both mathematical analysis and performance evaluations of our proposed D-LEACH in view of energy consumption of network. By computer simulation, we showed the ratio of preclustering executions, which is an additional stage to existing LEACH, to total rounds over whole network lifetime. Also, we investigate the progress of live nodes according to changes in local node density, and the progress of active nodes considering local node density in D-LEACH. Keywords: Clustering Scheme, D-LEACH, WSN, Local Node Density.

1

Introduction

Wireless sensor networks (WSNs) contain hundreds or thousands of sensor nodes equipped with sensing, computing and communication abilities. Each node has the ability to sense elements of its environment, perform simple computations, and communicate among its peers or directly to an external base station. Deployment of a sensor network can be performed in a random fashion (e.g. by airplane) or manually. These networks promise a maintenance-free, fault-tolerant platform for gathering different kinds of data. Because a sensor node needs to operate for a long time on a tiny battery, innovative techniques to eliminate energy inefficiencies that would shorten the lifetime of the network must be used. The networking principles and protocols for WSNs are currently being investigated and developed [1-9]. Some application examples of WSNs include target field imaging, intrusion detection, weather monitoring, security and tactical surveillance, distributed computing, detecting ambient conditions such as temperature, movement, sound, light, or presence of certain objects, inventory control, etc. In situations where *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 320–329, 2011. © Springer-Verlag Berlin Heidelberg 2011

A Performance Evaluation of a Novel Clustering Scheme

321

manual, optimal deployment of the sensors is not possible, the most common technique is to spread the sensors over the target area by dropping them form a plane. In this process it might happen that some areas end up more densely populated than others, thus increasing the amount of nodes that sense the same data. The energy used to gather duplicate data and to transmit it is wasted. In this paper, we propose a new energy efficient clustering scheme to prolong the network lifetime by reducing energy consumption at the sensor node. Every node will determine whether to participate in clustering with a certain probability, based on the local sensor density. By adjusting dynamically the probability of participating in clustering, the energy consumption of the network is reduced and the lifetime of the network is extended. The remainder of this paper is organized as follows: Section 2 shows related work. Section 3 mathematically analyzes the energy consumption of our proposed clustering method. Further, simulation results are given in Section 4. Section 5 contains the conclusions.

2

Previous Work

Heinzelman et al. [4] introduced a hierarchical clustering algorithm for sensor networks called low energy adaptive clustering hierarchy (LEACH). LEACH allowed for a randomized rotation of the cluster head’s role in the objective of reducing energy consumption (i.e., extending network lifetime) and to distribute the energy load evenly among the sensors in the network. The authors also made use of a TDMA/CDMA MAC to reduce inter- and intra-cluster collisions. Because data collection is centralized and performed periodically, this protocol is most appropriate when constant monitoring by the sensor network is needed. A user may not need all the data immediately. Thus, periodic data transmissions, which may drain the limited energy of the sensor nodes, are unnecessary. An adaptive clustering is introduced in LEACH, i.e., re-clustering after a given interval with a randomized rotation of the energy-constrained cluster head so that energy dissipation in the sensor network is uniform. Authors also found, based on their simulation model, which only 5% of the nodes need to act as cluster heads. The operation of LEACH is separated into two phases: the setup phase and the steady state phase. In the setup phase, the clusters are organized and cluster heads are selected. In the steady state phase, the actual data transfer to the base station takes place. The duration of the steady state phase is longer than the duration of the setup phase in order to minimize overhead. In D-LEACH[9], each cluster is organized as in LEACH, i.e. there is a Cluster Head (CH) and many member nodes, whose cluster membership is determined in a distributed manner. However, D-LEACH is different from existing LEACH in selection of the nodes participating in the clusters at each round. D-LEACH dynamically adjusts the probability of each node joining a certain cluster by taking into account the local node density. By preventing transmission of similar or redundant data from all of the nodes in dense cluster, D-LEACH can generally increase the lifetime of WSN by reducing the energy consumption of each node.

322

J.-S. Kim and T.-Y. Byun

The operation of D-LEACH is separated into three phases: the pre-clustering phase, the setup phase, the steady state phase. . For each round, in the pre-clustering phase, every node computes its own probability of joining a certain cluster in a distributed manner. In the setup phase, the clusters are organized and cluster heads are selected. In the steady state phase, the actual data transfer to the base station takes place. The duration of the steady state phase is longer than the duration of the setup phase in order to minimize overhead. In D-LEACH, each node i computes Di, the local density of node i, Di indicates that how many nodes exist around node i. To compute Di, each node i broadcasts a discovery-packet to another nodes in its own transmission radius or R. By counting the discovery-packets from adjacent nodes, each node can know the local density of node around itself. If the local density around a certain node is relatively high, it indicates that there are many nodes which measure same or similar values. Thus, it is needed to determine which nodes will join in current round based on the local node density. If the node density is high, it is desirable that nodes have low joining probability. Otherwise, nodes have relatively high joining probability. D-LEACH sets the recommended number of nodes per a cluster or D as N / k  , where N is total number of nodes in the area, k is the number of cluster. If the local node density around any node i, named Di is less than D, it indicates that the node i should always join cluster for current round. Otherwise, the node i should join the cluster with a probability of D/Di. The cluster joining probability, Pjoin(i) at current round is expressed by Equation (1).

Pjoin(i )

D ­ °D u D ® i °1 ¯

: if Di > D : otherwise

(1)

where a is an adjusting factor.

3

Mathematical Analysis

3.1

Energy Consumption Model

For mathematical analysis of energy consumptions of both LEACH and D-LEACH, we use energy consumption mode1 in LEACH, where the radio dissipates Eelec = 50 nJ/bit to run the transmitter or receiver circuitry and eamp = 100 pJ/bit/m2 for the transmit amplifier. We make the assumption that the radio channel is symmetric so that the energy required to transmit a message from node A to node B is the same as the energy required to transmit a message from node B to node A for a given SNR. For our mathematical analysis and computer simulation, we also assume that all sensors are sensing the environment at a fixed rate and thus always have data to send to the end-user.

A Performance Evaluation of a Novel Clustering Scheme

323

Fig. 1. Energy consumption model

3.2

Computation of Energy Consumption in LEACH

_ nonCH The amount of energy consumption allocated to setup stage is denoted by E(setup round , i , j ) _ nonCH includes for each non-CH(cluster head) in a cluster as equation (2). E(setup round , i , j )

several operations such as decision of non-CH as itself, receiving CH-Advertizing message from another CHs, selection of a proper CH for itself, sending Join-Cluster message and receiving a CH-Schedule message from preferred CH. _ nonCH _ nonCH E(setup = E(Decision + round,i, j ) round,i )

#ofCHs CH_ ADV k k =1

l

2 ⋅ Eelec + ESelect_ CH + l Join ⋅ε amp ⋅ distnonCH + l Schedule⋅ Eelec i ,CHj

(2)

Also, we can calculate the amount of energy consumption for each CH in setup _ CH as equation (3). stage is denoted by E(setup round , i , j ) _ CH Decision_ CH 2 E(setup + lCH _ ADV ⋅ ε amp ⋅ distMAX + round,i , j ) = E( round, j )

#ofJoin

l

Join i

2 ⋅ Eelec + ESchedule + l Schedule ⋅ ε amp ⋅ distMAX

(3)

i=1

This calculation covers major operations such as decision of CH as itself, transmitting CH-Advertizing message to non-CHs in its cluster, receiving Join-Cluster messages from non-CHs, preparing TDMA scheduling and sending a CH-Schedule messages to non-CHs. So, the amount of energy consumption in setup stage may be written as setup _ nonCH _ CH E(setup + E(setup round ,i , j ) = E( round ,i , j ) round ,i , j )

(4)

where round ¥ 1, 1§i, j§N and N is the number of initial live nodes over WSN. _ cluster The amount of energy consumption in steady-state stage, E(steady , which i, j ) considers major operations of both non-CHs and CH in a cluster may be written as _ nonCH 2 E(steady = l ⋅ Eelec + l ⋅ ε amp ⋅ dist nonCH l ,i , j ) i , CH j

  N '   N ' _ CH 4 E(steady =    − 1 ⋅ l ⋅ Eelec +   ⋅ l ⋅ Edata _ aggr. + Eelec ⋅ l '⋅ε amp ⋅ distCH l ,l ', j ) j ,SINK  k'    k'  

(5) (6)

324

J.-S. Kim and T.-Y. Byun

§ ª N ' º · steady_ nonCH _ cluster _ CH E(steady  E(steady ¨¨ « »  1¸¸ ˜ E(l ,i , j ) i, j ) l, j) © « k' » ¹ K'

4 | (2 N ' k ' ) ˜ l ˜ Eelec  N '˜l ˜ Edata_ aggr.  l '˜Eelec ˜ H amp ˜ ¦ distCH j , SINK j 1

ªN' º 1 k ' «« k ' »»

 ( N 'k ' ) ˜ l ˜ H amp ˜ ¦ j 1

(7)

¦ dist i 1

2 nonCHi ,CH j

where 1§i, j, k’§N’ and N’ is the number of live nodes in current round, k’ is the number of clusters in current round and l’>l, l’ is aggregated data size for transmitting from a cluster head j to sink node. Finally, total energy consumption in a round in LEACH may be written as setup steady _ cluster E(total round ,i , j ) ≈ E( round ,i , j ) + E(i , j )

3.3

(8)

Computation of Energy Consumption in D-LEACH

In case of D-LEACH, total amount of energy consumption may be calculated as a similar manner in LEACH except for pre-clustering stage and setup stage. First, the amount of energy consumption for pre-clustering stage which determines participating nodes for the current round may be written as   N   ing E(precluster = N (l disco. ⋅ Eelec + l disco. ⋅ ε amp R2 ) +    − 1 ⋅ l disco. ⋅ Eelec + Elocal_ density  round)  k    

(9)

where round¥1, k§N, N is the number of initial live nodes, and k is the recommended number of clusters in LEACH. Also, ldisco. is the data size of Discocery message from a node i to neighbors in a specific range R. A node i can calculate the local density of itself by counting how many messages are arrived from neighbors. The amount of energy consumption in setup stage may be written as _ nonCH _ nonCH E(setup = E(Decision + round,i , j ) round,i )

#ofCHs CH _ ADV k k =1

l

2 ⋅ Eelec + ESelect_ CH + l Join ⋅ ε amp ⋅ distnonCH + l Schedule⋅ Eelec i ,CH j

_ CH Decision_ CH 2 E(setup + lCH _ ADV ⋅ ε amp ⋅ distMAX + round,i , j ) = E( round, j )

#ofJoin

l

Join i

2 ⋅ Eelec + ESchedule + l Schedule⋅ ε amp ⋅ distMAX

(10) (11)

i =1

setup _ nonCH _ CH E(setup + E(setup round ,i , j ) = E( round ,i , j ) round ,i , j )

(12)

_ cluster , which The amount of energy consumption in steady-state stage, E(steady i, j )

considers operations of both non-CHs and CH in a cluster may be written as _ nonCH 2 E (steady = l ⋅ E elec + l ⋅ ε amp ⋅ dist nonCH l ,i , j ) i , CH j

(13)

A Performance Evaluation of a Novel Clustering Scheme

  N"   N" _ CH 4 E(steady =    − 1 ⋅ l ⋅ Eelec +   ⋅ l ⋅ Edata _ aggr . + Eelec ⋅ l '⋅ε amp ⋅ distCH l ,l ', j ) j , SINK k ' k '     

  N "  _ cluster _ nonCH _ CH =    − 1 ⋅ E(steady + E(steady E(steady i, j ) l ,i , j ) l, j) k '    k'

4 ≈ (2 N "− k ' ) ⋅ l ⋅ Eelec + N "⋅l ⋅ Edata _ aggr. + l '⋅Eelec ⋅ ε amp ⋅  dist CH j , SINK

325

(14)

(15)

j =1

 N"  −1 k '  k ' 

+ ( N "− k ' ) ⋅ l ⋅ ε amp ⋅  j =1

 dist i =1

2 nonCHi ,CH j

where 1§i, j, k’§N” and N” is the number of active or participating nodes after preclustering stage in current round, k’ is the number of clusters in current round and l’>l, l’ is aggregated data size for transmitting from a cluster head j to sink node. Total energy consumption in a round in D-LEACH may be written as preclustering steady _ cluster E(total + E(setup round ,i , j ) ≈ E( round ) round ,i , j ) + E( i , j )

4

(16)

Performance Evaluation

We enumerate the details of simulation parameters in Table 1. Most parameters except node distribution, several types of packet sizes and the location of sink node are identical to those of LEACH. Table 1. Simulation parameters Parameters

Unit

Values

Area for WSN

Meter2

100m × 100m

The number of initial sensor nodes

Integer

100

The location of sink node

(x, y)

(50, 350)

Eelec, Elocal-density, ESchedule, Edata-aggr., EDecision-CH, EDecison-nonCH, Eselect-CH,

nJ/bit

50

eamp

pJ/bit/m2

100

l (normal packet size)

Byte

500

ldisco., lCH-ADV, lJoin, lSchedule

Byte

25

Initial residual energy

Joule

1

The ratio of CHs to all of the nodes

Percentage (%)

5

Sensor node distribution model

Percentage (%)

uniform distribution, changes from 30% to 90% by 10%

326

J.-S. Kim and T.-Y. Byun

It is important to adjust the density of node for measuring the energy consumption and the lifetime of network in D-LEACH and LEACH. In addition to uniform distribution of the nodes, we add a variety of node distribution scenarios, where some nodes among all of the nodes are uniformly distributed placed across area for WSN. To measure performances according to varying local node density, some nodes are intentionally placed in specific area on WSN, and a variety of the node distribution scenarios are used in our simulation. We evaluate some performance metrics in both D-LEACH and LEACH as follows • The number of executions as cluster head per node • The ratio of round of pre-clustering to total rounds over WSN lifetime • The progress of live nodes according to changes in local node density • The progress of active nodes considering local node density Fig. 2 shows the progress of the number of executions as cluster head per node in both D-LEACH and LEACH. As the figure implies, the gap between D-LEACH and LEACH significantly increases when nodes over 70% among total live nodes are excessively placed on specific area. This is caused by extended lifetime in D-LEACH compared with that in LEACH, so each node generally have more chance as cluster head in D-LEACH.

Fig. 2. The number of executions as cluster head per node over network lifetime according to varying nodes distribution

The ratio of round executing pre-clustering to total rounds is shown in Fig. 3. Preclustering stage is executed in 10% rounds to total rounds on an average regardless of various nodes distribution. In LEACH, as the local density increases, network energy is abruptly exhausted due to increase of energy consumption of cluster heads after a node firstly dies, thereupon, network lifetime may be shorten. However, network lifetime in D-LEACH can remarkably be extended in comparing with that in LEACH[9]. So, the period from first-node-die to last-nod-die is greatly extended in comparing in that in LEACH as shown in Fig. 4. In D-LEACH, as the number of participating nodes in area with higher node density decreases, each live node averagely consumes less energy, finally results in prolongation of network lifetime.

A Performance Evaluation of a Novel Clustering Scheme

327

Fig. 3. The ratio of round of pre-clustering to total rounds according to varying the node distribution

(a) LEACH

(b) D-LEACH Fig. 4. The progress of live nodes over network lifetime

328

J.-S. Kim and T.-Y. Byun

The progress of active or participating nodes according to a variety of nodes distribution scenarios is shown in Fig. 5, here, the number of active nodes is measured by 1 round and 50 rounds respectivly. As mentioned above, increase of local node density leads to extension of network lifetime. Especially, as nodes over 60% to total initial nodes are deployed on specific area, which means relatively high local density, prolongation of network lifetime increases in comparing with that in LEACH.

(a) The progress of active nodes in case of sampling by 1 round

(b) The progress of active nodes in case of sampling by 50 rounds Fig. 5. The number of active nodes in clusters according to changes in the node distribution in D-LEACH

A Performance Evaluation of a Novel Clustering Scheme

5

329

Conclusion

In this paper, we mathematically analyzed details of energy consumption in DLEACH, a novel clustering protocol for WSN that minimizes global energy usage by adjusting the number of active nodes considering local node density per each node in a cluster. Also, we showed some performance metrics related with energy consumption of both D-LEACH and existing LEACH. Due to nature of D-LEACH operation, which selects some active nodes among all live nodes by considering local node density, D-LEACH has advantages of prolonging network lifetime.

References 1. Basagni, S.: Distributed Clustering Algorithm for Ad-hoc Networks. In: International Symposium on Parallel Architectures, Algorithms, and Networks, I-SPAN (1999) 2. Kwon, T.J., Gerla, M.: Clustering with Power Control. In: Proceeding of MilCOM 1999 (1999) 3. Amis, A.D., Prakash, R., Vuong, T.H.P., Huynh, D.T.: Max-Min D-Cluster Formation in Wireless Ad Hoc Networks. In: Proceedings of IEEE INFOCOM. IEEE Press (2000) 4. Heinzelman, W.R., et al.: Energy-Efficient Communication Protocol for Wireless Microsensor Networks. In: Proceedings of the 33rd HICSS 2000. IEEE Press (2000) 5. Banerjee, S., Khuller, S.: A Clustering Scheme for Hierarchical Control in Multi-hop Wireless Networks. In: Proceedings of IEEE INFOCOM. IEEE Press (2001) 6. Chatterjee, M., Das, S.K., Turgut, D.: WCA: A Weighted Clustering Algorithm for Mobile Ad Hoc Networks. Cluster Computing, 193–204 (2002) 7. Bandyopadhyay, S., Coyle, E.: An Energy-Efficient Hierarchical Clustering Algorithm for Wireless Sensor Networks. In: Proceedings of IEEE INFOCOM (2003) 8. Younis, O., Fahmy, S.: Distributed Clustering in Ad-hoc Sensor Networks: A Hybrid, Energy-Efficient Approach. In: Proceedings of IEEE INFOCOM. IEEE Press (2004) 9. Kim, J.-S., Byun, T.-Y.: A Density-based Clustering Scheme for Wireless Sensor Networks. CCIS, vol. 195, pp. 267–276. Springer, Heidelberg (2011)

Performance Analysis of DRAM-SSD and HDD According to the Each Environment on MYSQL Hyun-Ju Song1, Young-Hun Lee1,*, and Seung-Kook Cheong2 1

Dept. of Electronic Eng., Hannam University, Ojeong -dong, Daedeok-gu, Daejon 306-791, Korea [email protected], [email protected] 2 Dept. Principal Member of Engineering Staff.. ETRI, 161, Gajeong-dong, Yuseong-gu, Daejeon, 305-700, Korea [email protected] Abstract. Recently, Users needed storage for processing high-capacity data. Meanwhile, the HDD was used primarily as storage device, but SSD was developed as a fast access device. So, the use of SSD increased a lot of capacity-process. Using a SAN Switch, we tested in order to resolve data process faster. So in this paper, DRAM-SSD and HDD’s performance ability in the data process will be confirmed using TPC-H Benchmark on MYSQL at SAN environment. From the performance analysis results, performance difference of HDD and DRAM-SSD is little when database size is low at san switch connecting with storage. But ad-hoc query process ability difference between DRAM-SSD and HDD increased. Based on these results in a SANbased, DRAM-SSD has a better performance than the HDD. Therefore San is judged to be more effective when dealing with the large amounts of data using a SAN to manage data. Keywords: SSD, HDD, Mysql, TPC-H, SAN.

1

Introduction

Recently, Users needed storage for processing high-capacity data. Meanwhile, the HDD was used primarily as storage device, but the development of SSD is faster access as storage devices and a lot of research is being accelerated. From the research of HDD and SSD, the difference of data I/O processing performance was progressed by comparing performance of storage device of each.[1-3] Basic storage device’s manage data and use of DBMS which provided efficient and convenient way increased. So in this paper, we will compare DRAM-SSD to HDD and confirm showing good performance in data processing, using TPC-H Benchmark through SAN Switch on MYSQL, The process of this paper is as follow. In chapter 2 we will introduce used technology that evaluates performance of HDD and SSD, in chapter 3 we will introduce analysis environment and condition by thinking each storage device and using tool. And in chapter 4 analyzes test results will be shown with conditions described in chapter 3, and chapter 5 is concluded. *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 330–337, 2011. © Springer-Verlag Berlin Heidelberg 2011

Performance Analysis of DRAM-SSD and HDD According to the Each Environment

2

Related Study

2.1

MYSQL

331

MYSQL is the relative database management system of open source that uses SQL which is standard database quality language, which is very fast, flexible, and easy to use. MYSQL offers client/server environment, and a server installed MYSQL has MYSQL daemon called mysqld, so client program connects via network by this daemon so that it can control data. [5]

mysqld Client Network

mysqld Database

Client mysqld

MySQL Server

Fig. 1. Performance structure of MYSQL

2.2

SAN

SAN stands for Storage Area network and say move available high-speed network for large-capacity between storage equipment which unrelated distributed sort of Host. i.e., say all together configured network to communication storage component and system component at fiber channel networks. SAN integrate and share storage, and supply high-performance link to data device, and add overlapping-link to storage, and speeding up data backup, and support available-high-clustered systems. SAN have advantage by building implementing a highly available data access, integration resources and management of storage, required window to backup and traffic reduce, occupation solution of host CPU cycle, data retention function through disaster Accepted techniques. [6]

Fig. 2. General

storage server VS SAN

structure

332

2.3

H.-J. Song, Y.-H. Lee, and S.-K. Cheong

TPC-H

TPC stands for Transaction processing performance and is non-profit organization that has multiple hardware and software companies and small number of user organizations. Usually TPC called transaction processing performance evaluation committee, but TPC-alphabet notation tells the benchmark model. TPC becomes standard evaluating processing performance of on-line transaction processing systems. TPC defines transaction processing and benchmarks of database and is used to performance measure of total system including disk I/O and s/w. As a benchmarking tool to measure that how fast it can handle complex SQL, and defines 22 SQL statements and DB schema, and set of data about 1GB. TPC-H benchmark is public performance test that is used SQL that Business-oriented ad-hoc Query and concurrent data modifications made by the combination about large data. Fig 3 showing business environment of TPC-H, impromptu Query and modification transactions was performed at table from multiple users, model situation entered data from database of decision support system from OLTP system. [7]

Fig. 3. TPC-H's business environment

3

Performance Analysis Environment and Conditions

3.1

Performance Test Environment

This paper comparative analyzed the performance of each, which the HDD as a storage device in modern mainly used and the SSD as a storage device in modern society used increasingly. HDD Storage and DRAM-SSD Storage were constructed in conjunction with a SAN switch for test Environment. Test Environment constituted as fig 4, and test subject server formed as Linux CentOS 5.3 version, performance measure tool used TPC-H Benchmark, to use a tool as the database was installed MYSQL 5.0.90 version.

Performance Analysis of DRA AM-SSD and HDD According to the Each Environment

333

Fiig. 4. Test environment configuration

3.2

Test Procedure and d Condition

Performance of each sto orage device comparative analyzed to configure SAN environment using TPC-H Benchmark. Test using the TPC-H Benchmark compaared Test to performance analysis thrrough total three step procedure with the results. Load T is the first step, which maake up database and store to generate data. Next stepp is power test, which was anallyzed by measuring the ability when Single active useer is run the Query. Final step is Throughput Test, which was analyzed by measuring the ability when Multi active user u is run Query at same time. The performance of eeach storage device analyzed thrrough Power TEST and Throughput Test combined results in SAN environment. Data generation and total of eight tables produced to sttore generated data in data storaage of Mysql. Read data in the created file store data whhich get by running stored-codee in the DBMS, confirm insert play time of stored dataa. A total of 22 Query given in n the TPC-H benchmark, which the play-time checkk by running through the Query Browser, total of 22 Query stored for multi active user test T installed the directory. After run total of 22 Querry a through vim command in TPC-H script run to delete set by checking the result. Result of insert execution time and a delete execution time calculate value of Power T Test 22Query execution time and apply in below expression 1. 1

(1) QI (i, 0) is the execution n time of I second query, RI (j, 0) as the running timee of insert and delete unit secon nd the data generated by the size of SF, i.e., the value off the database. When each of Query Q accordingly practice ability of multiple users at the same time make amount of data as table 1 number of users to differ.

334

H.-J. Song, Y.-H. Lee L , and S.-K. Cheong Ta able 1. Conditions for throughput test Datab base capacity

User number

1GB

2

5GB

2

10GB

3

The fixed users perform m as shown in above table1 parallel 22 Query of each. A And time check was expressed in units of sec from first user starts the first Query to last user finish last Query. By substituting s in Equation 2 is to evaluate performance tiime to check on the recipe. (2) In the above expression 2, is performance time, S is number of user, SF shhow amount of data. Through th he result of expression 1 and expression 2, expression 3 has produced result to get final result to analysis performance of each storage device. (3) QphH@Size which is a calculated value from Equation 3 is reflected value of the many features of Query processing p system, these features are performed Queery, Query processing ability when w performed Query, Query performed by multiple ussers simultaneously including alll the processing ability. Thus, QphH@Size shows ad-hoc query capabilities that to haandle hourly capacities of the database.

creation n

Tool (DBGen, QGen)

Database <Mysql> application

creation n

Queries

Fig. 5. TPC-H H applies to get the results the block Diagram

Performance Analysis of DRAM-SSD and HDD According to the Each Environment

4

335

Test Results

Test results using TPC-H Benchmark comparative analyzed performance at SAN environment in according to described result calculation method in Section 3.2. Some of total 22 Query for test when Database capacity 1GB, 5GB, 10GB, by modifying comparative analysis performance for each size Shorten the time the test case. Table 2. SAN environment HDD performance measurements

HDD

Database capacity

Test result

1GB

5GB

10GB

Power@size

1.0891E-09

3.8770E-30

3.6293E-31

Throughput@size

15.7283

7.372

7.0681

QphH@size

1.3088E-04

5.3461E-15

1.6016E-15

Table 3. Local environment DRAM-SSD performance measurements

DRAM-SSD

Database capacity

Test result

1GB

5GB

10GB

Power@size

9.7911E-09

4.6299E-20

8.8710E-22

Throughput@size

19.3928

34.7933

80.1214

QphH@size

4.3575E-04

1.2692E-09

2.6660E-10

As can be seen through above Table 2 and Table 3, the practical value for comparing the performance of DRAM-SSD Storage and HDD Storage in SAN environment is QphH@size, compared result to only database-capacity of a storage device of each same below picture. QphH@size value that can handle ad-hoc queries, the little difference find out between ad-hoc query handle-capabilities of HDD Storage and DRAM-SSD Storage in low carrying a load 1G, but DRAM-SSD storagere can be seen that much higher which per hour to handle ad-hoc query capabilities more than HDD storage when amounts of data is increasing.

336

H.-J. Song, Y.-H. Lee, and S.-K. Cheong

Fig. 6. SAN environment database capacities of each storage device QphH@size value

5

Conclusion

In this paper, we analyzed the data processing performance of a DRAM-SSD and HDD as a data processing storage using the TPC-H Benchmark on mysql in a san environment. From the performance analysis results, when database Size increased performance compared analysis value QphH@size was increased with TPC-H, but difference of the ad-hoc query processing ability of DRAM-SSD and HDD are increased when compared of each database capacity. Based on these results, using by connecting SAN Switch and DRAM-SSD storage is judged to be effective at large amounts of data I/O required field or applications. Also mentioned in the introduction, SSD has benefits if SSD price is stabled. Using SSD is thought to be effective in case of industry side which required large amounts of data I/O or Media Server, using storage device of computer and notebook computer. In the future, we will be analyzed the power consumption from the TPC-H Benchmark at SAN Environment. Also, SAN Environment and Local Environmental TEST results will be analyzed to process the data.

