of Vibration in
Mechanical Systems
K10933_FM.indd 1 3/16/10 10:42:08 AM
AUTOMATION AND CONTROL ENGINEERING
A Series of Reference Books and Textbooks
Series Editors
Modeling and Control of Vibration in Mechanical Systems, Chungling Du
and Lihua Xie
Analysis and Synthesis of Fuzzy Control Systems: A ModelBased Approach,
Gang Feng
LyapunovBased Control of Robotic Systems, Aman Behal, Warren Dixon,
Darren M. Dawson, and Bin Xian
System Modeling and Control with ResourceOriented Petri Nets,
Naiqi Wu and MengChu Zhou
Sliding Mode Control in ElectroMechanical Systems, Second Edition,
Vadim Utkin, Jürgen Guldner, and Jingxin Shi
Optimal Control: Weakly Coupled Systems and Applications,
Zoran Gajic´, MyoTaeg Lim, Dobrila Skataric´, WuChung Su,
and Vojislav Kecman
Intelligent Systems: Modeling, Optimization, and Control, Yung C. Shin
and Chengying Xu
Optimal and Robust Estimation: With an Introduction to Stochastic Control
Theory, Second Edition, Frank L. Lewis; Lihua Xie and Dan Popa
Feedback Control of Dynamic Bipedal Robot Locomotion,
Eric R. Westervelt, Jessy W. Grizzle, Christine Chevallereau, Jun Ho Choi,
and Benjamin Morris
Intelligent Freight Transportation, edited by Petros A. Ioannou
Modeling and Control of Complex Systems, edited by Petros A. Ioannou
and Andreas Pitsillides
Wireless Ad Hoc and Sensor Networks: Protocols, Performance,
and Control, Jagannathan Sarangapani
Stochastic Hybrid Systems, edited by Christos G. Cassandras
and John Lygeros
Hard Disk Drive: Mechatronics and Control, Abdullah Al Mamun,
Guo Xiao Guo, and Chao Bi
Autonomous Mobile Robots: Sensing, Control, Decision Making
and Applications, edited by Shuzhi Sam Ge and Frank L. Lewis
Neural Network Control of Nonlinear DiscreteTime Systems,
Jagannathan Sarangapani
FRANK L. LEWIS, PH.D.,
FELLOW IEEE, FELLOW IFAC
Professor
Automation and Robotics Research Institute
The University of Texas at Arlington
SHUZHI SAMGE, PH.D.,
FELLOW IEEE
Professor
Interactive Digital Media Institute
The National University of Singapore
K10933_FM.indd 2 3/16/10 10:42:09 AM
CRC Press is an imprint of the
Taylor & Francis Group, an informa business
Boca Raton London New York
Modeling and Control
of Vibration in
Mechanical Systems
Chunling Du
Data Storage Institute
Singapore
Lihua Xie
Nanyang Technological University
Singapore
K10933_FM.indd 3 3/16/10 10:42:09 AM
MATLAB® and Simulink® are trademarks of The MathWorks, Inc. and are used with permission. The
MathWorks does not warrant the accuracy of the text of exercises in this book. This book’s use or
discussion of MATLAB® and Simulink® software or related products does not constitute endorsement
or sponsorship by The MathWorks of a particular pedagogical approach or particular use of the MAT
LAB® and Simulink® software.
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 334872742
© 2010 by Taylor and Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business
No claim to original U.S. Government works
Printed in the United States of America on acidfree paper
10 9 8 7 6 5 4 3 2 1
International Standard Book Number: 9781439817988 (Hardback)
This book contains information obtained from authentic and highly regarded sources. Reasonable
efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors and
publishers have attempted to trace the copyright holders of all material reproduced in this publication
and apologize to copyright holders if permission to publish in this form has not been obtained. If any
copyright material has not been acknowledged please write and let us know so we may rectify in any
future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information stor
age or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, please access www.copy
right.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222
Rosewood Drive, Danvers, MA 01923, 9787508400. CCC is a notforprofit organization that pro
vides licenses and registration for a variety of users. For organizations that have been granted a pho
tocopy license by the CCC, a separate system of payment has been arranged.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
K10933_FM.indd 4 3/16/10 10:42:09 AM
Contents
Preface xi
List of Tables xiii
List of Figures xv
Symbols and Acronyms xxiii
1 Mechanical Systems and Vibration 1
1.1 Magnetic recording system . . . . . . . . . . . . . . . . . . . . . 1
1.2 Stewart platform . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Vibration sources and descriptions . . . . . . . . . . . . . . . . . 4
1.4 Types of vibration . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4.1 Free and forced vibration . . . . . . . . . . . . . . . . . . 6
1.4.2 Damped and undamped vibration . . . . . . . . . . . . . 6
1.4.3 Linear and nonlinear vibration . . . . . . . . . . . . . . . 6
1.4.4 Deterministic and random vibration . . . . . . . . . . . . 6
1.4.5 Periodic and nonperiodic vibration . . . . . . . . . . . . . 7
1.4.6 Broadband and narrowband vibration . . . . . . . . . . 8
1.5 Random vibration . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5.1 Random process . . . . . . . . . . . . . . . . . . . . . . 11
1.5.2 Stationary random process . . . . . . . . . . . . . . . . . 12
1.5.3 Gaussian random process . . . . . . . . . . . . . . . . . . 12
1.6 Vibration analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6.1 Fourier transform and spectrum analysis . . . . . . . . . . 13
1.6.2 Relationship between the Fourier and Laplace transforms . 14
1.6.3 Spectral analysis . . . . . . . . . . . . . . . . . . . . . . 14
2 Modeling of Disk Drive System and Its Vibration 17
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 System description . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 System modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.1 Modeling of a VCM actuator . . . . . . . . . . . . . . . . 19
2.3.2 Modeling of friction . . . . . . . . . . . . . . . . . . . . 23
2.3.3 Modeling of a PZT microactuator . . . . . . . . . . . . . 29
2.3.4 An example . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4 Vibration modeling . . . . . . . . . . . . . . . . . . . . . . . . . 39
v
vi Modeling and Control of Vibration in Mechanical Systems
2.4.1 Spectrumbased vibration modeling . . . . . . . . . . . . 39
2.4.2 Adaptive modeling of disturbance . . . . . . . . . . . . . 43
2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3 Modeling of Stewart Platform 53
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2 System description and governing equations . . . . . . . . . . . . 53
3.3 Modeling using adaptive ﬁltering approach . . . . . . . . . . . . . 55
3.3.1 Adaptive ﬁltering theory . . . . . . . . . . . . . . . . . . 55
3.3.2 Modeling of a Stewart platform . . . . . . . . . . . . . . 58
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4 Classical Vibration Control 63
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2 Passive control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2.1 Isolators . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2.2 Absorbers . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2.3 Resonators . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2.4 Suspension . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2.5 An application example − Disk vibration reduction via
stacked disks . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3 Selfadapting systems . . . . . . . . . . . . . . . . . . . . . . . . 82
4.4 Active vibration control . . . . . . . . . . . . . . . . . . . . . . . 83
4.4.1 Actuators . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.4.2 Active systems . . . . . . . . . . . . . . . . . . . . . . . 84
4.4.3 Control strategy . . . . . . . . . . . . . . . . . . . . . . . 86
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5 Introduction to Optimal and Robust Control 89
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2 H
2
and H
∞
norms . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2.1 H
2
norm . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2.2 H
∞
norm . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.3 H
2
optimal control . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3.1 Continuoustime case . . . . . . . . . . . . . . . . . . . . 92
5.3.2 Discretetime case . . . . . . . . . . . . . . . . . . . . . 94
5.4 H
∞
control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.4.1 Continuoustime case . . . . . . . . . . . . . . . . . . . . 96
5.4.2 Discretetime case . . . . . . . . . . . . . . . . . . . . . 99
5.5 Robust control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.6 Controller parametrization . . . . . . . . . . . . . . . . . . . . . 104
5.7 Performance limitation . . . . . . . . . . . . . . . . . . . . . . . 108
5.7.1 Bode integral constraint . . . . . . . . . . . . . . . . . . 108
5.7.2 Relationship between system gain and phase . . . . . . . 111
5.7.3 Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Table of Contents vii
5.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6 Mixed H
2
/H
∞
Control Design for Vibration Rejection 115
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.2 Mixed H
2
/H
∞
control problem . . . . . . . . . . . . . . . . . . 115
6.3 Method 1: slack variable approach . . . . . . . . . . . . . . . . . 116
6.4 Method 2: an improved slack variable approach . . . . . . . . . . 117
6.5 Application in servo loop design for hard disk drives . . . . . . . . 123
6.5.1 Problem formulation . . . . . . . . . . . . . . . . . . . . 123
6.5.2 Design results . . . . . . . . . . . . . . . . . . . . . . . . 128
6.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
7 LowHump Sensitivity Control Design for Hard Disk Drive Systems 133
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.3 Design in continuoustime domain . . . . . . . . . . . . . . . . . 137
7.3.1 H
∞
loop shaping for lowhump sensitivity functions . . . 137
7.3.2 Application examples . . . . . . . . . . . . . . . . . . . . 141
7.3.3 Implementation on a hard disk drive . . . . . . . . . . . . 148
7.4 Design in discretetime domain . . . . . . . . . . . . . . . . . . . 152
7.4.1 Synthesis method for lowhump sensitivity function . . . . 152
7.4.2 An application example . . . . . . . . . . . . . . . . . . . 153
7.4.3 Implementation on a hard disk drive . . . . . . . . . . . . 158
7.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8 Generalized KYP LemmaBased Loop Shaping Control Design 161
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
8.2 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . 162
8.3 Generalized KYP lemmabased control design method . . . . . . 163
8.4 Peak ﬁlter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
8.4.1 Conventional peak ﬁlter . . . . . . . . . . . . . . . . . . 166
8.4.2 Phase lead peak ﬁlter . . . . . . . . . . . . . . . . . . . . 168
8.4.3 Group peak ﬁlter . . . . . . . . . . . . . . . . . . . . . . 169
8.5 Application in high frequency vibration rejection . . . . . . . . . 169
8.6 Application in midfrequency vibration rejection . . . . . . . . . . 177
8.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
9 Combined H
2
and KYP LemmaBased Control Design 183
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
9.2 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . 184
9.3 Controller design for speciﬁc disturbance rejection and overall error
minimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
9.3.1 Qparametrization to meet speciﬁc speciﬁcations . . . . . 185
9.3.2 Qparametrization to minimize H
2
performance . . . . . 187
9.3.3 Design steps . . . . . . . . . . . . . . . . . . . . . . . . 188
viii Modeling and Control of Vibration in Mechanical Systems
9.4 Simulation and implementation results . . . . . . . . . . . . . . . 189
9.4.1 System models . . . . . . . . . . . . . . . . . . . . . . . 189
9.4.2 Rejection of speciﬁc disturbance and H
2
performance min
imization . . . . . . . . . . . . . . . . . . . . . . . . . . 190
9.4.3 Rejection of two disturbances with H
2
performance mini
mization . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
9.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
10 Blending Control for MultiFrequency Disturbance Rejection 197
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
10.2 Control blending . . . . . . . . . . . . . . . . . . . . . . . . . . 197
10.2.1 State feedback control blending . . . . . . . . . . . . . . 199
10.2.2 Output feedback control blending . . . . . . . . . . . . . 200
10.3 Control blending application in multifrequency disturbance rejec
tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
10.3.1 Problem formulation . . . . . . . . . . . . . . . . . . . . 203
10.3.2 Controller design via the control blending technique . . . 205
10.4 Simulation and experimental results . . . . . . . . . . . . . . . . 207
10.4.1 Rejecting highfrequency disturbances . . . . . . . . . . . 207
10.4.2 Rejecting a combined mid and high frequency disturbance 211
10.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
11 H
∞
Based Design for Disturbance Observer 215
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
11.2 Conventional disturbance observer . . . . . . . . . . . . . . . . . 216
11.3 A general form of disturbance observer . . . . . . . . . . . . . . . 217
11.4 Application results . . . . . . . . . . . . . . . . . . . . . . . . . 220
11.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
12 TwoDimensional H
2
Control for Error Minimization 227
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
12.2 2D stabilization control . . . . . . . . . . . . . . . . . . . . . . . 228
12.3 2D H
2
control . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
12.4 SSTW process and modeling . . . . . . . . . . . . . . . . . . . . 231
12.4.1 SSTW servo loop . . . . . . . . . . . . . . . . . . . . . . 232
12.4.2 Twodimensional model . . . . . . . . . . . . . . . . . . 233
12.5 Feedforward compensation method . . . . . . . . . . . . . . . . . 235
12.6 2D control formulation for SSTW . . . . . . . . . . . . . . . . . 243
12.7 2D stabilization control for error propagation containment . . . . 244
12.7.1 Simulation results . . . . . . . . . . . . . . . . . . . . . . 244
12.8 2D H
2
control for error minimization . . . . . . . . . . . . . . . 245
12.8.1 Simulation results . . . . . . . . . . . . . . . . . . . . . . 245
12.8.2 Experimental results . . . . . . . . . . . . . . . . . . . . 247
12.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
Table of Contents ix
13 Nonlinearity Compensation and Nonlinear Control 251
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
13.2 Nonlinearity compensation . . . . . . . . . . . . . . . . . . . . . 251
13.3 Nonlinear control . . . . . . . . . . . . . . . . . . . . . . . . . . 252
13.3.1 Design of a composite control law . . . . . . . . . . . . . 256
13.3.2 Experimental results in hard disk drives . . . . . . . . . . 257
13.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
14 Quantization Effect on Vibration Rejection and Its Compensation 261
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
14.2 Description of control system with quantizer . . . . . . . . . . . . 261
14.3 Quantization effect on error rejection . . . . . . . . . . . . . . . . 266
14.3.1 Quantizer frequency response measurement . . . . . . . . 266
14.3.2 Quantization effect on error rejection . . . . . . . . . . . 266
14.4 Compensation of quantization effect on error rejection . . . . . . . 269
14.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
15 Adaptive Filtering Algorithms for Active Vibration Control 275
15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
15.2 Adaptive feedforward algorithm . . . . . . . . . . . . . . . . . . 275
15.3 Adaptive feedback algorithm . . . . . . . . . . . . . . . . . . . . 277
15.4 Comparison between feedforward and feedback controls . . . . . 280
15.5 Application in Stewart platform . . . . . . . . . . . . . . . . . . . 280
15.5.1 Multichannel adaptive feedback AVC system . . . . . . . 280
15.5.2 Multichannel adaptive feedback algorithmfor hexapod plat
form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
15.5.3 Simulation and implementation . . . . . . . . . . . . . . 284
15.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
References 293
Index 305
Preface
This book is primarily intended for researchers and engineering practitioners in sys
tems and control, especially those engaged in the area of modeling and control of
vibrations in mechanical structures and systems. The book aims at empowering read
ers with a clear understanding of characteristics of various vibrations, their effects on
system stability and performance, and techniques for rejecting vibrations of different
frequency ranges and their limitations. Special attention is given to recently devel
oped vibration modeling and control techniques in high precision systems. Many
realworld examples are given to demonstrate the modeling and control techniques.
Vibration exists in a wide spectra of engineering systems such as hard disk drives,
automotives, aerospace and aeronautic systems, manufacturing systems, etc. Vibra
tion is undesirable in most engineering applications, lowering system performance,
wasting energy and creating unwanted noise. Although the problem of vibration
control has been studied for a long time, it remains and indeed becomes more chal
lenging in many applications such as precision engineering and hard disk drives,
where an extremely high positioning accuracy is required. Therefore, vibration con
trol has drawn more intensive efforts from researchers and engineering practitioners
in recent years. It is our intention in this book to present to readers some of the recent
developments in this ﬁeld.
The book presents the latest results in vibration modeling and advanced control
design for vibration attenuation in mechanical actuation systems to achieve high
precision positioning performance. It focuses on vibration and disturbance rejec
tions using recently developed control techniques for high precision positioning, and
demonstration of the beneﬁts gained from the applications of these techniques. The
theoretical developments and principles of control design are elaborated in detail so
that the reader can apply the techniques developed to obtain solutions with the help
of MATLAB
. Examples are presented throughout the book so that the subject can
be better understood. A number of simulation and experimental results with compre
hensive evaluations are provided in each chapter, except Chapters 1, 4, and 5, which
are dedicated to the review of related background knowledge.
The book summarizes a collective research effort which we have had the plea
sure to contribute to. Many results reported in the book are due to the collaboration
with Guoxiao Guo from Western Digital Corporation, Jianliang Zhang and Jul Nee
Teoh from Data Storage Institute (DSI) of Singapore, Youyi Wang from Nanyang
Technological University (NTU), and Frank Lewis from the University of Texas at
Arlington. The research work contained in this book was mainly performed at DSI
and the School of Electrical and Electronic Engineering (EEE) of NTU, Singapore.
xi
xii Modeling and Control of Vibration in Mechanical Systems
Algorithms applied in magnetic recording systems were implemented at DSI and
those in the Stewart platform at the School of EEE, NTU. We would like to express
our sincere appreciation to DSI for its supportive environment and vibrant research
atmosphere. We are also sincerely grateful to Dr. Ong Eng Hong and the colleagues
in Mechatronics and Recording Channel Division of DSI, and EEE, NTU for their
support.
Lihua Xie
Chunling Du
MATLAB
is a registered trademark of The MathWorks, Inc. For product in
formation, please contact: The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA
017602098 USA, Tel: 508 647 7000, Fax: 5086477001, Email: info@mathworks.
com, Web: www.mathworks.com.
List of Tables
2.1 σ values of the modeling error (
ˆ
d
1
−d
1
) for different p and Γ . . . 49
4.1 % reduction of σ values of PES, RRO and NRRO and disk vibration
amplitude with stacked disks compared with single disk . . . . . . . 81
6.1 Control performance comparison . . . . . . . . . . . . . . . . . . . 131
7.1 Control performance comparison. . . . . . . . . . . . . . . . . . . 146
9.1 Comparison of performance speciﬁcations . . . . . . . . . . . . . . 191
14.1 Quantization and friction effect . . . . . . . . . . . . . . . . . . . 272
xiii
List of Figures
1.1 The servo control loop of a hard disk drive. . . . . . . . . . . . . . 2
1.2 Hexapod from Micromega Dynamics. . . . . . . . . . . . . . . . . 3
1.3 Zoomedin view of the hexapod. . . . . . . . . . . . . . . . . . . . 4
1.4 An example of random vibration. . . . . . . . . . . . . . . . . . . . 7
1.5 Spectrum density of broadband vibration. . . . . . . . . . . . . . . 9
1.6 Spectrum density of narrowband vibration. . . . . . . . . . . . . . 10
1.7 Histogram of signal x. . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1 A read back signal of embedded servo. . . . . . . . . . . . . . . . . 20
2.2 Frequency responses of a second order transfer function. . . . . . . 21
2.3 Measured VCM Bode plots (straight lines: pure double integrator
k/s
2
). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Multiplicative uncertainty of a VCM. . . . . . . . . . . . . . . . . . 24
2.5 The operator z
r
versus x. . . . . . . . . . . . . . . . . . . . . . . . 26
2.6 Friction f versus actuator displacement x. . . . . . . . . . . . . . . 28
2.7 A PZT actuated suspension. . . . . . . . . . . . . . . . . . . . . . 28
2.8 Equivalent spring mass system of PZT microactuator. . . . . . . . . 29
2.9 A typical frequency response of the PZT microactuator. . . . . . . . 30
2.10 An opened 1.8inch hard disk drive. . . . . . . . . . . . . . . . . . 31
2.11 Measured and modeled frequency responses of the VCM actuation
system (LDV range 0.5 µm/V). . . . . . . . . . . . . . . . . . . . . 32
2.12 Closed control loop of a disk drive with a VCM actuator for friction
measurement via LDV. . . . . . . . . . . . . . . . . . . . . . . . . 33
2.13 Control signal u versus displacement x. . . . . . . . . . . . . . . . 34
2.14 VCM actuator modeling with friction nonlinearity model F(x). . . 34
2.15 Measured and modeled friction and error. . . . . . . . . . . . . . . 36
2.16 Actuator frequency response for sinusoidal reference with amplitude
of 1 and 3 V, respectively. . . . . . . . . . . . . . . . . . . . . . . . 37
2.17 Actuator frequency response for sinusoidal reference with amplitude
of 0.5 V. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.18 Preload and twoslope model for friction modeling. . . . . . . . . . 38
2.19 Plant input voltage u versus displacement x. . . . . . . . . . . . . . 38
2.20 Closedloop control system disturbances d
1
, d
2
and noise n. . . . . 39
2.21 Closedloop control system with disturbance and noise models. . . . 40
2.22 Sensitivity function S(z). . . . . . . . . . . . . . . . . . . . . . . . 41
2.23 PES NRRO spectrum. . . . . . . . . . . . . . . . . . . . . . . . . . 42
xv
xvi Modeling and Control of Vibration in Mechanical Systems
2.24 Threelayer RBF neural network. . . . . . . . . . . . . . . . . . . . 44
2.25 Original disturbance d
1
and the modeled
ˆ
d
1
. . . . . . . . . . . . . . 49
2.26 Power spectrum of
ˆ
d
1
. . . . . . . . . . . . . . . . . . . . . . . . . 50
2.27 NRRO power spectrum from measurement and disturbance models,
i.e., e = −P(s) ·
ˆ
d
1
−
ˆ
d
2
+ ˆ n. . . . . . . . . . . . . . . . . . . . . 51
2.28 Modeling error (d
1
−
ˆ
d
1
) for different Γ with p = 1. . . . . . . . . 52
3.1 Singleaxis system using piezoelectric stiff actuator. . . . . . . . . . 54
3.2 Linear discrete time adaptive ﬁlter. . . . . . . . . . . . . . . . . . . 55
3.3 Block diagram of the LMS adaptive ﬁlter. . . . . . . . . . . . . . . 57
3.4 System identiﬁcation using LMS adaptive ﬁlter. . . . . . . . . . . . 59
3.5 Frequency responses of a PZT actuator. . . . . . . . . . . . . . . . 60
3.6 Estimated and experimental frequency responses. . . . . . . . . . . 61
4.1 Disk and spindle motor assembly of the spin stand. . . . . . . . . . 67
4.2 Comparison of singledisk and dualdisk axial vibrations measured
via LDV at 7200 RPM. . . . . . . . . . . . . . . . . . . . . . . . . 69
4.3 Comparison of singledisk and dualdisk RRO power spectrum at
7200 RPM (23% improvement of σ value). . . . . . . . . . . . . . 70
4.4 Comparison of singledisk and dualdisk NRRO power spectrum at
7200 RPM (18% reduction of σ value). . . . . . . . . . . . . . . . 71
4.5 Comparison of singledisk and dualdisk PES in time domain at 7200
RPM (21% reduction of σ value). . . . . . . . . . . . . . . . . . . 72
4.6 Comparison of singledisk and dualdisk axial vibrations measured
via LDV at 8400 RPM. . . . . . . . . . . . . . . . . . . . . . . . . 73
4.7 Comparison of singledisk and dualdisk RRO power spectrum at
8400 RPM (41% improvement of σ value). . . . . . . . . . . . . . 74
4.8 Comparison of singledisk and dualdisk NRRO power spectrum at
8400 RPM (28% reduction of σ value). . . . . . . . . . . . . . . . 75
4.9 Comparison of singledisk and dualdisk PES in time domain at 8400
RPM (38% reduction of σ value). . . . . . . . . . . . . . . . . . . 76
4.10 Comparison of singledisk and dualdisk axial vibrations measured
via LDV at 10200 RPM. . . . . . . . . . . . . . . . . . . . . . . . 77
4.11 Comparison of singledisk and dualdisk RRO power spectrum at
10200 RPM (the 3rd and 5th harmonics reduced signiﬁcantly, 33%
improvement of σ value). . . . . . . . . . . . . . . . . . . . . . . . 78
4.12 Comparison of singledisk and dualdisk NRRO power spectrum at
10200 RPM (27% reduction of σ value). . . . . . . . . . . . . . . . 79
4.13 Comparison of singledisk and dualdisk PES in time domain at 10200
RPM (32% reduction of σ value). . . . . . . . . . . . . . . . . . . 80
4.14 Generic feedback control system. . . . . . . . . . . . . . . . . . . . 86
5.1 Conﬁguration of standard optimal control. . . . . . . . . . . . . . . 94
5.2 A closedloop system with uncertainty. . . . . . . . . . . . . . . . . 102
List of Figures xvii
5.3 A closedloop system with additive uncertainty for robust stability
analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.4 A closedloop system with multiplicative uncertainty for robust sta
bility analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.5 Control system structure for Youla parametrization. . . . . . . . . . 107
5.6 Sensitivity function for continuoustime system. . . . . . . . . . . . 110
5.7 Sensitivity function for discretetime system. . . . . . . . . . . . . 110
5.8 Sensitivity function in discretetime domain. . . . . . . . . . . . . . 113
6.1 Mixed H
2
/H
∞
control scheme for HDD servo loop with distur
bance models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.2 Frequency responses of the VCM actuator. . . . . . . . . . . . . . . 125
6.3 Multiplicative uncertainty of the VCM actuator. . . . . . . . . . . . 126
6.4 Frequency response of sensitivity functions. . . . . . . . . . . . . . 129
6.5 Frequency response of sensitivity functions. . . . . . . . . . . . . . 130
7.1 Parallel structure of a dualstage actuation system with disturbances
and noise injected. . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.2 Power spectrum of PES nonrepeatable runout in open loop. . . . . . 135
7.3 Decoupled structure of dualstage actuation systems. . . . . . . . . 136
7.4 Structure of H
∞
loop shaping. . . . . . . . . . . . . . . . . . . . . 136
7.5 Frequency responses of S
v
(s) and S
m
(s). . . . . . . . . . . . . . . 139
7.6 Frequency responses of P
v
(s)C
v
(s) (solid line) and P
m
(s)C
m
(s)(dotted
line). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.7 Frequency response of VCM actuator P(s). . . . . . . . . . . . . . 141
7.8 Frequency response of VCM controller C
v
(s). . . . . . . . . . . . . 142
7.9 Frequency response of microactuator controller C
m
(s). . . . . . . . 143
7.10 Sensitivity function S
m
(s) (solid) and its weighting function inverse
(dashed). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
7.11 Open loop frequency response of the dualstage system. . . . . . . . 144
7.12 Sensitivity and complementary sensitivity functions. . . . . . . . . 145
7.13 Microactuator frequency response. . . . . . . . . . . . . . . . . . . 146
7.14 Frequency response of microactuator controller C
m
(s). . . . . . . . 147
7.15 Open loop frequency response of the dualstage system. . . . . . . . 147
7.16 Sensitivity and complementary sensitivity functions. . . . . . . . . 148
7.17 Experimental structure. . . . . . . . . . . . . . . . . . . . . . . . . 149
7.18 Sensitivity function of the dualstage system (smooth line: simula
tion result; rough line: testing result; dotted line: PID design). . . . 150
7.19 Open loop frequency response of the dualstage system (smooth line:
simulation result; rough line: testing result.) . . . . . . . . . . . . . 150
7.20 Step response of the dualstage system. . . . . . . . . . . . . . . . 151
7.21 3σ of PES NRRO versus frequencies. . . . . . . . . . . . . . . . . 151
7.22 VCM controller C
v
(z). . . . . . . . . . . . . . . . . . . . . . . . . 155
7.23 Microactuator controller C
m
(z). . . . . . . . . . . . . . . . . . . . 155
7.24 Sensitivity function of the dualstage system. . . . . . . . . . . . . 156
xviii Modeling and Control of Vibration in Mechanical Systems
7.25 Sensitivity function. . . . . . . . . . . . . . . . . . . . . . . . . . . 156
7.26 Sensitivity function of the dualstage system. . . . . . . . . . . . . 157
7.27 Frequency responses of P
v
(z)C
v
(z) (solid curve) and P
m
(z)C
m
(z)
(dashed curve). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.28 Sensitivity function of the dualstage system. . . . . . . . . . . . . 159