References [1] Park, K.-H., Choe, J.-k., Lee, Y.-H., Cheong, S.-K., Kim, Y.-S.: Performance Analysis for DRAM-based SSD system. Korean Information Technical Academic Society Dissertation 7(4), 41–47 (2009)

Performance Analysis of DRAM-SSD and HDD According to the Each Environment

337

[2] Kang, Y.-H., Yoo, J.-H., Cheong, S.-K.: Performance Evaluation of the SSD based on DRAM Storage System IOPS. Korean Information Technical Academic Society Dissertation 7(1), 265–272 (2009) [3] Cheong, S.-K., Jeong, Y.-W., Jeong, Y.-J., Jeong, J.-J.: Input-Output Performance Analysis of HDD and DDR-SSD Storage under the Streaming Workload. Korean Information Technical Academic Society Dissertation, 322–325 (2010) [4] Cheong, S.-K., Ko, D.-S.: Technology Prospect of Next Generation Storage. Korean Information Technical Academic Society Dissertation Summer Synthetic Scientific Announcement Thesis, 137 (2008) [5] Kim, H.: Learn to MYSQL Database Programming, http://www.Young-jin.com [6] Judd, J., Beaucbamp, C.: Building SANs Witch Brocade. Syngess, 9–16 (November 1996) [7] http://www.tpc.org/tpch/spec/tpch2.13.0.pdf

Dynamic Channel Adjustable Asynchronous Cognitive Radio MAC Protocol for Wireless Medical Body Area Sensor Networks Byunghwa Lee1, Jangkyu Yun1, and Kijun Han2,

*

1 The Graduate School of Electrical Engineering and Computer Science, Kyungpook National University, 1370, Sankyuk-dong, Buk-gu, Daegu, 702-701, Korea 2 The School of Computer Seience and Engineering, Kyungpook National University, 1370, Sankyuk-dong, Buk-gu, Daegu, 702-701, Korea {bhlee,kyu9901}@netopia.knu.ac.kr, [email protected]

Abstract. Medical body area networks (MBAN) impose several requirements to the medium access control layer which have various contexts: energy efficiency, QoS providing, reliability. And a cognitive radio (CR) network should be able to sense its environment and adapt communication to utilize the unused licensed spectrum without interfering with licensed users. As CR nodes need to hop from channel to channel to make the most use of the spectrum opportunities, we consider asynchronous medium access control (MAC) protocols to be solution for these networks. The DCAA-MAC protocol presented in this paper has been designed has been designed specifically for wireless body area network with cognitive radio capability. The DCAA-MAC protocol has energyefficiency, low latency, and no synchronization overhead by provide asynchronous and fast channel switching. Analytical models are shown to perform at low energy consumption, to scale well with network size. Keywords: Cognitive radio, Wireless body area network, MAC.

1

Introduction

The convergence of all media and data services is disseminated throughout innovation of wireless communication technologies. In recent years there has been also increasing interest in implementing ubiquitous monitoring system in hospital and house for patient. These monitoring systems are normally observed by on-body sensors that are connected by a controller of the medical body area network (MBAN). The MBAN is a promising solution for eliminating wires, thus allowing sensors to reliably and inexpensively collect multiple parameters simultaneously and relay the monitoring information wirelessly so that clinicians can respond rapidly [1]. MBANs for wireless patient monitoring is an essential component to improving patient outcomes and lowering healthcare costs. *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 338–345, 2011. © Springer-Verlag Berlin Heidelberg 2011

Dynamic Channel Adjustable Asynchronous Cognitive Radio MAC Protocol

339

Quality of service is a key requirement for MBANs, and hence the importance of having a relatively clean and less crowed spectrum band. The 2.4Ghz industrial, scientific and medical (ISM) band is not suitable for life-critical medical applications due to the interference and congestion for wireless networks in hospital and house. Cognitive radio (CR) is a novel technology, which improves the spectrum utilization by seeking and opportunistically utilizing radio resources in time, frequency and space domains on a real time basis. Cognitive radio is an emerging technology to address critical challenges of spectrum efficiency, interference management, coexistence and interoperability associated with current and future wireless networks. Cognitive wireless communications and networks have high potentials to bring enormous economic benefits to both customer and operators. In recent years, cognitive wireless communications and networking attract wide interests from both academia and industry. These features and requirements raise the demand of CR for MBAN implementation. We explore novel technique to provide CR MAC protocol for WBAN.

2

The Dynamic Channel Adjustable Asynchronous MAC Protocol in Wireless Medical Body Area Sensor Networks as Enabler for CR (DCAA-MAC)

In this section, we will present a new CR MAC protocol for MBAN that enables efficient spectrum sharing by borrowing licensed spectrum. The protocol is designed to protect primary users (PUs) from CR devices’ interference since borrowing licensed spectrum has to be protected. Basically, the DCAA-MAC based on asynchronous MAC paradigm to provide energy efficiency, low latency, reconfigurability to MBAN. At initialization time, each node scans channel to join the network. If it fails to find the network, the device selects the channel which has best condition to make the network. (E.g. the least frequently used channel, SNR, etc.) To minimize listen cost, every node in the network wakes up only in check interval which sensing communication channel. Each node goes to sleep and wake up periodically and independently of the others. When a node wants to communicate with a neighbor, it first send a preamble which contains address of destination node that lasts an entire duty cycle, and receive the ACK message from the destination node which hears preamble, just after that, it sends the data. The receiver detects the preamble which has same address of its when it wakes up, it remains awake in order to receive the data. After data transmission is done, nodes which transmitted and received data go to the sleep mode. Advantage of this technique is that it works in a completely unsynchronized environment.

340

B. Lee, J. Yun, and K. Han

Start Check Interval Duration Time Over

CCA

Busy Target Address Checking Target Address == My Address

Clear Target Address != My Address Unidentified

Sleep Target Address != My Address Listen ≤ 1 Check Interval Unidentified

ACK Message Transmission

Control Channel Open

Data Reception

Listen ≤ 1 Check Interval

CSI Reception Failed CSI Reception Success

DACK Message Transmission

CSACK Message Transmission

CSACK Reception Failed CSPreamble Transmission

CSACK Reception Success Channel Switching

Fig. 1. Activity diagram of DCAA-MAC

If when a node detects interferences on its current channel, either due to PU or noise appearance, it switches to another one so that communications can be maintained. The operation over multiple communication channels allows to maintain communications even when facing PUs appearance or interference. When a node detects signal which has over the power of threshold during CCA with fast sensing period on its current channel, listen to signal to find the address of destination during maximum one check interval. If it finds the address of destination and the address is same with own address, it sends ACK message to source of preamble, else it sleep to next check interval. If it does not find any address type, it determines that PU or interference appears and try to switch to another channel for PU protection and QoS provision. That is provided CR capability which based on result of energy detection during CCA with fast sensing period in MAC layer. Spectrum sensing is an important requirement for the realization of CR networks. Feature detection and energy detection are the most commonly used for spectrum sensing in CR networks. Feature detection determines the presence of PU signals by extracting their specific features. Although feature detection is most effective for the nature CR networks, it is

Dynamic Channel Adjustable Asynchronous Cognitive Radio MAC Protocol

341

computationally complex and requires significantly long sensing time [2]. Energy detection is optimal to detect the unknown signal if the noise power is known. In the energy detection, CR users sense the presence/absence of the PUs based on the energy of the received signals. It is difficult to provide QoS with channel switching with result of feature detection because of feature detection’s these characteristic. Therefore, channel switching is performed with only result of energy detection in DCAA-MAC. After determine channel switching, node switch the current channel to control channel. And then listen during maximum one check interval because try to find another node which broadcast channel switching preamble (CSPreamble). If receive the preamble which have the channel switching information, node transmits the channel switching ACK (CSACK) message to the source node and switch the channel as follow the channel switching information. If there are not any signals on control channel during one check interval, nodes transmitCSPreamble that include channel switching information which has channel number to move and after that receive CSACK message from destination node, switch the channel. These fast channel switching mechanisms provideQoS for MBAN by low latency. Fig.2 shows the example of the proposed protocol.

Fig. 2. Example of the proposed MAC protocol for CR.

3

Performance Analysis

3.1

System Model

Performance analyses are based on some assumptions. We define that a radio can be in one of three steady states: Transmission, Reception and Sleep. Furthermore, four

342

B. Lee, J. Yun, and K. Han

transient states are defined: SetupTx, SetupRx, SwitchTxRx, SwitchRxTx. The time spent in a transient state is a TTrState, the power consumption in each state is PState and the energy cost of a transition from one steady state to another is ETrState. Table 1 shows the parameters we have used for the performance analysis in detail. Table 1. Parameters for performance analysis Parameter

Value

PTx, PSetupTx PRx, PSetupRx PSleep PSwTxRx, PSwRxTx TSwTxRx TSetpuTx TSetupRx TCCA TW TM TPreamble TACK TCS

56.4 mW 25.5 mW 0.06 mW 54.3 mW 160 μs 12μs 192μs 128μs 500 ms 6.9 ms 128 μs 1.02 ms 61.5 μs 250 kbps

Bit rate

We assume that a network has N devices. The network is fully connected and each node can directly communicate with all the others. Each node generates traffic according to an exponential distribution with parameter = 1/ : the mean interval between two packets is L seconds. Each node receives as much data as it is transmitting. 3.2

Power Consumption

In case of CSMA protocol, power consumption is evaluated as follows: =

1 (1)

On average, during a time L, a radio using the CSMA must send a packet, receive a packet and sleep the rest of the time.

Dynamic Channel Adjustable Asynchronous Cognitive Radio MAC Protocol

343

In DCAA-MAC makes use of wake-up preambles. A node using this protocol must perform one carrier sensing, transmit data, receive data and sleep in the rest time of one period. This shows to the following expression: ,

= where: = =

1

1 =

1

(2)

=

Traffic adaptation shows how power consumption evolves with an increase of data rate.

Fig. 3. Protocol power consumption depending on traffic rate

Fig.3 shows the power consumption and energy efficiencies of CSMA and our DCAA-MAC protocols as a function of the mean time L between packets. CSMA and DCAA-MAC exhibit the same qualitative behavior. DCAA-MAC scale remarkably well and never far from the CSMA power consumption.

344

B. Lee, J. Yun, and K. Han

Network density is more important when considering node mobility. This can be the case for body area networks. 3.3

Delay

A simply analysis of average delay of protocol can be made by using standard results of queueing theory. We assumed that packets are generated at each sensor device according to a Poisson process of parameter λs, and the sum of these N independent and identically distributed random point processes also follows a Poisson distribution, of parameter λ=λs, If packets are constant size, the service time is deterministic and the queueing model M/D/1 can be applied. The average delay is given by: = where

= ,

,

is the service time

(3)

For CSMA, -1=TM+2TSIFS is approximation of the service time. And for DCAAMAC, μ-1=TW/2because the packet can arrive at the MAC layer at any time.

Fig. 4. Average packet latency for DDCA-MAC and CSMA

Fig.4 shows bilogarithmic scale the average packet delay as a function of the average number of packets per node, for various network densities (from 5 to 100).

Dynamic Channel Adjustable Asynchronous Cognitive Radio MAC Protocol

4

345

Conclusion

In this Paper, we have proposed a dynamic channel adjustable asynchronous MAC protocol as enabler for CR in wireless medical body area sensor network. In the proposed approach, each node goes to sleep and wake up periodically and independently of the others. Channel switching is performed not with result of feature detection but result of energy detection because QoSproviding. Analytical models are shown to perform at low energy consumption, to scale well with network size and to allow coexistence with simultaneously operating independent networks. Acknowledgments * This work was supported by the second phase of the Brain Korea 21 Program in 2011. * This work was supported by National Research Foundation of Korea Grant funded by the Korean Government.

References 1. Petel, M., Wang, J.: Applications, Challenges, and Prospective in Emerging Body Area Networking Technologies. IEEE Wireless Commun. 17(1), 80–88 (2010) 2. Hur, Y., Park, J., Woo, W., Lee, J.S., Lim, K., Lee, C.-H., Kim, H.S., Laskar, J.: A Cognitive Radio (CR) System Employing A Dual-Stage Spectrum Sensing Technique: A MultiResolution Spectrum Sensing (MRSS) and A Temporal Signature Detection (TSD) Technique. In: Proceedings of the IEEE Globecom 2006, pp. 1–5 (2006)

A Multiple-Metric Routing Scheme for QoS in WMNs Using a System of Active Networks Jangkyu Yun1, Byunghwa Lee1, Junhyung Kim1, and Kijun Han2, 1

*

The Graduate School of Electrical Engineering and Computer Science, Kyungpook National University, 1370, Sankyuk-dong, Buk-gu, Daegu, 702-701, Korea 2 The School of Computer Seience and Engineering, Kyungpook National University, 1370, Sankyuk-dong, Buk-gu, Daegu, 702-701, Korea {kyu9901,bhlee,jhkim}@netopia.knu.ac.kr, [email protected]

Abstract. Wireless mesh networking is emerging as an important architecture for the future generation of wireless communications systems. The challenging issue in WMNs is providing Quality of Service (QoS). So, this paper proposes a multiple-metric routing scheme for QoS in WMNs using a system of active networks. The Active Network paradigm offers the attractive capability of being able to carry executable payloads that can change the characteristics of a given platform. In other words, network nodes not only forward packets, but also perform customized computation on the packets flowing through them. It provides a programmable interface to the user. Keywords: wireless mesh networks, active networks, AODV.

1

Introduction

Wireless mesh networks (WMNs) will play an increasingly important role in futuregeneration wireless mobile networks. A WMN normally consists of mesh routers and clients, and can be independently implemented or integrated with other communications systems such as conventional cellular networks [1]. WMNs are characterized by their dynamic self-organization, self-configuration, and self-healing to enable quick deployment, easy maintenance, low cost, great scalability, and reliable services, as well as enhanced network capacity, connectivity, and resilience[2]. A method of providing QoS is the key technology for traffic management in WMNs. There have been various researches conducted on how to provide QoS in WMNs. A previous research for providing QoS has been studied with routing metrics such as hop count, Expected Transmission Rate (ETX) and Expected Transmission Time (ETT). The hop count is the base metric and is a simple measure of the number of hops between the source and destination of a path [3]. The ETX is a measure of link and path quality and the ETT is the Expected Transmission Time of a data transmission in a direct link [3][4][5]. Although there have been many studies on how to provide QoS, it is not suitable to offer QoS in a dynamic environment. *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 346–353, 2011. © Springer-Verlag Berlin Heidelberg 2011

A Multiple-Metric Routing Scheme for QoS in WMNs

347

In this paper, an active network architecture is used to provide QoS in WMNs. Active Networking is primarily a DARPA funded project focusing on mechanisms, applications, and operating systems research to develop a reconfigurable network infrastructure. The active network paradigm offers the attractive capability of carrying executable payloads that can change the characteristics of a given platform. The key point of an active network architecture is the active node construction because the Active packets are operating on the active nodes. The active nodes architecture, which was put forward by DARPA, depicted the data processing flow. The logical constructions of active nodes involve three parts: Node OS, Execution Environment and Active Application. The Node OS is similar to the general operating kernel which through a fixed interface to provide resources and render services for the execution environment. The Execution Environment is a transparent, programmable space and is unrelated to the platform. It operates in each active node and user terminal node; a multi-execution environment could operate on the same active node at the same time. The execution environment provides various network application interfaces for higher level applications. The Active Application is a series of user-defined procedures. By executing the network API provided by the executive environment, the necessary resources can be obtained when running a program. Finally, a customized function can be realized. The active node (routers, etc.) in Active Networks are not only data forward packets, but also perform customized computation on the packet flowing through them. The active node operating codes are initially contained in the active node. This code can also be dynamically inserted into the forwarding packet to configure them according to the needs of the applications in execution. This way, packets have the capacity of carrying not only data but also the code to be executed in remote nodes. Therefore, the user has the possibility of “programming” the network, providing that the programs to be used by the routers and switches to execute their computations are available. In an active network, the difference between network internal nodes (routers, switches, etc.) and user nodes is tenuous, since both are able to perform the same computations. Hence, the user can view the network as a part of his/her application and can adapt the network to obtain the best performance of his/her application [6].

2

Proposed Scheme

It has already been mentioned that, many performance metrics, such as ETX and ETT so on have been considered for QoS in WMNs. Since each individual routing metric considers some features, it is difficult to satisfy all the requirements of WMNs by using a single metric. Therefore, it is proposed that a multiple-metric routing scheme be used. An active networking technique for QoS was grafted onto an existing routing protocol. The basic routing protocol of the proposed scheme is the Ad hoc Ondemand Distance Vector (AODV). In this paper, this protocol is referred to as the

348

J. Yun et al.

Active AODV. A routing table and a route discovery process make the difference between these two protocols. The modification performed is the addition of a QoS Type field to the routing table. It determines what type of QoS to use such as reliability, delay, interference and so on. And RREP, RREQ and RERR packets are also modified. A QoS field is added to reserve the fields of these three packets, since they have reserved fields. The QoS field also determines what type of QoS is required for the application. Nodes fill the QoS field when sending the route discovery message. Nodes compare just the destination ID field when the process route discovery is in pure AODV. However, nodes compare not only the destination ID but also the QoS field of the discovery message. Therefore, each routing metric is used individually to select the demanded path in the Active AODV. In other words, the Active AODV determines where the suitable path is. For example, it determines the best path for the highest success rate for a packet, which will result in providing successful service. Further, it determines the fastest path for the delay demanded packet. There are many QoS types for applications. Nevertheless, this paper considers only five metrics: HOP, ETX, ETT, AB and EI. AB and EI mean Available Bandwidth and Expected Interference, respectively. The AB of the link is defined as following equation 1. ABl = Bl − UBl

(1)

Bl and UBl denote link bandwidth and using bandwidth of link, respectively. Also, The EI of the link is defined as following equation 2. EI l = ETTl × N l

(2)

Nl denotes the neighbor number of the link. On the active protocol level, several solutions for active networks have been proposed so far. The base protocol of this scheme is the Active Network Encapsulation Protocol (ANEP). Further, payloads are encapsulated within ANEP packets and ANEP packets are encapsulated within an IP packet as shown in Fig. 1. 



+2*GCFGT 8GTUKQP #0'2*GCFGT





+2 *GCFGT 4QWVGT#NVGT1RVKQP (NCI

6[RG+&

#0'2*GCFGT .GPIVJ

#0'22CEMGV.GPIVJ 1RVKQP

#1&82CEMGV 44'3 44'24'44

Fig. 1. Structure of encapsulated packet



A Multiple-Metric Routing Scheme for QoS in WMNs

349

All active packets are demultiplexed by ANEP. The Type ID field of the ANEP header indicates the evaluation environment of the packet. The active node should evaluate the packet in the proper environment. Yet, traditional nodes cannot demultiplex active packets, so it just forwards the packet. This means this scheme can operate on both networks Fig. 2 shows the active node structure. When the node receives a packet, node OS it dispatched to a suitable EE. An active packet is dispatched to ANEP, but a normal packet is dispatched to IPv4 of IPv6. After that, an active packet is demultiplexed by the ANEP and the AODV packet is processed on the Active AODV

...

Active AODV AA

ANEP

IPv4

IPv6

...

Management EE

EE

Dispatch

Security Enforcement Engine

Transmit

Store ...

.. .

Node OS

Input

Output

Policy

channels

channels

DB

Fig. 2. Structure of active node

Fig. 3 shows an example of the proposed multiple-metric routing. B C

A

D Routing Table of node C Destination

Next hop

Metric

A

D

4

2

A

B

5

3

…

QoS Type

. . .

Fig. 3. An example of a routing table

This is the routing table of node C. There are two entries to node A. Hence, node C forwards a packet to node D for the packet success rate. On the other hand, node C forwards a packet to node B for a delay required packet. It means every node has many entries to the same destination. The QoS Type field makes it possible.

350

3

J. Yun et al.

Evaluation

A simulation was conducted based on ns-2 to evaluate the Active AODV, which uses multiple metrics (HOP, ETT, ETX, AB, EI). In addition, a pure AODV and an Active AODV was used as a routing protocol. The simulations tried two ways of measuring the performance of the Active AODV according to the bandwidth in the networks. The first simulation used set nodes with varied bandwidths (1~4 Mbps) and the second simulation had a fixed bandwidth (1.5 Mbps). The topology is a 4 x 4 Grid, the number of nodes with regular deployment is 16 and an IEEE 802.11 which is a MAC protocol, was used in the simulation. Table 1 shows the detailed parameters in the simulations.

Table 1. Parameters used for simulation Parameters

Values

Topology

4 x 4 Grid

The number of node

16

The distance between nodes

90m

Radio rage

100m

MAC Protocol

IEEE 802.11

Packet size

512 Byte

Bandwidth

1~4 Mbps

Time of total simulation

200s

In the first simulation, the routings used in each HOP, ETX, ETT, AB and EI were compared according to nodes with increasing bandwidths from 1 to 4 Mbps.

Fig. 4. Packet delivery ratio (1~4 Mbps)

A Multiple-Metric Routing Scheme for QoS in WMNs

351

Fig. 4 and Fig. 5 present the result ratio and the delay of packet delivery. It shows that the experimental result of routing with ETX, ETT, AB and EI is better than HOP when pairs of routing nodes were increased in Fig. 4. However, AB and EI obviously decrease when there are 11 pairs of routing nodes. It can be observed that AB and EI are worse than other metrics at high loads. Fig. 5 illustrates that the performance of routing with ETT is greatly enhanced and the packet delivery delay is significantly reduced. It can be seen that routing with ETT provides an efficient path because it considers the packet delivery delay of a packet.

Fig. 5. Packet delivery delay (1~4 Mbps)

In the second simulation, the performance of routing with nodes with fixed 1.5 Mbps was tested according to the pairs of routing nodes in Fig. 6 and Fig. 7.

Fig. 6. Packet delivery ratio (1.5 Mbps)

352

J. Yun et al.

Fig. 6 represents similar results in the performance when pairs of nodes are 10 but routing using AB and EI sharply declined. Fig. 7 shows that delay of all routing rose according to increasing loads. Also, routing using ETX and ETT showed improved performance when pairs of routing node are 8. It was observed, through this simulation, that routing is influenced by metrics. There is a noticeable difference in the routing performance when it is measured in metrics. It has been proven that efficient routing is possible when a property metric is used according to the QoS of data since there is a varied bandwidth in WMNs. Therefore, a path can be selected by considering the success rate when the data delivery ratio is important or a path can be selected by considering delay when data delivery delay is important.

Fig. 7. Packet delivery delay (1.5 Mbps)

4

Conclusion

In this paper, the multiple-metric routing scheme for QoS in WMNs has been proposed by using a system of active networks. An active networking technique was grafted for QoS onto the existing AODV routing protocol. Moreover, the routing table and the routing discovery process of the AODV were modified. This makes the Active AODV determine where a suitable path is for the required QoS application. Plus, the simulation result proved this. The proposed scheme is especially useful in multiple-metric required environments in WMNs. Acknowledgments * This work was supported by the second phase of the Brain Korea 21 Program in 2011. * This work was supported by National Research Foundation of Korea Grant funded by the Korean Government.

A Multiple-Metric Routing Scheme for QoS in WMNs

353

References 1. Akyildiz, F., Wang, X., Wang, W.: Wireless Mesh Networks: A Survey. Elsevier Comp. Networks 47(4), 445–487 (2005) 2. Ci, S., et al.: Self-Regulating Network Utilization in Mobile Ad-hoc Wireless Networks. IEEE Trans. Vehic. Tech. 55(4), 1302–1310 (2006) 3. Gowrishankar, S., Sarkar, S.K., Basavaraju, T.G.: Performance analysis of AODV, AODVUU, AOMDV and RAODV over IEEE 802.15.4 in wireless sensor networks. Computer Science and Information Technology, 59–63 (2009) 4. Yang, Y., Wang, J., Kravets, R.: Designing routing metrics for mesh networks. WiMesh (2005) 5. Jiang, W., Zhang, Z., Zhong, X.: High Throughput Routing in Large-Scale Multi-Radio Wireless Mesh Networks. In: Wireless Communications and Networking Conference 2007 (2007) 6. Lu, Q., Ma, Y., Zhang, J.: A study of the active router structure. Computing, Communication, Control and Management 1, 157–160 (2009)

Implementation of Log Analysis System for Desktop Grids and Its Application to Resource Group-Based Task Scheduling Joon-Min Gil1 , Mihye Kim2 , and Ui-Sung Song3, 1

School of Computer & Information Communications Eng., Catholic Univ. of Daegu 2 Dept. of Computer Science Education, Catholic Univ. of Daegu, 13-13 Hayang-ro, Hayang-eup, Gyeongsan-si, Gyeongbuk 712-702, S. Korea {jmgil,mihyekim}@cu.ac.kr 3 Dept. of Computer Education, Busan National University of Education, 24 Gyodae-ro, Yeonje-gu, Busan 611-736, S. Korea [email protected]

Abstract. It is important that desktop grids should be aggressively deal with the dynamic properties arisen from the volatility and heterogeneity of resources. Therefore, it is required that task scheduling should positively consider the execution behavior that is characterized by an individual resource. In this paper, we implement a log analysis system which can analyze the execution behavior by utilizing actual log data of desktop grid systems. To verify the log analysis system, we conducted simulations and showed that the resource group-based task scheduling, based on the analysis of the execution behavior, offers faster turnaround time than the existing one even if few resources are used. Keywords: Desktop grids, Execution behavior, Log analysis system, Resource group-based task scheduling.

1

Introduction

With the popular use of Internet and high-performance of PC resources, it is possible to build a desktop grid environment, in which enables to build a virtual computing environment by binding the unused PC resources that is connected to Internet [1]. An important aspect in desktop grids is that each resource has volatility property, due to free withdrawal from execution participation even in the middle of task execution. Moreover, each resource has heterogeneity property as it has totally different computing environment (e.g., CPU performance, memory capacity, network speed, etc) [1,2]. Due to these two properties, it is not possible to expect the completion time of entire tasks. Therefore, if tasks are allocated without any consideration for the dynamic execution features of resources, execution failures will occur frequently due to the volatility and heterogeneity, and 

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 354–363, 2011. c Springer-Verlag Berlin Heidelberg 2011 

Implementation of Log Analysis System for Desktop Grids

355

thus the turnaround time of entire tasks becomes longer. However, execution failures can be prevented by selectively allocating tasks to those resources that are suitable for the current execution environment of desktop grids. As a result, it is necessary to analyze the execution behaviors of resources as a way to provide a stable execution environment for desktop grids. In this paper, we implement the log analysis system that can extract the execution behavior of resources from actual log data in Korea@Home system. The analysis results of the execution behavior can be used as the task allocation information to minimize the waste of resources and the execution delay in task scheduling. In particular, this study can be very useful as background on deciding task scheduling policies for various desktop grid systems as well as Korea@Home. The rest of this paper is organized as follows. In Section 2, we provide a brief description of desktop grids used in this paper. In Section 3, we define the execution behavior of resources as availability and credibility. The log analysis system implemented in this paper is presented in Section 4. This section also presents in detail the modules of the log analysis system. In Section 5. we present log analysis results and their application to resource group-based task scheduling. Section 6 concludes the paper.

2

Desktop Grid Environment

The desktop grid environment assumed in this paper is physically composed of a client (or an application submitter), a central server, and resources. The client is an entity submitting its own application to the central server. The central server takes responsibility of mediating resources and tasks and performs the functions such as task management, resource management, task allocation, etc. Each resource acts as a resource provider and executes the tasks sent by the central server during its idle cycles. Once a task is finished, the task result is sent back to the central server. The detailed steps to execute the parallel tasks submitted by a client in desktop grids are described as follows: Step 1 (Resource registration): Each resource registers its own information, such as CPU performance, memory capacity, OS type, etc., to the central server. Step 2 (Application submission): A client submits its own application to the central server. Step 3 (Task allocation): The central server allocates tasks in task pool to resources. Step 4 (Task execution): The resource to which tasks are assigned executes the tasks during its idle cycles. As soon as finishing task execution, it sends back task results to the central server. Step 5 (Result collection): The central server collects the task results received from resources and records them to databases. Step 6 (Application completion): The central server checks if all tasks are completed, and sends final results to the client.