7.29 Open loop frequency responses of the dualstage system. . . . . . . 159
7.30 Step response of the dualstage system. . . . . . . . . . . . . . . . 160
7.31 3σ value of PES NRRO versus frequency. . . . . . . . . . . . . . . 160
8.1 Q parameterization for control design. . . . . . . . . . . . . . . . . 163
8.2 Peak ﬁlter F in the nominal feedback loop. . . . . . . . . . . . . . 167
8.3 Peak ﬁlter in the frequency domain. . . . . . . . . . . . . . . . . . 167
8.4 Sensitivity functions before and after group peak ﬁltering activated. 170
8.5 PZT microactuator attached to VCM actuator arm. . . . . . . . . . 170
8.6 Power spectrum of the position error before servo control. . . . . . 171
8.7 PZT micro actuator frequency response. . . . . . . . . . . . . . . . 172
8.8 Sensitivity functions before and after the KYP lemmabased design:
simulation result. . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
8.9 Openloop Bode plot before and after the KYP lemmabased design. 175
8.10 Structure of experimental setup. . . . . . . . . . . . . . . . . . . . 175
8.11 Sensitivity functions before and after the KYP lemmabased design:
experimental results. . . . . . . . . . . . . . . . . . . . . . . . . . 176
8.12 σ value of PES versus frequency. . . . . . . . . . . . . . . . . . . . 176
8.13 Frequency response of the PZT microactuator. . . . . . . . . . . . . 179
8.14 PES NRRO power spectrum calculated from measured PES signal
without servo control, reﬂecting the vibration distribution of the sys
tem (3σ = 21 nm including the noise 3σ = 15.2 nm). . . . . . . . . 180
8.15 Comparison of sensitivity functions. . . . . . . . . . . . . . . . . . 181
8.16 Open loop frequency responses (PLPF (GM: 6 dB, PM: 50 deg.,
Bandwidth 1.4kHz)); KYP(GM: 6 dB, PM: 34 deg., Bandwidth: 1.7
kHz))). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
8.17 NRRO power spectrum with PLPF and KYP (50% reduction before
1 kHz). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
9.1 H
2
control scheme with Q parametrization for controller design. . . 184
9.2 Frequency response of a PZT microactuator. . . . . . . . . . . . . . 190
9.3 Openloop frequency responses. . . . . . . . . . . . . . . . . . . . 191
9.4 Designed sensitivity functions. . . . . . . . . . . . . . . . . . . . . 192
9.5 Comparison of sensitivity functions obtained from experiment. . . . 192
9.6 NRRO power spectrum with KYP Lemmabased controller with and
without H
2
minimization. . . . . . . . . . . . . . . . . . . . . . . 193
9.7 PES NRRO spectrum without servo control. . . . . . . . . . . . . . 194
9.8 Openloop frequency response. . . . . . . . . . . . . . . . . . . . . 195
9.9 Resultant sensitivity function (Solid line: with Spec. (i), (ii) and (iii);
Dashed line: with Spec. (i) and (ii)). . . . . . . . . . . . . . . . . . 195
List of Figures xix
9.10 Resultant sensitivity function with all the three requirements ful
ﬁlled. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
9.11 NRRO power spectrum with rejection of two speciﬁc disturbances at
0.65 and 2 kHz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
10.1 Blending control scheme. . . . . . . . . . . . . . . . . . . . . . . . 198
10.2 Control loop with injected disturbances at different frequencies. . . 204
10.3 Control structure. . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
10.4 Frequency response of the VCM actuator. . . . . . . . . . . . . . . 207
10.5 Openloop frequency response with disturbance rejection at 4 and 8
kHz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
10.6 Simulated sensitivity function with disturbance rejection at 4 and 8
kHz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
10.7 Measured (solid curve) sensitivity function with disturbance rejec
tion at 4 and 8 kHz. . . . . . . . . . . . . . . . . . . . . . . . . . . 210
10.8 Openloop with disturbance rejection at 3, 6.5, and 10 kHz. . . . . . 210
10.9 Sensitivity function with disturbance rejection at 3, 6.5, and 10 kHz. 211
10.10 Openloop with disturbance rejections at 0.65 and 2 kHz. . . . . . . 212
10.11 Sensitivity function with disturbance rejections at 0.65 and 2 kHz. . 213
11.1 Block diagram of the control loop with a conventional disturbance
observer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
11.2 Block diagram of the control loop with a general disturbance ob
server. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
11.3 Frequency response of the designed Q(z). . . . . . . . . . . . . . . 222
11.4 The sensitivity functions without and with the general disturbance
observer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
11.5 The sensitivity function comparison with the general and the con
ventional disturbance observers. . . . . . . . . . . . . . . . . . . . 223
11.6 Disturbance d
1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
11.7 Error signal e. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
11.8 Measured sensitivity functions without and with the general distur
bance observer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
11.9 Comparison of T
EQ−OL
about the general disturbance observer. . . 225
11.10 Comparison of T
EQ−OL
about conventional disturbance observer. . 226
12.1 SSTW process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
12.2 SSTW servo loop with disturbances and noise models. . . . . . . . 232
12.3 PES NRROand its σ values versus track number during propagation.
(The time sequence and the σ value increase with the track number.) 233
12.4 SSTW servo loop modeling in two dimensions. . . . . . . . . . . . 234
12.5 SSTW servo loop. . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
12.6 Frequency response of a VCM actuator. . . . . . . . . . . . . . . . 237
12.7 Frequency response of the closedloop transfer function with the PD
controller, PID controller, and H
2
controller. . . . . . . . . . . . . 238
xx Modeling and Control of Vibration in Mechanical Systems
12.8 Φ versus frequency. . . . . . . . . . . . . . . . . . . . . . . . . . 239
12.9 3σ of PES NRRO. . . . . . . . . . . . . . . . . . . . . . . . . . . 240
12.10 Frequency response of the H
2
controller. . . . . . . . . . . . . . . 241
12.11 Frequency response of the openloop system with the H
2
controller. 241
12.12 Comparison of sensitivity functions. . . . . . . . . . . . . . . . . . 242
12.13 σ value of PES NRRO versus track number. . . . . . . . . . . . . . 243
12.14 2D controller for SSTW servo loop. . . . . . . . . . . . . . . . . 245
12.15 σ of PES NRRO versus track number. . . . . . . . . . . . . . . . . 246
12.16 Openloop frequency response with stabilization controller. . . . . 246
12.17 Sensitivity function with stabilization controller (A
c2
, B
c2
, C
c
, D
c
). 247
12.18 Frequency response of controller (A
c2
, B
c2
, C
c
, D
c
). . . . . . . . . 248
12.19 Openloop frequency response with controller (A
c2
, B
c2
, C
c
, D
c
). . 249
12.20 Sensitivity function with controller (A
c2
, B
c2
, C
c
, D
c
). . . . . . . 249
12.21 Step response (Channel 1/2/3: Reference/Output/Control signal). . 250
13.1 Friction compensation for the actuation system. . . . . . . . . . . . 252
13.2 Input u versus displacement x with and without compensation. . . . 253
13.3 Actuator frequency responses with and without friction compensa
tion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
13.4 Actuator frequency responses with friction compensation for differ
ent displacements in voltage with 0.5µm/V (Straight smooth lines:
the pure double integrator). . . . . . . . . . . . . . . . . . . . . . . 254
13.5 Actuator frequency responses with friction compensation for differ
ent displacements in voltage with 0.5 µm/V (Straight smooth lines:
the pure double integrator). . . . . . . . . . . . . . . . . . . . . . . 254
13.6 Sensitivity functions with and without friction compensation. . . . . 255
13.7 Control structure of a plant P(s) with Youla parametrization ap
proach and adaptive nonlinear compensation. . . . . . . . . . . . . 256
13.8 Comparison of error rejection frequency response without and with
u
N
of different p and Γ. . . . . . . . . . . . . . . . . . . . . . . . . 258
13.9 NRRO power spectrum with KYP lemmabased linear control and
nonlinear compensation (80% reduction before 400 Hz). . . . . . . 259
14.1 Experimental setup. . . . . . . . . . . . . . . . . . . . . . . . . . . 263
14.2 Frequency response of the VCM actuator measured by injecting a
swept sine wave of 5mV amplitude. . . . . . . . . . . . . . . . . . 263
14.3 The servo loop in experiment. . . . . . . . . . . . . . . . . . . . . 264
14.4 Frequency response of the controller C(z). . . . . . . . . . . . . . 264
14.5 Frequency response of the sensitivity function S(z) with different
reference levels (i.e., actuator moving ranges are different). . . . . . 265
14.6 Frequency response of the quantizer before and after compensation
(bit number n = 6). . . . . . . . . . . . . . . . . . . . . . . . . . . 267
14.7 Frequency response of the quantizer with compensation (bit number
n = 8). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
List of Figures xxi
14.8 Frequency response of the quantizer with compensation (bit number
n = 10). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
14.9 Measured sensitivity function S
Q
(z) with different bits n. . . . . . . 268
14.10 Sensitivity function S
Q
(f) with f = 10 Hz versus bit n. . . . . . 269
14.11 Compensation scheme of quantization effect. . . . . . . . . . . . . 270
14.12 Choosing the threshold δ and the scaling factor a. . . . . . . . . . . 270
14.13 Sensitivity function with quantization compensation. . . . . . . . . 271
15.1 Block diagram of FXLMS algorithm. . . . . . . . . . . . . . . . . 276
15.2 FilteredX LMS adaptive feedback algorithm. . . . . . . . . . . . . 279
15.3 Adaptive inverse control scheme. . . . . . . . . . . . . . . . . . . 279
15.4 Block diagram of 2 × 2 adaptive feedback algorithm. . . . . . . . . 282
15.5 Block diagram of 6 × 1 FXLMS adaptive feedback control system. 283
15.6 General layout of the experimental setup. . . . . . . . . . . . . . . 286
15.7 60 Hz error signal in dB unit. . . . . . . . . . . . . . . . . . . . . 287
15.8 210 Hz error signal in dB unit. . . . . . . . . . . . . . . . . . . . . 287
15.9 240 Hz error signal in dB unit. . . . . . . . . . . . . . . . . . . . . 288
15.10 270 Hz error signal in dB unit. . . . . . . . . . . . . . . . . . . . . 288
15.11 Simulink diagram of automatic gain control. . . . . . . . . . . . . 291
15.12 180 Hz error signal without automatic gain control. . . . . . . . . . 292
15.13 180 Hz error signal with automatic gain control. . . . . . . . . . . 292
Symbols and Acronyms
R
n
: ndimensional real Euclidean space
R
n×m
: set of n ×m real matrices
I
n
: n × n identity matrix
(A, B, C, D): statespace representation of a system
A B
1
B
2
C
1
D
11
D
12
C
2
D
21
D
22
: compact representation of system:
x(k + 1) = Ax(k) + B
1
w(k) + B
2
u(k)
z(k) = C
1
x(k) +D
11
w(k) +D
12
u(k)
y(k) = C
2
x(k) +D
21
w(k) +D
22
u(k)
diag{A
1
, A
2
, · · · , A
n
}: block diagonal matrix with A
j
( not necessarily
square), j = 1, 2, · · · , n, on the diagonal
X
T
: transpose of matrix X
X
∗
: complex conjugate transpose of matrix X
P ≥ 0: symmetric positive semideﬁnite matrix P ∈ R
n×n
P > 0: symmetric positive deﬁnite matrix P ∈ R
n×n
P ≥ Q: P −Q ≥ 0 for symmetric P, Q ∈ R
n×n
P > Q: P −Q > 0 for symmetric P, Q ∈ R
n×n
¯ σ(X): largest singular value of X
Trace(X): trace of X
· : Euclidean vector norm
w
2
: ℓ
2
norm of a signal {w(k)}, i.e.,
∞
k=0
w(k))
2
.
xxiii
xxiv Modeling and Control of Vibration in Mechanical Systems
ℓ
2
{[0, ∞)}: space of square summable sequences on {[0, ∞)}.
The signal {w(k)} is said to be from ℓ
2
{[0, ∞)} or simply ℓ
2
if w
2
< ∞.
G
2
: H
2
norm of transfer function G
G
∞
: H
∞
norm of transfer function G
Re( ): the real part of a complex number
Im( ): the imaginary part of a complex number
ρ( ): spectral radius
AGC : automatic gain control
AV C : active vibration control
deg: degree
det: determinant
DSA: Dynamic Signal Analyzer
FFT: fast Fourier transform
FXLMS: ﬁlteredX LMS
HDD: hard disk drive
LDV : Laser Doppler Vibrometer
LFT: Linear fractional transformation
LMI: linear matrix inequality
LMS: least mean square
LQG: linear quadratic Gaussian
LTR: loop transfer recovery
MEMS: micro electromechanical system
MSE: mean square error
NRRO: nonrepeatable runout
PES: position error signal
Symbols and Acronyms xxv
PID : proportionalintegralderivative
PLPF: phase lead peak ﬁlter
PZT: lead zirconate titanate/piezoelectric
RBF: radial basis function
RMS: root mean square
RPM: rotations per minute
RRO: repeatable runout
SSTW: selfservo track writing
STW: servo track writing
TMR: track misregistration
V CM: voice coil motor
1
Mechanical Systems and Vibration
When studying mechanical systems, we have to include the subject of dynamics and
vibration. Dynamics is a branch of mechanics that deals with motion and its effect on
a body. Unlike statics, which deals with bodies at rest, dynamics takes into account
the effect of velocities and accelerations on the forces acting on bodies. Vibration is
regarded as a branch of dynamics, since forces and masses are taken into account in
vibration analysis. Thus it is natural that both dynamics and vibrations of mechanical
systems are studied in this book.
A vibration may be a signal, force, or temperature variation that affects the re
sponse of a system in an unacceptable manner. If our analysis of system response to
a vibration shows that it regularly affects the system performance in an unacceptable
manner, we need to alter or control the vibration response and bring the response
within acceptable levels by adding appropriate forces, called control forces which
are functions of system response such as displacement. This may require a design
by using the value of the system response to generate additional forces according to
certain rules or laws such that the modiﬁed response behaves according to desired
performance and within certain bounds. This results in a closedloop system that
incorporates feedback controls. The purpose of designing a system with feedback
force is to minimize unwanted behavior. Examples include magnetic recording sys
tems, Stewart platforms, positioningstages [27], the atomic force microscope (AFM)
[28], industry robots [29], as well as some automotive systems [34]. The magnetic
recording system and Stewart platform are examples to be presented next to help
in ﬁxing these ideas more ﬁrmly. With such a grounding, more advanced problems
become accessible.
1.1 Magnetic recording system
Figure 1.1 shows a servo control loop of a hard disk drive (HDD) with a voice coil
motor (VCM) and a piezoelectric (PZT) actuated servo system. It consists of a stack
of ﬂat rotating disks with positioning information or servo information embedded in
their surfaces. The servo information is used to position the magnetic heads on the
disk surfaces. Position measurement of the magnetic heads is achieved by means of
analyzing the position error signal (PES) calculated from the read back signal. To
1
2 Modeling and Control of Vibration in Mechanical Systems
have a disk drive with a high storage capacity, the head positioning error with re
spect to the target track center needs to be as small as possible. The error is mainly
due to, (1) torque disturbances from spindle motor, (2) actuator pivot friction, (3)
airﬂowinduced nonrepeatable disk, suspension and slider vibrations, (4) mechani
cal resonance vibration, and (5) head sensing and electronic noises, media noise, and
quantization noises. Hence how to deal with the variety of the disturbances is crit
ical to the head positioning accuracy, and subsequently the track density for a high
capacity disk drive.
FIGURE 1.1
The servo control loop of a hard disk drive.
1.2 Stewart platform
The sixleg parallel linkage mechanism known as the “Stewart platform” was dis
covered as early as in 1965 [30]. It is designed according to a cubic conﬁguration,
consisting of two triangular parallel plates connected to each other by six active legs
orthogonal to each other. Each leg is equipped with a voice coil actuator, a force
sensor and two ﬂexible joints. The closed kinematical linkage structure of a Stewart
platform has major advantages over any serial link robots: great rigidity, high force
to weight ratio, six degrees of freedom (DOF), etc. [31] The Stewart platform is
Mechanical Systems and Vibration 3
widely used as space born structures, as well as a high precision pointing device and
vibration isolator.
Stewart platforms can be divided into two main classes according to the stiffness
of the legs: stiff type and soft type. For the soft design, each leg essentially acts as
an axial springing parallel with a voice coil actuator, while the stiff design involves
piezoelectric or magneto restrictive legs whose extensions can be controlled [32].
The Stewart platform has been widely used in active vibration control. It has an
important property for vibration control application: forces transmitted between the
mobile plate and the base plate are totally axial forces of actuators. This implies that
if the axial forces can be measured and eliminated, the vibration created by these
forces can thus be eliminated. Thus the Stewart platform has become one of the
most popular approaches for 6DOF active vibration control in precision systems
due to its attractive properties.
FIGURE 1.2
Hexapod from Micromega Dynamics.
The Stewart platform (Hexapod) from MicromegaDynamics used as a vibration
isolation device is shown in Figure 1.2. The hexapod has a cubic architecture and
consists of two parallel plates connected to each other by six active legs. The plates
are made of aluminum with a thickness of 20 mm and diameter of 250 mm, with
the weight of the mobile plate at 1 kg. Each leg of the active interface consists of
a linear piezoelectric actuator, a collocated force sensor, and ﬂexible tips to connect
the two end plates. Flexible tips are used in order to avoid the problem of friction and
backlash, which comes with the use of spherical joints. The hexapod can be used to
4 Modeling and Control of Vibration in Mechanical Systems
FIGURE 1.3
Zoomedin view of the hexapod.
actively increase the structural damping of ﬂexible systems attached to it.
Figure 1.3 shows the zoomedin view of the inside of the hexapod, revealing the
arrangement of the six collocated sensoractuator legs. The wires shown in Figure
1.2 are the outputs of the force sensors (each for one sensor) and inputs to the piezo
electric actuators (each for one actuator), respectively.
Each of the legs in the Stewart platform consists of a PZT force sensor and an
ampliﬁed PZT actuator. They form a collocated sensoractuator pair conﬁguration.
If an actuator and a sensor are collocated, the associated signals for the actuator
and sensor are power conjugated, i.e., the product of the actuated velocity and the
measured force represents the power that is extracted from the mechanical structure.
A collocated actuatorsensor pair thus enables the control of power that is supplied
to the mechanical structure. Collocated actuatorsensor pairs are suitable in active
vibration control applications in the sense that they guarantee damping and stability
robustness if designed properly [33].
In this book, the modeling and vibration controls for the magnetic recording sys
tem and the Stewart platform will be detailed.
1.3 Vibration sources and descriptions
Any oscillatory motion of bodies that repeatedly appears is called vibration or os
cillation. There are usually forces associated with vibrations. They can be induced
by various types of excitation. Some of them are: ﬂuid ﬂow; rotating unbalanced
machinery; structure ﬂexible modes; electrical torque; reciprocating machinery; mo
tion induced in vehicles traveling over uneven surfaces; and ground motion caused
by earthquakes.
Flow induced vibration is generated by the forces exerted on an object by ﬂuid
motion. Such situations can be complicated by the fact that the motion of the vibrat
Mechanical Systems and Vibration 5
ing object can alter the ﬂuid ﬂow conditions, thus changing the ﬂuid forces. Another
complicating factor is the mass of the ﬂuid, which increases the effective mass of the
system. Examples of vibration caused by ﬂuid motion include: wave action on struc
ture; vortexinduced vibration such as vibration of transmission cables, underwater
cables used for towing and structural support, and cooling towers and chimneys; vi
bration caused by internal ﬂows, such as air ﬂow in hard disk drives, ﬂow through
pipes and hoses having bends; structural vibration caused by ﬂuctuating aerodynamic
forces such as turbulence [1][56]. In certain situations, the steadystate excitation
force due to ﬂuid motion is sinusoidal with an amplitude proportional to the square
of the forcing frequency. A model of the system undergoing such excitation is
m¨ x +c ˙ x + kx = F
0
ω
2
sinωt, (1.1)
where m is the system mass, x is the system response, ω is the forcing frequency, k
and c are respectively stiffness and damping, and F
0
is the force coefﬁcient. Further
analysis of system response to ﬂow ﬂuid motion is quite complicated and requires
detailed consideration of ﬂuid mechanics of the system.
Some oscillatory systems have simple harmonic motion of the form
y(t) = Bsin(ωt) + Ccos(ωt), (1.2)
where y(t) is the displacement of a mass, and is equivalent to
y(t) = Asin(ωt + φ), (1.3)
A =
B
2
+C
2
, (1.4)
cos(φ) =
B
A
, sin(φ) =
C
A
. (1.5)
There occur some oscillations having exponential amplitude as follows:
y(t) = Ae
rt
sin(ωt + φ), (1.6)
where oscillation amplitude decays exponentially when r < 0, and grows indeﬁ
nitely when r > 0.
Many types of motion cannot be easily represented by simple functions because
they are essentially random. Examples include air turbulence to arm and suspension
in hard disk drives, ground motion due to an earthquake, and base motion that occurs
when a vehicle travels over an uneven surface. It is possible, however, to characterize
them by means of statistical averages and spectrum plots, in which Fourier analysis
is used to identify the major frequency components in the vibration and they will be
described in more detail in the later part of the chapter.
1.4 Types of vibration
There are several ways to categorize vibrations. Basically, they can be calssiﬁed as
follows.
6 Modeling and Control of Vibration in Mechanical Systems
1.4.1 Free and forced vibration
If a system vibrates on its own after an initial disturbance and no external force acts
on it, the ensuing vibration is known as free vibration. A direct example of free
vibration is the oscillation of a simple pendulum.
If a system vibration is due to an external force, the arising vibration is known as
forced vibration. The oscillation in machines such as diesel engines that results from
an external force is an example of forced vibration. If the frequency of the external
force coincides with one of the natural frequencies of the system, the phenomenon
known as resonance occurs, and the system undergoes oscillation. The occurrence of
that resonance causing large oscillation may lead to failures of some structures such
as buildings, bridges, turbines, and airplane wings.
1.4.2 Damped and undamped vibration
If during oscillation there is no energy lost or dissipated in friction or other resistance,
the vibration is known as undamped vibration. On the other hand, if there is energy
lost during oscillation, it is called damped vibration. When analyzing vibration near
resonance in physical systems, consideration of damping becomes extremely impor
tant.
1.4.3 Linear and nonlinear vibration
If all the basic components in a vibratory system such as spring, mass and damper
behave linearly, the resulting vibration is classiﬁed as linear vibration. On the other
hand, if any of the basic components behaves nonlinearly, the vibration is categorized
as nonlinear vibration. Linear and nonlinear differential equations are used to govern
the behaviors of linear and nonlinear vibratory systems, respectively. If a vibration is
linear, the principle of linear systems such as superposition holds, and there are well
developed mathematical tools for analysis. As for nonlinear vibration, the superposi
tion principle is not valid, and techniques of analysis are more complicated and less
well known. Since all vibratory systems tend to behave nonlinearly with respect to
amplitude level of oscillation, some knowledge of nonlinear vibration is desirable in
dealing with practical vibratory systems. As is known, a describing function is one
approximation method used to analyze nonlinear vibratory systems.
1.4.4 Deterministic and random vibration
A vibration is known as deterministic vibration if it results from an excitation with
value or amplitude known at any given time. In some cases, the excitation acting on
a vibratory system is nondeterministic or random, and the value of the excitation at
a given time cannot be predicted. In these cases, a large amount of excitation data
collected may exhibit some statistical regularity. Statistical methods can be used for
analysis, as it is possible to estimate averages such as the mean and variance values
of the random excitation. Examples of random excitations are air ﬂow inside hard
Mechanical Systems and Vibration 7
disk drives, road roughness and ground motion during earthquakes. If the excitation
is random, the induced vibration is called random vibration, such as that shown in
Figure 1.4. It can be described in terms of statistical quantities.
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
Time(sec)
V
i
b
r
a
t
i
o
n
s
i
g
n
a
l
FIGURE 1.4
An example of random vibration.
1.4.5 Periodic and nonperiodic vibration
Periodic vibration can be represented by Fourier series as superposition of harmonic
components of various frequencies. That is, if x(t) is a periodic function with period
τ, its Fourier series representation is given by
x(t) =
a
0
2
+ a
1
cos(ωt) +a
2
cos(2ωt) + · · · + b
1
sin(ωt) + b
2
sin(2ωt) +...
=
a
0
2
+
∞
n=1
(a
n
cos(nωt) + b
n
sin(nωt)), (1.7)
where ω = 2π/τ is the fundamental frequency, and a
0
, a
n
, b
n
are constant coefﬁ
cients given by
a
0
=
ω
π
2π/ω
0
x(t)dt =
2
τ
τ
0
x(t)dt, (1.8)
8 Modeling and Control of Vibration in Mechanical Systems
a
n
=
ω
π
2π/ω
0
x(t)cos(nωt)dt =
2
τ
τ
0
x(t)cos(nωt)dt, (1.9)
b
n
=
ω
π
2π/ω
0
x(t)sin(nωt)dt =
2
τ
τ
0
x(t)sin(nωt)dt. (1.10)
Let
c
0
=
a
0
2
, (1.11)
c
n
= (a
2
n
+b
2
n
)
1/2
. (1.12)
The root mean square (RMS) value of the periodic function x(t) can be determined
as
RMS(x(t)) =
∞
n=0
c
n

2
. (1.13)
Thus the mean square value of x(t) is given by the sum of the squares of the absolute
values of the Fourier coefﬁcients. Equation (1.13) is known as Parseval’s formula for
periodic functions.
Although the series in (1.7) is an inﬁnite sum, most periodic functions can be
approximated by only a few harmonic functions.
A nonperiodic vibration x(t) can be represented by the integral Fourier transform
pair:
x(t) =
1
2π
∞
−∞
X(ω)e
iωt
dω (1.14)
and
X(ω) =
∞
−∞
x(t)e
−iωt
dt. (1.15)
The RMS value of the nonperiodic function x(t) can be determined as
RMS(x(t)) =
∞
−∞
X(ω)
2
2πτ
dω. (1.16)
Equation (1.16) is known as Parseval’s formula for nonperiodic functions.
1.4.6 Broadband and narrowband vibration
A broadband vibration is a stationary random process whose spectral density func
tion has signiﬁcant values over a range or band of frequencies which is approximately
of the same order of magnitude as the center frequency of the band. The density in
Mechanical Systems and Vibration 9
Figure 1.5 describes a broadband vibration, which is composed of components con
taining frequencies over a wide or broad frequency range. A narrowband random
vibration is a stationary vibration whose spectral density function has signiﬁcant val
ues only in a range of frequency whose width is small compared to the magnitude
of the center frequency. Figure 1.6 shows a vibration containing frequencies over a
narrow band.
10
1
10
2
10
3
10
4
0
2
4
6
8
10
12
14
16
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
FIGURE 1.5
Spectrum density of broadband vibration.
Arandomprocess whose power spectral density is constant over a frequency range
is called white noise. It is called ideal white noise if the band of frequencies is
inﬁnitely wide. An ideal white noise is physically unrealizable, since the variance of
such a random process would be inﬁnite because the area under the spectrum would
be inﬁnite. It is called bandlimited white noise if the band of frequencies has ﬁnite
cutoff frequencies ω
1
and ω
2
. The variance of a bandlimited white noise is given
by the total area under the spectrum, namely, 2S
0
(ω
2
− ω
1
), where S
0
denotes the
constant value of the spectral density.
10 Modeling and Control of Vibration in Mechanical Systems
10
2
10
3
0
1
2
3
4
5
6
7
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
FIGURE 1.6
Spectrum density of narrowband vibration.
Mechanical Systems and Vibration 11
1.5 Random vibration
1.5.1 Random process
In contrast to deterministic excitations, in many applications the inputs are not well
known and can be described only in terms of statistical measures such as mean value,
variance and standard deviation.
Mean value
The average is called the mean or expected value, often represented by µ. For dis
crete values, such as those produced by digital data acquisition, the mean is deﬁned
as
E(x) =
1
n
n
j=1
x
j
. (1.17)
For a continuous function x(t), the mean is
E[x(T
d
)] =
1
T
d
T
d
0
x(t)dt (1.18)
where the time duration of the data sample is T
d
.
Variance
Two signals may have the same mean value but one may ﬂuctuate with greater
amplitude about the mean. So we also need a measure of the range of ﬂuctuation.
Simply specifying the minimumand maximum values is insufﬁcient because a single
large ﬂuctuation (above or below the mean) can be misleading. A measure that
indicates the spread about the mean is the variance, which is deﬁned as the average
value of the square of the difference between the signal and its mean. The variance
is calculated for a discrete signal as follows.
var(x) = σ
2
=
1
n
n
j=1
(x
j
−µ)
2
, (1.19)
where σ
2
is the variance and σ is called the standard deviation.
For a signal continuous in time, the variance is calculated from
var(x, T
d
) =
1
T
d
T
d
0
[x(t) −µ]
2
dt. (1.20)
The meansquare value of x is the expected value of x
2
and is denoted E(x
2
). The
RMS value of x is
E(x
2
). The relation between the variance, the mean square,
and the mean is as follows.
σ
2
= E(x
2
) − [E(x)]
2
. (1.21)
12 Modeling and Control of Vibration in Mechanical Systems
If the mean of x is zero, the standard deviation σ
x
of x is calculated from
σ
x
=
E(x
2
), (1.22)
and is thus the same as the RMS value.
A random input will generate a random response, and the output signal given by
its sensor will appear to be random. Random signals, like the example shown in
Figure 1.4, have no apparent pattern and never repeat. Another characteristic of a
random signal is that it is impossible to predict what the signal will be in the future,
even if we have the past values of the signal.
1.5.2 Stationary random process
A special case of a random process is one that is stationary, which means that its
statistical properties (such as its mean and variance) are timeindependent.
If x(t) is stationary, its mean and covariance will be independent of t:
E[x(t)] = E[x(t +τ)] = µ, (1.23)
and
E[x(t)x(t + τ)] = σ
xx
(τ). (1.24)
1.5.3 Gaussian random process
A Gaussian or normal random process has a number of remarkable properties that
permit the computation of random vibration characteristics in a simple manner. The
probability density function of a Gaussian process x(t) is given by
p(x) =
1
√
2πσ
x
e
−
1
2
(
x−¯ x
σ
x
)
2
, (1.25)
where ¯ x and σ
x
denote the mean value and standard deviation of x. The mean ¯ x and
standard deviation σ
x
of x(t) vary with t for a nonstationary process but are constants
for a stationary process. A very important property of a Gaussian process is that the
form of its probability distribution is invariant with respect to linear operations. This
means that if the excitation of a linear system is a Gaussian process, the steadystate
response is generally a different random process, but still a normal one. The only
changes are that the magnitude of the mean and standard deviation of the response
are different from those of the excitation.
The graph of a Gaussian probability density function has a bellshaped envelope
as seen in Figure 1.7, and is symmetric about the mean value; its spread is governed
by the value of the standard deviation.
Mechanical Systems and Vibration 13
−4 −3 −2 −1 0 1 2 3 4
0
20
40
60
80
100
120
140
160
180
Signal x
P
r
o
b
a
b
i
l
i
t
y
d
e
n
s
i
t
y
FIGURE 1.7
Histogram of signal x.
1.6 Vibration analysis
1.6.1 Fourier transform and spectrum analysis
We know that a periodic signal can be expressed as a Fourier series of harmonic
functions. The Fourier series coefﬁcients, when plotted versus frequency, gives a
plot of the spectrum of the signal. The spectrum graphically displays the frequency
content of the signal.
A nonperiodic function is expressed with the following Fourier transform pair:
x(t) =
∞
−∞
a(ω)cos(ωt)dω +
∞
−∞
b(ω)sin(ωt)dω, (1.26)
where
a(ω) =
1
2π
∞
−∞
x(t)cosωtdt, (1.27)
b(ω) =
1
2π
∞
−∞
x(t)sinωtdt. (1.28)
14 Modeling and Control of Vibration in Mechanical Systems
The Fourier transform of x(t) is
X(ω) = a(ω) − ib(ω). (1.29)
An equivalent form of the Fourier transform is given by
x(t) =
∞
−∞
X(ω)e
iωt
dω, (1.30)
and
X(ω) =
1
2π
∞
−∞
x(t)e
−iωt
dt. (1.31)
The spectrum of a nonperiodic signal is the magnitude of its Fourier transform,
that is, X(ω). Equation (1.31) implies that the transform is symmetric about ω = 0,
that is
X(−ω) = X(ω). (1.32)
The transform of a timeshifted signal x(t − d) is X(ω)e
−iωd
and its spectrum is
the same as the spectrum of x(t), while the time shift d affects only the phase angle
of the transform.
1.6.2 Relationship between the Fourier and Laplace transforms
The Fourier transform X(ω) of x(t) is related to the Laplace transform X(s) as
follows.
X(ω) =
1
2π
X(s)
s=iω
. (1.33)
1.6.3 Spectral analysis
The spectral density, or power spectral density, S
xx
(ω), is one of the most useful
functions in vibration testing. It is deﬁned as
S
xx
(ω) =
1
2π
∞
−∞
R
xx
(τ)e
−iωτ
dτ, (1.34)
where
R
xx
(τ) = lim
T→∞
1
T
T/2
−T/2
x(t)x(t + τ)dt (1.35)
is the autocorrelation function and T is the time duration.
Mechanical Systems and Vibration 15
R
xx
(τ) can be written as the inverse Fourier transform of S
xx
(ω):
R
xx
(τ) =
∞
−∞
S
xx
(ω)e
iωτ
dω. (1.36)
The crossspectral density is written as
S
xy
(ω) =
1
2π
∞
−∞
R
xy
(τ)e
−iωτ
dτ, (1.37)
where
R
xy
(τ) = lim
T→∞
1
T
T/2
−T/2
x(t)y(t +τ)dt (1.38)
is the crosscorrelation function.
A very useful relation is that
R
xx
(0) =
∞
−∞
S
xx
(ω)dω = E(x
2
), (1.39)
which means that the meansquare value can be computed from the spectral density.
2
Modeling of Disk Drive System and Its
Vibration
2.1 Introduction
Modeling plays an important role in any control system analysis and design. A
physical system has to be modeled in order to design a control system. Physical
systems are more or less nonlinear and may vary with time. Researchers have made
many attempts to deal with a physical system: approximating it with a linear model
at an operating point, ﬁnding a control strategy that is robust and adaptable to the
changes in the physical system.
The system modeling in this chapter is discussed from the ﬁrst principle of actu
ators. System dynamics measurement in the frequency domain is used to determine
the parameters of a model in the form of transfer function. To have a more realistic
model, uncertainties, especially due to high frequency unmodeled dynamics, have to
be involved inherently. Moreover, nonlinearity induced by friction is measured under
the condition that the actuator is controlled to avoid unsteady signal measurement.
The hysteresis of friction versus actuator position is then obtained from the measure
ment in closedloop. An operator based modeling approach is adopted to model the
hysteresis, and an optimal model is obtained by minimizing the energy gain between
the position and the modeling error.
In this chapter, vibration modeling is based primarily on the spectral decomposi
tion of the error signal measured in a closedloop system. A decoupling procedure
is proposed which leads to approximate disturbance and noise models. Particularly,
lowfrequency disturbances are modeled as the output of an adaptive nonlinear mech
anism with the error signal as the input.
2.2 System description
A hard disk drive as shown in Figure 1.1 includes ﬁve major parts: baseplate and
cover, spindle and motor assembly, actuator assembly, disk, head/suspension assem
bly, and electronics card.
17
18 Modeling and Control of Vibration in Mechanical Systems
The spindle and motor assembly includes disk clamps to clamp disks. The actuator
assembly contains an actuator driven by voicecoil motor (VCM) and mounted via
ballbearing at each end of a pivot shaft, ﬂex cable carrying head and VCMleads, and
arms to support suspension/head extension between the disks. In the head/suspension
assembly, an airbearing surface is created on surface next to the rotating disk, the
slider carrying heads ﬂies on top of the disk surface, and a gimbal attaches the slider
to the suspension. The electronics card involves drivers for the spindle motor and
VCM, read/write (R/W) electronics, a servo demodulator, and micro processors for
servo control and control of interface to host computer.
The actuator servo channel consists of a demodulator producing position informa
tion from the servo burst read from the disk during seeking and following; a servo
controller to control the position of the R/W head during reading, writing and seek
ing; spindle control to keep the spindle rotating at a speciﬁc speed with a minimum
speed ﬂuctuation and a power driver to drive the spindle motor and the VCM actu
ator. The servo or position information is used to position the magnetic head on the
disk surfaces. Position measurement of the magnetic head is achieved by means of
analyzing the position error signal (PES) calculated from the read back signal.
The headpositioning servomechanism is a control system that positions the R/W
head from one track to another in minimum time, and repositions the R/W head over
a desired track with minimum statistical deviation from the track center. A settling
controller is used in between the above seeking and following modes.
The seek time is a measure of how fast the disk drive actuators can move the R/W
head to a desired location. The seek time is limited by the actuator behavior, acceler
ation current level and the control algorithm. The major requirement in the seeking
process is fast and smooth seeking with small or even no overshoot. Dualstage ac
tuation with a VCM as primary actuator and a microactuator as secondary actuator
works as one way to achieve fast seeking and settling due to higher bandwidth.