356

J.-M. Gil, M. Kim, and U.-S. Song

The above described steps are based on the client-server architecture which has been typically used in the most desktop grid systems (e.g., BOINC [3], XtremWeb [4], Korea@Home [5], etc). The application, that is submitted to the central server by clients, is divided into hundreds of thousands of tasks, each of which is small enough to be executed in one resource. In addition, each task is mutually independent without any dependency between each others.

3

Execution Behavior of Resources

The performance of desktop grids is largely influenced by the dynamic features such as the execution join and withdrawn of resources [1,6]. Due to the task stops that can occur at any time even in the middle of task execution, task failures will be unavoidably encountered. Thus, the availability, that represents how much time each resource can spend executing tasks for a given time period, and the credibility, that determines the trustworthiness of task results in the present of failures, can be considered as an important factor to enhance the overall performance of desktop grids. Due to the dynamic features of desktop grids, it is strongly recommended that tasks should be allocated to the resources with high availability and credibility; these resources can return credible task results as many as possible within a given time period. In this paper, therefore, we classify resources by two factors, availability and credibility, and analyze the execution behavior of resources based on these two factors. Definition 1. Availability (A): a probability that can execute tasks in the presence of task failures. A=

M T T CF M T T CF + M T T CR

(1)

Definition 2. Credibility (C): a factor of how many task results can be returned after tasks are allocated to resources.  C=

r n,

0,

if n > 0 if n = 0

(2)

In Equation (1), MTTCF (Mean Time To Critical Failure) means the average failure time that is caused by execution failures or resource defects, and MTTCR (Mean Time To Correct and Recover) means the average execution time including recovery time after a failure occurs [7]. Thus, the availability defined as Equation (1) is calculated as the rate divided execution time by the total time that includes failure time and execution time. In Equation (2), n means total number of the allocated tasks to resources, and r means the number of task results for the n number of tasks allocated to resources. In this paper, we will implement the log analysis system that can classify resources by the availability and credibility presented in Definitions 1 and 2.

Implementation of Log Analysis System for Desktop Grids

ok

ok

start

Version

Login

joblist completion end

Logout Logout

357

Joblist

failure

ok

task execution

ok

StartWork StartWork

EndWork task completion

Fig. 1. Execution flow of each resource

4 4.1

Log Analysis System Execution Flow of Resources

Before implementing the log analysis system, it is important to understand first the execution flow of each resource to capture the dynamic features of desktop grids. Fig. 1 shows the execution flow of each resource in Korea@Home desktop grid system [5]. In this figure, a resource requests a function to the central server by message communications with an XML document and receives all the data and information needed to execute tasks from the central server. As we can see in Fig. 1, request-response functions include Login, Version, Joblist, StartWork, EndWork, and Logout, and are essentially required for analyzing the execution behavior of each resource. – Login: it is first phase to participate in the task execution of desktop grids. In this phase, an agent program installed in each resource performs the authentication function to check the credentials of the resource. – Version: it queries to the central server to check if there is newer agent program than one already installed in a resource. If so, the update process to reinstall the newer program is performed. – Joblist: it is a procedure to request tasks to be executed by a resource, and the central server provides the job list for the resource as a response. It also has the download function that sends the task program and data files needed for task execution from the central server to the resource. – StartWork: it is a function to inform the central server of task execution once a resource is ready to execute a task, – EndWork: it is a function to inform the central server of task completion after a resource finishes task execution. It also uploads task results to the central server. – Logout: it logs out of desktop grids. The time period between StartWork and EndWork can be seen as actual execution time of a task. As long as there is no failure during the time period,

358

J.-M. Gil, M. Kim, and U.-S. Song

<request> <startwork> <workid>

<request> <endwork> <workid> <worktime> (b) EndWork

(a) StartWork

Fig. 2. Example of XML messages Log Analysis System for Desktop Grids LogManager

Raw Log File

Log Data

LogParser

Filtered Log Data

XMLParser

LogAnayser

Information for Each Resource

Availability & Credibility

DBManager

Log DB

Fig. 3. Log analysis system for desktop grids

pure execution time is same to actual execution time. However, if there are failures, actual execution time is obtained by adding the pure execution time to the failure time. Fig. 2 shows an example of XML request messages for StartWork and EndWork. The XML messages of these types are exchanged between a resource and the central server for each request-response. In this paper, we extract the execution behavior of resources by parsing log information in the XML messages and analyze the availability and credibility of resources with the execution behavior. 4.2

Detail Modules of Log Analysis System

The log analysis system consists of five modules as shown in Fig. 3. The detailed function of each module is described as follows: – LogManager: it receives primitive log data from a desktop grid system as a type of files and performs preprocessing for the log data. – LogParser: it takes the responsibility of filtering unnecessary log from primitive log data so as to obtain only information relevant to the execution behavior of resources.

Implementation of Log Analysis System for Desktop Grids

359

Table 1. RESTful-based open APIs Name getAgent(agentid) getAvailability(agentid) getCredibility(agentid) getAgentByTask(taskid) getTaskByAgent(agentid) getCompletedTaskList() getCompletedTaskTime()

Description Get the information of a specific agent Get the availability of a specific agent Get the credibility of a specific agent Get agents having executed a specific task Get all of the tasks executed by a specific agent Get a list of the completed tasks Get total execution time of the completed tasks

Fig. 4. Example of log data fetched by open APIs

– XMLParser: it extracts meaningful information for each log through an XML parsing procedure and captures detailed information relevant to the execution behavior of resources. – LogAnalyzer: it takes responsibility of analyzing actual availability and credibility, utilizing the detailed information extracted by XMLParser. Besides, it takes responsibility of cooperating with DBManager to store and update the analysis results of availability and credibility in databases. – DBManager: it stores, updates, and retrieves the availability and credibility information of resources to/from databases. The log analysis system in this paper has been implemented with JDK 1.6.0, and JDOM 1.1 [8] has been used as an XML parser to extract task information from request-response XML messages. Meanwhile, we have used MySQL 5.5 [9] as the database to store and manage all the information relevant to the execution behavior of resources, including availability and credibility. The log analysis system also supports RESTful-based open APIs that allows users to access log analysis results more easily via web browsers. The users can simply acquire log analysis results using the open APIs presented in Table 1. These open APIs have been implemented by the Restlet [10] that supports REST web services based on Java. Fig. 4 shows an example of the log data fetched by the implemented open APIs in a web browser.

360

J.-M. Gil, M. Kim, and U.-S. Song

300000

250000 200000 150000 100000

50000

Mon

Tue

Thu

Wen

Fri

Sun

Sat

18:00:00

6:00:00

12:00:00

0:00:00

18:00:00

6:00:00

12:00:00

0:00:00

18:00:00

6:00:00

12:00:00

0:00:00

18:00:00

6:00:00

12:00:00

0:00:00

18:00:00

6:00:00

12:00:00

0:00:00

18:00:00

6:00:00

12:00:00

0:00:00

18:00:00

6:00:00

0:00:00

12:00:00

0

Fig. 5. Average execution time in a week basis

5 5.1

Log Analysis Results and Their Application to Task Scheduling Log Analysis Results

In this paper, log analysis results have been produced based on the primitive log data of the Korea@Home system which is the only desktop grid system in Korea. The used log data have been collected for one month (Dec. 2007). Fig. 5 shows the log analysis results acquired by analyzing these log data. The log analysis results are depicted in this figure as an average execution time in a week basis. In this figure, x- and y-axis represents the time in a unit of hour and total execution time by all the resources for one hour, respectively. On analyzing the results of this figure, we can find that execution behavior in weekday is different from that in weekend. Although execution behavior per day in weekday is not exactly same to each other, it has almost similar patterns; i.e., task execution time gradually increases from 9:00 to 12:00, and high execution time is maintained without any sharply fluctuation between 12:00 and 15:00. The task execution time from 15:00 to 24:00 steadily decreases. The task execution time from 0:00 and 9:00 is almost flat, but comparatively lower than that of other times. In fact, this tendency is due to the work style of white-colored workers. We can also observe from this figure that execution behavior in weekend is different from that of weekday; i.e., task execution time is kept at certain value or decreases steadily regardless of anytime for each day. Thus, we can see from this figure that different task scheduling policies should be applied according to weekday and weekend. Meanwhile, the availability and credibility of resources can have an uplifting effect on the performance of task scheduling in desktop grids. For example, resources with relatively high availability and credibility can execute more tasks for a given time period and moreover, return much credible task results to the central server. On the contrary, it is highly probable that resources with low availability and credibility will fail to return credible task results within a given time period. Therefore, if tasks are properly allocated according to the availability and credibility, resources will be efficiently utilized due to the reduction of task failures, resulting in shortening turnaround time for entire tasks.

1.0

1.0

0.8

0.8

0.6

0.6

Credibility

Credibility

Implementation of Log Analysis System for Desktop Grids

0.4

0.4

0.2

0.2

0

0 0

361

0.2

0.4

0.6

Availability

0.8

1.0

(a) Dec. 2007

0

0.2

0.4

0.6

0.8

1.0

Availability

(b) Jan. 2008

Fig. 6. Distribution of availability and credibility Table 2. Four sets of Resource Groups by Availability and Credibility Group R1 R2 R3 R4

Availability High High Low Low

Credibility High Low High Low

Fig. 6(a) shows the distribution of availability and credibility for the total 3,390 resources participating in Korea@Home system during one month (Dec. 2007). During this period, average availability and credibility of resources are 50.62% and 62.65%, respectively. On analyzing the distribution of availability and credibility in this figure, we can see that all resources are widely scattered and resources with more than average availability and credibility have much contributed to task execution than the others. 5.2

Application to Resource Group-Based Task Scheduling

Here, we apply the execution behavior of weekday and weekend to task scheduling. First, a set of resource groups is constructed with the availability and credibility of resources. Table 2 shows four sets of resource groups, R1 , R2 , R3 , and ˜ and credibility R4 . We used two kinds of thresholds, availability threshold (A) ˜ threshold (C) as a criterion to make resource groups. We determined the values ˜ considering the execution behavior presented in Fig. 5. Table 3 of A˜ and C, shows the values of A˜ and C˜ for weekday and weekend, respectively. Let us consider task scheduling based on the resource groups that are classified by the execution behavior of resources. In this paper, we use the task scheduling policy, in which a task is first allocated to a resource in the group with the highest availability and credibility (i.e., R1 ). If there are no more resources in R1 , the task is allocated to a resource in R2 . By a way of this process, tasks are allocated to resources in each group in order of R1 , R2 , R3 , and R4 .

362

J.-M. Gil, M. Kim, and U.-S. Song ˜ and C˜ Table 3. Values of A (a) weekday Time zone A˜ 00:00∼09:00 0.8 09:00∼12:00 0.65 12:00∼15:00 0.5 15:00∼24:00 0.6

˜ C 0.8 0.7 0.6 0.7

(b) weekend ˜ Time zone A 00:00∼09:00 0.8 09:00∼12:00 0.8 12:00∼15:00 0.8 15:00∼24:00 0.8

30000

60000

Resource Group-based Scheduling Existing Scheduling

Resource Group-based Scheduling Existing Scheduling

50000

Number of Used Resources

Turnaround Time (min.)

25000

20000

15000

10000

5000

0

C˜ 0.8 0.8 0.8 0.8

1000

2000

3000

4000

5000

6000

7000

8000

Number of Tasks

(a) Turnaround time

9000 10000

40000

30000

20000

10000

0

1000

2000

3000

4000

5000

6000

7000

8000

9000 10000

Number of Tasks

(b) Used resources

Fig. 7. Comparison of scheduling performance

We conducted the simulation to evaluate the performance of desktop grids when the execution behavior of resources is applied to task scheduling. The log data of Jan. 2008, that are different from ones that have been used to extract the availability and credibility of resources, were used in the simulation (see Fig. 6(b)). For comparison, the existing task scheduling is also simulated without any consideration for the execution behavior. We used turnaround time and the number of used resources as performance measure. Fig. 7 shows the average turnaround time and the average number of used resources when 100∼1,000 tasks are used. From the results presented in Fig. 7(a), we can observe that the resource group-based task scheduling has much lower turnaround time than the existing task scheduling, regardless of the number of tasks. These results are caused from the use of different criterion on availability and credibility for each time zone. Meanwhile, as the existing task scheduling does not consider the execution behavior of resources, it allocates tasks to resources regardless of the execution environment of desktop grids. Thus, if resources have high availability and credibility, the return of task results can be fast. Otherwise, the fast return of task results can be hardly expected. As a result, the resource group-based task scheduling will obtain faster turnaround time by allocating tasks to resources according to the execution behavior of resources, compared to the existing task scheduling.

Implementation of Log Analysis System for Desktop Grids

363

The results of Fig. 7(b) show that the resource group-based task scheduling can complete entire tasks in spite of using much less resources, compared to the existing task scheduling. This results from the fact that our task scheduling can allocate tasks to the resources with as high availability and credibility as possible by analyzing the execution behavior of resources and thus task failures can be handled at a low level. On the contrary, in the existing task scheduling, tasks can be allocated to the improper resources for the execution environment of each time zone and hence the chances of task failures will increase dramatically. Once a task fails, it should be allocated to other resources again. Task failures will induce the waste of resources, resulting to low resource utilization. In conclusion, it is seen that fast turnaround time can be achieved with less resources if task scheduling is operated with the execution behavior of resources extracted from the log analysis system of this paper.

6

Conclusion

In this paper, we implemented a log analysis system which can extract the execution behavior of resources in desktop grid systems. Actual log data in Korea@Home systems were used to analysis the execution behavior. We observed from analysis results that there is a large difference between execution behaviors in weekday and weekend for each time zone. In the simulation to verify the applicability of the implemented log analysis system, resource group-based task scheduling was applied to actual desktop grid systems after grouping resources according to availability and credibility. Simulation results indicated that this scheduling can achieve faster turnaround time in spite of using less resources than the existing one. Therefore, we do expect that the log analysis system can be usefully utilized in improving the performance of new task scheduling policies or existing ones by analyzing task execution environment in advance before these policies are directly applied to actual desktop grid systems.

References 1. Kacsuk, P., Lovas, R., N´emeth, Z.: Distributed and Parallel Systems in Focus: Desktop Grid Computing. Springer, Heidelberg (2008) 2. Al-Azzoni, I., Down, D.G.: Dynamic Scheduling for Heterogeneous Desktop Grids. Journal of Parallel and Distributed Computing 70(12), 1231–1240 (2010) 3. BOINC: Berkeley Open Infrastructure for Network Computing, http://boinc.berkeley.edu 4. XtremWeb, http://www.xtremweb.net 5. Korea@Home, http://www.koreaathome.org 6. Kando, D., Fedaka, G., Cappello, F., Chien, A.A., Casanova, H.: Characterizing Resource Availability in Enterprise Desktop Grids. Future Generation Computer Systems 23, 888–903 (2007) 7. Shooman, M.L.: Reliability of Computer Sysytems and Networks. John Wiley & Sons Inc. (2002) 8. JDOM 1.1, http://www.jdom.org 9. MySQL 5.5, http://www.mysql.com 10. Lightweight REST framework for Java, http://www.restlet.org

FM Subcarrier Multiplexing Using Multitone Modulation for Optical Coherent Communications Hae Geun Kim and Ihn-Han Bae School of Computer and Information Communication Catholic University of Daegu 330 Kumrak-ri, Hayang-up, Kyungsan-si, 712-702, Korea [email protected] Abstract. FM (Frequency Modulation) subcarrier multiplexing using multitone modulation for optical coherent communication system have been firstly introduced to maximize the bandwidth utilization, where the optimized number of bits allocated in each subcarrier is defined. Each subcarrier transmitting different number of data bits is modulated by the M-QAM. When we implement FDM to divide the given bandwidth within FM channel, the spectrum of the subcarrier signals in the given FM bandwidth is the quadratic noise. We perform the numerical analysis the multicarrier modulation for only 4 subcarriers to prove applicability to FM subcarrier modulation, so the bit loading algorithm is not used. Firstly, we calculate the SNRs to obtain the BER of 10-9 for the cases of 4, 16, 64, 128-QAM at the given 4 subcarrier frequencies, 0.1, 0.15, 0,2, and 0.25 MHz. Secondly, we make the optimum choice the group composed of 20.7 dB for fsc=0.1 MHz, 18.4 dB for fsc=0.15 MHz, 20.3 dB for fsc=0.20 MHz, and 19.1 dB for fsc=0.25 MHz, where the SNR difference is lass than 2.3 dB. Hence the optimum modulators of FM subcarrier modulation are QPSK for fsc=0.1 MHz, 16-QAM for fsc=0.15 MHz, and 64QAM for both fsc=0.20 MHz and fsc=0.25 MHz. Keywords: Multitone Modulation, Narrowband Frequency Modulation Subcarrier Modulation, Optical Coherent Transmission.

1

Introduction

Frequency division multiplexing (FDM) with multitone modulation have been achieved the maximized reliable data rates over bandlimited communication channels [1]. Mostly, multitone modulation (MC) has been used in improving the performance of high bit rate digital subscriber lines. The main advantage of Multitone modulation maximizes the bandwidth utilization to optimize the number of bits allocated in each subcarrier. Besides, the MC employs bandwidth efficient modulation schemes such as QAM (quadrature amplitude modulation) [3][4]. A coherent FM-SCM optical communication system has been one of the essential solutions to transmit high speed optical data, because of recent development of high T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 364–370, 2011. © Springer-Verlag Berlin Heidelberg 2011

FM Subcarrier Multiplexing Using Multitone Modulation

365

speed electronics. Multichannel data can be also handled in the radio frequency (RF) domain with one optical carrier in a coherent SCM system [2]. In a FM SCM receiver, the noise characteristic is not uniform, so we can not transmit the same data bits for all subcarriers. Hence, we need to decide the optimum number of data bits for each modulator [5]. In this paper, we propose the FM-SCM combined with MC where FM subcarrier multiplexing using multitone modulation for an optical coherent communication system. In multicarrier modulation, the number of transferable bits by each modulator is proportion to SNR of the channel, so the number of bits depending on the power of FM quadratic noise. Consequently, the BERs of all modulators in the FM band become equal. In chapter 2, we describe the FM subcarrier multiplexing using multitone modulation system and the spectrum of the subcarrier signals in the given FM bandwidth with the FM quadratic noise. In chapter 3, numerical results of the noise variance within the FM bandwidth of the SCM system is performed, then

2

FM Subcarrier Multiplexing Using Multitone Modulation

Multicarrier modulation with N subcarrier frequencies is shown in Fig. 1 [1]. Serial input bits are converted to parallel, then divided into n groups for FDM with subcarrier frequencies, fsc1, fsc2,…… fscN. When we let the frequency difference between fscN and fscN-1 be Δf, the total bandwidth is nΔf.

m1 bits Serial-toInput bits

m2 bits

fsc1 fsc2 Σ

Parallel

To FM Modulator

Modulators Converter mn bits

f

Fig. 1. Multicarrier modulation with N subcarrier frequencies

The outputs of modulators are summed for FM modulation, where each modulator is transmitting different bits depend on the noise power within the narrowband FM band. Because the spectrum of narrowband FM is similar to AM, we can multiplex the signal in FDM form as shown in Fig. 2(a). The subcarriers are modulated by the M-QAM, so the spectrum of each data is sinc function. When we implement FDM to divide the given bandwidth within FM channel, the spectrum of the subcarrier signals in the given FM bandwidth with the FM quadratic noise is shown in Fig. 2(b).

366

H.G. Kim and I.-H. Bae

In multicarrier modulation, the number of transferable bits by each modulator is proportion to SNR of the channel. Hence, if we can choose appropriate number of bits depending on the power of FM quadratic noise for assigned band, the BERs of all modulators in the FM band become equal. In Fig. 2(b), we assigned the bits per signal, which is proportion to the noise power. Hence, for example, the number of bits, 8, 6, 4, and 2, per symbol are transmitted with 256-QAM, 64-QAM, 16-QAM, and QPSK, respectively.

fcarrier BPF

fsc1 fsc2 fsc3

fscN

FM Carrier

…. Bandwidth W (a) Amplitude

Bits per Signal

FM quadratic Noise spectrum 8

8

8

8 6

6

… (b)

Fig. 2. The spectrum of the subcarrier signals in the given FM bandwidth with the FM quadratic noise

FM Subcarrier multiplexing using multitone modulation for an optical coherent communication system is shown in Fig. 3 [5]. Here, a single mode fiber is used in transmitting the FM modulated signal. In the transmitter, the output signal of a multicarrier modulator in Fig. 1 with N subcarrier frequencies is fed to an FM modulator, and then the FM signal modulates transmitting laser signal. Fig. 3 depicts the transmitter and receiver of the SCM system using FM as a principal modulation scheme where N SCM with electrically modulated signals are summed and enter into an FM modulator. Then the laser diode signal at λi, where i = 1,…, k, is FM modulated for transmission through the single mode fiber.

FM Subcarrier Multiplexing Using Multitone Modulation

367

At the optical receiver, the input optical signal is combined with the local laser signal. A PD (photodiode) transfers to the microwave signal consisting N channel multicarrier signals. The BPF produces an FM signal which is multiplied by an integration and dumpG detector with subcarrier frequencies, fsc1, fsc2,…… fscN.

G G G G G XG G G G G G G G G

Receiver 1

Multicarrier Transmitter 1 skG G ɅXG

tœ“›Šˆ™™Œ™G

šŠXG s–Šˆ“G skG G ɅXG

v—›Šˆ“G j–œ—“Œ™G

”–‹œ“ˆ›–•G

šŠYG

ΣG UG UG

UG

XG

mtG t–‹œ“ˆ›–™G

•GmŽUGXG

Multicarrier Transmitter k skG G Ʌ’G

tœ“›Šˆ™™Œ™G ”–‹œ“ˆ›–•G

kG UG tG lG UG UG |G tG UG UG G z•Ž“ŒG |G UG t–‹ŒG UG m‰Œ™G G UG UG

šŠuG

mtG kŒ›ŒŠ›–™G

iwmG

wkG

Receiver k

šŠXG s–Šˆ“G skG G Ʌ’G

v—›Šˆ“G j–œ—“Œ™G

šŠYG ΣG

’G mtG t–‹œ“ˆ›–™G

G

wkG

iwmG

tœ“Šˆ™Œ™G kŒ›ŒŠ›–™GXG

mtG kŒ›ŒŠ›–™G

UG UG UG

tœ“Šˆ™Œ™G kŒ›ŒŠ›–™G’G

šŠuG

Fig. 3. Multicarrier transmission using a FM channel with a single mode fiber

An integration and dump detector extracts a single channel SCM signal with the noise, After the signal with the bandwidth W. the noise autocorrelation becomes [5]

Rn (τ ) =

An

π

W

2  ω cos ωτ dω = 0

2 An

π

cos Wτ −

4 An

πτ 3

sin Wτ +

2 AnW 2 sin Wτ 4πτ

(1)

where the input signal of baseband filter has the cut-off frequency, W, and the anvelop An.

3

Numerical Results and Analysis of the Noise Calculation

Calculation of the noise variance within the FM bandwidth of the SCM system is performed using (1) which can be described as [5]

T T  E{N c2 } = E   Rn (t − t ) cos(ωsc t ) cos(ωsct )dt dt  0 0 

(2)

where T is the period of the integration-and dump filter. When we assume there are only 4 subcarriers within the FM bandwidth, the noise power with a bandwidth, W, for each subcarrier is clearly defined as shown in Fig.4.

368

H.G. Kim and I.-H. Bae

The calculation results of the noise power of (1) at the stable bandwidth of 0.6 GHz in nW/Hz for the subcarrier frequency at 0.1, 0.15, 0.2, and 0.25 GHz are shown in Fig. 4. The relative gains between 0.1 GHz and 0.15 GHz, 0.2 GHz, and 0.25 GHz are - 2.68, - 4.58, and – 6.57 dB, respectively. Higher subcarrier frequency generates more noise power. So, we can allocate more data bits at the lower subcarrier frequency, typically known as bit loading. S (f) n

-B

-W

Sn(f)

W

0

B

f

f

FM quadratic Noise Calculated within Bandwidth W

fcarrier fsc1=0.1GHz

fsc2=0.15GHz

fsc4=0.25GHz fsc3=0.2GHz

f Calculated Noise Power [nW/Hz]

10.70

Gain [dB] Margin

0

19.83 - 2.68

32.68 - 4.58

48.56 - 6.57

Fig. 4. The calculated noise variance in nW/Hz and the relative gain in dB of the narrowband FM bandwidth

Bit loading divided into two categories: rate-adaptive and margin adaptive is a critical issue in multitone modulation. But, in this paper, since we perform the numerical analysis the multicarrier modulation for only 4 subcarriers to prove applicability, the bit loading algorithm is not used. First, we assume BER is set to 10-9, then the number of data bits for each subcarrier are defined. The bit error probability of M-ary QAM transmitting k bit/symbol within the Gaussian channel is denoted as [6]

pb =

 1 1  1 −  erfc  k M  

(M

3kEb 2 − 1 N0

)

   

where the Eb/N0 is the SNR and the erfc(x) is the complementary error function.

(3)

FM Subcarrier Multiplexing Using Multitone Modulation

369

When we consider the calculation results of [5], the SNRs required to achieve the BER = 10-9 in Figs. 6-9 for the QPSK is 12.5, 14.5, 16.5, and 19.1 dB, respectively. The SNRs to obtain the BERs of 10-9 for the cases of 16, 64, 128-QAM are also calculated in Table 1. Table 1. The BER calculation results to obtain the BER of 10-9 for the cases of 4, 16, 64, 128QAM at the given subcarrier frequencies, 0.1, 0.15, 0,2, and 0.25 MHz

Modulation Scheme QPSK 16-QAM 64-QAM 128-QAM

Subcarrier Frequency 0.1 MHz 12.5 dB 16.3 dB 20.7 dB 25.4 dB

0.15 MHz 14.5 dB 18.4 dB 22.6 dB 27.3 dB

0.2 MHz 16.5 dB 20.3 dB 24.5 dB 29.2 dB

0.25 MHz 19.1 dB 22.9 dB 27.1 dB 31.8 dB

In Table 1, we can make the group which consists of the the SNRs with about the same values in dB for each subcarrier frequency to optimize the number of bits loaded for the 4 modulation schemes. For example, when we group 25.4 dB for fsc=0.1 MHz, 22.6 dB for fsc=0.15 MHz, 24.5 dB for fsc=0.20 MHz, and 22.9 dB for fsc=0.25 MHz, the SNR difference is 2.8 dB. So, this case is still needed to diminish the SNR differences. With our FM channel case, the best choice is the group composed of 20.7 dB for fsc=0.1 MHz, 18.4 dB for fsc=0.15 MHz, 20.3 dB for fsc=0.20 MHz, and 19.1 dB for fsc=0.25 MHz, the SNR difference is lass than 2.3 dB. Hence the optimum modulators of FM subcarrier modulation are QPSK for fsc=0.1 MHz, 16-QAM for fsc=0.15 MHz, and 64-QAM for both fsc=0.20 MHz and fsc=0.25 MHz.

4

Conclusions

We introduce FM subcarrier multiplexing using multitone modulation for optical coherent communication system, where the optimized number of bits allocated in each subcarrier is defined. Multicarrier modulation is usually used in digital subscriber loop and its main advantage is to allocate the number of bits per subscriber based on the corresponding SNR. We successfully prove applicability to FM subcarrier modulation. In this paper we did not use the bit loading algorithm because only 4 channels are considered. We calculate the BERs to obtain the BER of 10-9, then make the optimum choice the number of bits for each channel. In later time, we plan to use the bit loading algorithm for multiple channels more than 100. Acknowledgments. This research was supported by the Research Grants of Catholic University of Daegu in 2011.