Once the actuator is regulating the position of the R/W head at the desired track,
the smaller the head position deviates from the desired track center, the closer the
tracks can be put together and the higher the track density becomes. In this stage,
the servo performance is limited by mechanical factors in the actuator, disk platter,
spindle motor, etc. An improved mechanical design is supposed to present less dis
turbance, causing less offtrack. On the other hand, a good closedloop servo system
is expected to reject the disturbances. The error transfer function must be well de
signed to yield a sufﬁciently small closedloop nonrepeatable runout. This typically
requires a satisfactory servo loop based on the disturbance spectrum. Generally, it
demands a high servo bandwidth, a high 0dB crossover frequency and a low hump
of error rejection transfer function. A secondary microactuator activated together
with the VCM primary actuator is generally used to produce a higher bandwidth
closedloop system.
In a disk drive, the positioning information or servo information (“servo bursts”)
is embedded in each disk surface. Servo bursts are conventionally written by costly
dedicated servo writing equipment external to the disk drive, which uses a laser
guided pushpin mechanism to position the write head on the disk surface until the
servo burst information is written on the disk completely [53] [54]. The defects such
Modeling of Disk Drive System and Its Vibration 19
as noncircularity caused by the spindle motor vibration in the servo bursts will make
servo tracks more difﬁcult to followin disk drives. The demand of big storage capac
ity in hard disk drives requires high track density. More accurate placement of servo
bursts is thus required accordingly. As technologies such as servo mechanism and
head and media technology advance, the capability of writing and reading narrower
tracks is improved, resulting in an increased track density.
Conventional servo writers require a clean room environment because the disk
and head will be exposed to the environment to allow access to the external head and
actuator. A selfservo track writer regenerates timing and radial information from
previously written tracks using the existing R/W head [135]. The external equip
ment is no longer needed in servo pattern writing and thus no open access is required
for the head disk assembly and servo track writing does not have to be carried out
in a clean room environment. However, the selfservo track writing creates the ra
dial error propagation problem, which will hinder the whole process of servo track
writing if not properly solved [140].
In an embedded servo, data and servo information are written on all tracks in an
interleaved manner. All tracks are divided into a ﬁxed number of radial sectors. Each
sector starts with a servo pattern followed by user data information. The pattern is
repeated for each sector. The ﬁelds in an embedded servo pattern include Pream
ble, AGC ﬁeld, sector or index mark, track address in gray code and servo bursts.
The position feedback information for the disk drive servo mechanism consists of
two components: gray code and position error signal. The gray code track num
ber provides coarse position information. It determines the absolute position of the
read/write head during seeking and tracking. Position error signal, or PES, is the
relative displacement of the R/W head from the track center. When PES and gray
code are combined together, the position of the read/write head is obtained. Figure
2.1 shows one read back signal of the embedded servo. The A, B, C, D bursts give
ﬁne position error quantifying the amount of track misregistration, i.e., the deviation
of the R/W head from the center of the track.
2.3 System modeling
2.3.1 Modeling of a VCM actuator
A linear VCM actuator moves in and out along a disk radius in one direction. It
contains a coil which is rigidly attached to the structure to be moved and suspended
in a magnetic ﬁeld created by permanent magnets. When a current passes through the
coil, a force is produced which accelerates the actuator radially inward or outward,
depending on the direction of the current. The produced force is a function of the
current i
c
. Approximately,
f
m
= k
t
i
c
, (2.1)
20 Modeling and Control of Vibration in Mechanical Systems
FIGURE 2.1
A read back signal of embedded servo.
where k
t
is a linearized nominal value called torque constant.
The resonance of the actuator is mainly due to the ﬂexibility of the pivot bearing,
arm, suspension, etc. When the bandwidth of a control loop is very low and the
resonance may not be a limiting factor to the control design, the actuator model
can be considered as the simpliﬁed and rigid one which is a double integrator with
transfer function k/s
2
, i.e.,
y =
k
s
2
u, (2.2)
or
˙ y = k
y
v, (2.3)
˙ v = k
v
u, (2.4)
where u is the input to the actuator, y and v are the displacement and the velocity of
the read/write head, k
y
is the position measurement gain, and k
v
= k
t
/m with the
actuator mass m.
With higher bandwidth, the actuator resonances have to be considered in the con
trol design, since the ﬂexible resonance modes will reduce the system stability and
affect control performance if ignored. Then the actuator model becomes
y =
k
v
k
y
s
2
P
r
(s)u, (2.5)
which includes the resonance model P
r
(s). Let ω
n
= 2πf
n
correspond to a single
resonance frequency f
n
, and ξ
n
be the associated damping coefﬁcient. A second
order transfer function can be used to represent the resonance, i.e.,
P
r
(s) =
ω
2
n
s
2
+ 2ξ
n
ω
n
s + ω
2
n
. (2.6)
Modeling of Disk Drive System and Its Vibration 21
Different ξ
n
gives different frequency responses, shown in Figure 2.2. The peak
of magnitude is higher when ξ
n
decreases.
M
a
g
n
i
t
u
d
e
(
d
B
)
Frequency(Hz)
P
h
a
s
e
(
d
e
g
)
0
0
FIGURE 2.2
Frequency responses of a second order transfer function.
Other forms of P
r
(s) include
P
r
(s) =
b
1
ω
n
s + b
0
ω
2
n
s
2
+ 2ξ
n
ω
n
s + ω
2
n
(2.7)
and
P
r
(s) =
b
2
s
2
+ b
1
ω
n
s + b
0
ω
2
n
s
2
+ 2ξ
n
ω
n
s + ω
2
n
(2.8)
with zeros included to facilitate a phase lift which is usually associated with reso
nance modes.
For lightly damped resonance, 0.005 ≤ ξ
n
≤ 0.05 is typical.
22 Modeling and Control of Vibration in Mechanical Systems
FIGURE 2.3
Measured VCM Bode plots (straight lines: pure double integrator k/s
2
).
Modeling of Disk Drive System and Its Vibration 23
Figure 2.3 shows the measured frequency response of a VCM actuator system. By
curve ﬁtting each resonant mode, one can obtain the parameters of the transfer func
tion. The deviation from the double integrator model in low frequencies in Figure
2.3 is due to the pivot friction and other nonlinearities. The nonlinearity modeling
will be discussed in detail in the next section.
Consider the resonance model in (2.7) ((2.6) and (2.8) can be handled similarly).
˙ y = k
y
v, (2.9)
˙ v = k
v
b
1
ω
n
s + b
0
ω
2
n
s
2
+ 2ξ
n
ω
n
s + ω
2
n
u. (2.10)
Deﬁne
x
1
=
ω
2
n
s
2
+ 2ξ
n
ω
n
s + ω
2
n
u, x
2
=
sω
n
s
2
+ 2ξ
n
ω
n
s + ω
2
n
u, (2.11)
i.e.,
˙ x
1
= ω
n
x
2
. (2.12)
Then, we have a statespace description
_
¸
¸
_
˙ y
˙ v
˙ x
1
˙ x
2
_
¸
¸
_
=
_
¸
¸
_
0 k
y
0 0
0 0 k
v
b
0
k
v
b
1
0 0 0 ω
n
0 0 −ω
n
−2ξ
n
ω
n
_
¸
¸
_
_
¸
¸
_
y
v
x
1
x
2
_
¸
¸
_
+
_
¸
¸
_
0
0
0
ω
n
_
¸
¸
_
u. (2.13)
More resonance modes can be modeled similarly.
Assume that nonlinearities such as friction and bias force and resonances are prop
erly compensated, the actuator model is a double integrator model. It is written as
_
˙ y
˙ v
_
=
_
0 k
y
0 0
__
y
v
_
+
_
0
k
v
_
u. (2.14)
On the other hand, to make the mathematical models more realistic, we need to
consider uncertainties inherent in the plant model. Note that it is difﬁcult to involve
all highfrequency dynamics in a modelbased servo control design. In this case, an
uncertainty is introduced in the plant model. For instance, a model with multiplica
tive uncertainty can be given by P = (1+∆)P
n
. Figure 2.4 shows the multiplicative
uncertainty ∆ of a VCM. To meet the robustness requirement against this unmodeled
high frequency dynamics, we need to properly handle the uncertain system so that it
can perform under stricter margins.
2.3.2 Modeling of friction
In this section, we focus on friction modeling. There are basically two kinds of
methodologies for friction control in the literature: Modelbased friction compen
sation and nonmodel based friction control. Friction models for the former can be
24 Modeling and Control of Vibration in Mechanical Systems
10
2
10
3
10
4
−60
−50
−40
−30
−20
−10
0
10
20
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
FIGURE 2.4
Multiplicative uncertainty of a VCM.
Modeling of Disk Drive System and Its Vibration 25
roughly classiﬁed into two categories: static models and dynamic models. There
are various static models for friction, for example, the Coulomb model, the viscous
model, and friction models with the Stribeck effect, etc. However, a static friction
model cannot capture observed friction phenomena like hysteresis, position depen
dence, and variations in breakaway forces. Therefore, a friction model involving
dynamics is necessary to describe friction phenomena accurately. A relatively new
dynamic friction model proposed in [43] combines the Dahl stiction behavior with
arbitrary steady state friction characteristics, which is able to include the Stribeck
effect. A nonlinear friction observer is then required for position control because
the involved interim state is not measurable and has to be observed in order to es
timate the friction force. Later, to overcome the limitations of the above model, an
integrated model is proposed in [44], which is used in [41] for VCM pivot friction
modeling in HDDs. The resultant friction model needs to be iteratively improved and
veriﬁed using the measured and the simulated responses. Another dynamic model
used in VCM pivot friction modeling is the preload and twoslope model, which is
detailed in [40, 42] respectively in the frequency domain and the time domain. How
ever, although the timedomain approach provides a good match between the time
domain response of the model and the data collected, it cannot guarantee a good
match in the frequency domain, and vice versa.
The nonmodel based approaches include the neural network method [46][47] and
the disturbance observer method [45]. The neural network method does not require
full knowledge of the nonlinearity model, but its implementation in real disk drives
seems difﬁcult because of slow convergence [48]. In [45], a novel method for the
cancelation of pivot nonlinearities is proposed and it consists of an accelerometer
and a disturbance observer. The accelerometer is employed to linearize the dynam
ics from the desired input signal to carriage angular acceleration, and the observer
estimates the nonlinear disturbances due to pivot friction for disturbance cancelation.
In this section, a mathematical model will be developed to closely describe the
friction hysteresis behavior. Among existing hysteresis models in the literature, the
Prandtl model [50] is less complex and more attractive in realtime applications. The
elementary operator in the Prandtl hysteresis model is a rateindependent backlash
or linear play operator, deﬁned by p
r
(π
0
, x(t)), where x(t) is the actuator response
and π
0
∈ R is usually initialized to 0. Hysteresis nonlinearity can be modeled by a
linearly weighted superposition of many backlash operators with different threshold
r > 0 and weight values w
b
, i.e.,
F
h
(x(t)) =
_
∞
0
w
b
(r)p
r
[π
0
, x(t)]dr, (2.15)
where the weight w
b
deﬁnes the ratio of the backlash operator, as seen in Figure 2.5.
In order to have an accurate mathematical model for the hysteresis, the creep model
proposed in [49] is also incorporated. Hence we consider the operator model given
by
F(x(t)) = ax(t) +
_
∞
0
w
b
(r)p
r
[π
0
, x(t)]dr +
_
∞
0
w
c
(λ)l
λ
[ξ
0
, x(t)]dλ,
26 Modeling and Control of Vibration in Mechanical Systems
(2.16)
where t ∈ [0, T], a, w
b
(r) and w
c
(λ) are parameters to be determined, p
r
and l
λ
are
the elementary hysteresis and linear creep operators, and are deﬁned as follows.
The elementary hysteresis operator p
r
with threshold r is deﬁned as the solution
operator p
r
[π
0
, x(t)] = z
r
(t) of the rate independent hybrid differential equation
˙ z
r
(t) =
_
_
_
˙ x(t), if x(t) = z
r
(t) − r
0, if z
r
(t) − r < x(t) < z
r
(t) + r
˙ x(t), if x(t) = z
r
(t) + r
with the initial value
z
r
(0) = max{x(0) − r, min{x(0) + r, π
0
(r)}}. (2.17)
FIGURE 2.5
The operator z
r
versus x.
Deﬁne the linear creep operator l
λ
with λ > 0 as the solution operator l
λ
[ξ
0
, x(t)] =
z
λ
(t) of the differential equation
1
λ
˙ z
λ
(t) + z
λ
(t) = x(t) (2.18)
with the initial value equation
z
λ
(0) = ξ
0
(λ).
Modeling of Disk Drive System and Its Vibration 27
The explicit integral formula for the linear creep operator l
λ
is as follows.
l
λ
[ξ
0
, x(t)] = e
−λt
ξ
0
(λ) + λ
_
t
0
e
λ(τ−t)
x(τ)dτ. (2.19)
For numerical implementation of the operatorbased modeling, a discretetime
model F(x(k)) of the operator F(x(t)) in (2.16) is developed as follows.
F(x(k)) = ax(k) +
n
i=1
w
bi
p
r
i
[π
0
, x(k)] +
m
j=1
w
cj
l
λ
j
[ξ
0
, x(k)], (2.20)
where
1) the output sequence of the discrete hysteresis operator is calculated by
p
r
i
[π
0
, x(k)] = z
r
i
(k), (2.21)
z
r
i
(k) =
_
_
_
x(k) + r
i
, if z
r
i
(k − 1) − r
i
≥ x(k)
z
r
i
(k − 1), if z
r
i
(k − 1) − r
i
< x(k) < z
r
i
(k − 1) + r
i
x(k) − r
i
, if z
r
i
(k − 1) + r
i
≤ x(k)
with the initial value z
r
i
(0) = max{x(0) − r
i
, min{x(0) + r
i
, π
0
(r
i
)}};
2) the discrete counterpart to the continuous elementary creep operator is given by
l
λ
j
[ξ
0
, x(k)] = z
λ
j
(k) (2.22)
with
z
λ
j
(k + 1) = e
−λ
j
T
s
· z
λ
j
(k) + (1 − e
−λ
j
T
s
) · x(k) (2.23)
and the initial value z
λ
j
(0) = ξ
0
(λ
j
).
Consider the hysteresis curve of the friction f versus the displacement x in Figure
2.6. Let f
e
= F(x(k)) be the approximated friction, then the approximation error
e = f −f
e
. We deﬁne the energy gain between the actuator position and the error as
T
ex
∞
=
¸
¸
¸
_
L
k=1
e
T
(k)e(k)
L
k=1
x
T
(k)x(k)
, (2.24)
where L is the number of data points.
Denote w
b
= (w
b1
, w
b2
, · · · , w
bn
), w
c
= (w
c1
, w
c2
, · · · , w
cm
), and Λ = (λ(1),
λ(2), · · ·, λ(m)). We aim to ﬁnd optimal parameters a, w
b
, w
c
, and Λ in (2.20) so
that (2.24) is minimized, and thus a model (2.20) can be obtained to approximate the
friction f with the displacement x as the input.
Note that T
ex
∞
is a function of a, w
b
, w
c
, and Λ, and is denoted as
T
ex
∞
= ℓ (a, w
b
, w
c
, Λ) . (2.25)
The MATLAB function fminsearch can be used to minimize ℓ (a, w
b
, w
c
, Λ) with
respect to (a, w
b
, w
c
, Λ).
28 Modeling and Control of Vibration in Mechanical Systems
FIGURE 2.6
Friction f versus actuator displacement x.
FIGURE 2.7
A PZT actuated suspension.
Modeling of Disk Drive System and Its Vibration 29
2.3.3 Modeling of a PZT microactuator
A piezoelectricbased microactuator located on the suspension as shown in Figure
2.7 is considered in this section. The mechanical operation of the microactuator
can be understood via an equivalent springmass system. The compliance of the
base plate is simpliﬁed as a single spring K
b
, and the compliance of the ﬂex hinge
elements is simpliﬁed as a single rotational spring K
r
.
FIGURE 2.8
Equivalent spring mass system of PZT microactuator.
An important point for PZT microactuator modeling is that the PZT element acts in
series with the base plate springs. Thus the displacement of the PZT element results
in displacements of the springs. The PZT and the base plate with spring constants
K
m
and K
b
can be equivalent to a single spring with constant
K
T
=
2
1
K
m
+
1
K
b
. (2.26)
The model is derived by applying forces at the interface of the piezo element and
the base plate spring and by summing moments about the pivot point. The free
expansion of the piezo element is expressed as
θ
f
=
L
m
d
exp
V
cl
1
, (2.27)
30 Modeling and Control of Vibration in Mechanical Systems
where L
m
is the piezo length, d
exp
is the piezo expansion coefﬁcient, V is the volt
age, c is the piezo thickness, and l
1
is the length as indicated in Figure 2.8.
The following second order differential equation can be derived to capture the
dynamic behavior of the microactutor:
K
d
2
θ
dt
+ C
dθ
dt
+ (K
r
+ K
T
l
2
1
)θ =
K
T
L
m
d
exp
l
1
c
V, (2.28)
where K is the torsional inertia, K
r
is the torsional spring rate. A typical frequency
response of the PZT microactuator from voltage input to position output is shown in
Figure 2.9.
10
1
10
2
10
3
10
4
−30
−20
−10
0
M
a
g
n
i
t
u
d
e
(
d
B
)
10
1
10
2
10
3
10
4
−200
−160
−120
−80
−40
0
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
FIGURE 2.9
A typical frequency response of the PZT microactuator.
2.3.4 An example
2.3.4.1 Dynamics modeling
The hard disk drive under consideration is shown in Figure 2.10. It includes the VCM
actuator mounted with the arm and the suspension/head. The actuator is driven by a
driver which converts voltage differences into current differences linearly.
The actuator dynamics measurement is taken using a Laser Doppler Vibrometer
(LDV) and a Dynamic Signal Analyzer (DSA). The displacement y is measured via
Modeling of Disk Drive System and Its Vibration 31
FIGURE 2.10
An opened 1.8inch hard disk drive.
the LDV, and the frequency response is measured by using the DSA to generate a
swept sine signal to excite the actuator. The measured frequency response is shown
in Figure 2.11, where the modeled one is plotted from the following transfer function
P(s) obtained by curve ﬁtting to the measured frequency responses.
P(s) = KP
f
(s)P
res
(s), K = 5.3290 × 10
17
,
P
f
(s) =
1
s
2
+ 2 × 0.25 × 2π120s + (2π120)
2
,
P
res
(s) =
s
2
+ 1081s + 7.3 × 10
8
(s
2
+ 1056s + 6.964 × 10
8
)(s
2
+ 6032s + 2.527 × 10
9
)
. (2.29)
The resonance of P
f
(s) at 120 Hz is due to the nonlinearity of actuator pivot
friction. When the friction nonlinearity is neglected, it is replaced with the pure
double integrators model, i.e.,
P(s) =
K
s
2
P
res
(s), K = 4.8209 × 10
17
, (2.30)
which is plotted as the dotted curves in Figure 2.11. However, the friction in the
actuator pivot [39] [40] is known to limit the low frequency gain of the control loop.
Translated to the error rejection function or sensitivity function, it lifts the magnitude
32 Modeling and Control of Vibration in Mechanical Systems
10
1
10
2
10
3
10
4
−40
−20
0
20
40
60
80
100
M
a
g
n
i
t
u
d
e
(
d
B
)
10
1
10
2
10
3
10
4
−600
−400
−200
0
200
Frequency(Hz)
P
h
a
s
e
(
d
e
g
)
Measured
Modeled
FIGURE 2.11
Measured and modeled frequency responses of the VCM actuation system (LDV
range 0.5 µm/V).
Modeling of Disk Drive System and Its Vibration 33
of the sensitivity function at low frequencies, and thus reduces the ability of the
control loop to reject lowfrequency vibrations and affects the positioning accuracy.
Therefore, it is necessary to compensate the friction impact.
2.3.4.2 Friction measurement and modeling
Due to the ﬂuctuation of the head when the disk is rotating (rotational speed is 4200
RPM), it is difﬁcult to have a steady displacement signal of the head. Thus the
friction measurement is carried out under the closed loop control as shown in Figure
2.12. The controller C(z) is a PID controller combined with a notch ﬁlter and is
expressed in (2.31). The sampling time T
s
is 83.3 ms. With the controller, the open
loop 0 dB crossover frequency is 945 Hz, the gain margin is 8.7 dB, and the phase
margin is 49 deg.
C(z) = k
c
(k
p
+ k
d
z − 1
T
s
z
+ k
i
T
s
z − 1
) ×
0.9023z
2
+ 0.9467z + 0.7242
z
2
+ 0.929z + 0.6442
,
k
c
= 0.0625, k
p
= 0.8, k
d
= 400 × 10
−6
, k
i
= 400. (2.31)
FIGURE 2.12
Closed control loop of a disk drive with a VCM actuator for friction measurement
via LDV.
A 10 Hz sinusoidal signal with increasing amplitude of 0.5, 1, and 3 V is re
spectively used as the reference signal in Figure 2.12. The control signal u and
displacement x are measured, and shown in Figure 2.13.
The VCM actuator model with consideration of nonlinearity F(x) is shown in
Figure 2.14, which includes two pure integrators, the resonance modes P
res
(s) and
the gain K given in (2.30).
With the measured u and x, u
a
can be obtained from x, and thus f = u −u
a
. The
relation between x and f can be obtained and shown as hysteresis curves in Figure
2.6.
34 Modeling and Control of Vibration in Mechanical Systems
−4 −3 −2 −1 0 1 2 3 4
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
x 10
−3
Displacement (V, 0.5 µm/V)
C
o
n
t
r
o
l
s
i
g
n
a
l
u
(
V
)
FIGURE 2.13
Control signal u versus displacement x.
FIGURE 2.14
VCM actuator modeling with friction nonlinearity model F(x).
Modeling of Disk Drive System and Its Vibration 35
In what follows, we shall ﬁnd F(x) to model the relationship between f and x.
The above mentioned operator based method to approximate hysteresis will be ap
plied to model the hysteresis of f and x as shown in Figure 2.6.
r
i
in (2.21) can be chosen as the amplitude of x. Since with the chosen peakto
peak values of x = 0.5, 1 and 3V, the r
i
are respectively
r
1
= 0.25; r
2
= 0.5; r
3
= 1.5. (2.32)
With m = 3, n = 3 and the initial values π
0
= 0, ξ
0
= 0, after 2000 iterations, a
minimum error e is achieved and the optimal parameters are obtained as
a = 9.0024,
w
b1
= −1.5783, w
b2
= 0.1667, w
b3
= −0.0655,
w
c1
= −0.1431, w
c2
= −7.3341, w
c3
= 0.4403,
which gives the minimal T
ex
∞
= 0.08.
With these parameters, f
e
= F(x(k)) can be calculated from (2.20). The time
traces of f
e
and f are compared in Figure 2.15. It is observed that the time trace
from the model (2.20) can closely track f, and the error f −f
e
is small. The modeled
hysteresis from x to f
e
is drawn and compared with the measured one in Figure 2.6.
It is seen that the modeled hysteresis and the measured one are close to each other.
If the creep term in the model (2.20) is removed, i.e., w
cj
= 0, j = 1, · · · , m, the
modeling accuracy decreases. Indeed, in this case, the minimal T
ex
∞
= 0.1293 >
0.08.
Figures 2.16 and 2.17 show that the quality of the agreement in the frequency
domain between 70 Hz and 150 Hz decreases with lower excitation amplitude espe
cially for the 0.5V case.
Note that the operator model f
e
= F(x(k)) describes the hysteretic characteristics
in Figure 2.6 as a mapping between the actuator position x and the friction force
f. It turns out that the model makes it possible to approximate hysteretic transfer
characteristics without modeling the underlying physics. This is different from the
friction models such as the preload and twoslope model in [40][42]. Compared to
the model in [40] [42], an advantage of our operator based model is that the frequency
response of the actuator can also ﬁt well to the measured one, as shown in Figures
2.16 and 2.17 .
For comparison, we also apply the preload and two slope model [40] for the hys
teresis, as shown in Figure 2.18. The preload model for velocity v and the twoslope
model for position x are given by
f
v
= k
v
v + k
s
sgn(v), (2.33)
and
f
x
=
_
k
a
x, x ≤ s
x
k
b
x + (k
a
− k
b
)s
x
, x > s
x
.
36 Modeling and Control of Vibration in Mechanical Systems
Using the data with the amplitude of x(t) of 0.25V, we obtain that
k
s
= 1.85e − 4, k
v
= 0.48, k
a
= 0.0145, k
b
= 0.0014, s
x
= 0.009.
Using this model, a comparison with the measured data in the time domain is
shown in Figure 2.19 for the amplitude of x(t) of 0.5 and 1 V, respectively. It can
be observed that the plant input u versus position x ﬁts reasonably well to the mea
sured results in the time domain. However, in the frequency domain, the magnitude
response for the case of the amplitude of x(t) of 0.5V, plotted as the dotted curve in
Figure 2.17 deviates signiﬁcantly from the measurement results.
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
x 10
−3
Time(sec)
F
r
i
c
t
i
o
n
a
n
d
e
r
r
o
r
From measurement (f)
From modeling(f
e
)
Error (f−f
e
)
FIGURE 2.15
Measured and modeled friction and error.
Modeling of Disk Drive System and Its Vibration 37
FIGURE 2.16
Actuator frequency response for sinusoidal reference with amplitude of 1 and 3 V,
respectively.
FIGURE 2.17
Actuator frequency response for sinusoidal reference with amplitude of 0.5 V.
38 Modeling and Control of Vibration in Mechanical Systems
FIGURE 2.18
Preload and twoslope model for friction modeling.
−0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
x 10
−3
Displacement (µm)
I
n
p
u
t
v
o
l
t
a
g
e
u
Measured
Modeled
1V
0.5V
FIGURE 2.19
Plant input voltage u versus displacement x.
Modeling of Disk Drive System and Its Vibration 39
2.4 Vibration modeling
Vibration in disk drives causes the deviation of the R/Whead positioningfromthe de
sired track center. It is the combination of the repeatable runout which is synchronous
with the spindle revolution and the nonrepeatable runout. Figure 2.20 shows a sim
pliﬁed blockdiagram of disk drive servo loop. y is the position of the R/W head and
e is the position error signal. The signal d
1
represents all the torque disturbances
to the system. Such disturbances include any torque due to airturbulence force to
the actuator, the suspension and the slider. The effects of the torque disturbances are
dominant at frequencies that are relatively low when compared to the servo band
width. The signal d
2
represents disturbances that are due to nonrepeatable motions
of the disk and motor, suspension and slider vibrations, which directly add to the
relative position of the R/W head and the servo track. The noise signal n includes
media and head sensing noises and also represents the effects of the PES demodu
lation noise which includes actual electrical noise and A/D quantization noise. The
noise signal n is thus reasonably modeled as a broadband white noise.
2.4.1 Spectrumbased vibration modeling
FIGURE 2.20
Closedloop control system disturbances d
1
, d
2
and noise n.
Since a controlled closedloop can provide steady signals, the vibration source
analysis is based on the signal that is collected from the closedloop system. From
Figure 2.20,
e(k) = −P(z)S(z)d
1
(k) − S(z)d
2
(k) + S(z)n(k), (2.34)
40 Modeling and Control of Vibration in Mechanical Systems
FIGURE 2.21
Closedloop control system with disturbance and noise models.
where P(z) is the transfer function of the discretized plant model P(s) and the sen
sitivity function or error rejection function is given by
S(z) =
1
1 + P(z)C(z)
. (2.35)
Figure 2.22 shows the sensitivity function S(z) of a closedloop control system.
Assume that d
1
, d
2
, and n are uncorrelated. The power spectrum denoted by S
e
of
the error signal e is given by
S
e
= P(z)S(z)
2
d
1
(k)
2
+ S(z)
2
d
2
(k)
2
+ S(z)
2
n(k)
2
. (2.36)
Figure 2.23 shows the spectrum of NRRO component in the error signal e. Two
humps are obviously observed in the baseline curve. One is in the frequency range
lower than 100 Hz, the other one is after 500 Hz. Considering (2.36) and Figure
2.22, the second hump is caused by S(z) through S(z)
2
n(k)
2
, and the ﬁrst hump
is due to d
1
through P(z)S(z) with a hump in a lower frequency range. Hence the
disturbance and noise modeling can be carried out as follows.
In Figure 2.21, models D
1
(s), D
2
(s) and N(s) are used to describe d
1
(s), d
2
(s)
and n(s) respectively, and w
i
(i = 1, 2, 3) are independent white noises with
variance 1. Eq. (2.36) becomes
S
e
= P(z)S(z)
2
D
1
(z)
2
+ S(z)
2
D
2
(z)
2
+ S(z)
2
N(z)
2
, (2.37)
where D
1
(z), D
2
(z) and N(z) are the discrete forms of D
1
(s), D
2
(s) and N(s),
respectively. D
1
(z) and N(z) can be determined by ﬁtting weighted versions of
P(z)S(z) and S(z) to the baseline curve of the spectrum and the spikes are consid
ered as the effect of D
2
(z). Hence the steps to obtain D
1
, N and D
2
are
Modeling of Disk Drive System and Its Vibration 41
FIGURE 2.22
Sensitivity function S(z).
42 Modeling and Control of Vibration in Mechanical Systems
FIGURE 2.23
PES NRRO spectrum.
Step 1). Find S
b
(j),
S
b
(j) =
jq
min
i=1+(j−1)q
S
e
(i), j = 1, 2, · · · , L/q, (2.38)
where L is the length of S
e
and q is as small as possible.
Step 2). Compute D
1
(z),
D
1
(z)
2
= W
L
(z)S
b
/P(z)S(z)
2
, (2.39)
where W
L
is a lowpass weighting function used to select S
b
in low frequency range.
Step 3). Compute N(z),
N(z)
2
= W
H
S
b
/S(z)
2
, (2.40)
where W
H
is the highpass weighting function to select S
b
in high frequency range.
Step 4). Till now, the baseline curve S
b
can be ﬁt well by the identiﬁed D
1
and N.
The remaining part of the spectrum is then regarded as D
2
. Thus,
D(z)
2
= {S
e
− [P(z)S(z)
2
D
1
(z)
2
+ S(z)
2
N(z)
2
]}/S(z)
2
. (2.41)
Modeling of Disk Drive System and Its Vibration 43
D
1
(s), D
2
(s) and N(s) are then obtained as follows from D
1
(z), D
2
(z) and
N(z) using the bilinear approximation method [81].
D
1
(s) =
1.3916 × 10
−5
(s + 575.8)(s + 575.6)(s
2
+ 0.04389s + 161.6)
(s
2
+ 315.5s + 8.178 × 10
4
)(s
2
+ 315.4s + 8.178 × 10
4
)
, (2.42)
D
2
(s) =
0.701588(s + 1.271 × 10
4
)
2
(s
2
+ 6.125 × 10
−5
s + 4.373 × 10
8
)
(s + 708.4)
2
(s
2
+ 0.0001317s + 4.376 × 10
8
)
,
(2.43)
N(s) =
1.1695(s + 1.431 × 10
4
)(s + 766.2)(s
2
+ 8609s + 4.672 × 10
7
)
(s + 4603)(s + 1538)(s
2
+ 4451s + 1.507 × 10
7
)
. (2.44)
With the disturbance and noise models, the feedback controller C(z) can be opti
mized to minimize the error due to w = [w
1
w
2
w
3
]
T
by using the H
2
optimal control
method or other advanced control methods.
2.4.2 Adaptive modeling of disturbance
2.4.2.1 Neural network approximation
A neural network usually consists of a large number of simple processing elements
called nodes. The nodes are interconnected by weighted links with weight param
eters adjustable. The different arrangement of the nodes and the interconnections
deﬁnes various architectures of neural networks [73][74], which are suitable to dif
ferent kinds of applications. In control engineering, a multilayer neural network
is usually used to generate the mapping from input to output since it can approxi
mate any function under mild assumption with any desired accuracy. The function
approximation is deﬁned as follows.
Deﬁnition 2.1 [75] If f(x) : R
n
→ R
m
is a continuous vector function deﬁned
on a compact set Ω, and any y(W, x) : R
t
× R
n
→ R
m
is an approximating
function that depends continuously on W and x, then, the approximation problem is
to determine the optimal W denoted by W
⋆
, for some index d such that
d(y(W
⋆
, x), f(x)) ≤ ε, (2.45)
for an acceptable small ε.
There are a number of neural networks studied for function approximation such as
multilayer perceptron networks, radial basis function (RBF) networks, and higher
order neural networks. The RBF network is suitable for online nonlinear adaptive
modeling and control, because it is a linearly parameterized network, has spatially
localized learning capability and thus has better memory during learning, and ex
hibits a fast initial rate of learning convergence.
The threelayer neural network shown in Figure 2.24 is a RBF network, where
x ∈ R
n
, y ∈ R
m
, and s ∈ R
p
are respectively the input, the output, and the
activation function vectors, and w
ij
is the second to the third layer interconnection
44 Modeling and Control of Vibration in Mechanical Systems
FIGURE 2.24
Threelayer RBF neural network.
Modeling of Disk Drive System and Its Vibration 45
weights. The output y
i
is given by
y
i
=
p
j=1
w
ij
s
j
(x −c
j
2
), i = 1, 2, · · · , m, (2.46)
or equivalently in a matrix form,
y(x) = W
T
s(x), (2.47)
where p is the number of nodes, x = [x
1
x
2
· · · x
n
]
T
∈ R
n
is the network
input vector, • denotes the Euclidean norm, c
j
∈ R
n
is the center vector, and
W
T
= [w
ij
]
T
.
The approximation of a general function f(x) : R
n
→ R
m
can then be expressed
as
f(x) = W
T
s(x) + ε(x) (2.48)
where ε(x) ∈ R
m
is the reconstruction error vector.
Several functions such as Gaussian, Hardy’s multiquadric and inverse Hardy’s
multiquadric functions have been used as activation functions. They are separately
written as
s
j
(x) = exp[
−(x −c
j
)
T
(x −c
j
)
σ
2
j
], (2.49)
s
j
(x) =
_
σ
2
j
+ (x −c
j
)
T
(x −c
j
), (2.50)
and
s
j
(x) =
1
_
σ
2
j
+ (x −c
j
)
T
(x −c
j
)
. (2.51)
For a mechanical system with dynamics given as a function of position, velocity
and acceleration, a dynamic neural network is needed in order to fully emulate the
dynamics. In a dynamic neural network, dynamic variables such as velocity and
acceleration are involved in the input x. In discretetime neural networks, past infor
mation such as x(k − 1) and x(k − 2) is used as input.
2.4.2.2 Disturbance modeling
In some mechanical motion systems, disturbances in frequency range lower than a
few hundred Hz are quite dominant and they may be due to torque disturbances,
nonlinear or unknown vibration sources. Here we use d
1
to represent these dis
turbances. This section aims to model the lowfrequency disturbance d
1
with an
46 Modeling and Control of Vibration in Mechanical Systems
adaptive nonlinear strategy based on the measured error signal e. In the openloop
without controller C(z),
e = −P(s) · d
1
− d
2
+ n. (2.52)
The disturbance d
1
is modeled as
ˆ
d
1
= f(Φ(e(k))), (2.53)
where f(·) is an unknown function of the bounded vectorvalued function Φ(e(k))
with
e(k) = [e(k) e(k − 1) · · · e(k − l)]
T
, (2.54)
where l is to be determined. Thus (2.53) is dependent on some time history of the
measured signal e.
For brevity, f(Φ(e(k))) is denoted as f(Φ
k
). As stated in the previous section,
the unknown function f can be approximated using neural networks. Here we se
lect a Gaussian RBF based neural network. The Gaussian RBF network has some
attractive properties and thus has been widely used in nonlinear control and signal
approximation. The properties are, (1) it is bounded and strictly positive, and (2)
it possesses a localized response. Based on the Gaussian RBF network, the low
frequency disturbances can be approximated as
ˆ
d
1
= f(Φ
k
) = W
T
(k)s(Φ
k
) + ε, (2.55)
where ε is the modeling error, s(Φ
k
) ∈ R
p
is a basis vector function with a suitable
number of nodes p. The radial basis function s(Φ
k
) is written as
s
i
(Φ
k
) = e
−
(e(k)−c
e
i
)
2
σ
2
e
i
· e
−
(e(k)−e(k−1)−c
∆e
i
)
2
σ
2
∆y
i
,
s(Φ
k
) = [s
1
(Φ
k
) s
2
(Φ
k
) ... s
p
(Φ
k
)]
T
, (2.56)
where σ
e
i
and c
e
i
(σ
∆e
i
and c
∆e
i
) are the variance and center position of the mea
surement e (velocity ˙ e).
The weight vector W(k) in (2.55) can be calculated iteratively according to the
following weight update law:
W(k + 1) = (1 − δ)W(k) − Γs(Φ
k
)e(k) (2.57)
with the adaptation gain matrix Γ ∈ R
p×p
being diagonal and satisfying Γ > 0 and
the forgetting factor 1 > δ > 0. The forgetting factor δ is to ensure the bounded
ness of W(k) when the system is subjected to bounded disturbances. The speed of
learning rate is related to the chosen Γ matrix.
Next, the power spectrum in Figure 2.27 of a position error signal e will be used to
verify the adaptive modeling scheme. It is noted that the contribution of disturbances
and noise to e is described as in (2.52).
Modeling of Disk Drive System and Its Vibration 47
A low frequency signal with the spectrum as in the low frequency range in Figure
2.27 is generated and injected as disturbance d
1
in Figure 2.21. The disturbance d
1
affects the error signal or the output e through the plant P(s), thus the nonlinear
model from e to
ˆ
d
1
aims to generate a cancelation signal of the disturbance d
1
. Since
the low frequency disturbances due to torque and bias are generally nonlinear, the
model from e to d
1
is chosen to be nonlinear.
The veriﬁcation is implemented with the sampling period T
s
= 1/45000sec. For
p = 1 with zero center positions c
y
1
and c
∆y
1
, after some trials, it was found that
when σ
2
y
1
= σ
2
∆y
1
= 10, δ = 0.5, and
Γ = T
s
· 10
6
· 220, (2.58)
p = 1, l = 1, (2.59)
the time trace
ˆ
d
1
calculated from (2.55) gives the best approximation of the distur
bance d
1
. The effect of different p and Γ on the modeling accuracy will be evaluated
later. Figure 2.25 shows the simulated time trace comparison of d
1
and
ˆ
d
1
from the
nonlinear model (2.55). It is observed that the time trace from the model (2.55) can
give a close tracking of the original one. The spectrum of
ˆ
d
1
is seen in Figure 2.26.
Moreover, from the spectrum in Figure 2.27 with the component of disturbance
P(s) ·
ˆ
d
1
removed and considering that noise n is a white noise, the disturbance d
2
and the noise n are represented approximately by
ˆ
d
2
= D
2
(s)w
2
, (2.60)
D
2
(s) =
0.0019(s
2
+ 3329s + 1.695 × 10
7
)(s
2
+ 3340s + 5.61 × 10
8
)
(s
2
+ 245s + 1.668 × 10
7
)(s
2
+ 477.5s + 5.701 × 10
8
)
,
and
ˆ n = N(s)w
3
= 0.005 w
3
. (2.61)
The NRRO spectrum obtained by combining the nonlinear and linear models is
compared with the measured one in Figure 2.27. It is found that the adaptive nonlin
ear modeling method can indeed be used to model the disturbance d
1
that is dominant
in low frequency range.
Let the modeling error d
e
= d
1
−
ˆ
d
1
. To evaluate the effects of different values of
p on the modeling error, two more cases with p = 5, 9 and the following parameters
for Γ are investigated.
Γ = T
s
· 10
6
· diag{33, 66, 220, 66, 33}, p = 5; (2.62)
Γ = T
s
· 10
6
· diag{38.5, 88, 88, 38.5, 220, 38.5, 88, 88, 38.5},
p = 9. (2.63)
The σ value of the modeling error e can be seen in Table 2.1. With p = 1, 5, 9,
σ increases from 4.43 to 4.60, which means higher p may not give a better result
48 Modeling and Control of Vibration in Mechanical Systems
in this practical application. Compared with p, Γ inﬂuences the modeling accuracy
more signiﬁcantly. For p = 1, with Γ being changed to 10% of the value in (2.58),
the σ value of the error e increases from 4.43 to 11. The time sequence comparison
of d
1
−
ˆ
d
1
is shown in Figure 2.28, where the difference is clearly seen. The other
two cases are similar.
2.5 Conclusion
In this chapter the models of a VCM actuator and a PZT microactuator have been
derived on the basis of their physical operations. The models have been veriﬁed
with the measured frequency responses from voltage input to displacement output.
To have a complete model of an actuation system, this chapter has also addressed
the modeling of uncertainties in high frequency and nonlinearities such as actuator
pivot friction. A Prandtl operator based model has been used to model the friction
nonlinearity, and the optimal model parameters have been obtained by minimizing
the energy gain between the actuator position and the modeling error. It turns out that
the derived model matches well the measured model not only in the time domain, but
also in the frequency domain.
The developed vibration and noise modeling approaches are based on the error
signal power spectrum of the closedloop system. In particular, for low frequency
disturbance modeling, an adaptive nonlinear scheme with system measurement as
input has been applied to approximate the original disturbance for realtime com
pensation.
Modeling of Disk Drive System and Its Vibration 49
TABLE 2.1
σ values of the modeling error (
ˆ
d
1
− d
1
) for different p and Γ
p 1 5 9
σ values of (
ˆ
d
1
− d
1
) for Γ and 0.1Γ.(×10
−4
) 4.43/11 4.55/11 4.60/10
0 0.05 0.1 0.15 0.2 0.25
−4
−3
−2
−1
0
1
2
3
4
5
x 10
−3
Time(sec)
O
r
i
g
i
n
a
l
d
i
s
t
u
r
b
a
n
c
e
a
n
d
t
h
e
m
o
d
e
l
e
d
(
µ
m
)
Original disturbance d
1
d
1
from nonlinear modeling
FIGURE 2.25
Original disturbance d
1
and the modeled
ˆ
d
1
.
50 Modeling and Control of Vibration in Mechanical Systems
0 1000 2000 3000 4000 5000 6000
0
0.5
1
1.5
2
2.5
3
3.5
4
x 10
−3
Frequency(Hz)
N
R
R
O
m
a
g
n
i
t
u
d
e
(
µ
m
)
FIGURE 2.26
Power spectrum of
ˆ
d
1
.
Modeling of Disk Drive System and Its Vibration 51
0 1000 2000 3000 4000 5000 6000
0
0.5
1
1.5
2
2.5
3
3.5
4
x 10
−3
Frequency(Hz)
N
R
R
O
m
a
g
n
i
t
u
d
e
(
µ
m
)
Modeled
Measured
FIGURE 2.27
NRROpower spectrum frommeasurement and disturbance models, i.e., e = −P(s)·
ˆ
d
1
−
ˆ
d
2
+ ˆ n.
52 Modeling and Control of Vibration in Mechanical Systems
0 0.05 0.1 0.15 0.2 0.25
−4
−3
−2
−1
0
1
2
3
4
x 10
−3
Time(sec)
M
o
d
e
l
i
n
g
e
r
r
o
r
(
µ
m
)
Γ
0.1Γ
FIGURE 2.28
Modeling error (d
1
−
ˆ
d
1
) for different Γ with p = 1.
3
Modeling of Stewart Platform
3.1 Introduction
An adaptive control system is able to adapt to the changes of a physical system. In
order to obtain more accurate models for physical systems, adaptive identiﬁcation
algorithms can be used. Indeed, adaptive ﬁltering is often adopted in control and
signal processing ﬁelds, for the modeling of an unknown system.
In this chapter, the governing motion equation of a piezoelectric Stewart platform
will be obtained ﬁrst, followed by the derivation of a transfer function fromthe actua
tor force to the sensor output based on frequency response data. An adaptive ﬁltering
approach will then be introduced and subsequently used to model the platform and
verify the transfer function obtained.
3.2 System description and governing equations
The study of active damping of ﬂexible structures has been induced by the require
ment of structural stability. Most solutions to the active damping problem rely on
the integration of smart actuators and sensors in the structure itself. Referring to the
discussion in Chapter 1, each leg of the hexapod Stewart platform consists of an am
pliﬁed piezoelectric actuator, a force sensor and two ﬂexible joints. This piezoelec
tric Stewart platform can be used as a precision pointing device, a vibration isolation
mount, or an active damping mount.
Shown in Figure 1.2 and Figure 1.3, the hexapod consists of two parallel plates
connected to each other by six active legs. The legs are mounted in such a way
to achieve the geometry of cubic conﬁguration. Each active leg consists of a force
sensor, an ampliﬁed piezoelectric actuator, and two ﬂexible joints.
Figure 3.1 shows an equivalent system of each leg connecting two rigid bodies:
the disturbance source m and the sensitive payload M that are connected by a force
sensor and a piezoelectric actuator represented by its elongation δ and spring stiffness
k. The Laplace form of the governing equation of the motion in this system is
53
54 Modeling and Control of Vibration in Mechanical Systems
FIGURE 3.1
Singleaxis system using piezoelectric stiff actuator.
Ms
2
x
p
= −ms
2
x
d
= k(x
d
−x
a
) = y (3.1)
and
δ = x
p
− x
a
. (3.2)
The transfer function between the extension δ of the piezo stack and the sensor
output y is
y
δ
= k
Mms
2
Mms
2
+ k(M +m)
. (3.3)
Consider the piezoelectric hexapod integrated in a structure. Let M and K be the
mass and stiffness matrices of the global passive system, i.e., structure and hexapod.
The dynamic equation governing the system is written as
Ms
2
x +Kx = Bf, (3.4)
where B is the force Jacobian matrix, x represents payload frame displacement, and
f = kδ represents the equivalent piezoelectric force in the leg. A force sensor is lo
cated in each leg of the hexapod and collocated with the actuator. The corresponding
sensor output is
y = k(q − δ), (3.5)
where y is the force sensor output and q is the leg extension from the equilibrium
position. Taking into account the relationship between the leg extension and the
payload frame displacement
q = B
T
x, (3.6)
Modeling of Stewart Platform 55
the sensor output equation becomes
y = k(B
T
x −δ). (3.7)
The sensor output y is to be used as the input to the controller.
Next an adaptive ﬁltering approach will be used to model the hexapod Stewart
platform and verify the dynamics equation (3.3).
3.3 Modeling using adaptive ﬁltering approach
3.3.1 Adaptive ﬁltering theory
LMS algorithm
FIGURE 3.2
Linear discrete time adaptive ﬁlter.
In Figure 3.2, the statistically stationary time sequence of input signal, x(0), x(1),
· · ·, is applied to the linear discrete time ﬁlter whose coefﬁcients are W0, W1, · · ·.
The ﬁlter output at time k, y(k), is to be as close as possible to the desired response,
d(k). The difference between y(k) and d(k) is deﬁned as the estimation error, e(k)
that is then applied back to the ﬁlter to adjust the ﬁlter weights so as to minimize
the estimation error in the statistical sense. y(k) = W
T
(k)X(k), where W
T
(k) =
[W0 W1 W2 · · ·], and X
T
(k) = [x(k) x(k −1) x(k − 2) · · ·].
In most practical instances the adaptive process aims to minimize the meansquare
value, or average power of the error signal [37]. Optimization under this criterion is
56 Modeling and Control of Vibration in Mechanical Systems
most effective in statistical sense. So the performance index of the adaptive ﬁlter is
determined by the meansquared error (MSE) criterion in which the cost function is
deﬁned as follows.
ξ(k) = E[e
2
(k)], (3.8)
where E[ ] denotes the expected value.
e(k) = d(k) − y(k)
= d(k) − W
T
(k)X(k). (3.9)
ξ(k) = E[(d(k) − W
T
(k)X(k))
2
]
= E[d
2
(k)] − 2W
T
(k)E[d(k)X(k)] +W
T
(k)E[X(k)X
T
(k)]W(k)
= R
dd
−2W
T
(k)R
dx
+ W
T
(k)R
xx
W(k), (3.10)
where R
dx
is the cross correlation function between d(k) and x(k), R
dd
and R
xx
are
respectively the autocorrelation functions of d(k) and x(k).
In many practical applications, the statistics of d(k) and x(k) are unknown; the
exact knowledge of gradient vector is not available. Thus the method of steepest de
scent [35] cannot be implemented in practical situations. Therefore another method
called the Least Mean Square (LMS) algorithm was introduced by Bernard Widrow
[37]. In the LMS algorithm, instantaneous squared error is used as an estimation
of the mean squared error. The mathematical derivation of the LMS algorithm is as
follows.
ˆ
ξ(k) = e
2
(k),
∇
ˆ
ξ(k) = ∇e
2
(k) = 2[∇e(k)]e(k),
e(k) = d(k) −W
T
(k)X(k),
∇e(k) =
∂e(k)
∂W(k)
= −X(k),
∇
ˆ
ξ(k) = −2X(k)e(k). (3.11)
The updating equation of the adaptive ﬁlter coefﬁcient:
W(k + 1) = W(k) −
µ
2
∇
ˆ
ξ(k),
becomes
W(k + 1) = W(k) +µX(k)e(k). (3.12)
Modeling of Stewart Platform 57
FIGURE 3.3
Block diagram of the LMS adaptive ﬁlter.
Equation (3.12) is the wellknown LMS algorithm that is suitable for practical
signal processing applications because of its simplicity and the availability of the in
stantaneous error in real time. The block diagram of the LMS algorithmis illustrated
in Figure 3.3.
Normalized LMS algorithm
The LMS adaptation process is very much dependant on the step size µ and the
reference signal power. The step size determines the system convergence rate and
stability [36][37]. The maximum stable step size is inversely proportional to the ﬁlter
order L and the power of the reference signal x(k). A technique used to optimize
the convergence speed, independent of the reference signal power, is known as the
normalized LMS algorithm (NLMS).
The NLMS algorithm is given as follows.
W(k + 1) = W(k) +µ(k)X(k)e(k), (3.13)
where each variable is identical to the one for the LMS algorithm except for µ(k)
that is an adaptive step size computed as
µ(k) =
α
L
ˆ
P
x
(k)
, (3.14)
where α is a normalized step size that satisﬁes 0 < α < 2, L is the ﬁlter length, and
ˆ
P
x
(k) is the estimated power of the reference signal x(k).
The estimation of
ˆ
P
x
(k) can be done using the mean square value of the reference
signal as follows.
ˆ
P
x
(k) =
X
T
(k)X(k)
L
. (3.15)
58 Modeling and Control of Vibration in Mechanical Systems
Then Equation (3.14) reduces to
µ(k) =
α
X
T
(k)X(k)
. (3.16)
This step size is the most widely used for the NLMS algorithm. A variant of the
NLMS algorithm uses a small constant ε as follows.
µ(k) =
α
ε +X
T
(k)X(k)
. (3.17)
This constant ensures that the step size does not tend to inﬁnity in the case of a
zero input signal.
The NLMS algorithm guarantees an attenuation level of γ ≤ 1, where γ is the
inducednorm from noise or signal inputs to the ﬁltering error. Therefore, a salient
feature of this algorithmis that it lowers the inﬂuence of the input signal on the noise
ampliﬁcation effect, especially when the input signal x(k) is large.
3.3.2 Modeling of a Stewart platform
Adaptive identiﬁcation is a technique that uses an adaptive ﬁlter to model an un
known system. An adaptive identiﬁcation method can be applied either online along
with the vibration control or ofﬂine prior to the vibration control. In an online identi
ﬁcation system, the number of adaptive ﬁlters required for an adaptive control system
will be increased. Furthermore, convergence of the adaptive ﬁlter used in the identi
ﬁcation section of the system can be affected by a large amount of primary vibration
signal [36]. Therefore an ofﬂine identiﬁcation technique will be applied for the sys
tem modeling.
The block diagram of the adaptive identiﬁcation is illustrated in Figure 3.4. P(z)
is the transfer function of the system to be identiﬁed. W(z) is a digital ﬁlter used
to model P(z) based on the LMS error minimization algorithm. Both systems P(z)
and W(z) are excited by a band limited white noise. The difference between the
two outputs d(k) and y(k) is fed back into the LMS algorithm as error signal e(k).
The LMS algorithmwill adaptively adjust the coefﬁcients of ﬁlter W(z) to minimize
e(k) based on the least MSE criterion. When the error signal reaches the minimum
level, the ﬁlter W(z) represents a model of P(z).
The LMS adaptive ﬁlter approach is applied for the modeling of the Stewart plat
form. A Simulink program is developed for ofﬂine identiﬁcation of the platform.
A 16th order LMS adaptive ﬁlter with adaptation step size of 0.01 is chosen. A
bandlimited white noise is used as the training signal. Sampling frequency of the
white noise generator is set at 1 kHz so that the response of the PZT actuator will be
conﬁned to the bandwidth of 500 Hz (half of sampling frequency). But the sampling
frequency of the adaptive ﬁlter is set to 10 kHz to give enough time for adaptive ﬁlter
to compute the ﬁlter weights within each sample of the training signal.
The Simulink program for the identiﬁcation is downloaded into a dSPACE real
time interface board DS1104. Identiﬁcation process for each PZT actuator is per
formed alternatively. Filter tap values are recorded through the dSPACE manager
Modeling of Stewart Platform 59
FIGURE 3.4
System identiﬁcation using LMS adaptive ﬁlter.
software during the identiﬁcation process. From the results of the identiﬁcation of
six PZT actuators of the Stewart platform in Figure 1.2, one of the resonance fre
quencies of the smart structure is found to be around 240 Hz. Figure 3.5 is the phase
and magnitude responses of one of the six PZT actuators. The result shows that there
is a resonance peak at about 240 Hz. The phase response between 60 Hz and 200
Hz is approximately linear. Experimental results of the active vibration control sys
tem also indicate that the frequency region that can attenuate the vibration signal is
between 60 Hz and 220 Hz.
By applying a frequency domain modeling approach, an approximate model in
terms of a transfer function for each leg can be obtained. Figure 3.6 shows the fre
quency responses of an PZT actuator, including the estimated and the experimental
ones. It is clearly shown that the estimated transfer function is well ﬁtted to the real
one. It is also noted that the shape of the frequency responses in Figure 3.6 agrees
with that of the model in (3.3).
The transfer function of the PZT actuator is obtained as:
P(s) =
−190.9s
4
−1.499 ×10
4
s
3
−5.703 × 10
8
s
2
− 2.507 × 10
10
s − 5.624 × 10
11
s
5
+ 3232s
4
+ 5.378 × 10
6
s
3
+ 1.621 ×10
10
s
2
+ 6.82 ×10
12
s + 1.943 ×10
16
.
(3.18)
60 Modeling and Control of Vibration in Mechanical Systems
FIGURE 3.5
Frequency responses of a PZT actuator.
Modeling of Stewart Platform 61
FIGURE 3.6
Estimated and experimental frequency responses.
62 Modeling and Control of Vibration in Mechanical Systems
3.4 Conclusion
This chapter has studied the modeling of an active piezoelectric Stewart platform.
The LMS adaptive ﬁltering approach has been adopted for the system identiﬁcation.
The model obtained via the adaptive identiﬁcation or the experimental testing has
shown consistency with the governing motion equation.
4
Classical Vibration Control
4.1 Introduction
The presence of vibration often leads to undesirable effects such as structural or
mechanical failure, frequent and costly maintenance of machines, worsening posi
tioning performance, and human pain and discomfort. Vibrations can sometimes be
eliminated theoretically. However, due to the high manufacturing cost that may be
involved in eliminating vibration, reduction of vibration is preferred so as to achieve
a compromise between an acceptable amount of vibration and a reasonable manu
facturing cost. Various classical techniques of vibration control for the purpose of
vibration reduction have been presented, such as balancing of rotating and recipro
cating machine, control of natural frequency, damping and stiffness modiﬁcation,
isolators, and absorbers. Some of the techniques will be introduced brieﬂy in this
chapter.
An active vibration control is required for a system if it needs an external power
to perform its function. Examples of some active vibration control include, (1) using
hybrid mass dampers to apply a control force to a movable mass so as to reduce
building sway caused by wind and seismic waves, (2) reducing aircraft cabin noise
by attenuating the vibration of the large panels of thin metal that form the cabin
walls, (3) damping out vibrations using piezoelectric devices installed on the trailing
edge of helicopter blades, etc.
4.2 Passive control
4.2.1 Isolators
Vibration isolation methods are used to reduce the undesired effects of vibration. It
involves the insertion of a resilient member called an isolator between the vibrating
mass and the source of vibration so that a reduction in the dynamic response of
the system is achieved under speciﬁed conditions of vibration excitation [4]. An
isolation system is said to be active or passive depending on whether or not external
power is required for the isolator to perform its function. A passive isolator consists
63
64 Modeling and Control of Vibration in Mechanical Systems
of a resilient member, i.e., stiffness, and an energy dissipator, i.e., damping. Typical
examples of passive isolators include metal springs, cork, pneumatic springs, and
rubber springs.
4.2.2 Absorbers
If a system is acted upon by a force whose excitation frequency nearly coincides with
its natural frequency, the vibration of the system can be reduced by using a dynamic
vibration absorber, which is simply a springmass system. We consider an auxiliary
mass m
2
attached to a machine of mass m
1
through a spring of stiffness k
2
. The
motion equations of the masses m
1
and m
2
are written as
m
1
¨ x
1
+k
1
x
1
+ k
2
(x
1
−x
2
) = F
0
sinωt, (4.1)
m
2
¨ x
2
+k
2
(x
2
− x
1
) = 0, (4.2)
where x
1
and x
2
represent displacements of m
1
and m
2
, respectively. The steady
state amplitudes of the mass m
1
and m
2
are given by
X
1
=
(k
2
−m
2
ω
2
)F
0
(k
1
+k
2
−m
1
ω
2
)(k
2
−m
2
ω
2
) −k
2
2
, (4.3)
X
2
=
k
2
F
0
(k
1
+k
2
−m
1
ω
2
)(k
2
−m
2
ω
2
) −k
2
2
. (4.4)
Equation (4.3) implies that if
ω
2
=
k
2
m
2
,
the amplitude X
1
of the machine m
1
will be zero. Consider the machine operat
ing near its resonance ω
2
≃ k
1
/m
1
before the addition of the dynamic vibration
absorber. If the absorber is designed such that
ω
2
=
k
2
m
2
=
k
1
m
1
, (4.5)
the vibration amplitude of the machine, while operating at its original resonant fre
quency, will be zero.
The above mentioned dynamic vibration absorber removes the original resonance
peak in the response of the machine, but introduces two new resonance peaks. If
it is necessary to reduce the amplitude of vibration of the machine over a range of
frequencies, a damped dynamic vibration absorber can be used [5].
4.2.3 Resonators
Resonators are used in certain cases where it is easier and more efﬁcient to locate
antivibration systems near the vibration source. The main idea is to create another
source of vibrations which will cancel out the original vibration if it is correctly
tuned. The principle is to use the kinetic energy in the resonant mass system, where
Classical Vibration Control 65
the mass is coupled to the vibrating source either by kinematic coupling or via a very
ﬂexible link [8].
For example, in a rotating structure the stator is often excited by the rotor via
inertial effects such as imbalance or other effects such as the aerodynamic asymmetry
of the blades or ﬂaps. Positioning a resonator on the rotor is to generate loads that
oppose the excitation loads. In the case of a helicopter, the resonant mass is placed
on the axis of the rotor hub. It is supported by three springs which allow it to vibrate
in a plane perpendicular to the axis of rotation. The motion of the counterweight is
mainly inplane motion so that the system will generate inplane loads to eliminate
the inplane vibration.
4.2.4 Suspension
Suspension acts as the link between two structures that is designed in order to isolate
one of the structures from the other. This is a concept for structure isolation that is
achieved by canceling the variable loads transmitted to the support structure. The
characteristics of the link are determined by analyzing the dynamic behavior and vi
brations without modifying its main function in static characteristics. In practice, the
link characteristics can be modiﬁed by varying its stiffness or appropriate positioning
of the system natural frequencies and its damping in terms of energy transfer. An ad
ditional technique consists of introducing a ﬂapping mass whose inertial effects will
neutralize the excitation inputs. Next we will brieﬂy introduce stiffness modiﬁcation
and damping modiﬁcation.
Stiffness modiﬁcation
The dynamic behavior of structures results from the exchange and dissipation of
energy. Dynamic forces transfer their energy to the structure, which then responds
via several mechanisms, such as bending or extension. Dynamic behavior can be
modeled in several ways. The best known is Newton’s second law of motion. If
the external force is static or quasistatic, structural stiffness forces develop to create
an equilibrium. External dynamic forces are balanced in a more complex way with
inertial and damping forces.
Stiffness, often schematically and conceptually represented by a spring, denotes
the capacity of a system to store strain energy. The stiffness force follows Hooke’s
law, F
s
= k
s
∆x, where the stiffness constant k
s
is expressed in the unit of force per
unit length. This is a linear model where the spring displacement is measured from
the rest length. More complicated laws exist, for example, the nonlinear relation
F = k(x)∆x
a
, where the parameter a would depend on the particular material being
modeled, and the stiffness parameter k(x) is a function of how much the spring has
been elongated.
Damping modiﬁcation
Damping deﬁnes the ability of a structure to dissipate energy. For an oscillatory
system, damping is a measure of howmuch energy is dissipated by the system during
an oscillation cycle. For example, structural connections between components add
damping to a structure.
66 Modeling and Control of Vibration in Mechanical Systems
Most systems possess damping to some extent, which is helpful in vibration con
trol. If a system undergoes a forced vibration, its response or amplitude of vibration
near resonance tends to become large if there is no damping. The presence of damp
ing always limits the amplitude of vibration. Damping can be modiﬁed in the system
to control its response, by the use of structural materials having high internal damp
ing, such as laminated or sandwiched materials. An example is the use of viscoelastic
materials.
We know that the response amplitude of a system at resonance ω = ω
n
under
harmonic excitation F(t) = F
0
e
iωt
is given by
F
0
kη
=
F
0
αEη
, (4.6)
where η is the energy loss factor, and the stiffness k is proportional to the Young’s
modulus, i.e., k = αE with a constant α.
When viscoelastic materials are used for vibration control, they are subjected to
shear or direct strains. A simplest arrangement is that a layer of the viscoelastic
material is attached to an elastic one. Another arrangement is that a viscoelastic
layer is sandwiched between the elastic layers. The material with the largest loss
factor will be subjected to the smallest stress, while the stress is proportional to
the displacement. Hence viscoelastic materials having large loss factor are used to
provide internal damping for vibration control. Damping tapes, consisting of thin
metal foil covered with a viscoelastic adhesive, are used on vibrating structures.
Another application example is presented as follows with detailed discussion and
evaluation [14].
4.2.5 An application example − Disk vibration reduction via stacked
disks
To support higher track per inch (TPI) density hard disk drive, the positioning accu
racy of test equipment such as spin stand and servo track writer has to be increased.
The Spin stand is commonlyused equipment for magnetic recording component test
ing [10]. A servo track writer (STW) [25] is used to pattern the recording media with
servo information for the HDD servo system. In both cases, spindle motor and disk
vibrations [23] affect the positioning accuracy and limit how close adjacent tracks
can be placed together, and thus restrict the TPI number that can be achievable. In
[11], damped laminated disks were used to reduce the amplitude of rocking modes
by sandwiching a viscoelastic layer in between two aluminum layers to increase disk
damping. Reference [24] investigated the effect of suppressing resonance amplitude
of disk vibrations by applying a squeeze ﬁlm damping. In addition to the laminated
disks and the squeezed air bearing plate methods, in this section we discuss an ap
proach with minimal mechanical alternation. That is to stack more than one normal
recording media together for reading and writing on the disk surfaces.
The Guzik spin stand shown in Figure 4.1 is used to evaluate the effectiveness of
dualdisk stack in position accuracy improvement. The disk vibrations in the axial
Classical Vibration Control 67
direction are also measured at the outer diameter (OD) region of the disk via LDV.
In the experiment, the spindle motor is spinning at 7200, 8400 and 10200 RPM
respectively, and the disk is an aluminum disk of 3.5inch in diameter and 0.8 mm in
thickness. Two such disks are stacked together and mounted on the spindle motor of
the spin stand. The experimental process is described as follows.
Step 1: Stack two disks together and mount on the spin stand;
Step 2: Spin the motor and write servo information on the top disk surface;
Step 3: Read back the written information and obtain the PES.
To write servo information on the disk surface in between the two disks, the spin
dle motor is stopped and the written surface of the disk is ﬂipped over, following
which, steps 1, 2 and 3 are repeated.
FIGURE 4.1
Disk and spindle motor assembly of the spin stand.
The disk vibration in the axial direction of the single disk and the dualdisk stack
are measured via LDV. The position error signals are also collected for both the
single disk and the dualdisk stack cases. To evaluate the improvement of position
accuracy, 20 traces of the position error signal are collected and denoted by PES
i
68 Modeling and Control of Vibration in Mechanical Systems
(i = 1, · · · , N). N = 20. The repeatable runout (RRO) and the nonrepeatable
runout (NRRO) are given by
RRO =
1
N
N
i=1
PES
i
, NRRO
i
= PES
i
− RRO. (4.7)
A. Positioning accuracy improvement at 7200 RPM
Tested disk vibration results of the dualdisk and the single disk are shown in Fig
ure 4.2. Besides the harmonics with respect to the spindle speed, the disk vibration
modes can be seen clearly as denoted by T1T9. It is found that the frequencies of the
disk vibration modes T1T9 are shifted and their amplitudes are apparently reduced.