370

H.G. Kim and I.-H. Bae

References 1. Bingham, J.: Multicarrier Modulation for Data Transmission: An Idea Whose Time Has Come. IEEE Commu. Mag. 5 (May 5-14,1990) 2. Gross, R., Olshansky, R., Schmidt, M.: Coherent FM-SCM System Using DFB Lasers and a Phase Noise Cancellation Circuit. IEEE Photo. Tech. Let. 2 (January 1990) 3. Vaidyanadan, P., Lin, Y., Phoong, S.: Discrete Multitone Modulation With Principal Component Filter Banks. IEEE Trans. Circuit and Systems 49(10), 1397–1412 (2002) 4. Lee, S., Breyer, F., Randal, S., Cardenas, D., Boom, H., Koonan, A.: Discrete Multitone Modulation for High-Speed Data Transmission over Multimode Fibers using 850-nm VCSEL. In: OFC/NFOEC Conference (February 2009) 5. Kim, H.: High Speed Optical Coherent Transmission System using Narrowband FM Subcarrier Multiplexing. To appear in CCIS Proceedings of AST (September 2011) 6. Sklar, B.: Digital Communications. Prentice-Hall, Inc. (2001)

An Ontology-Based ADL Recognition Method for Smart Homes* Ihn-Han Bae and Hae Geun Kim School of Computer and Information Communication Engineering, Catholic University of Daegu, Gyeongbuk, Rep. of Korea {ihbae,kimhg}@cu.ac.kr Abstract. This paper presents a method for recognition of Activities of Daily Living (ADLs) in smart homes. Recognition of activities of daily living and tracking them can provide unprecedented opportunities for health monitoring and assisted living applications, especially for elderly and people with memory deficits. We present ARoM (ADL Recognition Method) that discovers and monitors patterns of ADLs in sensor equipped smart homes. The ARoM is consists of two components: smart home management monitoring and ADL pattern monitoring. This paper studies on the ontology base and the reasoning that are main parts of ADL pattern monitoring. The ontology base supports the semantic discovery for location, device, environments domains in smart homes. The reasoning system discovers the activity for a person and the appropriate service for a present situation. On detection of significant changes of context, the reasoning is triggered. We design the ontology model for ARoM and implement the prototype system of ARoM by using Protege and Jess tools. Keywords: Activities of daily living, ontology, semantic reasoner, smart home, web ontology language.

1

Introduction

Smart Homes, also known as automated homes, intelligent buildings, integrated home systems or domestics, are a recent design development. Smart homes incorporate common devices that control features of the home. Originally, smart home technology was used to control environmental systems such as lighting and heating, but recently the use of smart technology has developed so that almost any electrical component within the house can be included in the system. Moreover, smart home technology does not simply turn devices on and off, it can monitor the internal environment and the activities that are being undertaken whilst the house is occupied. The result of these modifications to the technology is that a smart home can now monitor the activities of the occupant of a home, independently operate devices in set predefined patterns or independently, as the user requires. Recently, research of intelligent spaces that grasp actions of living people inside and help them has been activated. Based on this liveliness, more and more information has become able to be handled quickly by development of information *

This work was supported by research grants from the Catholic University of Daegu in 2011.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 371–380, 2011. © Springer-Verlag Berlin Heidelberg 2011

372

I.-H. Bae and H.G. Kim

technology. We can accumulate and deal with data of life behaviors or change of surroundings as time series data based on various sensors. There can be many supports based on the time series data acquired by sensor. For example, there are direct helps such as carrying loads when getting home or preparing ingredients in meals, and indirect helps such as batching our daily life and informing or warning in emergencies [1]. In this paper, we present ARoM that discovers and monitors patterns of ADLs in sensor equipped smart homes. The ARoM is consists of two components: smart home management monitoring and ADL pattern monitoring. This paper studies on the ontology base and the reasoning that are main parts of ADL pattern monitoring. The ontology base supports the semantic discovery for location, device, environments domains in smart homes. The reasoning system discovers the activity for a person and the appropriate service for a present situation. On detection of significant changes of context, the reasoning is triggered. We design the ontology model for ARoM and implement the prototype system of ARoM by using Protege and Jess tools. The remainder of this paper is organized as follows. Section 2 reviews OWL ontology and smart homes on related works. Section 3 designs the ontology model for ARoM using by Protege tool. Section 4 presents the ADL reasoning process with Jess, OWL reasoning engine. Section 5 concludes the paper and discusses future work.

2

Related Works

2.1

OWL Ontology

The term "ontology" originates from philosophy and refers to the discipline that deals with existence and the things that exist. Studer et al. [2] merged these two definitions stating that: "An ontology is a formal, explicit specification of a shared conceptualization." In computer science, an ontology is standardized representation of knowledge as a set of concepts within a domain, and the relationships between those concepts. It can be used to reason about the entities within that domain, and may be used to describe the domain [3]. Typical elements of ontologies are: (a) concepts and its attributes; (b) taxonomies to categorize concepts by generalization and specification; (c) relations between concepts; (d) axioms to define statements which are always true; and (e) individuals (or facts) are instances of concepts and its relations [4]. Ontology languages allow users to write explicit, formal conceptualization of domain models. The main requirements are: (a) a well-defined syntax; (b) a welldefined semantics; (c) efficient reasoning support; (d) sufficient expressive power; and (e) convenience of expression. OWL has been designed to meet this need for a Web Ontology Language. OWL is part of the growing stack of W3C (World Wide Web Consortium) recommendations related to the Semantic Web [5]. OWL is divided into three increasingly expressive sub-languages OWL-Lite, OWL-DL and OWL-Full. OWL-DL is most often used because it provides maximum expressiveness.

An Ontology-Based ADL Recognition Method for Smart Homes

2.2

373

Smart Homes

The recent emergence of ubiquitous environments, such as smart homes, has enabled the housekeeping, assistance and monitoring of chronically ill patients, persons with special needs or elderly in their own home environments in order to foster their autonomy in the daily living life by providing the required service when and where needed [6, 7]. By using such technology, we can reduce considerably costs, and alleviate healthcare systems. However, many issues related to this technology were raised such as activity recognition, person identification, assistance and monitoring. Activity recognition in smart environments is gaining increasing interest among researchers in ubiquitous computing and healthcare. Automatic recognition of activities is an important and challenging task. One of the typical applications in healthcare systems is the assistance and monitoring of the ADLs for persons with special needs and elderly to provide them with the appropriate services [8]. Several research works have been done, and several models are proposed to recognize activities for smart environments. B. Chikhaoui et al. [8] propose a new approach based on frequent pattern mining principle to extract frequent patterns in the datasets collected from different sensors disseminated in a smart environments. This approach adopt a hierarchical representation of activities, and generate patterns for each activity model. In order to recognize activities, a mapping function is used between the frequent patterns and the activity models. X. Hong et al. [9] presented homeADL that addressed the issues associated with the heterogeneous nature of storage and distribution of the data within the smart environment community. J. Xu [10] proposed an ontology-based framework to facilitate the automatic composition of appropriate applications. The system composed appropriate services depending upon the available equipments in each individual household automatically. P. A. Valiente and A. Lozano-Tello [11] presented IntelliDomo, an ontologies-based AmI system for the control of domestic systems. It used domestic components and its state values represented as instances of an ontology, and taken advantage of the power of the production rules specified by user in order to change the state of the system components in real time.

3

Ontology Model for ARoM

The use of Ambient Intelligence (AmI) is one of the areas which are rapidly gaining importance in the application of intelligent systems in companies and homes. Ubiquitous computing was suggested that computer and electric systems should be integrated into a physical environment and form part of it. AmI systems have sensors able to collect information in the environment. These systems are usually knowledge-based systems containing the specification of domestic elements and they are based on production rules that represent the system's reasoning elements. A correct way of representing domestic systems and behaving rules is throughout the use of ontologies and production rules based on the concepts established on these ontologies. In this paper, ARoM, the AmI system for smart homes is based on the apartments for elders or persons who live alone in the South Korea. Each apartment provides a means of independent living. A typical layout of an apartment is shown in Fig. 1.

374

I.-H. Bae and H.G. Kim

Position sensors in each room have the ability to monitor the movement of a person through the home environments, contact sensors detect if the window or the door has been opened/closed, and device sensors within electronic devices detect if the electronic devices have been turned on/off. Given the vast amount of information which may be generated from these sensors, it is necessary to discriminate between normal and abnormal situations. S2

S3

S1

Back Terrace S13

S8

Bathroom

Kitchen

S9

Dinngroom

Studyroom

S12

Front Door

S10

Dressroom

S5

S6

S11

Bedroom 1

Livingroom

S7

Bedroom 2

Front Terrace S4 S1-S4: Contact sensors, S5-S10: Position sensors, S11, S12: Temperature sensors, S13: Gas sensor

Fig. 1. Layout of space with various sensors to support independent living

We present ARoM, an ontology-based smart home system for discovers and monitors patterns of ADLs. Fig. 2 shows the overall architecture of ARoM system. A typical smart home environment contains sensors to detect the context and services to be invoked by users. Users interact with the system via user devices such as smart phones. The abstraction component maps the sensor data from various sensors to context information with a mapping function such as fuzzy membership functions. The context information is processed by the context provision component, which is a complex event processing system, producing formatted context events. The service management stores service descriptions, monitors and invokes home services. Service registration and un-registration are also performed by this component. The change detection decides when to trigger activity or service discovery. The ontology base component is introduced to support the semantic discovery for location, device, environments domains in smart homes. The reasoning component discovers the activity for a person and the appropriate service for a present situation. Sometimes the

An Ontology-Based ADL Recognition Method for Smart Homes

375

reasoning component forwards a ADL message to children or close relatives who live in elsewhere through mobile networks. After user identity certification process, the resident or the person who received the message can invoke a home service by sending a command to the service management component.

Fig. 2. System architecture of ARoM.

In order to infer ADL of a smart home from temporal contexts, ARoM is defined in this paper. In ARoM, the semantic web is used to represent the temporal contexts. The context model for the smart home is defined by OWL ontology, and the model is implemented by Protege tool, a graphical editor. Protege is a free, open-source platform that provides a growing user community with a suite of tools to construct domain models and knowledge-based applications with ontologies. At its core, Protege implements a rich set of knowledge-modeling structures and actions that support the creation, visualization, and manipulation of ontologies in various representation formats [12]. The ontology model for ARoM is shown in Fig. 3. It is composed of seven domains: AmIApplication, Device, DeviceStatus, Location, Person, Sensor and SensorStatus to represent the knowledge base in smart home systems. AmIApplication ontology describes the concepts related to the ambient intelligent applications for the smart home. AmIApplication class is consisted of two sunclasses: Activity and Service. Also, Service subclass is consisted of two subclasses: DailyLifeService and SafetyService. Sensor ontology describes the concepts related to various sensors which are installed in the smart home. Sensor class is consisted by three subclasses: ContextSensor, DeviceSensor and Timer. ContextSensor subclass has four individuals that are instances of concepts: Humidity, Light, Noise and Temperature. Fig. 4 shows the structure of AmIApplication class and it's OWL coding.

376

I.-H. Bae and H.G. Kim

Fig. 3. OWL-based context ontology for ARoM

... ... ... ...

Fig. 4. Structure of AmIApplication class and it's OWL coding

An Ontology-Based ADL Recognition Method for Smart Homes

377

...

Fig. 4. (continued)

The OWL structure of ARoM is shown in Fig. 5. Fig. 5(a) and Fig. 5(b) represent the classes and the properties of ARoM, respectively. The OWL model for ARoM has 75 individuals, but Fig. 5(c) represents the individuals of the subclass DailyLifeService of AmIAppication class in ARoM.

Fig. 5. OWL structure for ARoM. Table 1. Object properties in the OWL-based context ontology for ARoM Object property calls has hasDevice hasDuration hasSensing hasSensor hasStatus hasTime locatedIn uses

Domain Person Person Sensor Timer Sensor Device Device Timer Person Person

Range Service Activity Device TimeDuration SensorStatus Sensor DeviceStatus TimeInstance Location Device

Property characteristics Functional Functional InverseFunctional Functional Functional InverseFunctional Functional Functional Functional Functional

378

I.-H. Bae and H.G. Kim

Properties are binary relations on individuals i.e. properties link individuals together. Table 1 shows the object properties in the OWL-based context ontology for ARoM. For example, the property locatedIn might link the domain class Person to the range class Location. Properties can be limited to having a single value i.e. to being functional. They can also be either transitive or symmetric. For example, the inverse of hasDevice is hasSensor.

4

ADL Reasoning

A semantic reasoner, reasoning engine, rules engine, or simply a reasoner, is a piece of software able to infer logical consequences from a set of asserted facts or axioms. The notion of a semantic reasoner generalizes that of an inference engine, by providing a richer set of mechanisms to work with. The inference rules are commonly specified by means of an ontology language, and often a description language. A large number of rule engines work well with Java, and many are available as open source software. Some of the most popular engines include Jess, Algernon and SweetRules. We chose Jess as the semantic reasoner for ADL reasoning.

Fig. 6. OWL/Jess integration structure for ADL reasoning.

Jess (Java Expert System Shell) system [13] consists of a rule base, and an execution engine. Fig. 6 shows the OWL/Jess integration structure for ADL reasoning. Once the ARoM OWL concepts have been represented in Jess, the Jess execution engine can perform inference. As ADL rules fire, new Jess facts are inserted into the fact base. Those facts can be used in further inference. When the inference process completes, these facts must be transformed into OWL knowledge. Fig. 7 shows an example of ADL rule reasoning. Jess system has currently five facts that are represented in Fig. 7(a), Jess engine performs the ADL rule that is represented in Fig. 7(b), and Jess system displays the inference result message for the ADL rule, Fig. 7(c). After the ADL rule performed, new fact is generated by Jess engine, the new fact 'has Bae Sleeping' (the part of rectangular box in Fig. 8(a)) is inserted in fact base, and the short message that is generated by the inference is sent to children or close relatives who live in elsewhere through wireless networks (Fig. 8(b)).

An Ontology-Based ADL Recognition Method for Smart Homes

379

Fig. 7. Example of an ADL rule reasoning.

Fig. 8. Results of an ADL rule reasoning.

5

Conclusions and Future Work

Smart home provide a platform that allow us to monitor and assist with activities of daily living. By monitoring and recording such activities, automatic detection of changes in patterns of behaviors becomes possible. This information can subsequently reveal a decline in health, potential risks in the surrounding environment, and emergency situations.

380

I.-H. Bae and H.G. Kim

In this paper, we present ARoM that discovers and monitors patterns of ADLs in sensor equipped smart homes. The ARoM is consists of two components: smart home management monitoring and ADL pattern monitoring. This paper studies on the ontology base and the reasoning that are main parts of ADL pattern monitoring. The ontology base supports the semantic discovery for location, device, environments domains in smart homes. The reasoning system discovers the activity for a person and the appropriate service for a present situation. On detection of significant changes of context, the reasoning is triggered. We design the ontology model for ARoM and implement the prototype system of ARoM by using Protege and Jess tools. Our future works include studying on an fault tolerant ADL recognition method on the basis of variable precision rough sets and realization on the full version of proposed ARoM.

References [1] Mori, T., Fujii, A., Shimosaka, M., Noguchi, H., Sano, T.: Typical Behavior Patterns Extraction and Anomaly Detection Algorithm Based on Accumulated Home Sensor Data. Future Generation Communication and Networking 2, 12–18 (2007) [2] Studer, R., Benjamins, R., Fensel, D.: Knowledge engineering: Principles and methods. Data & Knowledge Engineering 25, 161–198 (1998) [3] Viinikkala, M.: Ontology in Information Systems, http://www.cs.tut.fi/~kk/webstuff/Ontology.pdf [4] Ay, F.: Context Modeling and Reasoning using Ontologies. University of Technology, Berlin (2007) [5] Antoniou, G., van Harmelen, F.: Web Ontology Language: OWL. In: Handbook on Ontologies in Information Systems, pp. 67–92. Springer, Heidelberg (2003) [6] Qin, W., Shi, Y., Suo, Y.: Ontology-Based Context-Aware Middleware for Smart Space. Tsinghua Science and Technology 2, 707–713 (2007) [7] Rashidi, P., Cook, D., Holder, L., Schmitter-Edgecombe, M.: Discovering Activities to Recognize and Track in a Smart Environment. IEEE Transactions on Knowledge and Data Engineering 23, 527–539 (2010) [8] Chikhaoui, B., Wang, S., Pigot, H.: A Frequent Pattern Mining Approach for ADLs Recognition in Smart Environments. In: International Conference on Advanced Information Networking and Application, pp. 248–255 (2011) [9] Hong, X., Nugent, C.D.: HomeADL for Adaptive ADL Monitoring within Smart Homes. In: Annual International IEEE Engineering in Medicine and Biology Society Conference, pp. 3324–3327 (2008) [10] Xu, J.: Ontology-based Smart Home Solution and Service Composition. In: Int. Conf. of Embedded Software and Systems, pp. 297–304 (2009) [11] Valiente-Rocha, P.A., Lozano-Tello, A.: Ontology-based expert system for home automation controlling. In: 23th International Conference on Industrial, Engineering & Other Applications of Applied Intelligent Systems, pp. 661–670 (2010) [12] Horridge, M.: A Practical Guide to Building OWL Ontologies using Protégé 4 and COODE Tools (2007), http://owl.cs.manchester.ac.uk/tutorials/protegeowltutorial/ resources/ProtegeOWLTutorialP4_v1_1.pdf [13] Mei, J., Bontas, E.P.: Reasoning Paradigms for OWL Ontologies. Technical Report B-0412, Department of Information Science. Peking University, p. 24 (2004)

Analysis of User Preferences for Menu Composition and Functional Icons of E-Book Readers in a Smartphone Environment Mihye Kim1, Joon-Min Gil2, and Kwan-Hee Yoo3 1 Department of Computer Science Education, School of Computer & Information Communication Engineering, Catholic University of Daegu, 330 Hayangeup Gyeonsansi Gyeongbuk, South Korea 3 Department of Computer Education and Information Industrial Engineering, Chungbuk National University, 410 Seongbongro Heungdukgu Cheongju Chungbuk, South Korea {mihyekim,jmgil}@cu.ac.kr, [email protected] 2

Abstract. With the rapid growth of the electronic-book (e-book) market, various types of e-book readers, such as dedicated e-book reading devices and e-book reader applications, have been released. However, the user interfaces of these e-book readers are highly diverse, which is becoming a major problem regarding usability. In this paper, user preferences for the menu composition of an e-book reader are analyzed via a survey, and an ideal menu composition for e-book readers is proposed on the basis of the survey results. The functional icons used in the menus of e-book readers are also analyzed, to investigate the necessity to standardize these icons. Keywords: E-books, E-book readers, Menu composition of e-book readers, Functional icons of e-book readers.

1

Introduction

The electronic book (e-book) industry is growing rapidly, far beyond expectations. The major online booksellers and publishers have already rushed into e-book production, and are leading the growth in the e-book market. Price Waterhouse Coopers (PwC) has forecast that the world e-book industry will reach a total of US$82 billion in 2013 (up from $24 billion in 2009), with the average annual growth of 27.2% [1], [2], and that 90 percent of reading material will be published in digital form by 2020 [3]. According to the Association of American Publishers [4], monthly e-book sales exceeded paper book sales in the United States in February 2010. In May 2011, Amazon.com also announced that they sold more e-books than printed books [5]. The increasing use of tablet PCs, such as iPad, Galaxy Tab, and PlayBook, as well as smartphones, seems to be accelerating this rapid growth in the e-book market. An e-book is a traditional printed book published in digital form “produced on, published by, and readable on computers or other electronic devices” [6, p.164]. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 381–392, 2011. © Springer-Verlag Berlin Heidelberg 2011

382

M. Kim, J.-M. Gil, and K.-H. Yoo

E-books can provide integrated resources with many additional functions, such as images, voices, videos, hyperlinks, searching and navigation, and dictionaries, while performing the traditional role of printed books without time or space limitations, using wired and wireless networks [7]. Users can easily read not only books, but also newspapers, magazines, and blogs on a screen enhanced with rich, riveting digital media technology, in much the same way they read printed books. Despite the remarkable growth of the e-book market, several fundamental issues exist. The major issues relevant to e-books are copyright problems pertaining to digital rights, interoperability problems associated with the current variety of exclusive e-book formats, and some negative psychological perceptions of e-books [3]. In an attempt to resolve the interoperability issue, the International Digital Publishing Forum [1] released the Electronic Publication format (abbreviated EPUB, ePub, Epub, or epub) to establish a free and open e-book standard. Another major ebook issue is usability. This issue is caused by the present diversity of e-book formats, as well as the variety of e-book readers with different menu compositions and functional icons. As many new e-book publishers and sellers are emerging in the ebook market, they provide their own dedicated e-book reading devices or e-book reader applications. The user interfaces of these e-book readers, including the menu composition for the main features and the functional icons displayed on these menus, are configured in various ways. In this paper, user preferences are analyzed for the menu composition and functional icons of the five most commonly used e-book readers: iBook, Kindle, Kyobo eBook, Stanza, and Wattpad. The analysis is based on the main features and functional icons of these e-book readers, with the objective of determining an ideal menu composition for an e-book reader in terms of usability, on the basis of user preferences. The necessity to standardize functional icons is also analyzed but limited to smartphones, which are emerging as universal e-book reader terminals. However, the menu structures of dedicated e-book readers and universal terminals are nearly identical in most cases. The remainder of this paper is organized as follows. Section 2 describes the theoretical background of this study, including the definition of an e-book and overviews of the five e-book readers, together with their main features and functional icons. Section 3 presents and discusses the method and questions of the survey, as well as the survey results. Section 4 summarizes the paper and concludes with possible directions for future research.

2

Theoretical Background

2.1

Definition of an E-book

A variety of definitions of e-books have appeared in the literature. The online Oxford Dictionary of English [8] defines an e-book as “an electronic version of a printed book which can be read on a computer or a specifically designed handheld device.” Rao defines an e-book “as a text in digital form; as a book converted into digital form; as digital reading material; as a book in a computer file format or electronic file of words; or as images with unique identifiers” [9, p. 364]. Gardiner and Musto define

Analysis of User Preferences for Menu Composition and Functional Icons

383

an e-book as follows: “An electronic book (also e-book, ebook, digital book) is a book-length publication in digital form, consisting of text, images, or both, and produced on, published through, and readable on computers or other electronic devices” [6, p. 164], [10]. Nelson [11] defines an e-book as an electronic publication that can be read with the aid of a computer, a personal digital assistant (PDA), a special e-book reader, or even a mobile phone. Similarly, Chrystal defines an e-book as follows: “An e-book is an electronic or digital representation of a given text, which scanned, typed, or programmed (for example, using HTML), having virtual “pages”, that are read using e-book reading software, either on a personal computer (PC), a PDA, a smartphone, or on a dedicated e-book reading device” [3, p. 2]. Accordingly, an e-book can be defined as an electronic or digital book that can be read on various types of devices, including desktop computers, notebooks, netbooks, tablet computers, smartphones, or dedicated devices using e-book reader applications. The term “e-book” refers not only to digital books, but also to digital newspapers, magazines, other digital publications, and all content in digital form, including blogs. 2.2

E-Book Readers

There is a variety of e-book-related terminology, such as e-book content, format, reader software, reader applications, and reading devices. The terms e-book (reader) software, e-book reader application, and e-book reading device are often considered interchangeable with the term e-book reader or simply e-reader. E-book devices can be classified into two types: dedicated e-book terminals and universal terminals [2]. Dedicated e-book terminals include e-reader devices such as the Amazon Kindle [12], the Barnes & Noble Nook [13], and the Sony Reader Touch Edition [14], and they are usually constructed with e-reader software. For example, Kindle is an e-book reading device with an embedded e-book reader application. On the other hand, to read e-books on a universal terminal (such as a PC, a tablet PC, or a smartphone), only an e-book reader application (i.e., software) is required. Thus, e-book sellers and publishers who produce dedicated e-book devices also support several different versions of their e-book applications, to enable users to read e-books on various types of general-purpose terminals. For instance, the version of Kindle that runs on PCs or smartphones is only an e-book application. Here, the terms e-book reading device and e-book reader application (or software) are both e-book readers. There are also many different types of e-book applications for general-purpose terminals. These include iBooks [15], Stanza [16], Wattpad [17], and Kyobo eBook [18]. The terms e-book reader application and e-book reader software are also referred to as e-book readers or simply e-books. In this paper, we analyze user preferences for the menu composition and functional icons of the five most commonly used e-book readers: iBooks, Kindle, Kyobo eBook, Stanza, and Wattpad. Fig. 1 shows screenshots of these e-book readers. Barnes & Noble, one of the world largest publishers, has its own e-book reading device, known as the Nook, but also uses iBooks and Stanza for universal terminals. The Nook is not included in the present analysis, because this study is limited to e-book readers for smartphones. Kyobo is included because it is the largest e-book seller in Korea.

384

M. Kim, J.-M. Gil, and K.-H. Yoo

iBooks. iBooks is an application for downloading and reading e-books on the iPad, iPod Touch, and iPhone by Apple, Inc. [15], [19]. iBooks uses the ePub format, and also supports PDF files via data synchronization with iTunes [19]. Users can download e-books from the iBookstore, as well as through Safari or Apple Mail. The downloaded books will then be available on the bookshelf (library) of the device. Kindle. Kindle is an electronic device for downloading and reading e-books, developed by the Amazon.com subsidiary Lab126. In other words, Kindle is an e-book reader serviced by Amazon.com, which allows users to read e-books, newspapers, magazines, blogs, and other digital media on its website [20], [21]. However, Kindle can be used to read e-books from other sellers or websites, as long as they are compatible. Some file formats such as HTML, JEPG, BMP, and PDF can also be converted into the Kindle format [21]. Amazon has released a number of versions of Kindle, from the initial version of 2007, to Kindle 2 in 2007 and Kindle DX in 2009, to Kindle DX Graphite and Kindle 3 in 2010, with enhanced features and technologies, including a longer-lasting battery, more storage capacity, converting function from text to speech, 3G and Wi-Fi connectivity, and new E-Ink Pearl display [12], [22]. Amazon has also released a number of e-book reader applications for universal terminals, such as Kindle for PCs and many different types of smartphones [22]. To read Kindle e-books on their devices, users must first register for an Amazon account, and then download and install an appropriate Kindle reader application. Kyobo eBook. Kyobo eBook is an e-book reader application developed by Kyobo Book Center Co., LTD, Korea’s largest bookstore chain. It supports desktop computers, notebooks, tablet PCs (iPad and Galaxy Tab), PDAs, Apple iPhones, and Android smartphones, as well as a number of dedicated e-book devices, including Samsung’s SNE-60k and iRiver’s Story [18]. It uses the ePub format, but also supports PDF and TXT file formats. Stanza. Stanza is an e-book application for downloading and reading e-books, and is compatible with a number of devices, including Apple iPad, iPod Touch, and iPhone, as well as desktop computers for Windows and Mac [16], [23]. It supports various e-book formats, including ePub, eReader, MS LIT, Amazon Kindle, Mobipocket, and PalmDoc, as well as general document formats such as HTML, PDF, MS Word, and Rich Text Format [23]. With its rich features, Stanza was the most popular e-book reader for the iPhone prior to the announcement of iBooks. Wattpad. Wattpad is one of the most popular e-book reading applications, and can run on most mobile phones, as well as on desktop computers [24]. Wattpad.com is also an e-book community for writers wishing to publish and share their content for free distribution [24]. After creating a free account, users can download almost all of the e-books for free, and can also upload their content to Wattpad without converting their file formats, and without any compatibility issues [25]. Wattpad is the only e-book reader that supports all major mobile phones [24], including Alcatel, Apple, BenQ-Siemens, BlackBerry, HTC, LG, Motorola, Nokia, O2, Panasonic, Pantech, Sagem, Samsung, Sharp, Sony Ericsson, Toshiba, VK Mobile and others [25].