The vibration modes are all shifted by 40 Hz to higher frequencies. The power spec
tra of the RRO and NRRO, computed from the collected PES data, are shown in
Figure 4.3 and Figure 4.4. As seen in Figure 4.3, except for the 1st harmonic, other
dominant harmonics before the 11th harmonic are all reduced. In Figure 4.4, we
can observe that almost all the reduced disk vibration modes T1T9 in Figure 4.2
are reﬂected in NRRO. The other peaks indicated by S1S4 can be identiﬁed to be
caused by slider vibrations. They are both lowered signiﬁcantly, which may be due
to better sliderdisk interaction. It is noted that the peak S2 for the single disk case is
shifted to S3 in the dualdisk case. This randomly happens and the writing process
at a different time may cause such a shifting. This phenomenon can be seen in other
testing results which will be shown later. The standard deviations (or σ values) of
RRO, NRRO and PES are obtained. It turns out that the σ value of RRO is reduced
by 23%, NRRO by 18%, and the PES σ value is reduced by 21%. Figure 4.5 shows
the time sequences of PES in both the cases and the amplitude reduction is seen
apparently.
B. Positioning accuracy improvement at 8400 RPM
The disk or the dualdisk stack is rotating at 8400 RPM, and the disk vibrations
in the axial direction are shown in Figure 4.6. It can be seen that the amplitudes of
the vibration modes are all reduced with shifts in frequencies, as indicated by T1T6.
The power spectra of the RRO and NRRO are calculated from the tested PES and
shown in Figure 4.7 and Figure 4.8. Figure 4.7 shows that almost all the ﬁrst seven
harmonics are decreased and as a result, the σ value of RRO is reduced by 41%.
In Figure 4.8, the corresponding T1T6 in Figure 4.6 can be found and they are all
suppressed in the case of the dualdisk stack. Notice that S1 is reduced signiﬁcantly,
but can not be traced in the tested disk vibrations in Figure 4.6. This further veriﬁes
that S1 is caused by the diskslider interaction. The corresponding S4 cannot be
found in this case, while T1 appears very close to S4. The resultant σ value of
NRRO is improved by 28%, and PES by 38%. Figure 4.9 shows the comparison of
singledisk and dualdisk PES in time domain, and the amplitude of PES in the case
of the dualdisk stack approach is much lower as compared to conventional single
disk approach.
Classical Vibration Control 69
0 500 1000 1500 2000 2500 3000
−9
−8.5
−8
−7.5
−7
−6.5
−6
−5.5
−5
−4.5
Frequency(Hz)
D
s
p
l
a
c
e
m
e
n
t
a
m
p
l
i
t
u
d
e
(
l
o
g
1
0
(
m
)
)
2−disk
1−disk
T5
T6 T7
T8
T9
T4
T3
T2
T1
FIGURE 4.2
Comparison of singledisk and dualdisk axial vibrations measured via LDV at 7200
RPM.
70 Modeling and Control of Vibration in Mechanical Systems
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.05
0.1
Frequency(Hz)
F
F
T
(
R
R
O
)
(
V
r
m
s
)
1−disk
2−disk
FIGURE 4.3
Comparison of singledisk and dualdisk RRO power spectrum at 7200 RPM (23%
improvement of σ value).
Classical Vibration Control 71
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Frequency(Hz)
F
F
T
(
N
R
R
O
)
(
V
r
m
s
)
2−disk
1−disk
T8
T9
T7
T6
T4
T3
T2
T1
S1
S2
S3
S4
FIGURE 4.4
Comparison of singledisk and dualdisk NRRO power spectrum at 7200 RPM (18%
reduction of σ value).
72 Modeling and Control of Vibration in Mechanical Systems
FIGURE 4.5
Comparison of singledisk and dualdisk PES in time domain at 7200 RPM (21%
reduction of σ value).
Classical Vibration Control 73
0 500 1000 1500 2000 2500 3000
−9
−8.5
−8
−7.5
−7
−6.5
−6
−5.5
−5
Frequency(Hz)
D
s
p
l
a
c
e
m
e
n
t
a
m
p
l
i
t
u
d
e
(
l
o
g
1
0
(
m
)
)
2−disk
1−disk
T1
T2
T3
T4
T5
T6
FIGURE 4.6
Comparison of singledisk and dualdisk axial vibrations measured via LDV at 8400
RPM.
74 Modeling and Control of Vibration in Mechanical Systems
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.05
0.1
Frequency(Hz)
F
F
T
(
R
R
O
)
(
V
r
m
s
)
1−disk
2−disk
FIGURE 4.7
Comparison of singledisk and dualdisk RRO power spectrum at 8400 RPM (41%
improvement of σ value).
Classical Vibration Control 75
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Frequency(Hz)
F
F
T
(
N
R
R
O
)
(
V
r
m
s
)
1−disk
2−disk
T6
T4
T3
T2
T1
T5
S1
S2(S3)
FIGURE 4.8
Comparison of singledisk and dualdisk NRRO power spectrum at 8400 RPM (28%
reduction of σ value).
76 Modeling and Control of Vibration in Mechanical Systems
FIGURE 4.9
Comparison of singledisk and dualdisk PES in time domain at 8400 RPM (38%
reduction of σ value).
Classical Vibration Control 77
C. Positioning accuracy improvement at 10200 RPM
A higher rotational speed of 10200 RPM is performed on both the dualdisk stack
and single disk cases. Figure 4.10 shows the obtained disk axial vibration by LDV.
The modes T1T6 are quite obvious with the amplitude reduced and the frequencies
shifted higher. Figure 4.11 and Figure 4.12 show the power spectrum of RRO and
NRRO. The 3rd and 5th harmonics are signiﬁcantly reduced as seen in Figure 4.11.
Figure 4.12 reﬂects the corresponding disk vibration modes T1T6 in Figure 4.10.
As for the slider related vibrations, S1 is reduced signiﬁcantly again and S2 and S3
are shifted, while S4 cannot be seen at all in Figure 4.12. As a result, the σ value of
RRO is improved by 33%, NRRO by 27%, and PES by 32%, compared with that of
the single disk approach. Figure 4.13 shows the comparison of PES in time domain,
and it is observed that the amplitude of PES in the case of the dualdisk stack is
decreased.
0 500 1000 1500 2000 2500 3000
−9
−8.5
−8
−7.5
−7
−6.5
−6
−5.5
−5
−4.5
Frequency(Hz)
D
s
p
l
a
c
e
m
e
n
t
a
m
p
l
i
t
u
d
e
(
l
o
g
1
0
(
m
)
)
2−disk
1−disk
T3
T5
T4
T6
T2
T1
FIGURE 4.10
Comparison of singledisk and dualdisk axial vibrations measured via LDV at
10200 RPM.
The reduction of disk vibration amplitude is evaluated from Figures 4.2, 4.6 and
4.10 and tabulated in Table 4.1 for different rotational speeds. The improvements of
78 Modeling and Control of Vibration in Mechanical Systems
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.05
0.1
0.15
0.2
Frequency(Hz)
F
F
T
(
R
R
O
)
(
V
r
m
s
)
1−disk
2−disk
FIGURE 4.11
Comparison of singledisk and dualdisk RRO power spectrum at 10200 RPM (the
3rd and 5th harmonics reduced signiﬁcantly, 33% improvement of σ value).
Classical Vibration Control 79
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Frequency(Hz)
F
F
T
(
N
R
R
O
)
(
V
r
m
s
)
1−disk
2−disk
T1
T2
T3
T4
T6
T5
S1
S2
S3
FIGURE 4.12
Comparison of singledisk and dualdisk NRRO power spectrum at 10200 RPM
(27% reduction of σ value).
80 Modeling and Control of Vibration in Mechanical Systems
FIGURE 4.13
Comparison of singledisk and dualdisk PES in time domain at 10200 RPM (32%
reduction of σ value).
Classical Vibration Control 81
TABLE 4.1
% reduction of σ values of PES, RRO and NRRO and disk
vibration amplitude with stacked disks compared with single disk
Speed(RPM) RRO NRRO PES Disk vibration amplitude
7200 23 18 21 24
8400 41 28 38 45
10200 33 27 32 35
the σ values of RRO, NRRO and PES are also summarized in Table 4.1. Remarkable
improvement is achieved for every rotational speed. It can be seen clearly from the
table that the best improvement is obtained when operating at 8400 RPM. The second
best result obtained is performed at 10200 RPM, where the dualdisk stack approach
leads to an improvement of 32% in positioning accuracy.
For all the cases studied, we have observed a remarkably reduced amplitude of the
disk vibration and its reﬂection in positioning accuracy improvement via the dual
disk stack. This indicates that the stacked disk is a simple way of reducing the disk
vibrations in spin stand tests, particularly when a single disk surface is accessed.
In Figures 4.2, 4.6, and 4.10, we have also observed the frequency shifting of disk
vibration modes.
Theoretically the maximum amplitude of disk vibration is given by [22]
V ∝
ρR
4
ω
2
t
3
Ed
, (4.8)
where R is the disk outer radius, ρ is the gas density, ω is the angular speed, t is
the disk thickness, E is the Young’s modulus of the disk substrate, and d is the
disk substrate damping. Equation (4.8) indicates that the disk vibration is inversely
proportional to t
3
. In the dualdisk case, the disk stack thickness
˜
t = 2t. Hence
assuming that the change of E is negligible, the amplitude of disk vibration
˜
V ∝
ρR
4
ω
2
8t
3
E
˜
d
, (4.9)
which means that
˜
V < V when the damping
˜
d >
1
8
d. The experimental results in
Figures 4.2, 4.6 and 4.10 veriﬁed the amplitude reduction of the disk vibration.
On the other hand, it is known that the natural frequency of disk vibration mode is
deﬁned by [23]
f
mn
=
λ
2
mn
2πR
2
Et
3
12γ(1 − ν
2
)
1/2
, (4.10)
where f
mn
is the natural frequency of the (m, n) mode, γ is the mass per unit area of
the disk, ν is the Poisson’s ratio, and λ
mn
is the dimensionless frequency parameter
which is generally a function of the boundary conditions on the plate, the ratio of the
82 Modeling and Control of Vibration in Mechanical Systems
inner to outer diameter, and Poisson’s ratio. Denoting Ω as the rotational speed of
the disk, the frequency of the disk vibration mode is given by [23]
f = f
mn
±nΩ. (4.11)
Assuming that the change of ν is also negligible, and considering that
˜
t = 2t,
˜
f
mn
= 2
√
2
˜
λ
2
mn
2πR
2
Et
3
12γ(1 − ν
2
)
1/2
, (4.12)
which implies some shifting of the disk vibration modes, as can be seen in the ex
perimental results in Figures 4.2, 4.6 and 4.10.
In addition to the doubled thickness, the mechanical properties such as d and λ
mn
[22][23] are certain to change correspondingly after stacking two disks. Further
research for details on how the stacked disks change these properties would be sig
niﬁcant.
To summarize, the proposed process is indeed an approach to reduce the disk
vibration on both disk surfaces for reading and writing, such as multidisk servo
track writers [25]. Thus such a disk vibration reduction approach can be used as an
alternative or complementary technique to the air shroud design [24] and the active
control approaches to further improve positioning accuracy.
4.3 Selfadapting systems
In certain applications, system dynamic parameters vary with time. For example,
the mass of a system like a car, airplane or helicopter decreases as it consumes fuel.
Distinct ﬂight conditions such as ﬂight level, maneuvers, or landing produce different
types of excitation. When the initial characteristics of the system no longer meet the
requirements of the system’s working conditions, the systemcharacteristics therefore
need to adapt to the parameter variations of the excitation and the system itself. Two
ways are possible: one is to change the stiffness of the structure; the other is to add
moving masses so that their inertial effects counteract the effects of the excitation
variation. Some selfadapting systems in use are brieﬂy introduced as follows [8].
Selftuning suspension
Selftuning suspension systems are equipped suitably for slowvariation of the sys
tem characteristics. One method, for example, used in selftuning systems involves
analyzing the system’s vibration frequency and its timewise variation. The tuning
is usually performed by an actuator. After the system status is measured, the tuning
system will modify the feature of the system, such as stiffness, damping and position
of a mass.
The selftuning system is tuned by a designed control algorithm. Simulations
using the control algorithm are required to identify the setting parameters of the
Classical Vibration Control 83
algorithm and to verify the reliability of the system. Subsequently, the algorithm
must be validated by experimental tests on the real structure. The car suspension
is an example of selftuning suspensions. It is able to adapt to various operating
conﬁgurations depending on the type of driving, the drivingconditions, or the desired
comfort. One method to achieve this is to modify the suspension stiffness or its
damping using various techniques.
Selfadjusting absorbers
The mechanical resonator, which is basically a massspring system, is used to con
trol certain vibration sources. It is placed where the vibrations are to be reduced. The
principle is that the resonance frequency of the massspring system as a resonator is
adjusted with the excitation frequency in operation, and as the mass vibrates, the
vibratory level is decreased.
In certain structures, the excitation frequency evolves with time gradually. These
changes take place over a long period, compared to the excitation frequency. As a
result, it is necessary to create a system whose resonance frequency would adjust
automatically to the variations of the excitation frequency. For a 20 Hz excitation
frequency, this variation usually takes place between 1 and 3 seconds. This kind of
system actually works as the case of selfadjusting absorbers.
Selfadapting resonator
Selfadapting resonators are capable of adjusting to changes in the excitation fre
quency. The hub resonator used in a helicopter is an example. It consists of a ﬂapping
mass that vibrates in a plane perpendicular to the rotation axis of the rotor. The mo
bile ﬂapping mass is supported by three ﬂexible elements indexed at 120
◦
, and it
slides along the rotor axis. The stiffness of the three elements is designed so that
the resulting antiresonance corresponds to the excitation frequency. The required
position of the mobile mass is determined by control algorithms through an actuator
and the mobile mass is moved by the actuator. The z position of the mass makes the
variation of the inertia of the assembly and thus the antiresonance frequency varies
accordingly.
4.4 Active vibration control
The selfadapting systems presented previously will not be sufﬁcient in the cases
where the source characteristics vary too fast for the involved algorithms or the re
quired level of performance is too high. Active methods should be used to decrease
the vibrations.
A vibration control system is called active if it uses external power to perform
its function. It is comprised of a servomechanism with an actuator, a sensor, and a
microprocessorbased system. The actuator applies a force to the mass whose vi
bration is to be reduced. The sensor measures the motion of the mass in terms of
displacement, velocity, or acceleration, depending on the application. The micropro
84 Modeling and Control of Vibration in Mechanical Systems
cessor based system consists of analogtodigital converters to process sensor inputs
and digitaltoanalog converters to convert the microprocessor’s output command
into an input signal to the actuator. The control logic, called the control algorithm,
programmed in the computer uses sensor measurement to decide how much force
the actuator should apply.
4.4.1 Actuators
Piezoelectric materials can convert electrical current into motion, and vice versa.
They change shape when an electric current passes through them, and they generate
an electric signal when they ﬂex. Thus they can be used as actuators to create force
or motion, and as sensors to sense motion.
The materials used for highprecision actuation include electrostrictive and mag
netostrictive materials, which are similar to piezoelectric materials. These are ferro
magnetic materials that expand or contract when subjected to an electric or a mag
netic ﬁeld.
The previously mentioned voice coil motor (VCM) actuator in hard disk drives is
a linear actuator moving in one direction. Because of its similarity to a loudspeaker,
it is referred to as a voice coil motor. A VCM actuator has a coil of wire rigidly
attached to the structure and suspended in a permanent magnetic ﬁeld. It is driven
when a force is produced to accelerate it radially as a current is passed through the
coil.
Some actuators cannot provide enough force for larger applications such as vehi
cle suspension and control of building motions. In building application, hydraulic
cylinders are usually used. With a working hydraulic pressure of commonly 2000
psi, a cylinder containing a piston whose area is only 1 square inch will generate 1
ton of force. Active vehicle suspensions use hydraulic devices, electric motors, and
magnetorheological ﬂuid dampers.
4.4.2 Active systems
An active system is applied to decrease vibrations by introducing dynamic loads
locally in the structure [8]. The dynamic loads are controlled by a processor in order
to minimize the vibratory level. The technology of generating loads is fundamental
in control strategy.
Active suspension
Several active principles are used for suspensions to isolate one structure from an
other. One of them is internal load control to modify the distribution of internal loads
in the structure. Its principle is to inject a set of dynamic loads into the structure in
order to minimize its vibratory response. The loads depend on the vibratory condi
tion of the structure, which is sensed with the help of a set of accelerometers or strain
gauges. Hydraulic or electrodynamic actuators suitably located on the structure are
used to inject the loads. The role of the actuators is to modify the distribution of
vibratory energy for different modes to minimize the structure vibrations, instead of
dispersing the vibratory energy of the structure.
Classical Vibration Control 85
Active resonator
An active resonator works by displacing a mass with an actuator. It uses the dy
namic ampliﬁcation of the mass to generate high loads using minimal energy. Differ
ent types of actuator technologies such as hydraulic, electromagnetic, or piezoelec
tric actuators, have been used to match the ﬁelds of application in terms of forces and
frequencies. We take an electromagnetic actuator as an example. The type of single
stage resonator uses the principle of single stage mechanical systems with displace
ment operated by an electromagnetic force. The motion equation of the mass M of
a permanent magnet for small motion around its static position is given by
M¨ x
1
+c
1
˙ x
1
+ k
1
x
1
= F
V
(t). (4.13)
With F
V
(t) = F
0
cos(ωt), the Laplace transform of the displacement produced by
the mass is
X
1
(s) =
1
Ms
2
+ c
1
s + k
1
F
0
, (4.14)
where k
1
and c
1
are respectively the stiffness and damping of the mass.
The load transmitted to the structure equals to
F
T
(t) = c
1
˙ x
1
− k
1
x
1
−F
V
(t). (4.15)
The mechanical parameters such as mass and stiffness are designed so that the
magnetic loads remain small and compatible with an acceptable energy consump
tion. The natural frequency of the massspring system must be designed as close as
possible to the frequency of the vibration that needs to be controlled.
A twostage electromagnetic resonator uses two masses called stages [8]. The
control force F
V
is introduced between the two masses by an electromagnetic load.
It is produced froma voltage V that is fed to a power ampliﬁer to generate a magnetic
force via the current in the coil. The role of the ampliﬁer is to ensure the law between
the generated force F
V
and the control voltage V is linear. The function of control
consists of tuning the system parameters such as mass and stiffness so that at the
control frequency, the force generated on the structure is ampliﬁed with respect to
the force F
V
, while maintaining a small relative displacement of the two stages.
The motion equations of the two masses are written as
M
1
¨ x
1
+ c
1
( ˙ x
1
− ˙ x
2
) +k
1
(x
1
−x
2
) = F
V
(t), (4.16)
M
2
¨ x
2
+ c
1
( ˙ x
2
− ˙ x
1
) +c
2
˙ x
2
+ k
1
(x
2
−x
1
) +k
2
x
2
= −F
V
(t), (4.17)
with the primary stage M
2
and the secondary stage M
1
.
The load transmitted to the structure is
F
T
= c
2
˙ x
2
+ k
2
x
2
. (4.18)
86 Modeling and Control of Vibration in Mechanical Systems
4.4.3 Control strategy
When the system response is not acceptable, the value of the response is used to
generate additional forces according to rules or laws such that the modiﬁed response
behaves according to the design and within certain bounds. This results in a closed
loop system that incorporates feedback control, where the response is evaluated by a
sensor and is fed back to an actuator that generates a force or motion. The purpose of
designing a system with feedback force is to minimize unwanted behaviors and elim
inate the effect of vibration on system performance in a desired manner. Feedback
control provides a mechanism for tailoring system behavior to speciﬁc standards and
needs. The block diagram depicting feedback control is in Figure 4.14. One can see
that the signal from a block goes concurrently to other blocks or a summing point.
The summing point indicates that the input and output signals are compared. The
difference is an error that generates a control action. The disturbance or forcing is
also shown being input directly to the plant or being injected to the output of the
plant.
FIGURE 4.14
Generic feedback control system.
In order to properly design a feedback control system, performance must be de
ﬁned in terms of system speciﬁcations. Standard performance measures are usually
deﬁned in terms of the step response. The general design objectives are the speed of
response, stability, and accuracy or allowable error. The ﬁrst objective implies that it
is desirable for a control system to respond to a reference command rapidly. Stability
is the prerequisite of any active control systemdesign. The allowable error represents
how close to the desired response the control force must bring the structure and how
much the control system can reject vibrations.
The controller produces control action or signal by using the comparison signal
of plant output with the desired reference value. There are a number of control
Classical Vibration Control 87
methods used by people to design control actions. Classical control is represented
by the wellknown proportionalintegralderivative (PID) control. Advanced control
techniques include robust and optimal controls, which will be introduced brieﬂy in
the next chapter.
An adaptive control system is one that can change or adapt either its gain values
or even its control algorithm to accommodate changing conditions. In mathematics
how the control gains or the control algorithm should adapt to changing conditions
is difﬁcult. There are thus fewer practical applications of adaptive controls.
4.5 Conclusion
Various classical techniques of vibration control were discussed in this chapter and
some typical techniques such as using isolators, absorbers, a resonator, and suspen
sion were introduced individually. Speciﬁcally, disk vibration reduction via stacked
disks as an application example was presented with detailed discussion and evalu
ation. Subsequently, active vibration control was introduced with actuators, active
systems, and control strategies.
5
Introduction to Optimal and Robust Control
5.1 Introduction
This chapter intends to brieﬂy review the optimal and robust control. H
2
and H
∞
norms and their calculation are ﬁrst introduced since they are two commonly used
measures of performance speciﬁcations of a control system. What follow are the
H
2
and H
∞
control problems and controller design via a linear matrix inequality
approach. Robust control of systems with multiplicative uncertainty and additive un
certainty, and the parametrization of all stable controllers, i.e., the socalled Youla
parametrization, will be investigated. Moreover, performance limitation of a feed
back control system is discussed, which is necessary to help understand the vibration
control schemes in the later chapters.
5.2 H
2
and H
∞
norms
5.2.1 H
2
norm
The H
2
norm of a matrix transfer function G(s) analytic in Re(s) > 0 (open right
half plane) is deﬁned as
G
2
:=
¸
sup
σ>0
{
1
2π
_
+∞
−∞
Trace[G
∗
(σ + jω)G(σ + jω)]dω}, (5.1)
or equivalently [2]
G
2
=
¸
1
2π
_
+∞
−∞
Trace[G
∗
(jω)G(jω)]dω. (5.2)
Although G
2
can be computed from its deﬁnition, there are some simple alter
natives taking advantage of a statespace representation of G(s).
LEMMA 5.1
[6] Consider a system G(s) with a statespace representation (A, B, C, D). If
89
90 Modeling and Control of Vibration in Mechanical Systems
A is stable and D = 0, then we have
G
2
2
= Trace(B
T
Y
2
B) = Trace(CX
2
C
T
) (5.3)
where X
2
and Y
2
are the controllability and observability Gramians that can
be obtained from the following Lyapunov equations:
AX
2
+ X
2
A
T
+ BB
T
= 0, (5.4)
A
T
Y
2
+ Y
2
A+C
T
C = 0. (5.5)
We also consider a discretetime linear timeinvariant system G(z) with the fol
lowing statespace representation:
x(k + 1) = Ax(k) +Bw(k), (5.6)
z(k) = Cx(k) + Dw(k), (5.7)
where x ∈ R
n
x
is the state, z ∈ R
n
z
is the controlled output, w ∈ R
n
w
is the
disturbance input. Let T
zw
denote the transfer function from the input w to the
output z. Then the H
2
norm is deﬁned as
T
zw
2
=
¸
1
2π
Trace[
_
π
−π
T
∗
zw
(e
jω
)T
zw
(e
jω
)dω]. (5.8)
By Parseval’s theorem, T
zw
2
can equivalently be deﬁned as
T
zw
2
=
¸
¸
¸
_
Trace[
∞
k=0
g(k)g
T
(k)], (5.9)
where g(k) is the impulse response of T
zw
.
Let the input w to the system be a wideband stationary stochastic process. The
H
2
norm of T
zw
can also be interpreted as the RMS value of the output z(k) when
the system is subject to a white noise having zero mean and unit variance. That is
T
zw
2
=
_
E[z
T
z]. (5.10)
The H
2
norm for the discretetime system T
zw
can be computed as
T
zw
2
=
_
Trace(D
T
D +B
T
Y
2
B) =
_
Trace(DD
T
+CX
2
C
T
), (5.11)
where Y
2
and X
2
are the reachability and observability Gramians that can be ob
tained from the following Lyapunov equations:
AY
2
A
T
−Y
2
+ BB
T
= 0, (5.12)
A
T
X
2
A− X
2
+C
T
C = 0. (5.13)
Introduction to Optimal and Robust Control 91
The following theorem presents an alternative LMI condition for bounding the H
2
norm of the discretetime system T
zw
.
LEMMA 5.2
Consider a discretetime transfer function T
zw
of realization (A, B, C, D).
Given a scalar µ > 0, T
zw
2
2
< µ if and only if there exist X
2
= X
T
2
and
Y
2
= Y
T
2
such that Trace(Π) < µ and
_
_
Π CX
2
D
X
2
C
T
X
2
0
D
T
0 I
_
_
> 0, (5.14)
_
_
Y
2
AY
2
B
Y
2
A
T
Y
2
0
B
T
0 I
_
_
> 0. (5.15)
Observe that (5.14) and (5.15) are linear in X
2
and Y
2
, and hence can be solved
by employing the LMI Tool [69] in MATLAB. The H
2
norm of the system can
be computed by minimizing µ using the function mincx in MATLAB Optimization
Toolbox.
5.2.2 H
∞
norm
The H
∞
norm of a matrix transfer function G(s) that is analytic and bounded in the
open righthalf plane is deﬁned as [3]
G
∞
:= sup
Re(s)>0
¯ σ[G(s)] = sup
ω∈R
¯ σ[G(jω)] = sup
w =0
z
2
w
2
, (5.16)
where w and z are respectively the input and output of G(s). A control engineering
interpretation of the inﬁnity norm of a scalar transfer function G(s) is the distance in
the complex plane from the origin to the farthest point on the Nyquist plot of G, and
it also appears as the peak value on the magnitude plot of G(jω). Hence the H
∞
norm of a transfer function can, in principle, be obtained graphically.
In general, the H
∞
norm of a stable matrix transfer function can be read directly
from its singular value plots.
The H
∞
norm can also be computed in statespace.
LEMMA 5.3
[6] Let γ > 0 and G(s) : (A, B, C, D) with A stable. Then G(s)
∞
< γ if
and only if ¯ σ(D) < γ and the Hamiltonian matrix H has no eigenvalues on
the imaginary axis, where
H :=
_
A +BR
−1
D
T
C BR
−1
B
T
−C
T
(I + DR
−1
D
T
)C −(A +BR
−1
D
T
C)
T
_
, (5.17)
92 Modeling and Control of Vibration in Mechanical Systems
and R = γ
2
I − D
T
D.
The following socalled Bounded Real Lemma gives a matrix inequality condition
for the system G(s) to have a prespeciﬁed level of H
∞
norm.
LEMMA 5.4
Continuoustime Bounded Real Lemma Consider a continuoustime transfer
function G(s) of realization G(s) = C(sI − A)
−1
B + D, where A is a stable
matrix. Given a scalar λ > 0, the H
∞
norm G
∞
< λ if and only if there
exists X = X
T
> 0 such that
_
_
A
T
X +XA XB C
T
B
T
X − λI D
T
C D −λI
_
_
< 0. (5.18)
In the discretetime case, the H
∞
normof a stable transfer function matrix T
zw
(z)
with realization (A, B, C, D) is deﬁned as
T
zw
∞
= sup
ω∈[0,2π]
¯ σ[T
zw
(e
jω
)] = sup
w =0
z
2
w
2
, (5.19)
where w and z are respectively the input and output of T
zw
.
The following discretetime Bounded Real Lemma provides a linear matrix in
equality condition for the system T
zw
to have T
zw
∞
< γ.
LEMMA 5.5
Discretetime Bounded Real Lemma Consider a discretetime and a stable
transfer function matrix T
zw
(z) with a statespace realization (A, B, C, D).
Then, for some given λ > 0, T
zw
∞
< λ if and only if there exists X =
X
T
> 0 such that
_
_
A
T
XA −X A
T
XB C
T
B
T
XA B
T
XB −λ
2
I D
T
C D −I
_
_
< 0. (5.20)
5.3 H
2
optimal control
5.3.1 Continuoustime case
We consider the closedloop system described by the block diagram in Figure 5.1.
The continuoustime linear timeinvariant plant P(s) is described by the statespace
equations:
˙ x(t) = Ax(t) +B
1
w(t) +B
2
u(t), (5.21)
Introduction to Optimal and Robust Control 93
z(t) = C
1
x(t) +D
11
w(t) + D
12
u(t), (5.22)
y(t) = C
2
x(t) +D
21
w(t) + D
22
u(t), (5.23)
where x ∈ R
n
x
is the state, y ∈ R
n
y
is the measurement output, z ∈ R
n
z
is the
controlled output, w ∈ R
n
w
is the disturbance input, u ∈ R
n
u
is the control input,
and A, B
1
, B
2
, C
1
, D
11
, D
12
, C
2
, and D
21
are of appropriate dimensions. We
assume D
22
= 0 without loss of generality [89].
Introduce the following dynamic output feedback controller C(s):
˙ x
c
(t) = A
c
x
c
(t) +B
c
y(t), (5.24)
u(t) = C
c
x
c
(t) + D
c
y(t). (5.25)
Denote ξ = [x
T
x
T
c
]
T
. From (5.21)−(5.23) and (5.24)−(5.25), the closedloop
system is given by
˙
ξ(t) =
¯
Aξ(t) +
¯
Bw(t), (5.26)
z(t) =
¯
Cξ(t) +
¯
Dw(t), (5.27)
where
¯
A =
_
A+B
2
D
c
C
2
B
2
C
c
B
c
C
2
A
c
_
,
¯
B =
_
B
2
D
c
D
21
+ B
1
B
c
D
21
_
, (5.28)
¯
C = [C
1
+D
12
D
c
C
2
D
12
C
c
] ,
¯
D = D
12
D
c
D
21
+ D
11
. (5.29)
The continuoustime H
2
control problemis to ﬁnd a proper, real rational controller
C(s) that stabilizes P internally and minimizes the H
2
norm of the closedloop
transfer function matrix T
zw
from w to z in (5.26)−(5.27).
Assume that the system (5.21)−(5.23) satisﬁes the following conditions:
Assumption 5.1
(1). D
12
is of full column rank;
(2). The subsystem (A, B
2
, C
1
, D
12
) has no invariant zeros on the imaginary axis;
(3). D
21
is of full row rank;
(4). The subsystem (A, B
1
, C
2
, D
21
) has no invariant zeros on the imaginary
axis.
Let X
2
≥ 0 and Y
2
≥ 0 be the solutions of the following Riccati equations:
A
T
X
2
+ X
2
A−(X
2
B
2
+C
T
1
D
12
)(D
T
12
D
12
)
−1
(X
2
B
2
+ C
T
1
D
12
)
T
+C
T
1
C
1
= 0, (5.30)
Y
2
A
T
+AY
2
− (Y
2
C
T
2
+B
1
D
T
21
)(D
21
D
T
21
)
−1
(Y
2
C
T
2
+B
1
D
T
21
)
T
+B
1
B
T
1
= 0. (5.31)
According to the H
2
optimal control theory, an H
2
optimal controller can be ob
tained as [7]
A
c
= A+ B
2
F + KC
2
, B
c
= −K, C
c
= F, D
c
= 0, (5.32)
94 Modeling and Control of Vibration in Mechanical Systems
FIGURE 5.1
Conﬁguration of standard optimal control.
where
F = −(D
T
12
D
12
)
−1
(D
T
12
C
1
+ B
T
2
X
2
),
K = −(Y
2
C
T
2
+B
1
D
T
21
)(D
21
D
T
21
)
−1
. (5.33)
The minimal H
2
norm of the transfer function T
zw
is given by
T
zw
2
=
_
Trace(B
T
1
X
2
B
1
) +Trace[(A
T
X
2
+X
2
A+ C
T
1
C
1
)Y
2
].(5.34)
If the conditions (1)−(4) in Assumption 5.1 are not satisﬁed, the socalled pertur
bation method is applied [7] so that the above design method to ﬁnd an appropriate
controller is still applicable.