Analysis of User Preferences P for Menu Composition and Functional Icons

(a) iBook ks

(c) Kyobo eBo

385

(b) Kindle

(d) Stanza

(e) Wattpad

he function menus of five e-book readers displaying e-books Fig. 1. Screenshots of th

2.3

Main Features of E-Book E Readers

Most e-book readers havee various features that allow users to read books m more conveniently. Table 1 show ws the main features of the five e-book readers consideered here. These features are av vailable while reading an e-book, but not all of them are included in any one device. The symbol ‘O’ indicates that an e-book reader supporrts a dicates that it does not. given feature, while ‘X’ ind In general, users can org ganize a library (i.e., bookshelf) of e-books, and can brow wse and search the books in thee library from a list sorted by title, author, or genre. T They can enjoy reading books by y touching or tapping the book icons and buttons with thheir fingers, and by switching between b portrait and landscape view. Users can turn paages by sliding the left or right sides of the screen, or by pressing a button to go forwardd or pping or scrolling up or down through the pages. They can backward, as well as by flip also view and edit book infformation such as the title, authors, and abstract of a boook, and can share their e-books via Twitter or Facebook, or e-mail them to others.

386

M. Kim, J.-M. Gil, and K.-H. Yoo Table 1. Main features of the five e-book readers E-book Readers

Category Scroll

Up and down

Slide

Gestures

Right and left One Center of page short Left of page Touch touch Right of page (Tab) Two short touch One long touch

Book information Share

Page turning Page turning effect Move to a special page Information on current page Font setting Search

Etc.

View book information Edit book information Move to library (bookshelf) Twitter Facebook E-mail By touching or tapping Sliding right and left By flipping through the page Scrolling up and down Like slide transition Flip style like book By table of contents (TOC) By bookmarks Move to beginning By page numbers By page navigator Chapter Page number Percentage Change font size Change font face Change font color Word search within book In built-in dictionary Insert bookmark Insert memo/note/annotation Changing backgound color Lock auto-rotation Change day and light mode

iBooks

Kindle

Kyobo eBook

Stanza

Wattpad

Adjust Page Scrolling brightness Page turning X View function menu Backward page turning Forward page turning Word selection X X X Auto scrolling Select words with Same as Same as Word X a magnifying glass iBooks iBooks selection X X O O O X X X O X O O O O O X O X O X X O X O O X X X X O O O O O O O O O O X O O X X X X X X X O X O O O O O X X O X O O O O X O O O O X X O X X X X O X X X O O O O O X X O O X O O O O O X O X O X O O O O O O X X O O X O X O O O O X O X O O X O X O O O O X O O X O X O X O O O X X O O X X X O O O X

X

X

Most e-books automatically save the place where a user left off, so that the user can easily resume reading from that place later. E-book readers usually provide movement and navigation features that enable users to go to a particular page in a book using the table of contents, bookmarks, or a page navigator (i.e., page scroll bar) located at the bottom of the screen, or by entering a specific page number. Users can look up information on the current page, such as the chapter the current page belongs to, the page number, and the current page location expressed as a percentage. They can also change the text size and the size, face, and color of the font. Most e-book readers also provide search features that allow users to find a specific word or phrase in a single book or a whole library, or look up a selected word or phrase in a built-in dictionary or

Analysis of User Preferences for Menu Composition and Functional Icons

387

via Google or Wikipedia. Users can highlight and insert memos/notes/annotations for specific words, phrases, or sentences, and can also bookmark favorite pages. They can control the foreground and background color, and adjust the screen brightness and contrast, as well as change the text and background colors by switching from day mode to night mode, or vice versa. Other features not shown in Table 1 can be found in references [11], [15], [16], [17], and [18].

3

User Preferences for the Menu Composition and Functional Icons of E-Book Readers

3.1

Survey Method

The analysis of user preferences for the menu composition and functional icons of the five e-book readers was conducted via a survey of university students using Google Androidbased smartphones or Apple iPhones. The students who participated in the survey were all majoring in computer education. One month before conducting the survey, we provided the participants with detailed instructions on the five e-book readers, and guided them in the downloading and use of free e-books from each of the e-book reader websites. A total of 25 students participated in the survey. A number of iPhones and iPads were prepared for the students using Android-based smartphones, to give them the opportunity to use iBooks and Stanza (which only support the Apple devices). Among the participants, 52% used Android-based smartphones and 48% used Apple iPhones. 3.2

Design of the Survey Questions

Most e-book readers have a variety of features that combine the traditional advantages of printed books with the additional benefits of digital media. Some functions, such as page turning and word selection, are continuously available while reading a book, whereas other features must be manipulated using a function menu. Accordingly, most e-book readers provide many of their features via a function menu with icons. This menu is displayed at the top and bottom of the page, generally when the user taps or touches the text anywhere around the center of the page. The function menu is typically composed of three lines, one at the top and two at the bottom of the page. The screenshots of Fig. 1 show examples of the function menus for each of the five e-book readers. Because the screen of a smartphone is small, all the features supported by an e-book reader cannot be displayed on a menu. Thus e-book readers normally display only 8 to 10 features on their function menus, represented by icons, while the remaining features are accessed through a settings feature. As Fig. 1 indicates, the features displayed on the function menu vary significantly from one reader to the next. Question 1 therefore targeted which features are most desirable on a function menu, in terms of usability. Fifteen features were selected from among the main features of Table 1. We then listed these 15 features in alphabetical order, and asked each respondent to choose 10 that would be most useful to him/her as menu items. Some of the main features listed in Table 1, such as “slide,” “touch,” and “insert annotation after selecting a word or text,” were excluded from the 15 features because they are activated directly, rather than by functional icons. Table 2 lists the 15 features presented to the respondents.

388

M. Kim, J.-M. Gil, and K.-H. Yoo Table 2. Fifteen features presented to the respondents in Question 1

Features 1) Adjust brightness 2) Change background color 3) Change day and light mode 4) Change font color 5) Change font face 6) Change font size 7) Display page navigator (i.e., page scroll bar)

Selection(√) □ □ □ □ □ □ □

Features 8) Information on current page 9) Insert bookmark 10) Lock auto-rotation 11) Move to library 12) Search 13) Share (Facebook, Twitter, E-mail) 14) View book information 15) View table of contents

Selection(√) □ □ □ □ □ □ □ □

The second question concerned the icons used in the function menus. As Fig. 1 shows, the features in the menus are denoted by icons. Table 3 lists the icons utilized in the function menus of the five e-book readers. As can be seen, some of the icons are very similar, whereas others are very different from one another. When circumstances compel a user to employ more than one e-book reader, different functional icons for the same feature can be confusing and inconvenient unless the icons clearly indicate the feature they represent. However, the meanings of some icons are not readily apparent from their images or shapes. This is why we listed the e-book reader icons together with their corresponding features, and then asked the respondents to choose the icon that best represents each feature.

Table 3. Functional icons used in the function menus of the five e-book readers (Question 2) E-Book Readers Features Font setting

iBook

Kindle

Change font size Change font face Change font color

Word search in a book Word search in a builtSearch in dictionary

Share

Stanza

X

X X X

(dic.) Auto connects to dic. when select a word

X (TOC)

X X

X X

(dic.)

X

(TOC)

X

(book inf.) X

X

(my lib.)

Twitter

X

X

X

Facebook

X

O

X

X

E-mail

X

X

X

X

Insert bookmark Etc

Wattpad

(settings) X

By table of contents View book information Book Edit book information Inf. Move to the library

Kyobo eBook

X O X

Insert memo/note/annotation

(memo)

Set day and night mode

X

X X

(night mode)

(annot.)

X

Analysis of User Preferences for Menu Composition and Functional Icons

3.3

389

Survey Results

Table 4 summarizes the results of Question 1. The features chosen by the respondents are listed in decreasing order of frequency. Under the assumption that a response rate greater than 80% is significnat, we can say that items ~ of Table 4 are the features users preferred for inclusion in a function menu.

① ⑧

Table 4. Questionnaire results on the menu features the respondents selected from Question 1 No

Features

Information on current £ page ¤ Insert bookmark ¥ Search ¦ Adjust brightness § Change font size page navigator ¨ Display (i.e., page scroll bar) © Move to library ª View table of contents

Selection (n)

%

No

Features

Selection (n)

%

25

100% « Lock auto-rotation

17

68%

25 25 23 23

100% 100% 92% 92%

16 13 8 6

64% 52% 32% 24%

2

8%

1 25x10=250

4% -

¬ ­ ® ¯

View book information Share (Facebook, Twitter, …) Change day and light mode Change font face

23

92% ° Change font color

22 21

88% ± Change background color 84% Total

Table 5 shows the features provided in the function menu of each of the five ebook readers, in order of user preference according to Table 4. The symbol ‘∆’ represents a feature supported indirectly via a settings feature in the function menu. As Table 5 indicates, only iBooks provides all of the features that more than 80% of the respondents selected in Question 1 (i.e., features of ~ in Table 4), although it does not support the features “lock auto-rotation,” “view book information,” and “share,” which more than 50% of the respondents preferred (i.e., features of ~ in Table 4). All of the features included in the questionnaire are provided by Stanza in its function menu, either by direct icon or indirect icon (i.e., a settings icon). The menus provided by Kindle, Kyobo eBook, and Wattpad are not in close accord with user preferences, especially in the case of Wattpad.

① ⑧

⑨

Table 5. Features supported by the function menus of the five e-book readers No

Features

£ Information on current page ¤ Insert bookmark ¥ Search

% of selection iBooks 100% O 100% O 100% O

O X O

Kyobo eBook O O X

Kindle

Stanza

Wattpad O X X

¦ Adjust brightness

92%

O

X

X

§ Change font size ¨ Display page navigator (i.e., page scroll bar) © Move to library ª View table of contents « Lock auto-rotation ¬ View book information ­ Share (Facebook, Twitter, E-mail) ® Change day and light mode ¯ Change font face ° Change font color ± Change background color

92%

O

O

∆

O ∆ O By sliding up & down ∆

92%

O

O

O

O

O

88% 84% 68% 64% 52% 32% 24% 8% 4%

O O X X X X O X O

O ∆ X X ∆ X X O X

O O ∆ O X O X X ∆

O O ∆ O ∆ O ∆ ∆ ∆

O X X O O ∆ ∆ ∆ ∆

X ∆

390

M. Kim, J.-M. Gil, and K.-H. Yoo

From the survey results, we can determine an ideal function menu for an e-book reader, namely the iBooks menu with an added settings feature icon, as in Stanza’s function menu. In other words, the ideal menu composition for an e-book reader consists of a total of nine icons: eight for features ~ in Table 4, and one as a settings feature, through which features ~ or others are indirectly supported. Table 6 shows the user preferences for the functional icons used in the five e-book readers. For the settings icon, 80% of the respondents chose that of Stanza, and 20% selected that of Kyobo, which employs the same image as Stanza, with the word “settings” inserted below the icon. Nobody selected the Kindle or Wattpad icons. Among the “move to library” icons, 84% selected those of iBooks and Kyobo, and 12% chose that of Kindle. Only one respondent (4%) selected the icons of Stanza and Wattpad. Regarding the “view book information” icon, 80% chose that of Stanza and 20% chose those of Kyobo and Wattpad. Among the “view table of contents” icons, similar user preferences were noted for those of iBooks and Kyobo: 56% preferred that of iBooks, and 40% chose that of Kyobo. Only one respondent selected the icons of Kindle and Stanza. For the “change font size” icon, there were similar user preferences for three different shapes, among which the icons of iBooks and Kindle were most often selected (by 44%). For the “show amount of reading” icon, 76% of the respondents chose that of Kyobo, 16% chose that of Kindle, and 8% selected the icon of Stanza. Regarding the “insert bookmark” icon, 88% chose that of iBooks and only 12% chose those of Kyobo and Stanza. Nobody preferred the Kindle icon. For the “insert memo/note/annotation” icon, users exhibited similar preferences for those of Stanza and iBooks: 48% chose that of Stanza, and 40% preferred that of iBooks. Only 12% of the respondents selected the Kindle icon. Regarding the “set day and night mode” icon, 76% chose that of Kyobo, and 24% selected that of Stanza.

⑨ ⑮

① ⑧

Table 6. User preferences of functional icons No

¤ ¥ ¦ § ¨ © ª «

E-book n Reader

Feature

Icon

%

Settings



Stanza 20 80%

Move to library View book information



iBooks, 21 84% Kyobo

View TOC



Icon

%

Icon

20%



3

12%



Kyobo, 5 Wattpad

20%

-

 Kyobo settings

Stanza 20 80%

 

iBooks 14 56%



Change font iBooks,  size Kindle Show amount  Kyobo of reading Insert  iBooks bookmark Insert  memo/note/an Stanza notation (annotation) Set day and  Kyobo night mode (night mode)

E-book n Reader

Kindle

5

0%

-

-

4% -

9

36%

Kindle

5 20%

4

16%



Stanza

2

8%

12%



Kindle

0

0%

iBooks 10 40%



Kindle

3 12%

Stanza

-

-

Kindle





Wattpad 0 Stanza, 1 Wattpad

Stanza

Kyobo, Stanza 3

19 76%

0%





(memo)

0



19 76%



Kindle

Kyobo 10 40%

11 44%

12 48%

%

Kindle, 1 Stanza



22 88%

E-book n Reader

6

24%

-

4%

-

Analysis of User Preferences for Menu Composition and Functional Icons

391

As the results of Question 2 indicate, there were several icons that nobody or only a few respondents selected (e.g., Kindle and Wattpad icons of , Kindle, Stanza, and Wattpad icons of , Kindle and Stanza icons of , Stanza icon of , Kyobo, Stanza and Kindle icons of , and Kindle icon of ). This seems to imply that the images of these icons do not clearly represent the corresponding features. Some icons were preferred at similar levels (e.g., the “view TOC” icons of iBook and Kyobo, the “change font size” icons of iBooks, Kindle, and Stanza, and the “insert memo/note/annotation” icons of Stanza and iBooks). In such cases, it seems to be relatively unimportant which of the icons is used in a function menu. We also observed an interesting trend regarding the use of two icon types: (1) icons that display only an image (i.e., icons without words) and (2) icons that display both an image and a word (i.e., a word that indicates the meaning of the icon). For example, for the “view TOC” feature, the iBooks icon is of the first type, and the Kyobo icon is of the second type. For icons whose images clearly represent a feature, the respondents tended to prefer icons without words (e.g., Stanza icons of and ). On the other hand, for icons whose images do not clearly represent a feature, the and ) or respondents preferred icons with words more (e.g., Kyobo icons of similarly (e.g., iBooks and Kyobo icons of ). This seems to suggest that if an icon cannot clearly convey its meaning with only an image, it would be better to add a word to the image to help users understand the exact meaning of the icon. From the results of Question 2, we can also notice that all icons utilized by iBooks adapt to user preferences, but most of the icons used by Kindle and some icons by Stanza and Wattpad are incompatible with user preferences.

②

⑦

⑧

④

4

④

①

⑥

① ⑥

③ ⑨

Conclusion

This paper presented an analysis of user preferences for the menu composition and functional icons of the five most commonly used e-book readers, based on their main features and functional icons. The objective was to suggest an ideal menu composition for an e-book reader on the basis of user preferences. The necessity to standardize functional icons was also analyzed. The analysis of user preferences was conducted via a survey of university students using Google Android-based smartphones or Apple iPhones. A total of 25 students participated in the survey. The survey results indicated that the menu compositions of iBooks and Stanza accommodate user preferences to some extent, but those of Kindle, Kyobo eBooks, and Wattpad are comparatively incompatible with those preferences. From the user preference results on menu composition, we were able to determine an ideal function menu for an e-book reader, namely the iBooks menu with an added settings icon to support the features that are not presented in the function menu. Furthermore, we confirmed that the use of function icons needs to be standardized to some extent. The icons that nobody (or only a few respondents) selected should be replaced with the icons preferred by the majority of the respondents. In such cases, users were actually confused by the icons, and did not understand their meaning. The survey results also suggested that when it is difficult to determine the exact meaning of an icon with only an image, it would be better to add a word to the image. However, considering the small screens used in smartphones, the icons should be as simple as possible.

392

M. Kim, J.-M. Gil, and K.-H. Yoo

This paper explored user preferences for the menu composition and function icons of e-book readers in terms of their usability. Nevertheless, several issues related to usability remain. These are related to the use of actions such as scrolling, sliding, and touching or tapping the e-book screen or icons. These issues will be investigated in future research.

References 1. The International Digital Publishing Forum (IDPE), http://idpf.org/ 2. Kim, W.: The status and prospects of the e-book market. Telecommunication Association 52, 72–77 (2010) 3. Chrystal, R.: The Evolution of e-Books: Technology and Related Issues. Digital Libraries, INFO 653 (2010) 4. Association of American Publishers (AAP), http://www.publishers.org/ 5. Amazon.com: New Release, http://phx.corporate-ir.net/ phoenix.zhtml?ID=1565581&c=176060&p=irol-newsArticle 6. Gardiner, E., Musto, R.G.: The Electronic Book. In: Suarez, M.F., Woudhuysen, H.R. (eds.) The Oxford Companion to the Book. Oxford University Press (2010) 7. Kim, M., Yoo, K.-H., Park, C., Yoo, J.-S.: Development of a Digital Textbook Standard Format Based on XML. In: Kim, T., Adeli, H. (eds.) AST/UCMA/ISA/ACN 2010. LNCS, vol. 6059, pp. 363–377. Springer, Heidelberg (2010) 8. Oxford Dictionaries: “e-book”, http://oxforddictionaries.com/ (accessed on August 2011) 9. Rao, S.: Electronic book technologies: an overview of the present situation. Library Review 53(7), 363–371 (2004) 10. Wikipedia: “E-book”, http://en.wikipedia.org/wiki/E-book (accessed on August 2011) 11. Nelson, M.R.: E-Books in Higher Education: Nearing the End of the Era of Hype? Educause Review 43(2), 40–56 (2008) 12. The Kindle, http://www.amazon.com/ref=gno_logo 13. NOOK Tech Specs – Barnes & Noble, http://www.barnesandnoble.com/ nook/features/techspecs/?cds2Pid=30195 14. Sony Reader Touch Edition, http://ebookstore.sony.com 15. iBooks, http://www.apple.com/ipad/built-in-apps/ibooks.html 16. Stanza, http://www.lexcycle.com/ 17. Wattpad, http://www.wattpad.com/ 18. Kyobo eBook, http://digital.kyobobook.co.kr/digital/guide/ guideMain.ink?guidePage=guide01&guide_menuNo=1 19. Wikipedia: “iBooks”, http://en.wikipedia.org/wiki/IBook 20. Dudley, B.: Kindle hacking, iPod parallels and a chat with the Kindle director. Seattle Times (November 19, 2007), http://blog.seattletimes.nwsource.com/ brierdudley/2007/11/chatting_with_amazons_kindle_d.html 21. Brinlee, D.: What Is a Kindle? AskDeb, http://www.askdeb.com/technology/kindle/what/ 22. Wikipedia: “Amazon Kindle”, http://en.wikipedia.org/wiki/Amazon_Kindle 23. Wikipedia: “Lexcycle Stanza”, http://en.wikipedia.org/wiki/Lexcycle_Stanza 24. Wikipedia: “Wattpad”, http://en.wikipedia.org/wiki/Wattpad 25. Wattpad – Free Mobile E-Book Reader With Free E-Books, Wattpad – Free Mobile EBook Reader With Free E-Books

Dynamic Transmission Target Selection Scheme for Load-Balancing in WSN* Seok-Yeol Heo1, Wan-Jik Lee1, and Won-Yeoul Lee2,** 1

Department of Applied IT & Engineering, Pusan National University of Pusan, Korea 2 Department of Cyber Police & Science Youngsan University of Yangsan, Kyungnam, Korea {syheo,wjlee}@pusan.ac.kr, [email protected]

Abstract. We studied a dynamic transmission target selection scheme for loadbalancing in WSN. The goal is for the energy consumption of packet transmission for all nodes to be nearly same as possible. For a load-balancing, it should be needed to predict the scale of the transmission energy of all nodes. In multihop transmission mode, a fixed transmission path is necessary for the loadbalancing of nodes that are at the same hop distance from the sink node (we call as horizontal hopping transmission), and a variable range transmission mode is needed for the load-balancing of nodes that are at different hop distances from the sink node(we call as vertical hopping transmission). In this paper, we developed a dynamic transmission target selection scheme for horizontalhopping and vertical-hopping transmission. The performance evaluation results show that the energy consumption of all the nodes was nearly the same, and the performance achieved was superior to that of existing load-balancing schemes. Keywords: WSN, dynamic transmission target selection, load-balancing, horizontal-hopping transmission, vertical-hopping transmission.

1

Introduction

It is not possible to derive an accurate load-balancing in WSN. However, it is possible to derive a transmission technology that can almost realize load balancing. We studied a transmission target selection for load-balancing of WSN providing periodic monitoring. We define the load-balancing to be a method that has uniform transmission energy consumption if packets are transmitted without errors from all nodes on the predetermined path. Load-balancing transmission technology involves transmission target nodes selection such that the packet transmission energy of all sensor nodes is the same. We assume that the WSN providing periodic monitoring has a single fixed sink node, and all the sensor nodes periodically transmit sensing data to the sink node. Moreover, the nodes have an ability of switching the transmission mode between multihop and direct. We assume this application area because this is the most widely used WSN. * **

This paper is supported by YoungSan University Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 393–402, 2011. © Springer-Verlag Berlin Heidelberg 2011

394

S.-Y. Heo, W.-J. Lee, and W.-Y. Lee

In a multihop transmission environment, the energy consumption increases if the distance between the nodes and the sink node is small. In a direct transmission environment, the greater the distance between the nodes and the sink node, the higher is the energy consumption. For load-balancing of the WSN, we need to predict the energy consumption. To do this, a sink node must select the transmission target nodes of all nodes in advance. The sink node collects the location and neighbor information of all nodes when the network is initially organized. If the sink node calculates the target nodes of all nodes and delivers the calculation result to all nodes, the transmission target selection for load-balancing is complete. This load-balancing setting has to be reset only when there is a change in the network phase. By doing so, it is possible to reduce energy consumption due to control packet for collecting information of neighbor nodes, and excessive energy waste can be prevented. For load-balancing in multihop-transmission mode, if necessary, more than one node can be selected as the next nodes, and some packets must be transmitted directly to the sink node. That is, multi-path and direct transmission should be appropriately combined for load balancing. This Function for load balancing differs slightly from the legacy routing function. The routing function uses routing information and metric. But in transmission target selection function, packets are forwarded by the predetermined path which consists of the next nodes. And the next nodes can vary dynamically. We call this function as a dynamic transmission target selection function. We propose a dynamic transmission target selection and transmission methods for load balancing. We assessed the performance by comparing the proposed scheme with existing load balancing routing schemes and evaluated whether or not the hot spot problem was solved. Related research for equalizing the energy consumption of sensor networks is discussed in Section 2, and Section 3 presents our proposed scheme. The performance evaluation and the results are given in Section 4, and the conclusion and future work are presented in Section 5.

2

Related Researches

Many routing schemes have been proposed to increase the lifetime of sensor networks [1-6] by load-balancing. However, these schemes improve only partially the weighted energy consumption pattern. The LEACH (Low-Energy Adaptive Clustering Hierarchy) protocol tried to distribute the load of a head node that has excessive energy consumption [7]. However, this scheme cannot solve the imbalance problem of the entire network. The UCS (Unequal Clustering Size) scheme [8] and the energy aware routing scheme [9] proposed to directly solve the energy imbalance problem of sensor networks. The UCS tried to correct the load imbalance using cluster size. Nodes with above-average energy and mobility are moved to predetermined positions and become cluster head nodes. The cluster size is determined by the distance from the sink node. Cluster head nodes that are close to the sink node consume energy faster than those more distant because they receive more multihop transmission requests from distant clusters.

Dynamic Transmission Target Selection Scheme for Load-Balancing in WSN

395

UCS was shown to be about 10%–30% better than the ECS (Equal Clustering Size) scheme. It assumes that the cluster head nodes can be moved to predetermined positions. However, in practice this assumption is unrealistic. In addition, the difference between internal communication costs and communication costs between clusters was not considered. Each protocol has different between-cluster and internal communication costs, so the effectiveness of UCS seems to be low. The Energy Aware Routing scheme selects efficient paths in a multihop transmission environment. The basic idea is that the path that consumes the smallest amount of energy is not always the best path. This scheme instead uses the probability of the candidate paths. An energy-metric-based probability value is allocated to each path. When a packet is received, a path is randomly selected based on its probability value. In this scheme, it is possible to continuously assess and change paths. In sensor networks, this may be either an advantage or a disadvantage. The path determination is carried out at each packet transmission, which is an additional burden. Thus, this routing system cannot avoid unbalanced energy consumption.

3

Dynamic Transmission Target Selection for Load-Balancing

This paper assumes the WSN that all nodes have the variable transmission ranges. In multihop transmission, all nodes have the same transmission range. Moreover, all nodes except 1-hop distance nodes have the variable transmission range because the nodes have to transmit in both multihop and direct transmission mode. The network is divided into areas on the basis of the distance from the sink node as shown in figure 2. The load balancing of nodes in the same area is achieved by equalizing the number of received multihop transmission requests of each node from the previous area. Hence, we define horizontal-hopping transmission, which ensures that all nodes in the same area receive the same number of multihop transmission requests. Horizontal-hopping transmission ensures load balancing for nodes in the same area, but load imbalance with nodes in other areas can occur. Load balancing between different area nodes can be accomplished by a proper combination of direct and multihop transmission. If a distant node carries out direct transmission rather than multihop transmission, its own energy consumption increases but the number of multihop transmissions for nodes close to the sink node decreases. Thus, the transmission energy can be maintained at almost the same level by performing multihop transmission and direct transmission in rotation for the nodes located more than 2-hops from the sink node. We define the Vertical-hopping transmission as the method to have the balancing of energy consumption using multihop and direct transmission in rotation. The sink node executes selection the next nodes for multihop transmission and calculation of the vertical-hopping transmission ratio when network initiation and whenever the network status changes.

396

3.1

S.-Y. Heo, W.-J. Lee, and W.-Y. Lee

Dynamic Transmission Target Selection for Horizontal-Hopping Transmission

Figure 1 shows the horizontal-hopping transmission method. In Fig. 1(a), node 1 in area 1 receives two multihop transmission requests from area 2, and node 2 receives three such requests. This leads to a difference in the energy consumption of nodes 1 and 2. For load balancing, node 3 in area 2 transmits the packets to node 1 and node 2 by turns. We call these nodes as hopping node which has more than 2 multihop transmission target node. A node that is not a hopping node is called a designated node, and a designated node performs multihop transmissions with only one. In Fig. 1(b), node 3 is the hopping node. Hopping node requests in rotation to selected target nodes. In Fig. 1(b), if the nodes in area 2 send 10 packets each, 50 packets are delivered to area 1, and each node in area 1 receives 25 packets respectively.

Sink Node Area 1

Area 2

1

2

3

Sink Node 1

2

Area 1

3

Area 2

Horizontal-Hopping Transmission (a)target selection of legacy scheme

(b)target selection of our scheme

Fig. 1. Horizontal-hopping transmission method

Figure 2 shows our performance evaluation network model. Solid marked nodes are hopping nodes decided by sink node. The nodes connected through lines from the hopping nodes are target nodes of each hopping node. We can perceive from figure 2 that the transmission range of hopping node is practically acceptable. 3.2

Dynamic Target Selection for Vertical-Hopping Transmission

There is a difference in energy consumption between nodes in different areas. To eliminate this difference, we must increase the energy consumption for distant nodes and decrease the energy consumption for nodes close to the sink node. That is, distant nodes should transmit some packets directly to the sink node. This reduces the number of multihop transmissions, decreasing the energy consumption of forwarding nodes. Therefore, a dynamic transmission target for vertical-hopping is the next nodes or the sink node.