5.3.2 Discretetime case
Consider the discretetime linear timeinvariant systemP(z) with the followingstate
space representation:
x(k + 1) = Ax(k) +B
1
w(k) + B
2
u(k), (5.35)
z(k) = C
1
x(k) +D
11
w(k) +D
12
u(k), (5.36)
y(k) = C
2
x(k) +D
21
w(k) +D
22
u(k), (5.37)
where x ∈ R
n
x
is the state, y ∈ R
n
y
is the measurement output, z ∈ R
n
z
is the
controlled output, w ∈ R
n
w
is the disturbance input, u ∈ R
n
u
is the control input,
and A, B
1
, B
2
, C
1
, D
11
, D
12
, C
2
, and D
21
are of appropriate dimensions. D
22
= 0
is also assumed.
Introduction to Optimal and Robust Control 95
Introduce the following dynamic output feedback controller C(z):
x
c
(k + 1) = A
c
x
c
(k) +B
c
y(k), (5.38)
u(k) = C
c
x
c
(k) + D
c
y(k). (5.39)
Denote ξ = [x
T
x
T
c
]
T
. From (5.35)−(5.37) and (5.38)−(5.39), the closedloop
system is given by
ξ(k + 1) =
¯
Aξ(k) +
¯
Bw(k), (5.40)
z(k) =
¯
Cξ(k) +
¯
Dw(k), (5.41)
where
¯
A =
_
A+B
2
D
c
C
2
B
2
C
c
B
c
C
2
A
c
_
,
¯
B =
_
B
2
D
c
D
21
+ B
1
B
c
D
21
_
, (5.42)
¯
C = [C
1
+D
12
D
c
C
2
D
12
C
c
] ,
¯
D = D
12
D
c
D
21
+ D
11
. (5.43)
The discretetime H
2
control problem is to ﬁnd a proper, real rational controller
C(z) that stabilizes P(z) internally and minimizes the H
2
norm of the transfer func
tion matrix T
zw
(z) from w to z of the closedloop system (5.40)−(5.41).
The counterpart of the Riccati equations (5.30)−(5.31) for discretetime systems
is as follows.
A
T
X
2
A −(A
T
X
2
B
2
+C
T
1
D
12
)(D
T
12
D
12
+ B
T
2
X
2
B
2
)
−1
(A
T
X
2
B
2
+ C
T
1
D
12
)
T
+C
T
1
C
1
= 0, (5.44)
AY
2
A
T
− (AY
2
C
T
2
+B
1
D
T
21
)(D
21
D
T
21
+C
2
Y
2
C
T
2
)
−1
(AY
2
C
T
2
+B
1
D
T
21
)
T
+B
1
B
T
1
= 0. (5.45)
A discretetime H
2
optimal controller can then be obtained as (5.32). And the
minimal H
2
norm of the transfer function T
zw
is given by
T
zw
2
=
_
Trace(B
T
1
X
2
B
1
) +Trace[(A
T
X
2
A +C
T
1
C
1
)Y
2
]. (5.46)
A parametrization of all H
2
controllers is developed in terms of LMIs as in the
following theorem which linearizes the H
2
norm conditions (5.14)−(5.15) for syn
thesis.
THEOREM 5.1
[90] Consider system (5.35)−(5.37). There exists a controller (5.38)−(5.39)
such that T
zw
2
2
< µ if and only if the following linear matrix inequalities
and equality admit a solution:
Trace(Π) < µ, (5.47)
96 Modeling and Control of Vibration in Mechanical Systems
_
_
Π C
1
X +D
12
E C
1
+ D
12
D
c
C
2
∗ X + X
T
−P
2
I +Z
T
−J
∗ ∗ Y + Y
T
− H
_
_
> 0, (5.48)
_
¸
¸
¸
¸
_
P
2
J AX + B
2
E A+ B
2
D
c
C
2
B
1
+ B
2
D
c
D
21
∗ H U Y A+ WC
2
Y B
1
+ WD
21
∗ ∗ X +X
′
−P
2
I + Z
′
−J 0
∗ ∗ ∗ Y + Y
T
− H 0
∗ ∗ ∗ ∗ I
_
¸
¸
¸
¸
_
> 0, (5.49)
and
D
11
+D
12
D
c
D
21
= 0, (5.50)
where ∗ denotes an entry that can be deduced from the symmetry of the matrix,
the matrices X, E, Y , W, U, D
c
, Z, J, and the symmetric matrices P
2
, H
and Π are the variables. A feasible H
2
controller is obtained by choosing N
1
and M
1
nonsingular such that N
1
M
1
= Z − Y X and calculating
C
c
= (E −D
c
C
2
X)M
−1
1
, D
c
= D
c
, (5.51)
B
c
= N
−1
1
(W −Y B
2
D
c
), (5.52)
A
c
= N
−1
1
[U −Y (A+ B
2
D
c
C
2
)X − N
1
B
c
C
2
X − Y B
2
C
c
M
1
]M
−1
1
. (5.53)
5.4 H
∞
control
5.4.1 Continuoustime case
We consider the continuoustime systemP described by (5.21)−(5.23), and the class
of causal, linear, timeinvariant and ﬁnitedimensional controllers that internally sta
bilize P, or namely, all admissible controllers for P. Our aim is to ﬁnd an admissible
controller C such that the closedloop system T
zw
satisﬁes
T
zw
∞
< γ. (5.54)
Assumption 5.2
(1). The pair (A, B
2
) is stabilizable and the pair (A, C
2
) is detectable.
(2). The matrices D
12
and D
21
satisfy D
T
12
D
12
= I
n
u
and D
21
D
T
21
= I
n
y
.
(3).
rank
_
A− jωI B
2
C
1
D
12
_
= n
x
+n
u
, for all real ω. (5.55)
(4).
rank
_
A− jωI B
1
C
2
D
21
_
= n
x
+n
y
, for all real ω. (5.56)
Introduction to Optimal and Robust Control 97
The assumption that (A, B
2
, C
2
) is stabilizable and detectable is necessary and
sufﬁcient for the existence of admissible controllers. The full rank assumptions (3)
and (4) are necessary for the existence of stabilizing solutions to the Riccati equations
that are used to obtain the solution to the H
∞
control problem.
Deﬁne the matrices
ˆ
A,
ˆ
B,
˜
A and
˜
C as
ˆ
A = A −B
1
D
T
21
C
2
,
ˆ
B
ˆ
B
T
= B
1
(I − D
T
21
D
21
)B
T
1
, (5.57)
˜
A = A −B
2
D
T
12
C
1
,
˜
C
T
˜
C = C
T
1
(I −D
12
D
T
12
)C
1
. (5.58)
Suppose that U
y
maps y to u and has a minimal realization (A
u
, B
u
, C
u
, D
u
)
satisfying det(I − D
22
D
u
) = 0 for a wellposed closed loop. Then a realization in
terms of LFT is given by
LFT(P, U
y
) =
_
_
A +B
2
D
u
MC
2
B
2
(I + D
u
MD
22
)C
u
B
1
+B
2
D
u
MD
21
B
u
MC
2
A
u
+ B
u
MD
22
C
u
B
u
MD
21
C
1
+ D
12
D
u
MC
2
D
12
(I + D
u
MD
22
)C
u
D
11
+D
12
D
u
MD
21
_
_
, (5.59)
with M = (I − D
22
D
u
)
−1
.
THEOREM 5.2
[71] Consider the system (5.21)−(5.23) satisfying Assumption 5.2. There
exists an admissible C such that the closedloop system (5.26)−(5.27) satisﬁes
(5.54) if and only if
1. There is a solution X
∞
≥ 0 to the algebraic Riccati equation
X
∞
˜
A+
˜
A
T
X
∞
− X
∞
(B
2
B
T
2
− γ
−2
B
1
B
T
1
)X
∞
+
˜
C
T
˜
C = 0, (5.60)
such that
˜
A −(B
2
B
T
2
− γ
−2
B
1
B
T
1
)X
∞
is asymptotically stable.
2. There is a solution Y
∞
≥ 0 to the algebraic Riccati equation
ˆ
AY
∞
+Y
∞
ˆ
A
T
−Y
∞
(C
T
2
C
2
− γ
−2
C
T
1
C
1
)Y
∞
+
ˆ
B
ˆ
B
T
= 0, (5.61)
such that
ˆ
A −Y
∞
(C
T
2
C
2
−γ
−2
C
T
1
C
1
) is asymptotically stable.
3. ρ(X
∞
Y
∞
) < γ
2
.
In the case when these conditions hold, C is an admissible controller satis
fying (5.54) if and only if C is given by the LFT
C = LFT(C
a
, U
y
), U
y
∞
< γ, (5.62)
where U
y
is a stable transfer function. The generator C
a
is given by
C
a
=
_
_
A
k
B
k1
B
k2
C
k1
0 I
C
k2
I 0
_
_
, (5.63)
98 Modeling and Control of Vibration in Mechanical Systems
where
A
k
= A +γ
−2
B
1
B
T
1
X
∞
− B
2
F
∞
− B
k1
C
2z
, (5.64)
_
B
k1
B
k2
¸
=
_
B
1
D
T
21
+ Z
∞
C
T
2z
B
2
+ γ
−2
Z
∞
F
T
∞
¸
, (5.65)
_
C
k1
C
k2
_
=
_
−F
∞
−C
2z
_
, (5.66)
C
2z
= C
2
+γ
−2
D
21
B
T
1
X
∞
, F
∞
= D
T
12
C
1
+B
T
2
X
∞
, (5.67)
Z
∞
= Y
∞
(I −γ
−2
X
∞
Y
∞
)
−1
= (I − γ
−2
Y
∞
X
∞
)
−1
Y
∞
. (5.68)
A point given by Theorem 5.2 is that a solution to the H
∞
generalized regulator
problem exists if and only if there exist stabilizing, nonnegative deﬁnite solutions
X
∞
and Y
∞
to the algebraic Riccati equations associated with the full information
H
∞
control problemand the H
∞
estimation of C
1
x such that the coupling condition
ρ(X
∞
Y
∞
) < γ
2
is satisﬁed.
The optimal H
∞
control problem is to ﬁnd an internally stabilizing controller
C(s) such that T
zw
∞
of the closedloop system (5.26)−(5.27) is minimized.
However, in practice it is often not necessary to design an optimal controller, and
it is usually appropriate to obtain a controller that gives rise to an H
∞
norm of the
closedloop system less than a prescribed value. More speciﬁcally, a suboptimal H
∞
control problem is that given γ > 0, ﬁnd an admissible controller C, if there is any,
such that T
zw
∞
< γ.
The following theorem gives a design method for a suboptimal H
∞
output feed
back controller.
THEOREM 5.3
[90] Consider system (5.21)−(5.23). Given a scalar γ > 0, there exists an
output feedback controller (5.24)−(5.25) such that T
zw
2
∞
< γ if the following
LMI admits a solution (E, W, U, D
c
, X, Y ):
_
¸
¸
_
AX + XA
T
+B
2
E + (B
2
E)
T
U
T
+ A+B
2
D
c
C
2
∗ A
T
Y + Y A+ WC
2
+ (WC
2
)
T
∗ ∗
∗ ∗
,
B
1
+B
2
WD
c
D
21
(C
1
X + D
12
E)
T
Y B
1
+WD
21
(C
1
+ D
12
D
c
C
2
)
T
−γI (D
11
+ D
12
D
c
D
21
)
T
∗ −γI
_
¸
¸
_
< 0, (5.69)
_
X I
I Y
_
> 0. (5.70)
In this case, a feasible H
∞
controller is obtained from (5.51)−(5.53), where
N
1
M
1
= I − Y X.
Introduction to Optimal and Robust Control 99
5.4.2 Discretetime case
Assume that the timeinvariant discretetime system (5.35)−(5.37) satisﬁes:
Assumption 5.3
(1). (A, B
2
, C
2
) is stabilizable and detectable.
(2). D
T
12
D
12
> 0 and D
21
D
T
21
> 0.
(3).
rank
_
A− e
jθ
I B
2
C
1
D
12
_
= n
x
+ n
u
, for all θ ∈ (−π, π]. (5.71)
(4).
rank
_
A−e
jθ
I B
1
C
2
D
21
_
= n
x
+n
y
, for all θ ∈ (−π, π]. (5.72)
We seek a causal, linear, timeinvariant and ﬁnitedimensional controller C(z)
such that the closedloop system (5.40)−(5.41) is stable and
T
zw
∞
< γ, (5.73)
or equivalently, under zero initial conditions,
z
2
2
− γ
2
w
2
2
≤ −εw
2
2
, (5.74)
for all w ∈ ℓ
2
[0, ∞) and some ε > 0.
Let
B =
_
B
1
B
2
¸
, C
d
=
_
C
1
0
_
, (5.75)
D
d
=
_
D
11
D
12
I
n
w
0
_
, (5.76)
J
s
=
_
I
n
z
0
0 −γ
2
I
n
w
_
, (5.77)
J
t
=
_
I
n
w
0
0 −γ
2
I
n
u
_
. (5.78)
With the assumptions (1)−(4), a causal, linear, ﬁnitedimensional stabilizing con
troller that leads to T
zw
∞
< γ exists if and only if the following two conditions
hold [71].
1. There exists a solution to the Riccati equation
X
∞
= C
T
d
J
s
C
d
+A
T
X
∞
A− W
T
R
−1
W, (5.79)
with
R =
_
R
1
R
T
2
R
2
R
3
_
= D
T
d
J
s
D
d
+ B
T
X
∞
B, (5.80)
W =
_
W
11
W
21
_
= D
T
d
J
s
C
d
+B
T
X
∞
A, W
21
∈ R
n
u
×n
x
, (5.81)
100 Modeling and Control of Vibration in Mechanical Systems
such that A −BR
−1
W is asymptotically stable and
X
∞
≥ 0, (5.82)
R
1
−R
T
2
R
−1
3
R
2
< 0, R
1
∈ R
n
w
×n
w
, R
3
∈ R
n
u
×n
u
. (5.83)
Denote
∇ = R
1
− R
T
2
R
−1
3
R
2
, W
∇
= W
11
− R
T
2
R
−1
3
W
21
, (5.84)
and let E
1
be an n
u
×n
u
matrix such that
E
T
1
E
1
= R
3
, (5.85)
and E
2
be an n
w
× n
w
matrix such that
E
T
2
E
2
= −γ
−2
(R
1
−R
T
2
R
−1
3
R
2
) = −γ
−2
∇. (5.86)
Deﬁne the system
_
A
t
B
t
C
t
D
t
_
=
_
_
A −B
1
∇
−1
W
∇
B
1
E
−1
2
0
E
1
R
−1
3
(W − R
2
∇
−1
W
∇
) E
1
R
−1
3
R
2
E
−1
2
I
C
2
− D
21
∇
−1
W
∇
D
21
E
−1
2
0
_
_
. (5.87)
2. There exists a solution to the Riccati equation
Y
∞
= B
t
J
t
B
T
t
+A
t
Y
∞
A
T
t
−M
t
S
−1
t
M
T
t
, (5.88)
where
S
t
= D
t
J
t
D
T
t
+C
t
Y
∞
C
T
t
=
_
S
1
S
2
S
T
2
S
3
_
, (5.89)
M
t
= B
t
J
t
D
T
t
+A
t
Y
∞
C
T
t
=
_
M
t1
M
t2
¸
, (5.90)
such that A
t
− M
t
S
−1
t
C
t
is asymptotically stable and
Y
∞
≥ 0, (5.91)
S
1
−S
2
S
−1
3
S
T
2
< 0. (5.92)
Note that C
t
in (5.87) is partitioned as
C
t
=
_
C
t1
C
t2
_
=
_
E
1
R
−1
3
(W −R
2
∇
−1
W
∇
)
C
2
− D
21
∇
−1
W
∇
_
. (5.93)
A controller that achieves the objective (5.73) is given by
x
c
(k + 1) = A
t
x
c
(k) +B
2
u(k) + M
t2
S
−1
3
(y(k) − C
t2
x
c
(k)), (5.94)
E
1
u(k) = −C
t1
x
c
(k)1 −S
2
S
−1
3
(y(k) −C
t2
x
c
(k)). (5.95)
Introduction to Optimal and Robust Control 101
All controllers that achieve the objective (5.73) are generated by the LFT C =
LFT(C
a
, U
c
), where U
c
is a linear causal system such that U
c
∞
< γ , and the
generator C
a
is given by
_
_
I −B
2
−M
t2
(X
T
2
)
−1
0 E
1
S
2
(X
T
2
)
−1
0 0 X
2
_
_
_
_
x
c
(k + 1)
u
(
k)
η(k)
_
_
=
_
_
A
t
0 (M
t1
−M
t2
S
−1
3
S
2
)(γ
2
X
T
1
)
−1
−C
t1
0 X
1
−C
t2
I 0
_
_
_
_
x
c
(k)
y(k)
φ(k)
_
_
, (5.96)
with
X
2
X
T
2
= S
3
, (5.97)
X
1
X
T
1
= −γ
−2
(S
1
− S
2
S
−1
3
S
T
2
). (5.98)
The following theoremgives one parametrization approach of all suboptimal discrete
time H
∞
output feedback controllers.
THEOREM 5.4
[90] Consider system (5.35)−(5.37). Given a scalar γ > 0, there exists an
output feedback controller (5.38)−(5.39) such that T
zw
2
∞
< γ if the following
LMI admits a solution:
_
¸
¸
¸
¸
¸
¸
_
P
∞
J AX +B
2
E A +B
2
D
c
C
2
B
1
+ B
2
D
c
D
21
0
∗ H U Y A+ V C
2
Y B
1
+V D
21
0
∗ ∗ X +X
′
−P
∞
I + Z
′
− J 0 X
T
C
T
1
+E
T
D
T
12
∗ ∗ ∗ Y +Y
T
−H 0 C
T
1
+C
T
2
D
T
c
D
T
12
∗ ∗ ∗ ∗ I D
T
11
+D
T
21
D
T
c
D
T
12
∗ ∗ ∗ ∗ ∗ γI
_
¸
¸
¸
¸
¸
¸
_
> 0, (5.99)
where the matrices X, E, Y , V , U, D
c
, Z, J, and the symmetric matrices P
∞
and H are the variables. A feasible H
∞
controller is obtained from (5.51)
(5.53).
5.5 Robust control
The H
∞
norm is used to test robust stability of a nominally stable system under
unstructured perturbations. The following socalled small gain theorem is the basis
for robust stability analysis.
102 Modeling and Control of Vibration in Mechanical Systems
THEOREM 5.5
Small Gain Theorem Consider a proper and stable transfer function matrix
T(s). Suppose that a stable ∆(s) is connected from the output of T(s) to the
input of T(s) as shown in Figure 5.2. Then the closedloop system given in
Figure 5.2 is internally stable if
¯ σ[∆(jω)]¯ σ[T(jω)] < 1, ∀ω ∈ R
_
∞. (5.100)
FIGURE 5.2
A closedloop system with uncertainty.
The small gain condition is sufﬁcient to guarantee internal stability of the closed
loop system even if ∆ is nonlinear and timevarying. The small gain theorem tells us
that an H
∞
norm bound on T implies closedloop stability in the presence of certain
H
∞
norm bounded system uncertainties. The H
∞
norm bound implies a certain
stability robustness. Note from (5.100) that the size of tolerable uncertainty varies
inversely proportional to the H
∞
norm bound of T, which means that the robustness
increases as the H
∞
norm bound decreases.
In what follows, the small gain theorem will be used to test robust stability under
model uncertainties. The modeling error ∆ is assumed to be stable and suitably
scaled with weighting functions W
1
and W
2
, i.e., the uncertainty can be represented
as W
1
∆W
2
.
Additive uncertainty
We assume that the model uncertainty can be represented by an additive perturba
tion:
Π = P + W
1
∆W
2
, (5.101)
Introduction to Optimal and Robust Control 103
which is shown in Figure 5.3.
FIGURE 5.3
A closedloop system with additive uncertainty for robust stability analysis.
THEOREM 5.6
[70] Let Π = P + W
1
∆W
2
, and C be a stabilizing controller for the nominal
plant P. Then the closedloop system is wellposed and internally stable for
all ∆
∞
< 1 if and only if W
2
CSW
1
∞
≤ 1, where
S =
1
1 +PC
. (5.102)
Similarly, the closedloop system is stable for all stable ∆ with ∆
∞
≤ 1 if and
only if W
2
CSW
1
∞
< 1.
Multiplicative uncertainty
The system model is described by the following multiplicative perturbation:
Π = (I +W
1
∆W
2
)P (5.103)
where W
1
, W
2
and ∆ are stable. Consider the feedback system shown in Figure 5.4.
THEOREM 5.7
[70] Let Π = (I + W
1
∆W
2
)P, C be a stabilizing controller for the nominal
plant P, and T = 1 − S. Then
(i) the closedloop system is wellposed and internally stable for all stable
∆ with ∆
∞
< 1 if and only if W
2
TW
1
∞
≤ 1.
(ii) the closedloop system is wellposed and internally stable for all stable
∆ with ∆
∞
≤ 1 if W
2
TW
1
∞
< 1.
(iii) the robust stability of the closedloop system for all stable ∆ with
∆
∞
≤ 1 does not necessarily imply W
2
TW
1
∞
< 1.
104 Modeling and Control of Vibration in Mechanical Systems
FIGURE 5.4
A closedloop system with multiplicative uncertainty for robust stability analysis.
(iv) the closedloop system is wellposed and internally stable for all stable
∆ with ∆
∞
≤ 1 only if W
2
TW
1
∞
≤ 1.
(v) In addition, assume that neither P nor C has poles on the imaginary
axis. Then the closedloop system is wellposed and internally stable for all
stable ∆ with ∆
∞
≤ 1 if and only if W
2
TW
1
∞
< 1.
5.6 Controller parametrization
Consider the standard system block diagram in Figure 5.1 with
P(s) =
_
_
A
p
B
1
B
2
C
1
D
11
D
12
C
2
D
21
D
22
_
_
. (5.104)
Suppose (A
p
, B
2
) is stabilizable, (C
2
, A
p
) is detectable, and D
22
= 0. The problem
discussed here is that given a plant P, parameterize all controllers C that internally
stabilize P.
The parametrization for all stabilizing controllers is known as Youla parametriza
tion [68], as shown in Figure 5.5. The Youla parametrization starts with a nominal
controller that is an estimatedstate feedback. The estimated state feedback controller
is given by
u = −Kˆ x, (5.105)
where the state feedback gain K is some appropriate matrix and ˆ x is an estimate of
the component of x, governed by the observer equation
˙
ˆ x = A
p
ˆ x +B
2
u + L(y − C
2
ˆ x), (5.106)
Introduction to Optimal and Robust Control 105
where L is the estimator gain. The transfer function of the estimatedstate feedback
controller is thus
K
n
(s) = −K(sI −A
p
+B
2
K +LC
2
)
−1
L. (5.107)
The nominal controller K
n
(s) will stabilize P provided K and L are chosen such
that A
p
− B
2
K and A
p
− LC
2
are stable.
To augment the estimated state feedback controller, we inject v into u as shown in
Figure 5.5, meaning that (5.105) is replaced by
u = −Kˆ x + v, (5.108)
and therefore the signal v does not induce any observer error. For the signal e we
take the output prediction error:
e = y − C
2
ˆ x. (5.109)
In Figure 5.5, the observer based controller applies output prediction error processed
through a stable transfer function Q and added to the output of K
n
. This augmen
tation is able to yield every controller that stabilizes the plant, which means every
stabilizing controller can be realized as an observer based controller with some sta
ble transfer function Q. Thus in the sequel we can form simple statespace equations
for the parametrization of all controllers that stabilize the plant, and all closedloop
transfer matrices achieved by controllers that stabilize the plant.
The statespace equations of the augmented controller are given by
˙
ˆ x = (A
p
−B
2
K − LC
2
)ˆ x + Ly + B
2
v, (5.110)
u = −Kˆ x +v, (5.111)
e = y −C
2
ˆ x. (5.112)
If Q has a statespace realization
˙ x
q
= A
q
x
q
+ B
q
e, (5.113)
v = C
q
x
q
+D
q
e, (5.114)
then a statespace realization of the observer based controller by eliminating e and
v from the augmented controller equations (5.110)−(5.112) and the Q realization
(5.113)−(5.114) is obtained as:
˙
ˆ x = (A
p
−B
2
K −LC
2
− B
2
D
q
C
2
)ˆ x +B
2
C
q
x
q
+ (L +B
2
D
q
)y, (5.115)
˙ x
q
= −B
q
C
2
ˆ x +A
q
x
q
+ B
q
y, (5.116)
u = −(K + D
q
C
2
)ˆ x +C
q
x
q
+ D
q
y, (5.117)
or equivalently,
C(s) = C
c
(sI − A
c
)
−1
B
c
+ D
c
, (5.118)
106 Modeling and Control of Vibration in Mechanical Systems
where
A
c
=
_
A
p
− B
2
K − LC
2
−B
2
D
q
C
2
B
2
C
q
−B
q
C
2
A
q
_
, (5.119)
B
c
=
_
L +B
2
D
q
B
q
_
, (5.120)
C
c
=
_
−K − D
q
C
2
C
q
¸
, (5.121)
D
c
= D
q
. (5.122)
On the other hand, the statespace equations for the closedloop system with only
the augmented controller (5.110)−(5.112) are found as follows by eliminating u and
y from (5.110)−(5.112) and the plant equations in (5.104).
˙ x = A
p
x −B
2
Kˆ x + B
1
w +B
2
v, (5.123)
˙
ˆ x = LC
2
x + (A
p
− B
2
K − LC
2
)ˆ x + LD
1
w +B
2
v, (5.124)
z = C
1
x − D
12
Kˆ x + D
11
w + D
12
v, (5.125)
e = C
2
x − C
2
ˆ x +D
21
w, (5.126)
which are equivalently written as
_
T
11
(s) T
12
(s)
T
21
(s) 0
_
= C
T
(sI − A
T
)
−1
B
T
+D
T
, (5.127)
where
A
T
=
_
A
p
−B
2
L
LC
2
A
p
−B
2
K −LC
2
_
, (5.128)
B
T
=
_
B
1
B
2
LD
1
B
2
_
, (5.129)
C
T
=
_
C
1
−D
12
K
C
2
− C
2
_
, (5.130)
D
T
=
_
D
11
D
12
D
21
0
_
. (5.131)
It has been veriﬁed that by augmenting the stable transfer function Q, the closed
loop transfer function from w to z is simply an afﬁne function of Q and equals
T
11
+T
12
QT
21
.
Introduction to Optimal and Robust Control 107
FIGURE 5.5
Control system structure for Youla parametrization.
108 Modeling and Control of Vibration in Mechanical Systems
5.7 Performance limitation
5.7.1 Bode integral constraint
The block diagram in Figure 2.20 shows a typical closedloop control system. The
closedloop transfer function from r to y is given by
T =
PC
1 +PC
. (5.132)
The sensitivityfunction is also known as the disturbance rejection function or error
rejection function, and given by
S =
1
1 +PC
. (5.133)
Note that
S + T ≡ 1, (5.134)
hence T is also commonly called the complementary sensitivity function.
Denote T
yd
2
the transfer function from d
2
to y. In Figure 2.20 , note that
S = T
yd
2
= −T
ed
2
= T
en
. (5.135)
The sensitivity function is thus important, because it explains how disturbance d
2
goes through the closedloop system and shows up at the output y, or the error signal
e. It is also important to understand how noise n will be ﬁltered through the closed
loop system.
The Bode plot of sensitivity function in continuoustime domain is shown in Fig
ure 5.6. It can be explained by the following Bode integral theorem.
THEOREM 5.8
[67] Bode’s Integral Theorem for Continuoustime Systems For a stable, ra
tional P and C with P(s)C(s) having at least 2pole roll oﬀ,
_
∞
0
logSdω = 0. (5.136)
An implication of Theorem 5.8 is that if the system is made less sensitive to dis
turbance at some frequencies, it will be more sensitive at other frequencies. If the
plant P or compensator is not stable, i.e., if P(s) and /or C(s) have a ﬁnite number
of unstable poles p
k
, then (5.136) becomes
_
∞
0
logSdω = 2π
K
k=1
Re(p
k
) > 0, (5.137)
Introduction to Optimal and Robust Control 109
where K is the number of unstable poles. Equation (5.137) implies that any unstable
poles in the system only make it worse, in that more of the disturbance would have
to be ampliﬁed.
Figure 5.7 means that for a discretetime system, the main difference is the Nyquist
frequency limits the frequency range we have to work with. In both cases, if we
want to attenuate disturbances at one frequency, we must amplify some disturbance
at another frequency.
THEOREM 5.9
[108] Bode’s Integral Theorem for Discretetime Systems Given a stable closed
loop discretetime feedback system, its sensitivity function has to satisfy the
following integral constraint:
1
π
_
π
0
logS(e
jφ
)dφ =
K
k=1
lnβ
k
, (5.138)
where β
k
are the openloop unstable poles of the system, K is the total number
of unstable poles, and φ = T
s
ω with the sampling time T
s
and the frequency
ω in radians/sec.
Note that T
s
is the sampling period, and the upper limit of the frequency spectrum
is π/T
s
, the Nyquist frequency. Typical digital control systems assume PC is small
and S ≈ 1 at or above the Nyquist frequency, which is in general not practical for
a physical system. Further research beyond Nyquist frequency is needed to address
vibrations at frequencies above the Nyquist frequency, which strives to overcome the
limitation due to the integral theorem.
Theorem 5.9 implies that if for some frequency S < 1, then at some other fre
quency S > 1. Unlike the continuous time result, there is no inﬁnite bandwidth to
spread this over. Thus all S > 1 happen below the Nyquist frequency, and there
fore in a ﬁnite frequency range. Since the theorem is limited by frequencies up to the
Nyquist frequency, if the closedloop bandwidth is pushed up, better performance at
low frequency may result in worse performance at high frequency. In a word, with
the linear feedback control whenever we improve the disturbance rejection at one
frequency we pay for it at another. Nevertheless, if we have sufﬁcient knowledge
of system disturbance and place the disturbance ampliﬁcation at places where the
disturbance is negligible, then we succeed. Otherwise, most of the disturbances may
be ampliﬁed.
110 Modeling and Control of Vibration in Mechanical Systems
Frequency(Hz)
0
M
a
g
n
i
t
u
d
e
i
n
d
B
(
2
0
l
o
g
1
0

S

)
Area of vibration
rejection
Area of vibration
amplification
FIGURE 5.6
Sensitivity function for continuoustime system.
Frequency(Hz)
Nyquist frequency
Area of vibration
rejection
Area of vibration
amplification
0
M
a
g
n
i
t
u
d
e
i
n
d
B
(
2
0
l
o
g
1
0

S

)
FIGURE 5.7
Sensitivity function for discretetime system.
Introduction to Optimal and Robust Control 111
5.7.2 Relationship between system gain and phase
In the classical feedback theory, the Bode’s gainphase integral relation has been
used as an important tool to express design constraints in feedback systems. Let
L = PC denote the openloop system. It is noted that ∠L(jω
0
) will be large if
the gain L attenuates slowly near ω
0
and small if it attenuates rapidly near ω
0
. The
behavior of ∠L(jω) is particularly important near the crossover frequency ω
c
, where
L(jω
c
) = 1, and π +∠L(jω
c
) is the phase margin of the feedback system. Further
1 + L(jω
c
) = 1 +L
−1
(jω
c
) = 2sin
π +∠L(jω
c
)
2
 (5.139)
must not be too small for good stability robustness. If π + ∠L(jω
c
) is forced to be
very small by rapid gain attenuation, the feedback system will amplify disturbances
and exhibit little uncertainty tolerance at and near ω
c
. A nonminimum phase zero
contributes an additional phase lag and imposes limitations upon the roll off rate of
the openloop gain. Thus the conﬂict between attenuation rate and loop quality near
crossover is clearly evident.
In the classical feedback control theory, it has been common to express design
goals in terms of the shape of the openloop transfer function. A typical design
requires that the openloop transfer function has a high gain at low frequencies and
a low gain at high frequencies while the transition should be well behaved [70].
5.7.3 Sampling
Mathematical relations and operations can be handled by digital microprocessor only
when they are expressed as a ﬁnite set of numbers rather than as functions having
an inﬁnite number of possible values. Thus any measured continuous signal must
be converted to a set of pulses by sampling, which is the process used to measure a
continuoustime variable at separated instants of time. The inﬁnite set of numbers
represented by the smooth curve is replaced by a ﬁnite set of numbers. Each pulse
amplitude is then rounded off to one of a ﬁnite number of levels depending on the
characteristics of the converter. The process is called quantization. Thus a digital
device is one in which signals are quantized in both time and amplitude. In an analog
device, signals are analog; that is, they are continuous in time and are not quantized
in amplitude. The device that performs the sampling, quantization, and converting
to binary form is an analog to digital (A/D) converter.
The number of binary digits or bits generated by the device is its word length,
which is an important characteristic related to the resolution of the converter. The
resolution measures the smallest change in the input signal that will produce a change
in the output signal. An example is that if an A/D converter has a word length of 10
bits or more, an input signal can be resolved to 1 in 2
10
or 1024. If the input signal
has a range of 10 V, the resolution is 10/1024, or approximately 0.01 V. Thus in
order to produce a change in the output the input must change by at least 0.01 V.
A discretetime signal is extracted by sampling from a continuoustime signal. If
the sampling frequency is not selected properly, the resulting sampled sequence will
112 Modeling and Control of Vibration in Mechanical Systems
not accurately represent the original continuous signal. Aproper sampling frequency
is readily determined in many cases by means of the following sampling theorem.
THEOREM 5.10
Sampling Theorem A continuoustime signal y(t) can be reconstructed from
its uniformly sampled values y(KT
s
) if the sampling period T
s
satisﬁes
T
s
≤
π
¯ ω
(5.140)
where ¯ ω is the highest frequency contained in the signal, that is, Y (ω) = 0
for ω > ¯ ω.
If a system involves a sampling operation of continuoustime signals to gener
ate discretetime signals, a time delay may be induced. Any time delay added into
a closedloop control system will decrease the stability of the system and in some
cases may even cause system instability. The Nyquist rate shown in Figure 5.7 signi
ﬁes that the freedom to spread the ampliﬁcation area around is limited by the Nyquist
frequency, which is half of the sampling rate. Figure 5.7 implies that a certain rejec
tion amount S < 1 must be accompanied by a certain area of S > 1, which has
to occur before the Nyquist frequency. With a higher sampling rate and closedloop
system bandwidth being kept constant, the ampliﬁcation S > 1 will essentially
spread over a broader frequency band and the height of ampliﬁcation hump shrinks,
as shown in Figure 5.8. If we push the closedloop system bandwidth from f
1
to f
2
,
as seen in Figure 5.8, better performance in low frequency range may result in worse
performance in high frequency range.
5.8 Conclusion
Before presentating a series of advanced vibration control methodologies, this chap
ter has been used to recall some standard advanced control techniques, which helps
understand problems to be addressed in the later chapters and possible solutions. It
has reviewed H
2
and H
∞
performances, H
2
and H
∞
controls, robust control, con
troller parametrization, as well as performance limitation of linear feedback control
systems.
Introduction to Optimal and Robust Control 113
Frequency(Hz)
f
N1
0
M
a
g
n
i
t
u
d
e
i
n
d
B
(
2
0
l
o
g
1
0

S

)
2f
N1
f
2
f
1
FIGURE 5.8
Sensitivity function in discretetime domain.
6
Mixed H
2
/H
∞
Control Design for Vibration
Rejection
6.1 Introduction
The mixed H
2
/H
∞
control problemis concerned with the design of a controller that
minimizes the H
2
norm of a certain closedloop transfer function while satisfying a
H
∞
normconstraint on the same or another closedloop transfer function. One of the
important applications of this problemis to address the optimal nominal performance
subject to a robust stability constraint.
In this chapter, we employ two methods: the one described in Chapter 5 and an
improved method. The two methods are applied to the control design in disk drives.
In order to meet the robustness requirement against unmodeled high frequency dy
namics of the VCM actuator in disk drives, an H
∞
constraint is to be satisﬁed when
minimizing the H
2
performance of the nominal system. Hence, a mixed H
2
/H
∞
control can be formulated for disk drive control. Both the simulation and experiment
results demonstrate that the mixed H
2
/H
∞
control design gives a signiﬁcant perfor
mance improvement of TMR over the conventional method of PID control combined
with notch ﬁlters.
6.2 Mixed H
2
/H
∞
control problem
This section aims to derive an improved design method for the mixed H
2
/H
∞
control
which will give rise to an equal or better performance than the method in Chapter 5.
We consider the statespace representation for linear timeinvariant systems:
x(k + 1) = Ax(k) +B
1
w(k) +B
2
u(k), (6.1)
y(k) = C
2
x(k) + D
21
w(k), (6.2)
z
1
(k) = C
z1
x(k) + D
z11
w(k) + D
z12
u(k), (6.3)
z
2
(k) = C
z2
x(k) + D
z21
w(k) + D
z22
u(k), (6.4)
where x(k) ∈ R
n
x
is the state, y(k) ∈ R
n
y
is the measurement output, z
i
(k) ∈
115
116 Modeling and Control of Vibration in Mechanical Systems
R
n
z
(i = 1, 2) are the controlled outputs, w(k) ∈ R
n
w
is the disturbance input,
u(k) ∈ R
n
u
is the control input, and A, B
1
, B
2
, C
2
, D
21
, C
z1
, C
z2
, D
z11
, D
z12
,
D
z21
, D
z22
are of appropriate dimensions.
Let a controller of the same dimension as that of the system (6.1)−(6.4) be of the
form:
x
c
(k + 1) = A
c
x
c
(k) +B
c
y(k), (6.5)
u(k) = C
c
x
c
(k) + D
c
y(k), (6.6)
where the matrices (A
c
, B
c
, C
c
, D
c
) are to be determined.
First, denote ξ = [x
T
x
T
c
]
T
. It follows from (6.1)−(6.4) and (6.5)−(6.6) that
ξ(k + 1) =
¯
Aξ(k) +
¯
Bw(k), (6.7)
z
1
(k) =
¯
C
1
ξ(k) +
¯
D
1
w(k), (6.8)
z
2
(k) =
¯
C
2
ξ(k) +
¯
D
2
w(k), (6.9)
where
¯
A =
_
A+ B
2
D
c
C
2
B
2
C
c
B
c
C
2
A
c
_
,
¯
B =
_
B
2
D
c
D
21
+ B
1
B
c
D
21
_
, (6.10)
¯
C
1
= [C
z1
+ D
z12
D
c
C
2
D
z12
C
c
] ,
¯
D
1
= D
z12
D
c
D
21
+D
z11
, (6.11)
¯
C
2
= [C
z2
+ D
z22
D
c
C
2
H
22
C
c
] ,
¯
D
2
= D
z22
D
c
D
21
+ D
z21
. (6.12)
Denote T
z
1
w
the transfer function matrix from w to z
1
, and T
z
2
w
the transfer
function matrix from w to z
2
. The mixed H
2
/H
∞
control problem is then stated as:
Given a positive scalar γ, design a controller of the form (6.5)−(6.6) such that the
H
2
norm, T
z
1
w
2
is minimized subject to the constraint T
z
2
w
∞
< γ.
6.3 Method 1: slack variable approach
Given the solutions of H
2
control and H
∞
control in Chapter 5, the following solu
tion to the mixed H
2
/H
∞
control follows directly.
THEOREM 6.1
Consider the system (6.1)−(6.4). Given scalars µ > 0 and γ > 0, a controller
of the form (6.5)−(6.6) that solves the mixed H
2
/H
∞
control problem exists
if the following conditions are satisﬁed.
Trace(Π) < µ, (6.13)
_
_
Π C
z1
X + D
z12
E C
z1
+ D
z12
D
c
C
2
∗ X +X
T
− P
2
I +Z
T
− J
∗ ∗ Y + Y
T
− H
_
_
> 0, (6.14)
Mixed H
2
/H
∞
Control Design for Vibration Rejection 117
_
¸
¸
¸
¸
_
P
2
J AX +B
2
E A +B
2
D
c
C
2
B
1
+ B
2
D
c
D
21
∗ H U Y A +WC
2
Y B
1
+ WD
21
∗ ∗ X +X
T
− P
2
I +Z
T
− J 0
∗ ∗ ∗ Y +Y
T
−H 0
∗ ∗ ∗ ∗ I
_
¸
¸
¸
¸
_
> 0, (6.15)
D
z11
+D
z12
D
c
D
21
= 0, (6.16)
and
_
¸
¸
¸
¸
¸
¸
_
P
∞
˜
J AX +B
2
E A +B
2
D
c
C
2
B
1
+ B
2
D
c
D
21
0
∗
˜
H U Y A +WC
2
Y B
1
+ WD
21
0
∗ ∗ X + X
T
−P
∞
I + Z
T
−
˜
J 0 X
T
C
T
z2
+ E
T
D
T
z22
∗ ∗ ∗ Y +Y
T
−
˜
H 0 C
T
z2
+C
T
2
D
T
c
D
T
z22
∗ ∗ ∗ ∗ I D
T
z21
+ D
T
21
D
T
c
D
T
z22
∗ ∗ ∗ ∗ ∗ γI
_
¸
¸
¸
¸
¸
¸
_
> 0, (6.17)
where the matrices X, E, Y , W, U, D
c
, Z, J,
˜
J and the symmetric matrices
P
2
, P
∞
, H,
˜
H and Π are the variables. A feasible mixed H
2
/H
∞
controller
is obtained by choosing N
1
and M
1
nonsingular such that N
1
M
1
= Z − Y X
and calculating
C
c
= (E − D
c
C
2
X)M
−1
1
, D
c
= D
c
, (6.18)
B
c
= N
−1
1
(W − Y B
2
D
c
), (6.19)
A
c
= N
−1
1
(U − Y (A+ B
2
D
c
C
2
)X −N
1
B
c
C
2
X −Y B
2
C
c
M
1
]M
−1
1
. (6.20)
In the early development of the mixed H
2
/H
∞
control, to solve the problem in
terms of LMIs, a single Lyapunov matrix is adopted for both the H
2
and H
∞
perfor
mances, which is very conservative in general. A signiﬁcant improvement was made
in [90] where a slack variable technique is introduced which separates the Lyapunov
matrices from the controller parameters and hence allows them to be different for
the H
2
and H
∞
performances. This is observed in the above theorem where the
Lyapunov matrices P
2
, J and H are used for H
2
performance and different matrices
P
∞
,
˜
J and
˜
H are used for H
∞
performance.
6.4 Method 2: an improved slack variable approach
Recall fromChapter 5 that for a given controller that stabilizes the system(6.7)−(6.9),
the H
2
norm square, T
z
1
w
2
2
, can be computed by the following minimization [95]:
min
(Q=Q
T
,Π=Π
T
)
Trace(Π), (6.21)
118 Modeling and Control of Vibration in Mechanical Systems
subject to
¯
A
T
Q
¯
A−Q+
¯
C
T
1
¯
C
1
< 0, (6.22)
¯
B
T
Q
¯
B +
¯
D
T
1
¯
D
1
< Π. (6.23)
The following lemma leads to an alternative approach for computing the H
2
norm
of the system (6.7)−(6.8).
LEMMA 6.1
The minimization of the H
2
norm square in (6.21) is equivalent to the fol
lowing minimization:
min
(Q=Q
T
,Π=Π
T
,Σ)
Trace(Π), (6.24)
subject to
˜
A
T
diag{Q, I}
˜
A−
_
Q Σ
Σ
T
Π
_
< 0, (6.25)
where
˜
A =
_
¯
A
¯
B
¯
C
1
¯
D
1
_
.
Proof First of all, (6.25) can be rewritten as
_
¯
A
T
Q
¯
A −Q+
¯
C
T
1
¯
C
1
¯
A
T
Q
¯
B +
¯
C
T
1
¯
D
1
−Σ
¯
B
T
Q
¯
A+
¯
D
T
1
¯
C
1
− Σ
T
¯
B
T
Q
¯
B +
¯
D
T
1
¯
D
1
−Π
_
< 0. (6.26)
It is then clear that if there exists a solution (Q, Π, Σ) to (6.25), the same Q and Π
also satisfy (6.22) and (6.23), respectively. On the other hand, if there exist Q and
Π satisfying (6.22) and (6.23), then (6.25) is also met by the same Q and Π and
Σ =
¯
A
T
Q
¯
B +
¯
C
T
1
¯
D
1
.
REMARK 6.1 The characterization of the H
2
norm in the above lemma
has the advantage that it provides a uniﬁed treatment of H
2
and H
∞
designs
via an LMI approach, as seen later. Furthermore, the additional parameter
Σ in (6.25) will oﬀer an additional freedom in optimization of performance
when a mixed H
2
and H
∞
control design is concerned.
Without loss of generality, we shall assume n
w
= n
z
, i.e. the disturbance input
and the signal to be controlled have the same dimension. Note that, if this is not the
case, some simple modiﬁcation can render the requirement satisﬁed. For example,
if n
w
< n
z
, the matrices B
1
, D
21
, D
z11
and D
z21
can be augmented as
ˆ
B
1
=
[B
1
0
n
x
×(n
z
−n
w
)
],
ˆ
D
21
= [D
21
0
n
x
×(n
z
−n
w
)
],
ˆ
D
z11
= [D
z11
0
n
z
×(n
z
−n
w
)
] and
ˆ
D
z21
= [D
z21
0
n
z
×(n
z
−n
w
)
].
Mixed H
2
/H
∞
Control Design for Vibration Rejection 119
LEMMA 6.2
There exists a solution (Σ, Q, Π) with Q = Q
T
to (6.25) if and only if there
exist matrices (Σ, Π, Q, F, G) with Q = Q
T
and Π = Π
T
such that
_
_
−
_
Q Σ
Σ
T
Π
_
+
˜
A
T
F + F
T
˜
A −F
T
+
˜
A
T
G
−F +G
T
˜
A diag{Q, I} −(G+ G
T
)
_
_
< 0.
(6.27)
Proof First, if (6.25) holds for some Q > 0, by applying the Schur complement,
it is easy to know that (6.27) is satisﬁed with F = 0 and G
T
= G = diag{Q, I}.
On the other hand, if (6.27) holds for some (Σ, Q, F, G), multiplying (6.27) from the
left and from the right by Γ
T
and Γ, respectively, where
Γ =
_
I
˜
A
_
,
(6.25) follows.
REMARK 6.2 It should be pointed out that (6.21)−(6.23) and Lemma
6.2 give equivalent computations of the H
2
norm of the system. For systems
without uncertainty, it is well known that (6.21)−(6.23) can be applied to
derive the optimal H
2
controller [95]. Hence, there is no advantage of using
Lemma 6.2. However, as will be seen later, when additional performances such
as the H
∞
performance are to be met in addition to the H
2
performance, the
latter will result in a less or equally conservative design due to the additional
variables F and G. We observe that when F = 0 and Σ = 0, Lemma 6.2
reduces to the result in Theorem 6.1.
While Lemma 6.2 can be applied to compute the H
2
normof the system(6.7)−(6.8)
when a controller (6.5)−(6.6) is given, it may not be directly applicable to the H
2
control design problem due to the presence of the products of F with
˜
A and G with
˜
A. To overcome this difﬁculty, we specialize the matrices F and G as follows.
F =
_
λ
1
Φ 0
0 λ
2
I
_
, G =
_
Φ 0
0 λ
3
I
_
, (6.28)
where Φ ∈ R
2n×2n
and λ
i
, i = 1, 2, 3 are real scaling parameters. While this
specialization of F and G generally introduces some conservatism, it contains three
additional variables λ
i
, i = 1, 2, 3 as compared to the result of Theorem 6.1 which
help reduce the design conservatism in Theorem 6.1.
Substituting (6.28) into (6.27) leads to
_
¸
¸
_
−Q+ λ
1
¯
A
T
Φ + λ
1
Φ
T
¯
A λ
2
¯
C
1
T
+λ
1
Φ
T
¯
B − Σ
λ
2
¯
C
1
+ λ
1
¯
B
T
Φ − Σ
T
− Π + λ
2
_
¯
D
T
1
+
¯
D
1
_
−λ
1
Φ + Φ
T
¯
A Φ
T
¯
B
λ
3
¯
C
1
−λ
2
I +λ
3
¯
D
1
120 Modeling and Control of Vibration in Mechanical Systems
−λ
1
Φ
T
+
¯
A
T
Φ λ
3
¯
C
T
1
¯
B
T
Φ − λ
2
I + λ
3
¯
D
T
1
Q−(Φ + Φ
T
) 0
0 (1 −2λ
3
)I
_
¸
¸
_
< 0. (6.29)
A similar characterization for the H
∞
performance can be derived. Indeed, recall
that when the system (6.7) and (6.9) is known, it is stable with its H
∞
norm less than
γ if and only if there exists a matrix P = P
T
> 0 such that
˜
A
T
diag{P
∞
, I}
˜
A −diag{P
∞
, γ
2
I} < 0, (6.30)
where
˜
A =
_
¯
A
¯
B
¯
C
2
¯
D
2
_
.
Observe that (6.30) turns out to be a special case of (6.25) with Π = γ
2
I and Σ = 0.
Thus, following a similar procedure for deriving (6.29), it can be shown that the
system (6.7) and (6.9) has an H
∞
performance γ if and only if for some real scalars
ε
i
, i = 1, 2, 3, there exist matrices P
∞
= P
T
∞
> 0 and Φ such that
_
¸
¸
_
−P
∞
+ ε
1
(
¯
A
T
Φ + Φ
T
¯
A) ε
2
¯
C
2
T
+ε
1
Φ
T
¯
B
ε
2
¯
C
2
+ ε
1
¯
B
T
Φ −γ
2
I + ε
2
_
¯
D
T
2
+
¯
D
2
_
−ε
1
Φ + Φ
T
¯
A Φ
T
¯
B
ε
3
¯
C
2
−ε
2
I +ε
3
¯
D
2
−ε
1
Φ
T
+
¯
A
T
Φ ε
3
¯
C
T
2
¯
B
T
Φ −ε
2
I + ε
3
¯
D
T
2
P − (Φ + Φ
T
) 0
0 (1 −2ε
3
)I
_
¸
¸
_
< 0. (6.31)
REMARK 6.3 When setting ε
1
= ε
2
= 0 (i.e. setting F = 0) and ε
3
= 1
and by some rowcolumn exchanges, the above inequality reduces to
_
¸
¸
_
P
∞
− Φ − Φ
T
Φ
T
¯
A Φ
T
¯
B 0
¯
A
T
Φ −P
∞
0
¯
C
T
2
¯
B
T
Φ 0 −γ
2
I
¯
D
T
2
0
¯
C
2
¯
D
2
−I
_
¸
¸
_
< 0
which is the result in Chapter 5. Therefore, ε
i
, i = 1, 2, 3 are additional
variables which can be exploited to alleviate the conservatism in the mixed
H
2
/H
∞
design of Theorem 6.1.
Denote
Ω (Π, X, Y, E, U, W, Z, D
c
,
¯
Q
11
=
¯
Q
T
11
,
¯
Q
12
,
¯
Q
22
=
¯
Q
T
22
,
¯
Σ
1
,
¯
Σ
2
, λ
i
, i = 1, 2, 3) =
Mixed H
2
/H
∞
Control Design for Vibration Rejection 121
_
¸
¸
¸
¸
¸
¸
_
−
¯
Q
11
+λ
1
(AY +Y
T
A
T
+ E
T
B
T
2
+B
2
E)
−
¯
Q
T
12
+λ
1
(A
T
+C
T
2
D
c
T
B
T
2
) + λ
1
U
T
λ
2
(C
z1
Y +D
z21
E) +λ
1
(B
T
1
+D
T
21
D
c
T
B
T
2
) −
¯
Σ
T
1
−λ
1
Y
T
+AY + B
2
E
−λ
1
I + U
λ
3
(C
z1
Y + D
z21
E)
∗
−
¯
Q
22
+λ
1
(A
T
X + X
T
A) +λ
1
(C
T
2
W
T
+WC
2
)
λ
2
(C
z1
+ D
z21
D
c
C
2
) +λ
1
B
T
X + λ
1
D
T
21
W
T
−
¯
Σ
T
2
−λ
1
Z +A+ B
2
D
c
C
2
−λ
1
X + X
T
A+WC
2
λ
3
(C
z1
+D
z21
D
c
C
2
)
∗ ∗
∗ ∗
−Π +λ
2
(D
T
21
D
c
T
D
T
z21
+D
z21
D
c
D
21
+ D
T
z11
+ D
z11
) ∗
B
1
+ B
2
D
c
D
21
¯
Q
11
− (Y + Y
T
)
X
T
B
1
+ WD
21
¯
Q
T
12
− (I +Z
T
)
−λ
2
I + λ
3
(D
z11
+ D
z21
D
c
D
21
) 0
∗ ∗
∗ ∗
∗ ∗
∗ ∗
¯
Q
22
− (X +X
T
) ∗
0 (1 −2λ
3
)I
_
¸
¸
¸
¸
¸
¸
_
. (6.32)
We have the following solution to the mixed H
2
/H
∞
control.
THEOREM 6.2
Consider the system (6.1)−(6.4). A controller of the form (6.5)−(6.6) that
solves the mixed H
2
/H
∞
control problem exists if for some λ
i
, ε
i
, i = 1, 2, 3,
there exists a solution (
¯
P
11
,
¯
P
12
,
¯
P
22
,
¯
Q
11
,
¯
Q
12
,
¯
Q
22
,
¯
Σ
1
,
¯
Σ
2
, X, Y, Π, U, E, Z, W,
Ψ, D
c
) with
¯
Q = [
¯
Q
lj
] =
¯
Q
T
> 0 and
¯
P = [
¯
P
lj
] =
¯
P
T
> 0 to the following
optimization:
minTrace(Π)
subject to
Ω (Π, X, Y, E, U, W, Z, D
c
,
¯
Q
11
,
¯
Q
12
,
¯
Q
22
,
¯
Σ
1
,
¯
Σ
2
, λ
i
,
i = 1, 2, 3) < 0, (6.33)
Ω (γ
2
I, X, Y, E, U, W, Z, D
c
,
¯
P
11
,
¯
P
12
,
¯
P
22
,
¯
Σ
1
,
¯
Σ
2
, ε
i
,
i = 1, 2, 3) ¯
Σ
1
=
¯
Σ
2
=0
< 0. (6.34)
122 Modeling and Control of Vibration in Mechanical Systems
In this situation, a mixed H
2
/H
∞
controller is given by
A
c
= M
−T
1
(U −X
T
(A +B
2
D
c
C
2
)Y −X
T
B
2
C
c
N
1
−M
T
1
B
c
C
2
Y )N
−1
1
, (6.35)
B
c
= M
−T
1
(W − X
T
B
2
D
c
), (6.36)
C
c
= (E −D
c
C
2
Y )N
−1
1
, D
c
= D
c
, (6.37)
where M
1
, N
1
satisfy N
T
1
M
1
= Z −Y
T
X.
Proof First, observe from (6.29) that Φ is invertible since Φ + Φ
T
> Q > 0.
Denote
Φ =
_
X M
M
1
U
_
, Φ
−1
=
_
Y H
N
1
E
_
,
and
J =
_
Y I
n
N
1
0
_
, J
1
= diag{J, I, J, I}.
Further, denote
E = D
c
C
2
Y +C
c
N
1
, (6.38)
U = X
T
(A +B
2
D
c
C
2
)Y +
X
T
B
2
C
c
N
1
+ M
T
1
B
c
C
2
Y +M
T
1
A
c
N
1
, (6.39)
W = X
T
B
2
D
c
+M
T
1
B
c
, (6.40)
Z = Y
T
X +N
T
1
M
1
. (6.41)
Multiplying from the left and the right of (6.29) by J
T
1
and J
1
respectively and
applying (6.10), we obtain (6.33) where
¯
Q = [
¯
Q
lj
] = J
T
QJ and [
¯
Σ
T
1
¯
Σ
T
2
] = Σ
T
J.
By applying a similar procedure to (6.31), (6.34) can be obtained.
If there exists a solution for the LMIs (6.33) and (6.34), it is easy to see that
_
Y + Y
T
I +Z
I + Z
T
X + X
T
_
>
¯
Q > 0.
Multiplyingthe above fromthe left by [Y
−T
−I] and fromthe right by [Y
−T
−I]
T
,
we obtain that
(X −Y
−T
Z) + (X − Y
−T
Z)
T
> 0.
It is then clear that Z −Y
T
X is invertible. Hence, there exist invertible matrices M
1
and N
1
such that Z −Y
T
X = N
T
1
M
1
. Thus, it follows from (6.38)−(6.41) that the
controller parameters of (6.5)−(6.6) can be obtained as in (6.35)−(6.37).
REMARK 6.4 It is worth noting that when setting λ
1
= λ
2
= ε
1
=
ε
2
= 0, λ
3
= ε
3
= 1 and
¯
Σ
1
=
¯
Σ
2
= 0, Theorem 6.2 reduces to Theorem
6.1. Hence, by exploring the freedoms oﬀered by these parameters, a less or
equally conservative mixed H
2
/H
∞
design can be achieved. Certainly, the
Mixed H
2
/H
∞
Control Design for Vibration Rejection 123
controller from Theorem 6.1 is derived by a convex optimization and can be
further reﬁned by an iterative procedure. In the disk drive application to
be presented later, we demonstrate that Method 2 can give a much better
performance than Method 1 with Theorem 6.1 together with an iterative
reﬁnement in the latter design.
REMARK 6.5 Observe that for given λ
i
, ε
i
, i = 1, 2, 3, (6.33) and (6.34)
are linear in (
¯
P
11
,
¯
P
12
,
¯
P
22
,
¯
Q
11
,
¯
Q
12
,
¯
Q
22
,
¯
Σ
1
,
¯
Σ
2
, X, Y, Π, U, E, Z, W, D
c
), and
hence can be solved by employing the LMI Tool [69]. The problem of searching
for the optimal scaling parameters λ
i
, ε
i
, i = 1, 2, 3 in general may be nu
merically costly although the function fminsearch in MATLAB Optimization
Toolbox may be applied.
6.5 Application in servo loop design for hard disk drives
6.5.1 Problem formulation
A blockdiagram representation of a typical HDD servo loop is shown in Figure
6.1 with disturbances injected. P(z) and C(z) represent transfer functions of the
plant and controller, respectively. v represents all torque disturbances. d represents
disturbances that are due to nonrepeatable disk and suspension/slider motions. n
denotes the PES demodulation and measurement noise. z
1
is the true position error,
and e is the measured position error or measurement output y in (6.2). D
1
, D
2
, and
N are the disturbance and noise models, and w
1
, w
2
and w
3
are white noises of zero
mean and unit variance.
FIGURE 6.1
Mixed H
2
/H
∞
control scheme for HDD servo loop with disturbance models.
124 Modeling and Control of Vibration in Mechanical Systems
Through experiments, the frequency responses of the actual VCM is obtained and
is shown in Figure 6.2. A5th order model is used to approximate the actual frequency
responses of the VCM actuator and is given by
P(s) =
5.172 ×10
12
s
2
+ 1.82 ×10
17
s + 3.267 ×10
21
s
5
+ 2.117 ×10
4
s
4
+ 1.032 × 10
9
s
3
+ 1.906 × 10
13
s
2
+8.587 ×10
15
s + 7.345 ×10
18
. (6.42)
Figure 6.2 shows the comparison between the frequency responses of the actual data
and those of P(s). It is clear that their difference is more signiﬁcant for the frequency
range of over 4 kHz. To capture the unmodeled dynamics in high frequencies, dozens
of frequency response measurements are carried out and Figure 6.3 shows the multi
plicative uncertainty of the VCM actuator deﬁned by
∆(ω) =
N
mea
max
i=1
¸
¸
¸
¸
P
i
(jω) −P(jω)
P(jω)
¸
¸
¸
¸
, (6.43)
where N
mea
is the number of measurements and P
i
(jω) is the actual frequency
response of the plant in the ith measurement, and P(jω) is the frequency response
of the model in (6.42). An approximate bounding function W
u
(s), i.e., the smooth
line in Figure 6.3, is obtained as
W
u
(s) =
3s
2
+ 2.903 × 10
4
s + 1.433 ×10
8
s
2
+ 3.016 ×10
4
s + 1.421 ×10
9
. (6.44)
From Figure 6.3, it is clear that the uncertainty at frequency over 5 kHz is the major
concern. We observe that the actual uncertainties at some frequencies below 5 kHz
exceed the bounding function, which, however, will not cause any major problem.
In fact, we verify that the robust stability of our designed system is guaranteed even
with the worst case of uncertainty.
By discretization using the zeroorder hold, the corresponding zdomain models
of the VCM and the bounding function, i.e., P(z) and W
u
(z) can be obtained.
The disturbance and noise models D
1
(s), D
2
(s) and N(s) are given by (2.42)−(2.44),
and D
1
(z), D
2
(z) and N(z) are their discretetime forms.
As mentioned, one of the most important performance measures for HDDs is the
track misregistration or TMR, the total amount of random ﬂuctuation about the de
sired track location. TMR is used to judge the required accuracy of positioning and
thus to scale the disk capacity. To achieve a high capacity disk drive, one way in
servo control is to minimize TMR, which is given in terms of the standard deviation
of the true PES, i.e.,
3σ
z
1
= 3
¸
¸
¸
_
1
q −1
q
i=1
z
1
(i)
2
, (6.45)
where q is the number of true PES samples.
Mixed H
2
/H
∞
Control Design for Vibration Rejection 125
10
2
10
3
10
4
−40
−20
0
20
40
60
M
a
g
n
i
t
u
d
e
(
d
B
)
measured
modeled
10
2
10
3
10
4
−1200
−1000
−800
−600
−400
−200
0
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
FIGURE 6.2
Frequency responses of the VCM actuator.
126 Modeling and Control of Vibration in Mechanical Systems
10
2
10
3
10
4
−60
−50
−40
−30
−20
−10
0
10
20
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
VCM uncertainty
bounding curve
FIGURE 6.3
Multiplicative uncertainty of the VCM actuator.
Mixed H
2
/H
∞
Control Design for Vibration Rejection 127
Let ˜ w = [w
1
w
2
w
3
]
T
and T
z
1
˜
w
denote the transfer function matrix from ˜ w to z
1
.
When q is large enough, the H
2
norm of T
z
1
˜
w
is given by [88]
T
z
1
˜
w
2
≈
¸
¸
¸
_
1
q − 1
q
i=1
z
1
(i)
2
. (6.46)
Thus, the control design problem to minimize TMR can be treated as an H
2
optimal
control problem.
On the other hand, we need to ensure the system stability against the unmodeled
high frequency dynamics of the VCM actuator, i.e., the constraint TW
u
∞
< 1
is to be met, where T is the closedloop transfer function and W
u
is the bounding
function of the unmodeled dynamics which was derived earlier. Therefore, we have
the mixed H
2
/H
∞
control scheme as shown in Figure 6.1, where w
4
∈ ℓ
2
[0, ∞), a
disturbance input or a reference. Clearly, the transfer function from w
4
to z
2
is TW
u
.
We now derive the statespace representation (6.1)−(6.4) for the system in Figure
6.1 with
x =
_
x
T
p
x
T
d
1
x
T
d
2
x
T
n
x
T
u
¸
T
, (6.47)
w =
_
w
1
w
2
w
3
w
4
¸
T
, (6.48)
A =
_
¸
¸
¸
¸
_
A
p
B
p
C
d
1
0 0 0
0 A
d
1
0 0 0
0 0 A
d
2
0 0
0 0 0 A
n
0
B
u
C
p
0 B
u
C
d
2
0 A
u
_
¸
¸
¸
¸
_
, (6.49)
B
1
=
_
¸
¸
¸
¸
_
B
p
D
d
1
0 0 0
B
d
1
0 0 0
0 B
d
2
0 0
0 0 B
n
0
0 B
u
D
d
2
0 0
_
¸
¸
¸
¸
_
, B
2
=
_
¸
¸
¸
¸
_
B
p
0
0
0
0
_
¸
¸
¸
¸
_
, (6.50)
C
2
=
_
−C
p
0 −C
d
2
C
n
0
¸
, (6.51)
D
21
=
_
0 −D
d
2
D
n
1
¸
, C
z1
=
_
−C
p
0 −C
d
2
0 0
¸
,
D
z11
=
_
0 −D
d
2
0 0
¸
, D
z12
= 0, (6.52)
C
z2
=
_
D
u
C
p
0 D
u
C
d
2
0 C
u
¸
, (6.53)
D
z21
=
_
0 D
u
D
d
2
0 0
¸
, D
z22
= 0, (6.54)
and x
p
, x
d
1
, x
d
2
, x
n
, and x
u
are respectively the state variables of the VCM actua
tor P(z), the input disturbance model D
1
(z), the output disturbance model D
2
(z),
the measurement noise model N(z), and the uncertainty W
u
(z). (A
p
, B
p
, C
p
, D
p
),
(A
d
1
, B
d
1
, C
d
1
, D
d
1
), (A
d
2
, B
d
2
, C
d
2
, D
d
2
), (A
n
, B
n
, C
n
, D
n
) and (A
u
, B
u
, C
u
, D
u
)
are respectively the statespace models of P(z), D
1
(z), D
2
(z), N(z), and W
u
(z).
Note that for the servo control shown in Figure 6.1, n
z
= n
y
= 1.