Dynamic Transmission Target Selection Scheme for Load-Balancing in WSN

397

100 90 80 70 60 50 40 30 20 10 0 0

10

20

Sink

30

40

50

Sensor Node

60

70

80

90

100

Hopping Node

Fig. 2. An example of target selection of hopping nodes

By vertical-hopping transmission, all nodes except 1-hop distance nodes transmit some packets via multihop transmission and the rest via direct transmission. Figure 3 illustrates vertical-hopping transmission. Sink Node

Area 1

Area 2

1

2

3

Sink Node

1

2

3

Area 1

Area 2

Vertical-Hopping Transmission Fig. 3. Vertical-hopping transmission method

For load balancing through vertical-hopping transmission, we must calculate the direct-to-multihop transmission ratio as we call vertical-hopping ratio for each area. Figure 4 shows the vertical-hopping ratio analysis model. In figure 4, Nm_j and Nm_i are the number of multihop transmissions for the node in area j and area i, respectively. Nd_j and Nd_i are the number of direct transmissions for node in area j and area i, respectively. Nodes in the same area have the same number of transmissions and receptions via horizontal-hopping, i.e., Nm_i1 = Nm_i2 and Nd_i1 = Nd_i2. N is the number of packets for one round per node. Only Nd_j packets are transmitted via direct transmission; the remainder are transmitted via multihop transmission.

398

S.-Y. Heo, W.-J. Lee, and W.-Y. Lee

Node i1:N

multi-hop transmission = N + 2*Nm_i – Nd_j Node j:N

direct transmission = Nd_i

Node i2:N

direct transmission = Nd_j

direct transmission = Nd_i

Fig. 4. Vertical-hopping ratio analysis model

To calculate the vertical-hopping ratio, we need to know the transmission and reception energy of the sensor nodes. We applied the energy consumption pattern developed in [7]. For the calculation of the vertical-hopping ratio, we assumed that all transmission and reception packet sizes are the same. We did not consider reception energy because it cannot be compared with transmission energy. Table 1 shows the parameters for the calculation of the vertical-hopping ratio. Table 1. Parameters for calculation of vertical-hopping ratio parameter K H Np_i Nd_i r D Ei ET_d

meaning hop distance from sink node maximum hop distance number of nodes of area i number of direct transmission of area i diameter of area maximum multihop transmission range after decision of load balancing routing path transmission energy of area i transmission energy when the distance is d

Given the model of figure 4, equation 1, 2, and 3 give the energy consumption of a node in each area. E1 in equation 1 and E2 in equation 2 are the energy consumed by a single node in areas 1 and 2 respectively; Ek in equation 3 is the energy consumed by a single node in area k. H   H   N p _ i − (N d _ i ⋅ N p _ i )  i =1 i =1 E  E1 =  T_D  Np_i    

(1)

Dynamic Transmission Target Selection Scheme for Load-Balancing in WSN H   H   N p _ i − (N d _ i ⋅ N p _ i )  i =2 E E2 =  i = 2 + N d _ 2 ⋅ ET _ 2 r  T _D  Np_i     H H     N p _ i − (N d _ i ⋅ N p _ i )  i=k i =k E  Ek = + N d _ k ⋅ ET _ k⋅r  T_D  Np _i    

399

(2)

(3)

We set E1 = E2= ⋯ = Ek and, using equation 1, 2, and 3, obtain Nd_1, Nd_2, ⋯ , Nd_k. The vertical-hopping ratio is then the number of direct transmissions that equalizes the energy consumption of all the nodes. Equation 4 gives Nd_H, the number of direct transmissions in area H, and using this value, we can find the numbers of the remaining direct transmissions. Nd _ H

4

H   H −1    N p _ 1 −  N p _ i  ⋅ ET _ D +   N d _ i ⋅ N p _ i  ⋅E T _ D i =1   i =2  = N p _ 1 ⋅ (ET _ D − E T _ H ⋅r ) − N p _ H ⋅ ET _ D

(4)

Performance Assessments

We used simulation to assess the performance of the load-balancing routing system. We measured the number of surviving nodes for each round (specific period) and the remaining energy of each area. In addition, we compared our approach with the multihop transmission scheme, the direct transmission scheme, and the clustering scheme. The conditions for the performance assessment are given in Table 2. Table 2. Network model for performance assessment parameter network size number of node diameter of area(r) multihop transmission distance(D) processing energy(Eelec) amplitude energy(ε) initial energy packet size(k)

Meaning/value a quarter of circle with 100 meter radius 100 25 m Max_Tr 2.5 μJ/bit 1.8 μJ/bit/m2 15 kJ 400 bit

Figure 5 shows the number of surviving nodes for each round of our scheme. It can be seen that the energy of the nodes is depleted at nearly the same time. The first energy depletion occurs at about round 2800. The node energies are then depleted rapidly. Not all the nodes are depleted at the same round because of differences in the packet reception energy or differences introduced by approximation in the verticalhopping ratio.

400

S.-Y. Heo, W.-J. Lee, and W.-Y. Lee 100 90

Number of sensors still alive

80 70 60 50 40 30 20 10 0 0

500

1000

1500

2000

2500

3000

3500

4000

Round

Fig. 5. Number of surviving node by each round

% of residual energy

Figure 6 shows the remaining node energy of each area based on the distance from the sink node. The changes in the remaining energy are approximately constant in all the areas, confirming that the load balancing of each area is successful.

Area 4 Area 3 Area 2 Area 1

Round

Fig. 6. Remaining node energy by each area

Figure 7 shows the packet reception rates of the sink node. The packet reception rate is calculated as the total number of packets sent by sensor nodes versus the total number of packets received by the sink node. The packet reception rate drops rapidly if the connection is cut off because of a hot spot. For the multihop scheme, the packet reception rate drops rapidly because hot spots occur: the energy of nodes close to the sink node is quickly depleted. For the direct transmission, the packet reception rate drops because the energy of distant nodes is quickly depleted. The clustering scheme performs better than the multihop and direct transmission schemes, but it is not possible to avoid the hot-spot problem.

Dynamic Transmission Target Selection Scheme for Load-Balancing in WSN

401

The packet reception rate of the hopping routing scheme remains high for a longer period. That is, the network service is provided continuously without hot spots until the energy of each node is depleted. 100

Packet reception rate of sink node

90 80 70

호핑전송 proposed sheme direct trans. 직접전송 multihop trans. 멀티홉전송 clustering 클러스터링

60 50 40 30 20 10 0 0

1000

2000

3000

4000

5000

Round

Fig. 7. Packet reception rate of sink node

Number of sensors still alive

The network lifetime is an important performance indicator for sensor networks. Different applications define the lifetime differently. For periodic monitoring applications, FND (First Node Dies) is the reference point for the lifetime. Figure 8 shows the network lifetime of our scheme, the multihop transmission scheme, the direct transmission scheme, and the clustering scheme. The lifetime of our scheme is about 1.3~2 times longer than those of the other schemes.

proposed sheme Direct transmission Multihop transmission clustering

Round

Fig. 8. Performance comparison by the number of survival nodes

402

S.-Y. Heo, W.-J. Lee, and W.-Y. Lee

The transmission energy of all the sensor nodes is consumed almost equally when our scheme is applied. Moreover, the network lifetime is better than those of other systems, and our scheme nearly eliminates the hot-spot problem.

5

Conclusions and Future Work

We have proposed a dynamic transmission target selection scheme for load-balancing in WSN. Proposed scheme is possible to predict the energy consumption pattern of all the nodes of the network. The energy consumption of all the nodes is nearly equalized via horizontal-hopping transmission and vertical-hopping transmission. An increase of transmission range because of horizontal-hopping transmission can be avoided by selecting the proper hopping node. The vertical-hopping ratio is needed to find the number of direct transmissions in one round. We can get this ratio via the energy consumption model. Simulations showed that our proposed scheme is effective in solving the load imbalance problem.

References 1. Gomez, J., Campbell, A.T.: A case for variable-range transmission power control in wireless multihop networks. In: Proc. of the 23rd International Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM 2004), pp. 1425–1436 (2004) 2. Xu, Y., Heidemann, J., Estrin, D.: Adaptive energy-conserving routing for multihop ad hoc networks. Research Report 527, USC/ISI (2000), http://www.isi.edu/johnh/PAPERS/Xu00a.html 3. Heinzelman, W., Chandrakasan, A., Balakrishnan, H.: An application specific protocol architecture for wireless microsensor networks. IEEE Transactions on Wireless Communications 1(4), 660–670 (2002) 4. Perillo, M., Cheng, Z., Heinzelman, W.B.: On the problem of unbalanced load distribution in wireless sensor networks. In: IEEE GlobeCom Workshops 2004, November 29December 3, pp. 74–79 (2004) 5. Younis, M., Youssef, M., Arisha, K.: Energy aware routing in cluster based sensor networks. In: Proceedings of the 10th IEEE/ACM International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems (MASCOTS 2002), pp. 129–136 (2002); Fort Worth 6. Gao, J., Zhang, L.: Load-Balanced Short-Path Routing in Wireless Networks. IEEE Transactions on Parallel and Distributed Systems 17(4), 377–388 (2006) 7. Heinzelman, W., Chandrakasan, A., Balakrishnan, H.: An application specific protocol architecture for wireless microsensor networks. IEEE Transactions on Wireless Communications 1(4), 660–670 (2002) 8. Soro, S., Heinzelman, W.B.: Prolonging the Lifetime of Wireless Sensor Networks via Unequal Clustering. In: Proceedings of the 5th International Workshop on Algorithms for Wireless, Mobile, Ad Hoc and Sensor Networks (IEEE WMAN 2005), pp. 236–243 (April 2005) 9. Akyildiz, I.F., et al.: Wireless sensor networks: a survey. Computer Networks 38, 393–422 (2002)

Organizing Virtual Research Groups with Light Path Technology Min-Ki Noh1, Won-Hyek Lee1, Seung-Hae Kim1, and Joon-Min Gil2 1

Korea Institute of Science and Technology Information, 245 Daehangno, Yuseong-gu, Daejeon 305-806, Korea {mknoh,livezone,shkim}@kisti.re.kr 2 Catholic University of Daegu, 13-13 Hayang-ro, Hayang-eup, Gyeongsan-si, Gyeongbuk 712-702, Korea [email protected]

Abstract. Advances in network technologies lead not only the advent of new applications but also research form change from the individual research. Research and Development communities for specific cooperative workgroups over LP (Light Path) are required to support dedicated network resources, intelligent network management, and sophisticated access control and monitoring. In this paper, we introduce LP technologies and describe tightly remote cooperative work community, namely VRG (Virtual Research Group). Lastly, we address the function and framework of VRG with LP technologies Keywords: Optical network, Network resource allocation, Light path, Research network.

1

Introduction

Currently, global-scale collaborative researches have a feature that many researchers, organizations, and laboratory equipments require a tight cooperative work. In addition, the technology of network progress has influenced on research methods, types, and applications in mutual relevance. The network connected to each resource is the representative cutting-edge network services that can provide researchers and research groups with international collaboration works. Most researches on network activities has focused on providing the high quality of services (QoS) over networks such as bulk data transmission, resource sharing, real-time remote conferencing, and the remote control of equipment. In particular, the form of cooperative research is changed to the virtual expansion, and thus this research form needs more enhanced security and sensitive applications, sharing the resources between researchers. To support virtual environment, network resources are reserved to the researchers between end sites. As shown in Fig. 1, each site offers the required bandwidth through a dedicated path and thus can provide highquality and high-performance services on congestion-free networks. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 403–411, 2011. © Springer-Verlag Berlin Heidelberg 2011

404

M.-K. Noh et al.

Fig. 1. Optical network diagram creating E2E LP

However, most E2E (End-to-End) LP reservations consume lots of resources by static allocation. This results in considerable waste of resources. Therefore, there is a need of a capacity limitation when available resources are allocated to the E2E LP all at once [1]. In order to solve this problem, network technologies and researches on the virtualization of network resources and dynamic resource allocation have been variously presented in this literature. Virtualization network resources and dynamic resource allocation have been mainly used in such technologies and researches. The E2E LP, which is created over optical network, is suitable for the performance and quality of network required by research demands, but may suffer from the continuous waste of network resources. Therefore, it is necessary to manage E2E LP by efficient resource allocation. Moreover, it would be beneficial to create as newer LPs as possible under limited network resources [2] [3]. In this paper, we suggest the network design for VRG that can make with integrating the set of E2E LP technologies. In addition management of VRG for researchers which separated from Original Domain and support tight cooperative work. The rest of this paper is organized as follows. Section 2 describes the characteristics of E2E LP. In Section 3, we suggest the design of organizing the VRG that can be configured with optimal number of LPs. In this section, we also describe structure and management for VRG, and show the improvement of LP by testing on media transmission and data transmission. Lastly, we conclude the paper along with our plans for future work.

2

Characteristics of E2E LP

Typically, E2E LP is composed of several parts. These parts are connected by network equipments, such as an amplifier, DWDM (Dense Wavelength Division Multiplexing), borders, and cross links. Each of optical equipments makes the path in transport network part and connects to systems (e.g., users, equipments, servers, etc.) via end site network systems. The path established between end systems is LP. As the LP, EOS (Ethernet Over SONET/SDH) can efficiently manage network resources by encapsulation. The path, that is composed of Ethernet switching and optical signal

Organizing Virtual Research Groups with Light Path Technology

405

integration in transport layer, is conversed to Ethernet frame in consumer layer. The LP determines the path through the network composed of several nodes and it has the function of dividing and allocating the bandwidth of STM-N/OC-N (Synchronous Transport Module-N, Optical Carrier-N) units on E2E links. The E2E link configured by the LP manages end-to-end network bandwidth and meets the quality requirement of services such as High Speed Lambda Service (HSLS), On-Demand Communications Circuit (OCC), Optically Extended LAN (OEL), and so on [4]. The LP on circuit level was able to avoid congestion with other traffics and It could minimize the network delay (Serialized Delay) with away from network electronic element. The path constructed with reserved resources makes very stable network environment and provides the high available bandwidth and throughput.

Fig. 2. TCP throughput test on Light Path and IP network

Fig. 2 shows the throughput for LP (red line) of about 870Mbps on 1Gbps links. In this figure, we can see that the resources of more than 87% are available and thus throughput line is extremely stable than IP network (green line). Thus, we can confirm the performance and stability of LP, and the suitability of establishing virtual cooperative environment. Consequently, LP can be configured on distributed and long distance environment. Organizations, researchers, and communities can tightly connect and comprise VRG which close to the local network environment that not only connected stable network but also can provide excellent transmission performance while it is possible.

3

Related Work

We expect that E2E LP technologies will be employed in large-scale optical networks in the near future. To improve the flexibility and efficiency of the transport network resource utilization, virtualization technologies have been widely extended from the computer area to the network area [5]. Based on testing results presented in Section 2, the E2E LP can offer users the fully allocated bandwidth usage of wavelength by isolating resources in optical layer. We can assume that the E2E LP is an ideal service to the users who have a strong requirement on large and stable network resources on virtual research environment.

406

M.-K. Noh et al.

The researchers belonged to a VRG have increasingly used the LP to connect their partners or shared resources. However, network resource allocation and LP creation will meet to network resource limitation. One of the most important issue is the efficient management and allocation of the dedicate network resources and the creation of VRG. In this section, we address the issues on network organization and efficient management. 3.1

Design of Virtual Research Group with E2E LP

First of all, we have to define optical network elements composed of VRG. As shown in Fig. 3(a), VRG is comprised of 3 sites. In order to join in cooperative research community on E2E LP, they need several network elements. Core nodes (N) are connected other core nodes by WDM links. An important factor in optical network is the product of the data rate of single wavelength, and the number of wavelengths of a fiber line [3]. These factors mean that total capacity of fiber increases with the number of Lines and the number of direct connections to neighbor core nodes. In addition, a core node can provide interface to access links and can be extended to transport network layer (see Fig 3(b)).

(a) Links of VRG

(b) Network Components of E2E LP

Fig. 3. Network components of VRG

Fiber lines (L) in Fig. 3(b) are connected to each core node, allocating bandwidth access on the Fiber lines to reserve each LPs. Access Links are connected end sites electronic interface and interface of Multi Service Provision Platform (MSPP) on core node. In addition, each link consists of several Lambdas (λ) and the bandwidth allocated by control plan. Therefore, we can define components of LP which connect to VRG as each lines and lambda with bandwidth. Also, the number of lambdas can be allocated on WDM line. The bandwidth allocated on each lambda means the available bandwidth of total links and total lambda’s capacity.

Organizing Virtual Research Groups with Light Path Technology

407

Fig. 4. Connect to network node with E2E LP

Let us assume that the bandwidth of the access links (B) can serve to end-user. Given the above network diagram and network components, we can formulate the increasing number of the LP based on required bandwidth (see Fig. 4). (1)

In order to maximize the effectiveness of E2E LP classify researcher and organizations involved with the VRG. In order to increase the values of LP, it needs to remove ports or hops on nodes. To solve this problem, we installed the L1/L2 switch to reduce hop counts. By doing this, the total number of LPs assigning to the VRG can be increased. As shown in Fig. 5, the inputs are the physical topology of the above optical network and the number of λs on each node. Each λ can be more reserved than before when it has a bypass. The sum of paths between source (S) and destination (D) should be bigger than the sum of λ or same. |

|

(2)

The value of path number between destination and source is closed to the value “d-1”. |

|

1

(3)

Equation (3) indicates that the destination of the hop count information can be accessed in the same VRG.

408

M.-K. Noh et al.

Fig. 5. Components of optical network for creating LP

The value of latency should be considered and the information of node located in optimal region that establish for research purposes. Lastly, we have to consider capacity and fiber on each node. We can’t create anymore LP on this site if there is no more capacity on a fiber. Thus, we have to increase the number of network components related with the core node (N), it can make more bypaths between source and destination. Consequently, we establish more paths on above optical network. In the previous study, we can know that to maximize the effectiveness of LP classified researcher and organizations involved with VRG. It means that not only can control location of virtual node with minimum wastage WDM links but also provide environment guaranteed network qualities required as the same. In other words, logical configuration for each virtual environment beyond of the existing research in the local domain, considered submitted requirements, and optimize LPs regarded design the network suggested by this paper. • The optical links independently connect to researchers with each collaborative groups and these links connect to VRG than existing local network domain. - Estimate the resource transformed by optical network components - Set the bandwidth with links and wavelengths. - Calculate total of access link connected to VRG with network resource • Regard as same Domain node that is located in a single hop, and we organized the LP network that requires maximum hop is (N-1). • If there are additional maximum (N-1) hops between source and destination. - Choose the location of extend LP to destination. - Reduce hop count of maximum less than (N-1) - Terminate E2E Light Path and form VRG which connected to researchers and organizations with same purpose

Organizing Virtual Research Groups with Light Path Technology

409

Organize flow of Virtual Research Group are as follows.

Fig. 6. Flow chat of organization of Virtual Research Group

3.2

Experimental Install and Performance Test

We conducted performance tested for a network environment. As shown in Fig. 7, we constructed the experimental network services for medical research group over KREONET (Korea National Research and Science Network). In case of medical research group in the network environment, particular network resource is required according to applications types. For example, high-resolution image data transmission application needed available bandwidth, sensitive video conferencing including video stream of patients need an excellent quality network. First, Fig. 7 shows the network diagram designed for medical researchers. Each of organization alternately serves the multiple stream connection for other researcher and other member being a client receiving the stream and image data. The result of test is shown in Fig 7. Above all we have confidence that LP can provide more number of channels and higher resolution image transmission with guaranteed latency and jitter. Since Bytes In/Out amount of data is very similar with every other traffic, in spite of fifth LP, site (2) is created on other network region and connected with fiber channel node via L2 switch. Moreover, network of low value of jitter and RTT is to guarantee real-time multimedia stream transmission with a satisfactory delay and buffer requirements.

410

M.-K. Noh et al.

Fig. 7. Experimental Extension S/W install and result

4

Conclusion and Future Plan

We proposed a new optical network service for planning and allocation of E2E LP links and optical network resource under organizing node and using layer 2 and layer 1 extension switch. The approach efficient allocation and performance evaluation through a node and path increase network capacity for optical network and guarantee the network-level QoS such as latency, jitter, packet loss, etc. Test results indicate that our network design can efficiently allocate E2E LP for researchers, and show the network performance made up with network components. It is designed to be quite flexible in terms of create LP, to be suitable for research group have to be on multi-links [6]. In this paper, we composed the LPs operated on L1/L2 infrastructure through the Lambda networking technique. This technique can be used to compose the improved networks in terms of stability and performance. However, there are some disadvantages associated with restricting direct resource selection by users. To solve these problems, we have the plan to conduct the dynamic resource allocation that can automatically organize LP composition on a middleware through information exchange with resources [7].

References 1. Benjamin, D., Trudel, R., Shew, S.: Optical services over the intelligent optical network. IEEE Communication Magazine 39(9), 73–78 (2001) 2. Rajagopalan, et al.: IP Over Optical Networks: Architectural Aspects. IEEE Commun. Mag. 38(9), 94–102 (2001)

Organizing Virtual Research Groups with Light Path Technology

411

3. Ramaswami, R., Sivarijan, K.N.: Optical networks–a practical perspective. Morgan Kaufmann Publishers, San Francisco (1998) 4. Smit, M.K., Dam, C.: PHASER-based WDM-devices: principles, design, and application. IEEE Journal of Selected Topics in Quantum Electron 2(2), 236–250 (1996) 5. Haque, A., Ho, P.-H., Boutaba, R., Ho, J.: Group shared protection (GSP): a scalable solution for spare capacity reconfiguration in mesh WDM networks. In: Proceedings of the 47th IEEE Global Telecommunications Conference, Dallas, TX, vol. 3, pp. 2029–2035 (2004) 6. Ho, P.-H., Mouftah, H.T.: On optimal diverse routing for shared protection in mesh WDM networks. IEEE Transactions on Reliability 53(6), 216–225 (2004) 7. Ramamurthy, B., Ramakrishnan, A.: Design of virtual private networks (VPNs) over optical wavelength division multiplexed (WDM) networks. SPIE Optical Networks Magazine 3(1) (2002)

Remote Monitoring Information Management System for Preventing Performance Degradation of Database Myung-Ju Kim1, Un-Bai Lee2, and Kwang Sik Chung2,* 1

2

Service Development Team, TELCOIN Co., Ltd. Dept. of Computer Science, Korea National Open University

[email protected], {lub,kchung0825}@knou.ac.kr Abstract. In order to guarantee the stability and reliability of database performance, database monitoring system and program are necessary. The previous commercial database monitoring servers and programs make database overhead and degrade performance of database management system. And there were no alert message or warning message for database administrators in the previous commercial database monitoring system. We propose the remote database monitoring information management system without degrading performance of database. In the proposed system, remote monitoring information database is constructed by collecting monitoring information from each database server and the remote database monitoring information server deliver monitoring information and each database state to administrators. Thus additional overhead of database monitoring does not occur in the proposed system. The proposed system uses a smart phone as administrator terminal and send alert or warning message to administrators. It provides monitoring service without limitation of time and place of administrators. Keywords: database monitoring system, smart phone, monitoring server, database monitoring information.

1

Introduction

In order to manage Database stably, many database monitoring systems were developed and are now installed and used commercially in many fields. In the recent, previous commercial monitoring systems are directly connected with target database systems through sessions, and, in order to collect monitoring information about database, regularly and frequently send queries to the database. In previous database monitoring systems, input/output transactions of database, session of connection and query processing overhead increases and performance of database decreases, as the number of database monitoring session increases. Since terminal with previous database monitoring software only can monitor the target database, monitoring operation place and time is limited. While an administrator does not run database monitoring software, urgent alert or urgent changes can not be monitored by the administrator. *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 412–418, 2011. © Springer-Verlag Berlin Heidelberg 2011

Remote Monitoring Information Management System

413

Thus we propose new database monitoring system that can avoid the database monitoring overhead and monitoring place and time limitation. In the proposed database monitoring system, we use a smart phone as monitoring terminal, support real-time SQL processing state information for database tuning, and avoid performance degradation of database that comes from monitoring queries process.

2

Remote Monitoring Information Management System

In this chapter, we propose remote monitoring information management system architecture and implementation. Figure 1 is the proposed remote monitoring information management system architecture that can guarantee the stability of database performance that can be affected by monitoring query process.

Fig. 1. Remote Monitoring Information Management System Service Flow

In figure 1, the remote monitoring information server is connected with database server 1 ~ N with Java RMI mechanism. The remote monitoring information server protects database servers from malicious and abnormal access with unpermitted IP address and port number to database servers. A smart phone as administrator’s monitoring terminal uses Wifi networks or 3G networks for connection with the remote monitoring information server and can receive a warning message from the remote monitoring information server. And anywhere and anytime a database administrator can monitors state of database through the remote monitoring information server. Figure 2 is software architecture of the remote monitoring information system.

414

M.-J. Kim, U.-B. Lee, and K.S. Chung

Fig. 2. Remote Monitoring Information Management System Architecture

In figure 2, the remote monitoring information server is implemented with JSP and consists of MIP(Monitoring Information Provider), MIS(Monitoring Information Server), and MIA(Monitoring Information Analyst). MIP(Monitoring Information Provider) delivers monitoring information to a smart phone. MIS(Monitoring Information Server) collects monitoring information from database servers and stores collected monitoring information to monitoring information database. MIA(Monitoring Information Analyst) analyzes monitoring information to monitoring information database and decides whether current state is alerted to an administrator or not. MIC(Monitoring Information Client), monitoring information agent, in database servers regularly collects and processes monitoring information, and sends it to MIS. Lastly, on smart phone, MICA(Monitoring Information Client Application) has monitoring information display function and warning message receiving function are implemented with Android platform. MIA regularly analyzes the latest accumulated monitoring information on monitoring information database. If MIA detects abnormal state of database or abnormal monitoring information of database, then MIA makes alert message for an administrator. MIA is implemented with Java programming language and SQL. Measured monitoring information value is ‘S’, base standard value is ‘L’, an number of base standard excess is ‘C’, and delivery message is ‘M’. Figure 3 is an algorithm in which MIA detects base standard excess and makes and saves an alert message table.

Remote Monitoring Information Management System

415

After processing the request of MICA, MIP converts requests result into XML format file and sends it to MICA. MIP that is implemented with JSP, SQL, XML is run on Tomcat. MIP provides database monitoring information search service and alert message search service. Figure 3 and figure 4 are MICA request monitoring information with XML format from MIP.

Fig. 3. Database Network Traffic Request

Fig. 4. Database Session Request

3

Evaluation of the Remote Monitoring Information Management System

We measure database input/output traffic changes on proposed remote monitoring information management system and the previous system according to monitoring requests increase. Table 1 shows that measurement of database input/output traffic with 5 minutes interval changes according to the number of connection. We use Lab128 program for experiments, and, for accurate measurement, repeatedly measure database input/output traffic of the previous system and the proposed database monitoring system after 2 minute 36 seconds from the time of increasement of monitoring connection. Since connection session of Lab128 program is established for database input/output traffic measure, although the number of monitoring connection is 0, there is initiative traffic.

416

M.-J. Kim, U.-B. Lee, and K.S. Chung Table 1. Input/output Traffic Measurement Results Connection #

0

1

2

3

4

5

input (KB/s)

Previous system Proposed system

0.181 (0) 0.159 (0)

0.358 (0.177) 0.168 (0.009)

0.506 (0.148) 0.169 (0.001)

0.644 (0.138) 0.166 (-0.003)

0.774 (0.13) 0.168 (0.002)

0.931 (0.157) 0.167 (-0.001)

output (KB/s)

Previous system Proposed system

2.08 (0) 2.064 (0)

4.254 (2.174) 2.112 (0.048)

6.587 (2.333) 2.170 (0.058)

8.602 (2.015) 2.081 (-0.089)

10.856 (2.254) 2.101 (0.02)

13.337 (2.481) 2.311 (0.21)

In Table 1, database input traffic in the previous system increases by 0.15 KB per second and database output traffic in the previous system increases by 2.25 KB per second. In the previous system, as the number of monitoring connection increases, each connection session to the target database is respectively established which makes additionally input/output traffic. But, in the proposed monitoring information management system, only one connection session to the target database for monitoring information collection is needed. And although the number of administrator and query request increase, request reply traffic does not increase. The reason is that the proposed monitoring information management system needs only connection with the remote monitoring information server. Figure 5 shows the database input traffic changes on proposed remote monitoring information management system and previous system.

Fig. 5. Database Input Traffic Comparison

Remote Monitoring Information Management System

417

Figure 6 shows the database output traffic changes on the proposed remote monitoring information management system and previous system.

Fig. 6. Database Output Traffic Comparison

As the number of database monitoring connection and query increases, database input/output traffic and session management overhead increases in the previous system. But in the proposed remote monitoring information management system, instead of the direct connection with each database server, remote monitoring information server is needed and the number of administrator and connections session managements do not make additional overhead of each database server.

4

Conclusion and Future Work

The previous commercial database monitoring and tuning systems degraded performance of database system. And they limited place and time of administrator’s operation. Lastly they did not send any alert message to administrators while administrators are out of operation rooms. In order to decrease the overhead of database server occurred by database monitoring and alert administrators while out of operation room, we propose and develop the remote monitoring information management system and evaluation it. The remote monitoring information management system has several advantages as follows, from comparison with the previous database monitoring systems. First, since the remote monitoring information management system requests and collects database

418

M.-J. Kim, U.-B. Lee, and K.S. Chung

monitoring information from the remote monitoring information server, it degrades the performance of each database server. Second, since the remote monitoring information management system can use both of a desktop PC and a smart phone as a monitoring terminal of an administrator, there are no limitation for time and place of monitoring operation. Lastly, since the remote monitoring information management system uses a smart phone as a monitoring terminal, instant alert message can be sent to an administrator and seamless database monitoring operation is possible for an administrator. But the remote monitoring information management system can provide only accumulated database monitoring information on the remote monitoring information server to an administrator. And additional cost for developing the remote monitoring information management system is needed. We have a plan to study to avoid time gap between real-time database monitoring information and accumulated database information on the remote monitoring information server.

References 1. 2. 3. 4. 5.

Toad for Oracle, http://www.toad.co.kr/ Lab128, http://www.lab128.com/ Oracle Database, http://www.oracle.com/kr/index.html Android, Android Developers, http://developer.android.com/index.html Jang, S.: Monitoring System using Mobile Device for Database Administrators, Master Thesis, Graduate School of Computer Science and Technology, Korea University (2001) 6. An, T.: A study on establishment and management plan for database integration monitoring system in corporatio, Master Thesis, The Graduate School of International Affairs and Information, Dongguk University (2005)

Noise Reduction Scheme for Precise Indoor Localization Inseok Moon and Won-Kee Hong Dep. Of Information and Communication Engineering, Daegu University Gyeongsan, Gyeongbuk, Rep. of Korea [email protected], [email protected]

Abstract. The indoor localization is very useful and essential technique in several application area of the wireless sensor network. Various noises such as clock offset, clock drift, and other environmental noises are a hindrance to the accurate localization measurement of a mobile node. In this paper, a new indoor localization scheme is introduced that takes the Kalman Filter to reduce the noises and meausre distances between nodes accurately.The experimental results show that the proposed method improves the accuracy of distance measurement by 18%. Keywords: Localization, Distance Measurement, Kalman Filter, SDS-TWR.

1

Introduction

The wireless sensor network, where lots of tiny sensor devices monitor their surrounding environment continuously and in real time, demands the localization information of the sensor device. While the outdoor localization is easily resolved by the GPS [1], the indoor localization gives several challenging issues such as accurate measurement, real time computation and system load [2],[3],[4]. The indoor localization techniques can be categorized into range-based localization and rangefree localization depending on the requirement of distance measurement. In general, range-based localization provides an exact localization because of being based on the distance between nodes. The exact measurement of distance between two nodes is the basis of figuring out the position of a node. The IEEE 802.15.4a CSS (Chirp Spread Spectrum) was produced that is a standard protocol to measure the distance between nodes via wireless RF communication [5]. It allows the information of time to be sampled by the chirp signal. It uses TWR (Two Way Ranging) or SDS-TWR (Symmetric Double Sided Two Way Ranging) algorithm to compute the distance by using the sampled time. The TWR and SDS-TWR are the schemes to reduce the time error due to the time asynchronization between nodes [6]. However, they still suffer from the clock offset, clock drift and other environmental noise [7]. This is the significant problem in the indoor localization that only tolerates error with centimeters. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 419–428, 2011. © Springer-Verlag Berlin Heidelberg 2011

420

I. Moon and W.-K. Hong

In this paper, a new approach exploiting the Kalman Filter is presented to obtain an accurate localization. It is based on the IEEE 802.15.4a protocol and SDS-TWR for distance calculation. The Kalman Filter assumes any measured data on an object contains a probabilistic error and the current state of the object has a linear relation with its previous state. The Kalman Filter has the characteristics that it does not have to hold all the previous data. That is, it needs only current and previous data and to get a distance. The system values of the Kalman Filter are defined in this paper to reflect the noise in the distance measurement. Experiments are performed by comparing measurement of distance and location depending on the application of the proposed Kalman Filter based on the SDS-TWR scheme. The experimental results shows that the proposed method using the Kalman Filter improves the distance measurement accuracy by 18%. The organization of this paper is as follows. Section 2 introduces the related works about localization in WSN. In section 3, the proposed distance measurement scheme using Kalman Filter will be explained. The experimental results and analysis are presented in section 4. Finally, section 5 contains concluding remarks.

2

Related Works

In this section, researches on the localization in the literature are introduced, categorized with the range-based scheme and range-free scheme. Moreover, recent studies on the exact distance measurement in the ToA (Time of Arrival) based localization are described. The range-based scheme is a technique to figure out a node’s position based on the distance from beacon nodes which already know their position. The RF’s inherent characteristics such as time of flight, reception angle of antenna, or signal strength are exploited to get the distance between two nodes. The RSSI (Received Signal Strength Indicator) exploits the characteristics that as the distance is longer, RF’s signal strength will be weaker [8]. However, it has a difficulty to measure a distance exactly because the signal strength is very susceptible to the environmental noise and its fluctuation is very severe. The ToA (Time of Arrival) takes advantage of the time of flight to measure the distance [9],[10]. Basically, it performs the distance calculation with the time stamp received from the other node so that the time synchronization among nodes is assured to get the exact distance. The TDoA (Time Difference of Arrival) obtains the distance using the time difference between the different received signals. It requires time synchronization among beacon nodes and extra hardware cost due to the additional communication device [9],[10]. The AoA (Angle of Arrival) uses the directional angle of received signal and the crossing point between nodes [9],[10] requires at least 2 directional angles to calculate the distance. It has a difficulty in getting exact directional angles due to the multi-path fading. While the range-based scheme is a mathematical approach based on the distance between nodes, the range-free scheme is a parametrical approach to exploit surrounding information around a node such as surrounding images and neighbor nodes to estimate its position. The Centroid [11], APIT [12] and DV-Hop[13] are the

Noise Reduction Scheme for Precise Indoor Localization

421

representative range-free scheme. In the Centroid, there are fixed nodes that are deployed at equal interval. A mobile node gets the information packets, which holds the position coordinates of a fixed node, periodically from surrounding nodes. It calculates its location by averaging all the coordinates collected from the surrounding nodes. It requires high density of fixed nodes deployed to obtain an exact location, which results in high cost. The APIT is a method to figure out rough location of an object using surrounding node. It scans out neighboring nodes within a one-hop radius of a node requesting its location and then picks out three nodes forming the smallest area of triangle which includes the requesting node. The DV-hop calculates a node’s location with the number of hops assuming all the hop distances are the same. 2.1

Ranging Protocol in ToA Based Localization

In general, distance is obtained by multiplying time by velocity. Because the speed of light is a constant parameter, distance is determined by the time taken a message to be transferred between two nodes. That is, the transfer time can be a difference between sending time (t ) and arrival time (t ) of message shown in Fig. 1. Therefore, the distance dAB can be denoted by the following equation: (1)

Fig. 1. The basic distance calculation between node A and node B

However, the equation (1) does not consider the time error due to asynchrony between two nodes. Given that time error is ε, the distance dAB can be denoted by the following equation: (2) This is a one way ranging (OWR) scheme that requires time synchronization between nodes to improve distance accuracy. The two way ranging (TWR) scheme removes the requirement of time synchronization by transferring two types of messages back and forth as shown in Fig. 2. Using TWR, the time difference (tP ) can be obtained like following: 1 2

(3)

422

I. Moon and W.-K. Hong

Fig. 2. The communication flow between node A and node B in the TWR

In equation (3), the time error due to time asynchrony is removed but clock offset error within a node still remains. If the clock offset errors are εA and εB at node A and node B respectively, the time difference (t̃ P ) can be denoted by the following equation: 1 (4) t t t̃ t 1 ε t 1 ε P

A

2

B

Thus, time difference error can be defined by the following equation: 1 t εA t t εB t t̃ P t P (5) 2 1 t P εA t εB B εA 2 t is much larger than t P since it includes the time to prepare a reply message. In general, it takes a few milliseconds, while propagation time is only several nanoseconds. It implies that the error largely depends on the reply time at node B and the difference between clock offsets of node A and node B. In order to reduce the time difference error, Symmetric Double Sided-Two Way Ranging (SDS-TWR) sends a ranging message twice each other as shown in Fig.3. Given with no clock offset, the propagation time (t P ) in SDS-TWR can be denoted by the following equation: t

t 1 4

tP

t

t

A

t

A

t

4t P

B

t

t

(6)

B

When clock offsets are εA and εB , the propogation time (t̃ P ) can be defined by the following equation: t̃

1 4

t

t

A

1

εA

t

t

B

1

εB

(7)

Noise Reduction Scheme for Precise Indoor Localization

423

Fig. 3. The communication flow between node A and node B in SDS-TWR

The time difference error can be derived from the above two equations: t̃ P

tP

1 4

1 2

t tP tA

t tB

A

1 4

t

εA trelpyA

t trelpyB

B

εB

εB

(8)

εA

2t t 2t P t where t B, t B As shown in equation (8), SDS-TWR reduces the time difference error by more than a half of that of TWR. This equation also implies that the error gets reduced as the preparation times and clock offset of each node get converged at some point.

3

Ranging Compensation with Kalman Filter

The Kalman filter is a recursive filter to track along noisy status in a dynamic linear system. The ranging compensation scheme with the Kalman filter designed in this paper removes the error from the distance obtained by the SDS-TWR passing through the predict phase and the correct phase of the Kalman Filter as shown in Fig. 4. The proposed method takes a set of measured distance Z as an input and produces a set of estimated distance X as an output. |

0,1, … , |

1, … ,

424

I. Moon and W.-K. Hong

Fig. 4. Ranging compensation with Kalman Filter

What should be done to advance through the Kalman Filter is to model the localization system which consists of the system model and the measurement model. The system model and measurement model can be defined as follows respectively:

The A is a coefficient that relates the estimated value at the previous step with the one at the current step without considering noise in the system. It is assumed that A is one because the distance should always be consistent in case that the object does not move and there is no noise. In the system model, the is a noise in the system and it has a Gaussian distribution. Given the system noise covariance is , it is defined by the following: ~

0,

,

where N (a, b) is the Gaussian function with mean a and covariance b. In this paper, it can be set to 0. The H is is assumed that there is no noise in the system. Therefore, a coefficient that relates the estimated value at the current step with the measured value at the same step without considering noise in the system. The H is set to one because the measured value should be exactly the same as the estimated value if there is no noise.

Noise Reduction Scheme for Precise Indoor Localization

425

In the measurement model, the is a noise occurred in the real measurement and it has a Gaussian distribution. Given the measurement noise covariance is , it is defined by the following: ~ 0, The measurement noise is assumed to be the difference between the measured data at the current step and the estimated value at the previous step. |

|

where is a measured data at step i and is an estimated data at step i-1. Then, can be defined by the following: the measurement noise covariance |

| |

|

The initial estimated value x and covariance Matrix P is needed. In this paper, it is assumed that x is z and P is 1.

4

Experimental Results and Analysis

In this section, results of experiments are presented and analyzed comparing measurement of distance and location depending on the application of the proposed Kalman Filter based on the SDS-TWR scheme. KF_ST stands for the SDS-TWR with Kalman Filter and ST stands for the SDS-TWR without Kalman Filter. The three beacon nodes which are denoted by A, B, and C and already know their locations and a node which wants to figure out it location are placed in an indoor space as shown in Fig. 5. The real distance between the node and three beacon nodes are 3, 4 and 5 meters respectively.

Fig. 5. Deployment of beacon node (A, B, C) and mobile node used in the experiments for performance evaluation

426

I. Moon and W.-K. Hong

Fig. 6 shows the results of comparing the distance measurement of the KF_ST and then ST when the real distances are 3, 4, and 5 meters respectively. While the measured data by the ST shows lots of fluctuation, those measured by KF_ST converges into a specific value as the number of measurement goes by. It is because the KF_ST can reduce the noise effectively. Table 1 shows the average of 100 distance measurements of the ST and the KF_ST. While the average distance error of ST is 58cm that of KF_ST is shown to be 40cm.

Fig. 6. Distance measurement of ST and KF-ST Table 1. Average measured distance A node[m]

B node[m]

C node[m]

ST

3.55

4.77

5.44

KF_ST

3.27

4.67

5.27

Noise Reduction Scheme for Precise Indoor Localization

427

The node location (x, y) can be obtained by the trilateration using the measured distances from the three beacon nodes. The real node location is (290, 130). Fig. 7 shows the frequency of x-coordinate and y-coordinate of the node location calculated 100 times by distance values measured by the KF_ST and the ST. Fig. 7 (a) shows all the x-coordinates corresponds between 300 and 330, which are very close to the real x-coordinate in case of the KF_ST but the ST has a wide distribution of x-coordinates. Fig. 7 (b) shows all the y-coordinates correspond between 290 and 320, which are very close to the real y-coordinate. Therefore, these results demonstrate the KF_ST provides the position of a node more accurately than the ST. In the KF_ST, the error rate of x-coordinate is 21% and that of y-coordinate is 12%. This is x-coordinate accuracy by 18% and y-coordinate accuracy by 24% lower than that of ST.

Fig. 7. Distribution of measured location

5

Conclusions

The exact measurement of distance between nodes is very important in the localization. The distance measurement scheme through wireless message packet transmission should resolve the time synchronization problem to obtain an exact distance. The TWR and the SDS-TWR are the distance measurement methods to overcome the time error due to clock drift or clock offset of a node. However, the distance error still remains and it is fatal defect in indoor localization application that requires higher location accuracy. In this paper, a method is proposed to perform an exact distance measurement by removing noise using the Kalman Filter. The Kalman Filter is used to remove the noise produced in the SDS-TWR. The performance evaluation is conducted in terms of distance measurement and location measurement using Nanotron NanoLOC. The distance error in the KF_ST that adapts the Kalman Filter to the SDS_TWR is 18% lower than that in the ST that does not use the Kalman Filter. Consequently, the coordinates calculated by the KF_ST is shown to be deployed densely around the real coordinates.

428

I. Moon and W.-K. Hong

Acknowledgments. This research was supported by the Daegu University Research Grant, 2010.

References 1. Per, E., Todd, W., Sam, P., Changdon, K., Yi-Chung, C., Yeou-Jyh, T.: Wide AreaAugmentation of the Global Positioning System. Proceedings of the IEEE (1996) 2. Morellu, C., Nicoli, M., Rampa, V., Spagnolini, U., Alippi, C.: Particle Filters for rssbased localization in wireless sensor networks an experimental study. In: ICASSP, pp. 957–960 (2006) 3. Hyuntae, C., Yeonsu, J., Hyunsung, J., Ingu, P., Yunju, B.: Precision Time Synchronization System over Wireless Networks for TDOA-based Real Time Location Systems. Journal of Korean Information and Communications Society 34(1) (2009) 4. Ahmad, H.: Application of Channel Modeling for Indoor Localization Using TOA and RSS. Worcester polytechnic institute in partial fulfillment of the requirements for the Degree of Doctor of Philosophy (2006) 5. Hach, R.: Symmetric double sided two-way ranging. IEEE 802.15.4a standard, doc. IEEE P.802.15-05-0334-00-004a (2005) 6. Jiang, Y., Leung, V.: An Asymmetric Double Sided Two-Way Ranging for Crystal Offset. In: Int’l Symposium on Signals, Systesms and Electronics (ISSSE 2007), pp. 525–528 (2007) 7. Heidarian, F., Schmaltz, J., Vaandrager, F.: Analysis of a Clock Synchronization Protocol for Wireless Sensor Networks. In: Cavalcanti, A., Dams, D.R. (eds.) FM 2009. LNCS, vol. 5850, pp. 516–531. Springer, Heidelberg (2009) 8. Bahl, P., Padmanabahan, V.: RADAR: An In-Building RF-based User Location and Tracking System. In: Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies, vol. 2, pp. 775–784 (2000) 9. Boukerche, A., Oliceira, H.A.B., Nakamura, E.F., Loureiro, A.A.F.: Localization systems for wireless sensor networks. IEEE Wireless Communications 14, 6–12 (2007) 10. Sayed, A.H., Tarighat, Khajehnouri, N.: Network-based wireless location: challenges faced in developing techniques for accurate wireless location information. IEEE Signal Processing Magazine 22, 24–40 (2005) 11. He, T., Huang, C., Blum, B.M., Stankovic, J.A., Abdelzaher, T.F.: Range-Free Localization Schemes in Large Scale Sensor Networks. In: Mobicom 2003, pp. 81–95 (2003) 12. Roxin, A., Gaber, J., Wack, M., Nait-Sidi-Moh.: Survey of Wireless Geolocation Technoques. In: Globecom Workshops 2007 IEEE, pp. 1–9 (2007) 13. Niculescu, D., Nath, B.: Ad-Hoc Positioning Systems (APS). In: IEEE GLOBECOM 2001, pp. 2926–2931 (2001) 14. Kleinbauer, R.: Kalman Filtering Implementation with Matlab. Study Report in the Field of Study, Geodesy and Geoinformatics at Universitat Stuttgart

Development of a Korean Language-Based Augmentative and Alternative Communication Application Chang-Geol Kim1, Soo-Won Kwak2, Ryu Juang Tak1, and Byung-Seop Song1 1

Deptartment of Rehabilitation Science & Technology, 2 School of Electronics, Daegu University, Jillyang-eup, Gyeongsan-si, Gyeongbuk, Korea {chang014,jryu,bssong}@daegu.ac.kr, [email protected]

Abstract. Communication is an essential element of human interaction with a community. People with communication disorders that would otherwise impede this interaction can interface with their communities via augmentative and alternative communication. The Korean government supports these disabled people with augmentative and alternative communication devices. However, most devices require high cost and have only simple functions. Furthermore, they require space for storage. As an effort to relieve the difficulties of using such devices, this study develops an augmentative and alternative communication application that can be mounted on widely spreading smartphones and tablet PCs. Keywords: Augmentative and Alternative Communication, Tablet PC, Application

1

Introduction

Communication is one of the most fundamental elements by which a person contacts others and lives as a social creature. Lack of communication results in restricted social interaction, emotional withdrawal, anxiety, and frustration. Furthermore, it severely restricts the ability to learn new things [1-3]. For those who have communication disorders, a supplementary approach called Augmentative and Alternative Communication (AAC) has been designed to assist and improve independent communication in every situation. An AAC system consists of symbols, assistive aids, techniques, and strategies. Symbols include body languages, pictures, gestures, facial expressions, paintings, words, line drawings, and Blissymbols. Assistive aids include physical tools such as communication boards, communication books, and communication devices used for exchanging messages. Techniques such as direct selection, scanning, and encoding determine how to deliver messages. Strategies are plans to efficiently utilize the symbols, aids, and techniques to improve communication [1]. T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 429–436, 2011. © Springer-Verlag Berlin Heidelberg 2011

430

C.-G. Kim et al.

Another essential element of prompt interaction with a community in modern society is access to information, i.e. information technology equipment. People utilize information technology to learn skills for daily life and economic activities and to enjoy society and culture [4]. However, not everyone has access to this benefit. Those who cannot use information technology or access information due to their poverty or physical disabilities are alienated and isolated, resulting in yet another handicap in terms of social activity. In order to relieve these social disabilities via supporting disabled people’s access to information, the Korean government has been supplying information aids to disabled people since 2003. To some extent, governmental efforts can liberate the disabled from their isolation from information. However, living as an integrated member of a community and interacting with that community places a number of demands on a disabled person. These demands include purchasing new devices and learning new skills. Patients with complicated problems, such as physical disabilities or brain lesions, may need one or more items of supplementary equipment such as an electric wheel chair, AAC device, alternative input device for access to information, or environmental control unit (ECU). Most of this equipment demands familiarity with device-specific operation instructions. Furthermore, it comes in a variety of appearances and requires space for storage [5]. To make matters worse, these devices are expensive, so the user’s financial burden is significant. In an effort to relieve financial and spatial difficulties and to reduce the inconvenience of carrying large amounts of equipment, this study develops a Korean AAC application that can be mounted on smartphones and tablet PCs.

2

Survey of AAC Devices

This chapter outlines the AAC devices selected as the Korean government’s official supplies in 2011. It will give an overview of common AAC devices used in Korea. (Korea National Information Society Agency, 2011) 2.1

SuperTalker

The SuperTalker is a voice output device that holds a voice recording of up to 16 minutes. Its dimensions are 32 20 45 cm , and it has exchangeable panels of 2, 4, and 6 grids that correspond to the user's ability for direct selection. Up to 8 additional switches can be installed to further support direct selection. Although it is the product of a U.S. brand, Koreans can use it easily because the user can record its output voices instead of relying on pre-recorded machine voices. However, it is limited by the fact that it takes a maximum of only 8 words, requiring the user to rerecord the voices and change the words depending on the situation.

Development of a Korean Language-Based Augmentative

431

Fig. 1. SuperTalker

2.2

OK Toc Talk

The OK Toc Talk is a product of Korea, which holds up to 24 minutes of voice recording. Its dimensions are 28.2 14.8 2.8 cm . Its settings are versatile enough to accommodate 2 users, and it supports 6 channels. Its voices and images can be edited on a PC. When the user changes the picture panel, the corresponding voices can be turned on by changing the channel to the one that matches the picture card. However, this device is useful only when the user is capable of direct selection, and the sizes of the direct selection buttons are not changeable. Furthermore, the available words are limited to a total of 120, and the user needs additional assistance to change the picture card.

Fig. 2. OK Toc Talk

2.3

KidsVoice

KidsVoice is a TTS-based device developed in Korea. Its and 1Kg system is equipped with Korean-familiarized image symbols. Its word sets

432

C.-G. Kim et al.

Fig. 3. KidsVoice

are selectable according to the situation, and sentence outputs can be generated from combinations of words. The size of the button is flexible, and it supports direct selection and scanning via the USB switch as well. Furthermore, the vocabulary and symbols are editable so that various AAC strategies can be applied. However, its price is as high as 2,970,000 Won, and configuring its numerous functions is not trivial.

3

System Design

The conceptual diagram of the application described in this paper is shown in Fig. 4. The symbol system used in the application is flexible enough in choosing texts or image symbols to accommodate various strategies. The assistive aid employs an Android-based tablet PC, which is widely used in Korea. The technique of word selection supports both direct and indirect techniques, so that various AAC strategic choices are available. The size of the button for direct selection is automatically adjusted according to the user-specified number of cells displayed on the screen. If more words are stored than the screen can display on one page, another page is automatically added so that all the stored words are accessible. The user can explore the pages by either selecting the corresponding page button on the bottom of the screen or by dragging the page. In order to help the user construct sentences, words that can follow the selected word are suggested in a dynamic display. Indirect selection is designed to make efficient use of limited space by scanning words one by one when the user touches the screen. Furthermore, various pre-set word panels can be prepared in advance, which the user can choose according to the situation.

Development of a Korean Language-Based Augmentative

433

Fig. 4. Conceptual diagram of the system

4

Implementation of the System

The AAC application is constructed based on the Android OS and mounted on a Samsung Galaxy Tab as the platform. The application works as follows: First, tapping the icon loads the application. The first screen to display is the front page of situation options, shown in Fig. 5. Selecting a situation leads to the primary word set associated with that situation. Now the system is ready to use. When the user selects a word from the displayed list, the AAC application displays the corresponding text in the text box at the top of the screen and outputs the voice using the onboard TTS engine. The default words used in the application are stored in specified text files, which the user can edit. The words consist of 2-level sets: primary words, which are mainly used in normal situations, and secondary words used in conjunction with the primary words to make sentences. The primary words are registered as the user creates a text file in a specified folder and names the file to reflect the situation. The application recognizes the secondary words of a primary word as it finds words in the text file whose name is the same as the primary word. As the user selects the situation from the front page, the application automatically distributes the primary words onto pages depending on the number of primary words and the number of cells on the screen. The pages are generated in the order of the words recorded in the text file. The total number of pages and current page number are indicated in red at the top-right corner of the screen. The pages are turned as the user touches the right and left buttons at the bottom of the

434

C.-G. Kim et al.

screen. When the user selects a primary word, the application searches for the same file name as the selected primary word and, if it is found, the screen turns to the list of associated secondary words, and the system waits for the user’s selection. If no secondary word is found, the system only displays the text and outputs the TTS voice for the selected primary word. The user can return to the situation page by tapping the text box at the top of the screen or the information pad.

Fig. 5. Implemented AAC application

4.1

Symbol System

The symbol system of the AAC application employs both texts and images. Clicking the menu button of the tablet PC activates the menu of the application, as shown in Fig. 6. The configuration button at the right corner leads to the set-up menu, where the user decides whether or not to use images. Furthermore, different images can be used for words if the jpg file corresponding to the word, i.e. the file in the image folder under the same folder as the text file, is edited or switched.

Fig. 6. Process to change the symbol system

4.2

Strategy

To allow more choices for direct selection, the application employs direct communication with a screen keyboard for users who are able to manipulate their fingers to select fine print. Furthermore, as shown in Fig. 7, the number of cells in direct selection can be controlled depending on the user's manual dexterity. A setting that allows users to turn the page by dragging is also available for users with hand skills fine enough to drag. These settings are stored in separate files so that they can be customized for different users.

Development of a Korean Language-Based Augmentative

435

Fig. 7. Process to change the number of cells

In the configuration for indirect selection, both activation of indirect selection and the scanning period for each word can be specified, as shown in Fig. 8.

Fig. 8. Configuration and usage of scanning

5

Conclusion

An AAC application was developed using an Android OS–based tablet PC. The application supports text and image symbols to help the user access the AAC via suitable strategies and techniques. The application can accommodate both direct and indirect inputs. The size and the number of cells can be adjusted according to the user’s motor skills. The scanning speed for indirect selection can be controlled.

436

C.-G. Kim et al.

The AAC application described here may relieve some of the financial burden of people with communication disorders, and it may also reduce the inconvenience of carrying many devices. As a result, it will give people with disabilities the opportunity to participate more fully in the community. Furthermore, because the application has versatile function settings, those who want to learn how to use AAC devices can experience various strategies without changing the device itself. Therefore, it can be available at minimal cost, making it useful in the field of education as an exploratory alternative to expensive professional AAC devices for children. This application has the limitation that, as it is based on the tablet PC, the TTS voice options are limited. Acknowledgement. This work was supported by the Korea Health Industry Development Institute grant funded by the Korea government (A101986).

References 1. Jeong, H.D., Kim, J.Y., Park, E.H., Park, S.J.: AAC Education for students who have disabilities. Korea National Institute for Special Education, Anshan, Gyeonggi-do (1999) 2. Lee, M.H., Park, E.H.: Effect of Peer Training for Peer-Mediated Intervention with AAC on the Social Interaction of Children with Severe Physical Disabilities. The Korean Journal of Early Childhood Special Education 6(1), 109–127 (2006) 3. Park, E.H., Bak, S.H.: A Survey of Special Education Teachers’ Perceptions on Education for Students with Severe Disabilities. Korean Journal of Special Education 36(1), 29–55 (2001) 4. Kim, T.I., Do, S.G.: The Analysis of Digital Divide between Disabled and Non-disabled People. Social Welfare Policy 21, 341–365 (2005) 5. Kim, C.G., Song, B.S.: Development of Integrated Computer Interface for Power Wheelchair User. Journal of biomedical engineering research 31, 251–257 (2010)

Adaptive Power Management for Nanoscale SoC Design Jeong-Tak Ryu and Kyung Ki Kim* School of Electronic Engineering, Daegu University, Gyeongsan, South Korea {jryu,kkkim}@daegu.ac.kr

Abstract. The demand for power sensitive designs in system-on-chip (SoC) has grown significantly as MOSFET transistors scale down. Since portable battery powered devices such as cell phones, PDA's, and portable computers are becoming more complex and prevalent, the demand for increased battery life will require designers to seek out new technologies and circuit techniques to maintain high performance and long operational lifetimes. As process dimensions shrink further toward nanometer technology, traditional methods of dynamic power reduction are becoming less effective due to the increased impact of standby power. Therefore, this paper proposes a novel adaptive power management system for nanoscale SoC design that reduces standby power dissipation. The proposed design method reduces the leakage power at least by 500 times for ISCAS’85 benchmark circuits designed using 32-nm CMOS technology comparing to the case where the method is not applied. Keywords: Power management, Stand-by power, Leakage power, SoC.

1

Introduction

The demand for power sensitive designs in system-on-chip (SoC) has grown significantly as MOSFET transistors scale down. Since portable battery powered devices such as cell phones, PDA's, and portable computers are becoming more complex and prevalent, the demand for increased battery life will require designers to seek out new technologies and circuit techniques to maintain high performance and long operational lifetimes. Lowering power-supply voltage in the system is one of the most effective schemes to reduce the power dissipation. A number of methods have been proposed to scale down the power-supply voltage dynamically [1]-[5]. Even though they are effective in decreasing dynamic power dissipation, they do not help reduce leakage power effectively. As transistor geometries are scaled down aggressively, threshold voltage decreases to achieve high performance, resulting in the exponential increase in leakage current. Due to the continued scaling of the technology and supply/threshold voltage, leakage power has become a dominant portion in the power dissipation of nanoscale VLSI *

Corresponding author.

T.-h. Kim et al. (Eds.): FGCN 2011, Part II, CCIS 266, pp. 437–446, 2011. © Springer-Verlag Berlin Heidelberg 2011

438

J.-T. Ryu and K.K. Kim

systems. An analysis of trends based on the International Technology Roadmap for Semiconductors shows that the power lost to leakage is beginning to exceed the power spent on useful computation. Therefore, the leakage power is a very serious problem in portable electronic systems that operate mostly in standby mode. The impact of standby power is increasing steadily as process dimensions shrink. In nanometer MOSFET circuits, the main components of standby power are subthreshold, gate tunneling, and reverse-biased junction band-to-band-tunneling (BTBT) leakage current. The reduction of transistor geometries necessitates the reduction of the supply voltage to avoid an electrical break down and to obtain the required performance. However, to retain or improve performance, it is necessary to reduce the threshold voltage (Vth) as well, resulting in the exponential increase of sub-threshold leakage [6][7]. To control the short-channel effects and increase the transistor driving strength in deep sub-micron (DSM) circuits, gate-oxide thickness also becomes thinner as technology scales down. The aggressive scaling in the gate-oxide results in currents tunneling through the oxide, which is a strong exponential function of the oxide thickness and the voltage magnitude across the oxide. In scaled devices, the higher substrate doping density and the application of the "halo" profiles cause significantly large reverse biased junction band-to-bandtunneling (BTBT) leakage current because of drain-substrate and source-substrate junctions [6][7]. In order to minimize the leakage power dissipation, several circuit techniques have been proposed, such as multi-threshold voltage CMOS (MTCMOS) [8] and variable threshold voltage CMOS (VTCMOS) [9] using variable substrate bias voltage. To reduce the leakage power by increasing the threshold voltage of MOSFET transistors during standby mode, adaptive reverse body-biasing (ABB) technique has been proposed [10]. The ABB decreases the sub-threshold leakage current of the scaled MOSFET. However, it increases the depletion width of the MOSFET parasitic junction diode and rapidly increases the BTBT current between the substrate and source/drain, especially in halo implants. Recently, methods using forward body biasing in active mode have been introduced [11]. Forward biasing increases the dynamic range of the device threshold and improves the circuit performance by decreasing threshold voltage. In addition, a power-performance trade off methodology for microprocessors has been proposed [12]. Another research has shown that simultaneous supply voltage scaling and bidirectional body biasing (forward + reverse biasing) is more effective in achieving high performance in active mode and low-power dissipation in standby mode. Therefore, the optimal voltage scaling and bidirectional body biasing determine the optimal tradeoff between supply voltage and body-bias voltage [13]. However, these techniques require significant circuit modification and performance overhead for leakage reduction, and they have not been complete or robust enough to apply to VLSI systems since the process, voltage, and temperature (PVT) variations are not considered for leakage power, especially in nanometer technology.

Adaptive Power Management for Nanoscale SoC Design

439

Fig. 1. Graphical representation showing reduction of total active and standby leakage power

This paper proposes a novel power management system to achieve low power in standby mode by exploiting the supply-voltage scaling and body-bias-voltage scaling taking PVT variations into consideration.

2

Proposed Power Management System

As described in the previous section, leakage currents will become a large component of total power dissipation as technology scales down. Although total power dissipation (dynamic + leakage) during the active mode is reduced with scaling, further power gains can be achieved if leakage currents are controlled wherever possible because these currents will make up a larger percentage of overall power dissipation in future technologies. Furthermore, the overall idle power dissipation tends to increase during the idle mode or standby mode where no computation is taking place due to the large leakage currents as shown in Fig. 1 [14]. Therefore, power reduction in standby mode has to be aggressively controlled so that the aggregate power consumption for the circuit in standby mode can be minimized. In order to reduce the leakage power component during the standby period, this paper proposes a new power management system in standby mode as shown in Fig. 2. The proposed power management system is composed of the main control unit and several sub-control units. The main control unit monitors each functional unit and sends sleep and active signals to each sub-control unit. Also, the main unit controls the entire workload and clock synchronization. Each sub-control unit is contained in a functional unit (FU) and monitors the idle period and the PVT variation of the FU. The sub-control unit consists of an input pattern generator for minimum leakage power, a data retention latch, a body voltage biasing circuit, PVT monitoring circuits, and a VDD scaling circuit.

440

J.-T. Ryu and K.K. Kim

(a)

(b) Fig. 2. Block diagram of the proposed power management system: (a) Main control unit, (b) Sub-control unit

3

Protocol Design

To explain the proposed scheme, we define the following messages.  

SLEEP_IN : broadcast message which contains SLEEP_TABLE. SLEEP_FIND : request message of SLEEP_TABLE.

Before one terminal of the active mode enters the sleep mode, it broadcasts SLEEP _IN message. SLEEP FIND message is used when the terminal of the sleep mode wants to transmit the packet. Here, SLEEP_TABLE contains the sleep mode ID, the time to sleep, and sleep duration as shown in Fig. 3 (a). All the terminals should manage the SLEEP_TABLE independently. Moreover, all terminals have three modes as follows:  Active_mode : transmits or receives data packet.  Listen_mode: only listens through receiver and does not any transmit packet through transceiver.  Sleep_mode : neither transmits nor receives packet.

Adaptive Power Management for Nanoscale SoC Design

441

The transition among modes is shown in Fig. 3 (b)(c). The idea behind the power management of each sub-control unit is described with five operating regions depicted in the timing diagram and the state machine shown in Fig. 3. In the time diagram, the sleep signal is generated for clock gating at the first breakeven (Tsleep1). Throughout sleep #1 to #3 region, the clock gating state is stayed to disables the clock signal and save power consumption. The five operation regions are as follows: (1)

Active: In this state, the FU is being used, and the sub-control unit counts the idle time of the FU by monitoring the primary inputs/outputs of the FU.

(2)

Sleep#1: This state is reached if the FU stays in the idle state longer than the first breakeven point (Tsleep1). In this state, the sub-control unit starts to scale down the supply voltage of the FU primary input changes of the FU. The sub-control unit goes back to the active state if the primary inputs of the FU have new transitions or if the main control unit sends an active signal.

(3)

Sleep#2: This state is reached if the FU stays in the idle state of the sleep #1 longer than the second breakeven point (Tsleep2). In this state, the optimal body biasing is applied to the FU with the scaled supply voltage to reduce more leakage power consumption. The other operations are the same as the sleep #1 state.

(4)

Sleep#3: This state is reached if the FU stays in the idle state of the sleep #2 longer than the third breakeven point (Tsleep3). In this state, the lowleakage input pattern is used in the FU with the scaled supply voltage and the optimal body voltage to reduce much more leakage power consumption than previous states. Then, all the internal node states of the FU are changed based on the input pattern, unlike the sleep #1 and #2 states. However, the previous node states are held through data retention circuits. The sub-control unit goes back to the active state if the primary inputs of the FU have new transitions or if the main control unit sends an active signal with an enable command. The FU restores the previous node states after one clock period.

(5)

Wait: This state is reached if the main control unit sends an active signal without the enable command. In this state, the sub-control unit prevents the FU from sending its primary output signals and waits during a period requested from the main control unit. If the main control unit sends a reset signal to clear all the internal node states restored from the retention circuits, the sub-control unit sends a reset signal to each latch contained in the FU.

442

J.-T. Ryu and K.K. Kim

(a)

(b)

(c) Fig. 3. Power management in standby mode: (a) Sleep Table format, (b) Timing diagram (c), PMS management state machine

If one terminal neither transmits nor receives the packet during a certain period, it broadcasts the SLEEP_IN message and enters the sleep mode. If the terminals in the active mode receive the SLEEP_IN, they keep or update SLEEP_TABLE. The procedure that enters the sleep mode is shown in Fig. 4 (a). As shown in Fig. 4 (b), when the terminals in the sleep mode want to transmit the packet, they first enter the active mode and broadcast SLEEP_FIND message. The terminal in the active mode which receives SLEEP_FIND transmits SLEEP_TABLE including ACK signal after a random delay to avoid conflict. Then, after the source

Adaptive Power Management for Nanoscale SoC Design

443

terminal sees the information of destination in the SLEEP_TABLE, it determines whether it transmits right now or not. In case the destination terminal is in the sleep mode, the source terminal waits until the destination terminal in the sleep mode wakes. However, if the destination terminal is in the active mode (destination ID is not included in SLEEP_TABLE), the source terminal does not need to wait any longer and transmits the packet immediately. When the terminals in the active mode want to communicate with the terminal in the sleep mode to transmit the packet, they monitor the time to start to sleep and the sleep duration in the SLEEP_TABLE and wait until the terminal in the sleep mode wakes as shown in Fig. 4 (a). If the destination is active mode (destination ID is not included in SLEEP_TABLE), the packet is transmitted immediately. The terminals in the sleep mode periodically wake and enter the listen mode to check whether they need to receive the packet or not. If they do not need to receive any packet, they enter the sleep mode again, but if the receiving packet exists, they go to the active mode and receive the packet. There is a special case when all terminals except one terminal are in the sleep mode. In this case, the last active terminal does not enter the sleep mode although it neither transmits nor receives the packet during a certain time because at least one terminal should manage the SLEEP_TABLE. If all terminals do not have SLEEP _TABLE, they do not know when other terminals in the sleep mode wake and which terminal is in the sleep mode. Therefore, the last active terminal which wants the sleep mode waits until another terminal enters the listen mode or the active mode. If the terminal in the sleep mode enters the listen mode, the last active terminal sends SLEEP_IN to it and enters the sleep mode itself. The SLEEP TABLE is distributed to reduce the power consumption of a specific terminal to manage SLEEP_TABLE. If only one terminal manages SLEEP_TABLE, it should continually wake. Then, this terminal cannot use LPM and its power is soon exhausted.

4

Experimental Results

The proposed optimal control system using 32-nm MOSFET technology has been implemented and evaluated using ISCAS’85 benchmark circuits designed in the same technology. The number of transistors of the sub-controller is 851, and its power dissipation is 141 W. Table 1 shows the summary of the results for the proposed approach at 50oC and a typical corner. The average leakage power has been measured using random input test vectors at 0.9V supply voltage. As shown in Table 1, the new technique for the minimal standby power provides average 1000 times reduction in leakage power compared to the simulation results of benchmark circuits without any optimization techniques. In order to show the effects of temperature and process variations, three temperature conditions (-25, 50, 125±oC) and three process-corner conditions (slow, typical, fast) are considered in the ISCAS benchmark circuit simulation. The optimal VDD/VBody control technique gives at least 500 times reduction in leakage power dissipation compared to the simulation results of benchmark circuits

444

J.-T. Ryu and K.K. Kim

without any optimization techniques being used. Moreover, our simulation results presents that when the proposed technique is applied, the leakage power dissipation is far less sensitive to the temperature and process variations because the optimal supply voltage and body bias voltage are changed according to the temperature and process. All the simulation results demonstrate that the proposed system is very effective in reducing the standby power in the big circuits.

(a)

(b) Fig. 4. Flow chart of an active and a sleep mode operation: (a) An active mode operation of proposed system, (b) A sleep mode operation of proposed scheme

Adaptive Power Management for Nanoscale SoC Design

5

445

Conclusion

As technology scales down below 90 nm, leakage currents have become a critical issue. In the past, circuit techniques and architectures ignored the effects of these currents because they were insignificant compared to the switching currents and threshold voltages were high enough. However, in modern technologies, the role of the leakage currents cannot be ignored and becomes increasingly significant issue with further scaling. Therefore, new circuit techniques and design considerations must be developed to control leakage currents in standby mode in order to provide lowpower solutions. In order to reduce the standby power, this paper proposed a novel asynchronous power management system to reduce the leakage power dissipation during standby mode. The proposed system consists of main control units, a bus interface, and subcontrol units. To reduce the power consumption of each terminal, each terminal is allowed to enter the low power mode using Sleep_Table and has five operation regions. Also, the power management system includes a novel control system that uses an adaptive method to find the optimal scaling algorithm during standby mode. Based on the temperature and process conditions, the optimal supply voltages is generated to reduce the leakage power, and body-bias voltage is automatically adjusted continuously by the control loop to adapt to the PVT variations. The results show that the proposed control system is a viable solution for high energy reduction in nanoscale CMOS circuits. Table 1. Experimental results for the standby power

Circuit

# of gates

With optimization (Typical, T=85ºC)

Average leakage power measured in Hspice ( VDD=0.9V, T=85ºC)

C432

160

8.56 nW

10.14 μW

C499

202

29.48 nW

48.90 μW

C880

383

18.53 nW

23.92 μW

C1355

546

23.11 nW

31.63 μW

C1908

880

46.36 nW

69.86 μW

C2670

1193

74.34 nW

124.45 μW

C5315

2307

145.10 nW

180.08 μW

C6288

2388

97.36 nW

110.64 μW

Acknowledgments. This work was supported by IC Design Education Center (IDEC) – CAD Tools (Softwares) only.

446

J.-T. Ryu and K.K. Kim

References 1. Meijer, M., Pineda de Gyvez, J., Otten, R.: On-chip digital power supply control for system-on-chip applications. In: IEEE ISLPED, pp. 311–314 (August 2005) 2. Nakai, M., Akui, S., Seno, K., Meguro, T., Seki, T., Kondo, T., Hashiguchi, A., Kawahara, H., Kumano, K., Shimura, M.: Dynamic voltage and frequency management for a lowpower embedded microprocessor. IEEE J. Solid-State Circuits 40(1), 28–35 (2005) 3. Wang, W., Mishra, P.: System-Wide Leakage-Aware Energy Minimization Using Dynamic Voltage Scaling and Cache Reconfiguration in Multitasking Systems. IEEE Transactions on Very Large Scale Integration (VLSI) Systems 99, 1–9 (2011) 4. Nourivand, A., Al-Khalili, A.J., Savaria, Y.: Postsilicon Tuning of Standby Supply Voltage in SRAMs to Reduce Yield Losses Due to Parametric Data-Retention Failures. IEEE Transactions on Very Large Scale Integration (VLSI) Systems 99, 1–13 (2011) 5. Pavan, T.K., Jagannadha Naidu, K., Shekar Babu, M.: Implementation of delay and power monitoring schemes to reduce the power consumption. In: IEEE International Conference on Signal Processing, Communication, Computing and Networking Technologies (ICSCCN), pp. 459–464 (July 2011) 6. Elgharbawy, W.M., Bayoumi, M.A.: Leakage sources and possible solutions in nanometer CMOS technologies. IEEE Circuits Syst. Magazine 5(4), 6–17 (2005) 7. Agarwal, A., Mukhopadhyay, S., Raychowdhury, A., Roy, K., Kim, C.H.: Leakage power analysis and reduction for nanoscale circuits. IEEE Micro. 26(2), 68–80 (2006) 8. Anis, M., Elmasry, M.: Multi-Threshold CMOS Digital Circuits: Managing Leakage Power. Kluwer, Norwell (2003) 9. Inukai, T., Hiramot, T., Sakurai, T.: Variable threshold voltage CMOS (VTCMOS) is series connected circuits. In: Int. Symp. Low Power Electron. Des. (ISLPED), pp. 201–206 (2001) 10. Roy, K., Mukhopadhyay, S., Mahmoodi-Meimand, H.: Leakage current mechanisms and leakage reduction techniques in deep-submicrometer CMOS circuits. Proc. IEEE 91(2), 305–327 (2003) 11. Hokazono, A., Balasubramanian, S., Ishimaru, K., Ishiuchi, H., Liu, T.K., Hu, C.: MOSFET design for forward body biasing scheme. IEEE Electron Device Lett. 27(5), 387–389 (2006) 12. Kulkarni, M., Sheth, K., Agrawal, V.D.: Architectural power management for high leakage technologies. In: IEEE Southeastern Symposium on System Theory (SSST), pp. 67–72 (March 2011) 13. Ishibashi, K., Fujimoto, T., Yamashita, T., Okada, H., Arima, Y., Hashimoto, Y., Sakata, K., Minematsu, I., Itoh, Y., Toda, H., Ichihashi, M., Komatsu, Y., Hagiwara, M., Tsukada, T.: Low-voltage and low-power logic, memory, and analog circuit techniques for SoCs using 90 nm technology and beyond. IEICE Trans. Electon E89-C(3), 250–262 (2006) 14. Anis, M., Elmasry, M.: Multi-threhsold CMOS digital circuits. Kluwer Academic Publishers (2003)

Author Index

Gerardo, Bobby D. II-220, II-229, II-239 Gil, Joon-Min II-354, II-381, II-403

Abdullah, Jiwa II-301 Adorna, Henry N. II-208 Agustin, Oliver C. I-244 Ahmed, Sabbir I-67 Alisherov, Farkhod II-20 Bae, Ihn-Han II-364, II-371 Bae, Kyeong-ryeol I-136, I-147 Baek, Yeong-Jin I-277 Baguda, Yakubu S. I-188 Bojkovic, Zoran S. I-198 Byun, Tae-Young II-320 Byun, Yung-Cheol II-220, II-229, II-239 Cabioc, Mark Dominic II-229 Chen, Bo-Han I-300 Chen, Wei-Sheng I-300 Cheong, Seung-Kook II-330 Chiang, Meng-Shu I-283 Chien, Shih-Pei I-236 Chimeh, Jahangir Dadkhah I-59 Cho, Jin Haeng II-292 Cho, Moon-Taek II-43 Cho, Seongsoo I-15, II-57 Cho, Woong II-26 Choi, Jae-Hoon II-274 Choi, Sang-Min II-248 Choi, Seong Gon I-111 Choi, Seung Ho II-124, II-132, II-154 Choi, Yeonyi I-93 Choi, Young B. I-310 Chouta, Rajath I-310 Chowdhury, Nawshad U.A. I-103 Chun, Chan Jun II-114, II-124 Chung, Kwang Sik II-412 Darwish, Ashraf I-209 De Castro, Joel T. II-220 Eid, Heba F. I-209 Eun, Ae-cheoun II-179 Farooq, Muhammad Omer Fisal, Norsheila I-188

I-1

Ha, Young-guk II-179 Han, Kijun II-338, II-346 Han, Sunyoung II-179 Hassanien, Aboul Ella I-209 Heo, Seok-Yeol II-393 Hong, Bong-Hwa II-34, II-57, II-65, II-96 Hong, Bonghwa I-15 Hong, Seong-Sik II-83 Hong, Suck-Joo II-43 Hong, Won-Kee II-419 Hoq, Md.T. I-103 Hsu, Tz-Heng I-283 Huh, Eui-Nam II-1 Hur, Kyung Woo II-283 Jang, Jae Hyuck II-283 Jang, Sei-Jin II-114, II-124 Jang, Seok-Woo I-120 Jeong, Hwa-Young II-65, II-96 Jiang, Jing-Jing II-162 Jin, Ik Soo I-261 Jo, Sung Dong II-114 Joe, Inwhee I-93 Joo, Hae-Jong II-73 Joo, Kil Hong I-293 Jung, Ho Min II-292 Jung, Hyo-Young I-219 Kang, Bong-Soo I-53 Kang, Cheoul-Shin I-179 Kang, Chul-Ung II-199 Kang, Chul Uoong II-189 Kang, Jang-Mook II-34 Kang, Jin Ah II-132 Kang, Sung Woon II-292 Kawai, Makoto I-67 Kawser, Mohammad T. I-103 Khan, Jahangir I-198 Khanam, Solima I-120 Kim, Byung Ki II-283 Kim, Chang-Geol II-429

448

Author Index

Kim, Do-Hoon II-11 Kim, Dongik I-93 Kim, Dongkyun II-258 Kim, Eun-Kyoung I-129 Kim, Hae Geun II-364, II-371 Kim, Haeng-Kon I-166 Kim, Heemin II-179 Kim, Hong Kook II-104, II-114, II-124, II-132, II-143 Kim, Hye-Jin I-47, I-53 Kim, Hyun Jong I-111 Kim, Jeong-Sam II-320 Kim, Jin-Mook II-83 Kim, Junhyung II-346 Kim, Kyung Ki II-437 Kim, Mihye II-354, II-381 Kim, Myung-Ju II-412 Kim, Sang-Soo II-73 Kim, Seon Man II-104 Kim, Seung-Hae II-403 Kim, Sung-Gyu II-20 Kim, Tai-hoon I-209 Kim, Young-Choon II-43 Kim, Yun-Hyuk I-53 Ko, Daesik II-268 Ko, Seok-Jun II-199 Ko, Young Woong II-283, II-292 Kunz, Thomas I-1 Kwak, Ho-Young I-53 Kwak, Soo-Won II-429 Kwon, Dong Rak II-312

Lee, Seok-Pil II-114, II-124 Lee, Seongjun I-53 Lee, Sung Joo II-104 Lee, Un-Bai II-412 Lee, Wan-Jik II-393 Lee, Won-Hyek II-403 Lee, Won-Yeoul II-393 Lee, Yong-Hwan I-129, I-136, I-147 Lee, Young Han II-143 Lee, Young-Hun I-179, II-330 Lee, Youngkon I-23, I-31, I-39 Lee, Young-Wook II-90 Lee, Yun Keun II-104 Li, Yi-Da I-283 Lim, Jong Hwan II-189 Lim, Kyungshik II-258 Lin, Chu-Hsing I-77, I-82, I-87, I-236 Lin, Hung-Yan I-82 Liu, Jung-Chun I-77, I-82 Lu, Shu-Yuan I-236

La, Keuk-Hwan I-15, II-57 Lai, Shin-Pin I-87 Lee, Byunghwa II-338, II-346 Lee, Chen-Yu I-87, I-236 Lee, Chien-Hsing I-77 Lee, Chung-Sub I-219 Lee, Euy-Soo II-73 Lee, Gi Min II-189 Lee, Ho-Cheol II-312 Lee, Hyuek Jae I-269 Lee, Jae-Dong I-129 Lee, Jae-Won II-154 Lee, Jong-Heon I-53 Lee, Jongsup I-15, II-57 Lee, Jung Geun II-292 Lee, Junghoon I-47, I-53 Lee, Jun-Hyung II-1 Lee, Sang-Hoon I-277

Oh, Byung-Joo I-244 Oh, Hyun Seo II-26 Oh, Sang Yoon I-136 Oh, Tae Hwan I-310 Ok, Seung-Ho I-136, I-147 Osorio, Francisca D. II-220

Malinao, Jasmine A. II-208 Maravilla Jr., Reynaldo G. II-208 Marwat, Muhammad Imran Khan I-198 Matsuo, Tokuro II-169 Moon, Byung-Hyun II-248 Moon, Byungin I-129, I-136, I-147 Moon, Inseok II-419 Na, Sang-Ho II-1 Nguyen, Tien-Dung II-1 Noh, Min-Ki II-403

Paik, Woojin I-120 Pangapalan, Ana Rhea II-220 Parapari, Saeed Bashirzadeh I-59 Park, Byungjoo II-20 Park, Byung-Seok I-179 Park, Gyung-Leen I-47, I-53 Park, Hwase II-268 Park, Kun Hyun II-189 Park, Nam Hun I-293 Park, Nam In II-143 Pun, Chi-Man II-162

Author Index Rashid, Rozeha A. I-188 Ryu, Heung-Gyoon II-11, II-274 Ryu, Jeong-Tak II-437 Seo, Yong-Ho I-219 Shrestha, Bhanu I-15, II-57 Shuaibu, Dahiru S. I-188 Son, Hyeon-Sik I-136, I-147 Song, Biao II-1 Song, Byung-Seop II-429 Song, Ho-Bin II-43 Song, Hyun-Ju II-330 Song, ShiLi I-227 Song, Ui-Sung II-354 Su, Wei Wei I-227 Surendran, Purushothaman II-199 Syed, Sharifah H. I-188 Syfullah, Md.K. I-103 Tabanda, Elise A. II-208 Tak, Ryu Juang II-429

Takahashi, Satoshi II-169 Tang, Wei II-1 Tanguilig III, Bartolome II-239 Tsai, Sheng-Hsing I-82 Wahid, Abdul II-258 Wang, Yi Fan I-227, I-254 Wen, Hong I-227, I-254 Wu, Tang-Wei I-77 Yang, Chao-Tung I-300 Yang, Ling I-227, I-254 Yang, Tae-Kyu I-219 Yang, Tzu-Chien I-87 Yeom, Kiwon I-156 Yoo, Kwan-Hee II-381 Yun, Jangkyu II-338, II-346 Yusof, Sharifah K. I-188 Zhang, Gao Yuan I-254 Zhou, Liang I-254

449