128 Modeling and Control of Vibration in Mechanical Systems
6.5.2 Design results
In this section, we will apply the mixed H
2
/H
∞
control to hard disk drive servo
formulated previously. The sampling frequency being used is 20 kHz. By applying
Theorem 6.2 and searching for the optimal scaling parameters, we obtain λ
1
= 0.3,
λ
2
= 0.31, λ
3
= 0.9, ε
1
= 0.3, ε
2
= 0.28, ε
3
= 1.1 and the minimum H
2
norm,
i.e., the σ value of the true PES z
1
, of 0.00748 µm.
For the purpose of comparison, we also design a mixed H
2
/H
∞
controller for the
disk drive using the approach in Theorem 6.1. The minimum H
2
norm of 0.01013
µm is obtained. Starting from this controller, we carried out a further iterative pro
cedure between controller variables and Lyapunov parameters:
Step 1: obtain the closedloop system (
¯
A,
¯
B,
¯
C,
¯
D) with the controller parameters
(A
c
, B
c
, C
c
, D
c
), and let A =
¯
A, B =
¯
B, C =
¯
C, and D =
¯
D.
Step 2: solve LMIs (5.14)−(5.15), (5.20) for P and X, and minimize Trace(Π).
If Trace(Π) does not differ from the previous value, stop. Otherwise, go to step 3.
Step 3: With the obtained P and X, solve LMIs (5.14)−(5.15), and (5.20) for
(A
c
, B
c
, C
c
, D
c
).
Step 4: go to Step 1.
The iterative procedure gives a controller that produces a slight improvement of
the H
2
norm, i.e., 0.01002 µm. Hence, the improved approach represents about
25.3% more improvement on TMR than the design method in Theorem 6.1 together
with an iterative reﬁnement.
Figure 6.4 shows the comparison of sensitivity functions, where we can see that
the sensitivity function designed based on the improved design method is better than
that from Theorem 6.1, except that its hump is slightly higher. The comparison of
control performances obtained by the improved method and that of Theorem 6.1
is given as in Table 6.1, where the bandwidth from the improved method is much
higher. Although the H
∞
norm of TW
u
of the improved design is slightly higher
than that of the design using Method 1, it is below one as required, implying the
designed controller makes the closedloop system robustly stable in the presence of
uncertainty bounded by W
u
. Further, from Figure 2.23, the disturbance mainly con
centrates on frequencies below1 kHz, hence, the slight higher peak does not degrade
the disturbance rejection performance much. Figure 6.5 shows the testing result of
sensitivity functions, which is consistent with the simulation results in Figure 6.4.
REMARK 6.6 From Table 6.1, one may argue that a reduced H
2
norm
for the method of Theorem 6.1 may be obtained by a γ value of greater
than one as the actual H
∞
norm is 0.78, lower than that of the improved
method. However, based on our simulations, no obvious improvement on the
H
2
performance has been observed. For example, with γ = 1.5, a slightly
reduced H
2
norm of 10.11 nm is obtained whereas the actual H
∞
norm is
0.79. With γ = 5, the H
2
norm is reduced to 10.08nm with an unchanged
H
∞
norm of 0.78. This means that with a larger γ, the improvement on the
H
2
norm for the method of Theorem 6.1 is negligible.
Mixed H
2
/H
∞
Control Design for Vibration Rejection 129
The improved method is also compared with conventional PID design. As shown
in Figure 6.4 and Figure 6.5, the sensitivity function by PID control has a lower
bandwidth and almost the same peak value, which gives a higher TMR as listed in
Table 6.1.
The previous simulation is carried out with
¯
Σ
1
and
¯
Σ
2
as solutions to the LMI
(6.32). When (3,1) and (3,2) blocks in (6.32) are set to zeros, it is found that the
designed controller gives a 9% lower bandwidth for the closedloop system, leading
to a worse TMR. This demonstrates that the parameter Σ in Lemma 6.1 is useful in
achieving a better performance.
10
3
10
4
−30
−25
−20
−15
−10
−5
0
5
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
Method 1
Method 2
PID
FIGURE 6.4
Frequency response of sensitivity functions.
Next, we shall calculate the H
2
norm using the measured plant frequency response
as in Figure 6.3 and sensitivity function as in Figure 6.5. The spectrum of z
1
is given
by
z
1
(f
k
)
2
= P(f
k
)S(f
k
)
2
D
1
(f
k
)
2
+ S(f
k
)
2
D
2
(f
k
)
2
+1 −S(f
k
)
2
N(f
k
)
2
(6.55)
130 Modeling and Control of Vibration in Mechanical Systems
10
3
10
4
−45
−40
−35
−30
−25
−20
−15
−10
−5
0
5
10
M
a
g
n
i
t
u
d
e
(
d
B
)
Frequency(Hz)
Method 1
Method 2
PID
FIGURE 6.5
Frequency response of sensitivity functions.
Mixed H
2
/H
∞
Control Design for Vibration Rejection 131
TABLE 6.1
Control performance comparison
Method Method 2 Method 1 PID
Openloop crossover frequency (Hz) 1.4k 969 976
H
2
norm (σ (nm)) 7.48 10.13 12.11
TW
u
∞
0.89 0.78 0.27
where f
k
(k = 1, 2, 3..., K) are frequency points, P(f
k
) represents the measured
frequency response of the plant, and S(f
k
) is the frequency response of the sensi
tivity function. The resultant σ value of z
1
is 0.01155, better than 0.02714 with the
controller designed by the method in Theorem 6.1.
6.6 Conclusion
This chapter has presented two design methods for the mixed H
2
and H
∞
control.
One is a slack variable approach, and another one is a less or equally conservative
design in terms of LMIs that contain more free variables than the conventional ap
proaches. Those variables offer additional freedoms in optimization, resulting in a
less or equally conservative control design. The improved H
2
/H
∞
control design
has been applied to hard disk drives to minimize the track misregistration while guar
anteeing the system robustness in the presence of actuator uncertainties. Compared
with the slack variable method, the improved design result for hard disk drives has
indicated a marked improvement of 25.3%in the H
2
norm or the TMR performance
while guaranteeing the system robustness by satisfying the required H
∞
constraint.
The experimental result validates the advantage of the improved design.
7
LowHump Sensitivity Control Design for
Hard Disk Drive Systems
7.1 Introduction
In feedback control systems, sensitivity functions are critical to the determination of
their ability in disturbance and noise rejections. However, Bode has shown the limi
tation of using a feedback structure in terms of an integral constraint on the sensitivity
function, as discussed in Chapter 5. Brieﬂy speaking, the Bode integral theorem im
plies that we cannot have a sensitivity function less than unity at all frequencies using
output feedback with a ﬁnitebandwidth controller. Such a sensitivity function must
amplify the disturbances existing in frequencies higher than the system bandwidth.
In view of this, we shall employ the special structure of a secondary actuator system
and design appropriate controllers for primary and secondary actuators such that the
hump of the sensitivity function comes as low as possible without the cost of low
frequency performance. This optimized sensitivity function is expected to minimize
the ampliﬁcation of highfrequency disturbances while attenuating lowfrequency
and midfrequency disturbances. With this lowhump sensitivity, the dualstage con
trol system is able to reduce the contribution from all existing disturbances to the
error. Two types of microactuator models are considered in this chapter: a MEMS
actuator [102] and a PZT actuator [103]. The purpose is to design controls for the
primary and the secondary actuators such that a low hump of the sensitivity func
tion can be achieved with the help of secondary actuators. A comparison will be
made to evaluate the effectiveness of the proposed method for the two microactua
tors. Besides simulations, an implementation with a PZT microactuator veriﬁes that
the HDD servo loop design method leads to a lowhump sensitivity function.
7.2 Problem statement
Figure 7.1 shows a dualstage actuation system with one primary actuator P
v
(s) and
one secondary actuator P
m
(s), and two parallel controllers C
v
(s) and C
m
(s). With
disturbances and noise injected, the error is contributed by the disturbances and noise
133
134 Modeling and Control of Vibration in Mechanical Systems
in terms of
S
e
= P
v
(f
k
)S(f
k
)
2
d
1
(f
k
)
2
+ S(f
k
)
2
d
2
(f
k
)
2
+S(f
k
)
2
n(f
k
)
2
, (7.1)
where S
e
is the power spectrum of the error e, S is the sensitivity function, and
f
k
(k = 1, 2, · · · , N) are frequency points. Figure 7.2 shows an example of main
disturbances existing in disk drive systems, where the lowfrequency components are
represented by d
1
, the higher frequency portions such as disk vibration and windage
are lumped as d
2
, and the base line of the spectrum stands for the noise n.
FIGURE 7.1
Parallel structure of a dualstage actuation system with disturbances and noise in
jected.
Equation (7.1) implies that the sensitivity function S is important in determining
the disturbance rejection of the dualstage closed loop control system. In a conven
tional singlestage system using one primary actuator it is difﬁcult to have a low
hump sensitivity function. Thus the dualstage structure is applied to increase the
system bandwidth and lower the sensitivity function peak such that
sup S(jω) ≤ 1 + τ, ω = 0, · · · , ∞ (7.2)
where 0 < τ << 1 is a sufﬁciently small tolerance. The sensitivity function satisfy
ing (7.2) is called a nearly “ﬂat” sensitivity function.
A dualstage actuation system uses a microactuator to increase the system band
width. Among many control schemes for the dualstage control loop, Figure 7.1 is
one of the most popular ones [106]. The overall sensitivity function S(s) is
S(s) =
1
1 +P
v
(s)C
v
(s) +P
m
(s)C
m
(s)
. (7.3)
Another popular control scheme is the decoupled structure as shown in Figure 7.3
[101], where the overall sensitivity function of the closedloop system is given by
LowHump Sensitivity Control Design for Hard Disk Drive Systems 135
cascading each sensitivity function, i.e.,
S(s) = S
m
(s)S
v
(s),
S
v
(s) =
1
1 + P
v
(s)C
v
(s)
,
S
m
(s) =
1
1 + P
m
(s)C
m
(s)
. (7.4)
In this chapter we are mainly concerned with the parallel structure as in Figure 7.1
since it is a basic structure for dualstage control systems and can be converted to the
decoupled masterslave structure.
FIGURE 7.2
Power spectrum of PES nonrepeatable runout in open loop.
136 Modeling and Control of Vibration in Mechanical Systems
FIGURE 7.3
Decoupled structure of dualstage actuation systems.
FIGURE 7.4
Structure of H
∞
loop shaping.
LowHump Sensitivity Control Design for Hard Disk Drive Systems 137
7.3 Design in continuoustime domain
7.3.1 H
∞
loop shaping for lowhump sensitivity functions
The H
∞
loop shaping method is used to design controllers for the primary actuator
and the microactuator to achieve a lowhump sensitivity function for the dualstage
actuator system. The structure of the H
∞
loop shaping method is depicted in Figure
7.4, where W(s) is a weighting function of the desired sensitivity function. For
a plant model P(s), a controller C(s) is to be designed such that the closedloop
system is stable and
T
zw
∞
< 1 (7.5)
is satisﬁed, where T
zw
is the transfer function from w to z, i.e., S(s)W(s). Clearly,
(7.5) means that the sensitivity function S(s) can be shaped similarly to the inverse
of the chosen weighing function W(s). A simple form of W(s) is
W(s) =
1
M
s
2
+ 2ωζ
1
√
M
s + ω
2
s
2
+ 2ω
√
εs +ω
2
ε
, (7.6)
where ω is valued by the desired bandwidth, ε is to determine the lowfrequency
level of the desired sensitivity function, and ζ is the damping ratio.
Associated with the weighting function, Figure 7.4 can be formulated as follows.
˙ x(t) = Ax(t) +B
1
w(t) +B
2
u(t), (7.7)
z(t) = C
1
x(t) +D
11
w(t) + D
12
u(t), (7.8)
y(t) = C
2
x(t) +D
21
w(t) + D
22
u(t), (7.9)
where
A =
A
p
0
B
w
C
p
A
w
, B
1
=
0
B
w
, B
2
=
B
p
B
w
D
p
,
C
1
=
D
w
C
p
C
w
, D
11
= D
w
, D
12
= D
w
D
p
,
C
2
=
C
p
0
, D
21
= 1, D
22
= D
p
,
(A
p
, B
p
, C
p
, D
p
) and (A
w
, B
w
, C
w
, D
w
) are respectively the statespace realiza
tions of plant P(s) and weighting function W(s). Let (A
c
, B
c
, C
c
, D
c
) be the state
space description of C(s). Then (A
c
, B
c
, C
c
, D
c
) is to be designed such that (7.5)
is satisﬁed. An LMI approach stated in Theorem 5.4 is used to design the controller.
It is known that MATLAB functions, say “hinfsyn.m,” are available to design
the controller. However, numerical errors will occur due to the large gain of VCM
actuator and will be the hindrance for running the function. Thus, we would rather
use the LMI approach in our application to the VCM actuator. There always exists
a minimum level γ that makes the LMI (5.99) solvable, which gives a sensitivity
138 Modeling and Control of Vibration in Mechanical Systems
function closer to the inverse of its weighting function than that given by a larger γ.
Certainly, the solvability is also related to the chosen weighting W(s), which must
be realistic due to the Bode limitation.
Bode’s integral theorem allows the possibility of a “ﬂat” sensitivity function up to
a frequency of our concern in the continuous time domain, since the integral con
straint in (5.136) is deﬁned on the frequency range from 0 to inﬁnity. Equation (7.2)
can be achieved by choosing appropriate weighting functions for S
v
and S
m
sep
arately. One way is to put the peak of S
v
within the reduction band of S
m
, as in
Figure 7.5, and lower the highfrequency hump of S
m
. This would need to decrease
the bandwidth of the primary actuator loop and increase the bandwidth of the mi
croactuator loop, which can be realized by adjusting the weighting functions W
v
(s)
for the primary actuator and W
m
(s) for the microactuator. Here, the primary actuator
loop and the microactuator loop are designed separately for the dualstage parallel
structure. An additional reason for this is to meet the goals of ensuring the stability
of the separate primary and microactuator loops and as in Figure 7.6, letting the pri
mary actuator open loop have a higher gain at low frequencies and the microactuator
open loop have a higher gain at high frequencies. The dualstage parallel control
scheme could be formulated as an MIMO problem, however, to satisfy these speciﬁc
requirements, the controller design would be more complicated.
LowHump Sensitivity Control Design for Hard Disk Drive Systems 139
FIGURE 7.5
Frequency responses of S
v
(s) and S
m
(s).
140 Modeling and Control of Vibration in Mechanical Systems
FIGURE 7.6
Frequency responses of P
v
(s)C
v
(s) (solid line) and P
m
(s)C
m
(s)(dotted line).
LowHump Sensitivity Control Design for Hard Disk Drive Systems 141
7.3.2 Application examples
This section will apply the proposed control design method to a dualstage actuation
system in hard disk drives with the structure as in Figure 7.1 such that the designed
controllers can produce a lowhump sensitivity function. The dualstage HDD uses
a microactuator as a ﬁne positioner to increase the positioning accuracy. The mi
croactuator piggyback on a VCM actuator is driven jointly with the VCM actuator
through suspension, slider or head [101]. Due to various microactuator designs, two
different microactuator models will be studied and they are coupled separately with
one VCM actuator.
Figure 7.7 shows the frequency response of the VCM actuator and the poles, zeros
and gain of its model are listed in (7.10)−(7.11). With ω = 2π350, ε = 10
−3.2
,
ζ = 0.4 in (7.6), VCM controller C
v
is designed as in Figure 7.8 using the H
∞
loop
shaping method described in the previous section. The sensitivity function S
v
(s) is
shown in Figure 7.5.
FIGURE 7.7
Frequency response of VCM actuator P(s).
poles = [−2312.2122 ± 57759.0421j, −2670.3538 ± 53340.2745j,
−779.1150 ±38947.9570j, −942.4778 ± 31401.7862j,
−219.9115 ±588.5772j, −18849.5560]; (7.10)
142 Modeling and Control of Vibration in Mechanical Systems
zeros = −2356.1945 ±47064.9481j,
Gain = 2.127723028788067 ×10
40
. (7.11)
FIGURE 7.8
Frequency response of VCM controller C
v
(s).
Various microactuators have been proposed for HDD dualstage servo systems,
including the MEMS actuator and PZT actuator, which are designed by different
mechanism and possess different dynamics. They can generally be characterized by
a pade delay and a 2pole rolloff model:
PadeDelay ×O(s
2
) (7.12)
According to their physical behavior and the locations of the poles, microactuators
are categorized as two cases. The lowhump design method in Section 7.3.1 will be
applied to the two cases to investigate their effect on the ultimate dualstage sensi
tivity function.
Case 1: Poles are located at low frequency, such as a MEMS actuator [101]
P
m
(s) =
2 × 10
6
s
2
+ 282.7s + 8.883 ×10
7
. (7.13)
Equation (7.13) is a special case of (7.12) without delay. With ε = 10
−0.6
, ζ = 1
and ω = 2π2391 in (7.6), the designed controller C
m
(s) is shown in Figure 7.9.
LowHump Sensitivity Control Design for Hard Disk Drive Systems 143
Figure 7.10 implies that the resultant sensitivity function S
m
(s) follows its weighting
function closely. The dualstage servo systemhas a gain margin of 17 dB and a phase
margin of 92
◦
, and its open loop frequency response is shown in Figure 7.11. A
“ﬂat” sensitivity function is achieved as shown in Figure 7.12 even though the Bode
integral limitation (5.136) has to be fulﬁlled in this case.
FIGURE 7.9
Frequency response of microactuator controller C
m
(s).
144 Modeling and Control of Vibration in Mechanical Systems
FIGURE 7.10
Sensitivity function S
m
(s) (solid) and its weighting function inverse (dashed).
FIGURE 7.11
Open loop frequency response of the dualstage system.
LowHump Sensitivity Control Design for Hard Disk Drive Systems 145
FIGURE 7.12
Sensitivity and complementary sensitivity functions.
Case 2: Poles are located around 10 kHz, such as the active piezoelectric suspen
sion modeled by
P
m
(s) =
−1.438572836 ×10
9
(s − 6.157 ×10
5
)(s
2
+ 923.7s + 1.934 ×10
9
)
(s + 6.157 ×10
5
)(s
2
+ 791.7s + 1.517 ×10
9
)(s
2
+ 5089s + 7.195 ×10
9
)
(7.14)
with the frequency response in Figure 7.13.
For the control design of (7.14), ε is kept unchanged to have the same low
frequency level of the sensitivity function as in Case 1, and ω is adjusted to be
2π2107 to have a similar openloop bandwidth. The resultant C
m
(s) is shown in
Figure 7.14. The designed S
v
(s) and S
m
(s) are matched as in Figure 7.5, which
subsequently leads to the overall sensitivity as in Figure 7.16. We can observe that
a “ﬂat” sensitivity function has been achieved. The frequency response of the open
loop dualstage system is shown in Figure 7.15, where the gain margin is 31 dB and
the phase margin is 88
◦
. Note that a nonminimum phase zero is included in the
microactuator model and a nearly “ﬂat” sensitivity function is still achieved.
REMARK 7.1 The control design is based on the relevant sensitivity
function only and does not consider robustness to plant pole/zero variations.
This will, however, not hamper the practical usability of the resulting con
trollers due to the large gain and phase margins.
146 Modeling and Control of Vibration in Mechanical Systems
TABLE 7.1
Control performance comparison.
Microactuator model type Case 1 Case 2
Gain Margin (dB) 17 31
Phase Margin (deg) 90 88
Sensitivity function Flat Flat
Control performances are summarized in Table 7.1 for the two dualstage systems
consisting of different microactuators. It can be seen that they both achieve a “ﬂat”
sensitivity function with bandwidth of 2 kHz. Case 2, the PZT actuated suspension,
looks better than Case 1, the MEMS actuator, in the sense that its gain margin is much
higher and phase margin is slightly lower, whereas its behavior is to be improved to
possess a characteristic as simple as a padedelay which relies on its manufacturing
process. Also, froma manufacturing point of view, the MEMS actuator beneﬁts from
its mass production and low cost.
FIGURE 7.13
Microactuator frequency response.
LowHump Sensitivity Control Design for Hard Disk Drive Systems 147
FIGURE 7.14
Frequency response of microactuator controller C
m
(s).
FIGURE 7.15
Open loop frequency response of the dualstage system.
148 Modeling and Control of Vibration in Mechanical Systems
FIGURE 7.16
Sensitivity and complementary sensitivity functions.
7.3.3 Implementation on a hard disk drive
The implementation is carried out for Case 2 in the previous section, i.e., the dual
stage system consisting of the VCM in Figure 7.7 and the PZT microactuator in
Figure 7.13. The VCMand PZT controllers are obtained by discretizing the designed
controllers with frequency responses shown in Figure 7.8 and 7.14 and sampling rate
of 40 kHz.
The controllers are implemented with dSpace 1103 on TMS320C240 DSP board.
An LDV is used to measure the position of the dualstage actuator as the feedback
signal. The displacement range used is 2 µm/V. A DSA is used to measure the
frequency response of the sensitivity function by injecting a swept sine signal at
point A in Figure 7.17. The frequency response of points B over A is the sensitivity
function.
The resultant sensitivity functions are shown in Figure 7.18, where the rough line
is the tested result and the smooth line is the simulation result. We can observe that
the hump of the sensitivity function is lower than 3 dB, which is better than that with
the following PIDlike controllers designed by the leadlag method:
C
v
(z) =
0.9z
2
− 1.778z + 0.8788
z
2
−1.532z + 0.5326
, (7.15)
C
m
(z) =
0.4081z + 0.2627
z
2
−1.209z + 0.2631
. (7.16)
The tested and simulation results of the dualstage open loop are shown in Figure
LowHump Sensitivity Control Design for Hard Disk Drive Systems 149
7.19. The step response of the dualstage system is shown in Figure 7.20, where
Channel 1 is the output corresponding to the reference Channel 4, Channel 2 is the
control signal of VCM and Channel 3 is the control signal of PZT.
Note that the hump of the sensitivity function S(z) is tested to be 3 dB, instead of
0 dB as in Figure 7.16. This is mainly due to the locations of the microactuator poles
which are around 10 kHz and not sufﬁciently far away from the bandwidth of 2 kHz.
We know that a pole at 20 kHz may not affect 2 kHz bandwidth that much and thus
it is not needed to include a pole far beyond 20 kHz in the actuator modeling for 2
kHz servo bandwidth. Moreover, an actuator with highfrequency poles is helpful to
the lowhump sensitivity function design. It is expected that the microactuator could
be modeled as a padedelay within the frequency range of our interest. In the next
section we discuss the possibility of achieving lowhump sensitivity functions in the
discrete time domain.
According to (7.1), the sensitivity function as in Figure 7.18 must amplify the cor
responding highfrequency disturbances due to the hump above 0 dB, and certainly
the one with the lower hump will result in less ampliﬁcation.
The position error is evaluated according to (7.1) with the sensitivity function in
Figure 7.18 associated with the PES spectrum in Figure 7.2. The Parseval’s formula
stated in Chapter 1 addresses how to calculate σ value in the frequency domain. As
such, Figure 7.21 shows the 3σ value of position error versus frequencies. The low
hump design outperforms the PID design for disturbance rejection with almost the
same performance at frequencies lower than 1.5 kHz, which is consistent with Figure
7.18.
FIGURE 7.17
Experimental structure.
150 Modeling and Control of Vibration in Mechanical Systems
FIGURE 7.18
Sensitivity function of the dualstage system (smooth line: simulation result; rough
line: testing result; dotted line: PID design).
FIGURE 7.19
Open loop frequency response of the dualstage system (smooth line: simulation
result; rough line: testing result.)
LowHump Sensitivity Control Design for Hard Disk Drive Systems 151
FIGURE 7.20
Step response of the dualstage system.
FIGURE 7.21
3σ of PES NRRO versus frequencies.
152 Modeling and Control of Vibration in Mechanical Systems
7.4 Design in discretetime domain
This part is concerned with servo loop control design in the discretetime domain
that aims to achieve a lowhump sensitivity function for dualstage actuator systems.
Before the control design, sensitivity functions for a discretetime dualstage system
will be analyzed on the basis of the Bode integral theorem.
7.4.1 Synthesis method for lowhump sensitivity function
We recall a simpliﬁed Bode’s theorem which is helpful to analyze the sensitivity
function of the dualstage servo system.
THEOREM 7.1
Discrete Bode’s Theorem [104] Consider a discretetime SISO LTI system
whose openloop transfer function G(z) does not have unstable poles. The
sensitivity function of the system is given by S(z) = 1/(1 + G(z)). If the
closedloop system is stable and k
g
= lim
z→∞
G(z), then
π
−π
lnS(e
jω
)dω = 2π(−lnk
g
+ 1). (7.17)
Based on the discrete Bode’s theorem, we have the following statements.
a). When G(z) is strictly proper, then lim
z→∞
G(z) = 0 and
π
−π
ln S(e
jω
)dω = 0.
b). When the orders of the denominator and numerator of G(z) are the same and
k
g
< −2 or k
g
> 0, then
π
−π
ln S(e
jω
)dω < 0.
c). When the orders of the denominator and numerator of G(z) are the same,
−2 ≤ k
g
< 0 and k
g
= −1, then
π
−π
ln S(e
jω
)dω ≥ 0.
Thus, we can conclude that only if G(z) is nonstrictly proper and k
g
< −2 or
k
g
> 0, then
S(e
jω
) ≤ 1, ω ∈ [0, π],
i.e., the sensitivity function is bounded by 0 dB.
LowHump Sensitivity Control Design for Hard Disk Drive Systems 153
The openloop transfer function of the dualstage parallel system is the sum of
each path, i.e.,
G(z) = G
v
(z) + G
m
(z), (7.18)
G
v
(z) = P
v
(z)C
v
(z), (7.19)
G
m
(z) = P
m
(z)C
m
(z). (7.20)
Note that a VCM actuator can be approximately represented by a double integrator
(i.e., k/s
2
) combined with some resonance modes and is often described by a strictly
proper model. When G
m
(z) is strictly proper, G(z) is also strictly proper and when
G
m
(z) is nonstrictly proper, G(z) is nonstrictly proper. It can be concluded that
the Bode’s integral of the overall dualstage loop is determined by the microactuator
loop. Only a nonstrictly proper model of microactuator could possibly produce a
“ﬂat” sensitivity function for dualstage servo systems. An additional condition for
the “ﬂat” sensitivity is k
g
< −2 or k
g
> 0. These are necessary conditions and could
be used as a criterion to examine the closedloop design.
In what follows, the discrete H
∞
loop shaping method will be applied to the con
trol designs of the VCM loop and the microactuator loop such that the sensitivity
functions of the two loops are coupled and achieve a lowhump overall sensitivity.
The structure of the H
∞
loop shaping method is the same as in Figure 7.4. In the
discrete time case, Figure 7.4 is formulated as follows.
x(k + 1) = Ax(k) +B
1
w(k) + B
2
u(k), (7.21)
z(k) = C
1
x(k) +D
11
w(k) +D
12
u(k), (7.22)
y(k) = C
2
x(k) +D
21
w(k) +D
22
u(k), (7.23)
where
A =
A
p
0
B
w
C
p
A
w
, B
1
=
0
B
w
, B
2
=
B
p
B
w
D
p
C
1
=
D
w
C
p
C
w
, D
11
= D
w
, D
12
= D
w
D
p
,
C
2
=
C
p
0
, D
21
= 1, D
22
= D
p
,
(A
p
, B
p
, C
p
, D
p
) and (A
w
, B
w
, C
w
, D
w
) are respectively the statespace real
izations of plant P(z) and weighting function W(z). Let (A
c
, B
c
, C
c
, D
c
) be a
state space description of C(z). An LMI approach stated in Chapter 5 will be used
to design the controller C(z) : (A
c
, B
c
, C
c
, D
c
).
7.4.2 An application example
The VCM actuator and the microactuator are the same as those in Case 2 in Section
7.3.2. Notice that the PZT microactuator is represented using a Padedelay with two
2nd order resonance terms. The form (7.14) can be regarded as a general model of
PZT actuated suspensions. It is strictly proper and thus according to the analysis in
154 Modeling and Control of Vibration in Mechanical Systems
Section 7.4.1, a “ﬂat” sensitivity function is impossible. The used sampling rate for
the controller design is 40 kHz. The discretetime actuator models are obtained using
“zeroorderhold” method to ensure that the designed controllers are implementable.
With ω = 2π350, ε = 10
−3.2
, ζ = 0.4 in (7.6), a VCM controller C
v
(z) is de
signed as in Figure 7.22 using the H
∞
loop shaping method in the previous section.
Also applying the LMI approach, the controller C
m
(z) for the microactuator is de
signed as in Figure 7.23 with ω = 2π3700, ε = 10
−1.38
, ζ = 1, and M = 0.07
1/2
in (7.6). The sensitivity function of the dualstage system is shown in Figure 7.24,
where we can observe that the hump is below 3 dB. The openloop system has the
gain margin of 8.4 dB, the phase margin of 64
◦
and the bandwidth of 2.49 kHz.
REMARK 7.2 If manufacturing processes could produce an ideal actu
ator that can be modeled by a Padedelay only, which is nonstrictly proper,
it is possible to have a “ﬂat” sensitivity like that in Figure 7.25 with k
g
> 0.
An actual microactuator, which resembles the ideal case, is the PZT microactuator
in [107]. It can be described by a 1pole rolloff model
P
m
(s) =
1257
s + 1.257 × 10
5
, (7.24)
which is strictly proper and does not satisfy the necessary conditions in Section 7.3.1.
The sensitivity function with the hump lower than 2 dB, as shown as in Figure 7.26,
can be obtained with the same weighting function as that for (7.14).
LowHump Sensitivity Control Design for Hard Disk Drive Systems 155
FIGURE 7.22
VCM controller C
v
(z).
FIGURE 7.23
Microactuator controller C
m
(z).
156 Modeling and Control of Vibration in Mechanical Systems
FIGURE 7.24
Sensitivity function of the dualstage system.
FIGURE 7.25
Sensitivity function.
LowHump Sensitivity Control Design for Hard Disk Drive Systems 157
FIGURE 7.26
Sensitivity function of the dualstage system.
FIGURE 7.27
Frequency responses of P
v
(z)C
v
(z) (solid curve) and P
m
(z)C
m
(z) (dashed curve).
158 Modeling and Control of Vibration in Mechanical Systems
7.4.3 Implementation on a hard disk drive
The experimetal setup is the same as in Figure 7.17. An LDV with a range of 2
µm/V is used to measure the position of the dualstage actuator. Controllers are
implemented with dSpace 1103 on a TMS320C240 DSP board. When the dual
stage loop is closed and stabilized with the designed controllers, a swept sine signal
is injected at point A. A DSA is then used to measure the frequency response of
points B over A and obtain the sensitivity function.
The resultant sensitivity function is shown in Figure 7.28, where the rough line is
the testing result and the smooth line is the simulation result. We can observe that the
hump of the sensitivity function is lower than 3 dB, which is better than that by a PID
design as shown by the dotted line in Figure 7.28. The testing and simulation results
of the dualstage open loop system are shown in Figure 7.29. The step response in
Figure 7.30 shows that the system is stabilized and working in real time. Channel 2
in Figure 7.30 is the control signal of the VCM actuator and Channel 3 is the control
signal of the PZT microactuator.
The sensitivity function as in Figure 7.28 will amplify the corresponding high
frequency disturbances shown in Figure 7.2 due to the hump above 0 dB. The po
sition error is evaluated from (7.1) with the designed sensitivity functions. The 3σ
value of the position error versus frequencies is shown in Figure 7.31, where we can
see that the lowhump design outperforms the PID design for disturbance rejection
after 2.4 kHz, which is consistent with Figure 7.28.
The proposed control design is based on the sensitivity weighting function only
and does not consider robustness to plant parameter variations. This, however, will
not hamper the practical application of the resulting controllers due to the large gain
and phase margins. The system is veriﬁed to maintain stability in spite of the vari
ation of resonance frequency, e.g., ±5% shift of PZT resonance frequency around
13.5 kHz. The performance change with the presence of the system parameter un
certainty is illustrated in Figure 7.31, where we can see that the performance varies
slightly.
7.5 Conclusion
An H
∞
method has been proposed in both continuous and discrete time domains
to achieve a lowhump sensitivity function for dualstage HDD systems using an
LMI approach. Two different microactuator models have been studied, which are
represented by a MEMS actuator and a PZT actuated suspension. With the proposed
selection of sensitivity weighting functions, the sensitivity function with a hump
below 3 dB has been achieved in both simulations and experiments. Such a design
process can generate a robust servo controller with high disturbance rejection in a
low frequency range, and less vibration ampliﬁcation in a high frequency range, and
thus is effective in achieving higher positioning accuracy.
LowHump Sensitivity Control Design for Hard Disk Drive Systems 159
FIGURE 7.28
Sensitivity function of the dualstage system.
FIGURE 7.29
Open loop frequency responses of the dualstage system.
160 Modeling and Control of Vibration in Mechanical Systems
FIGURE 7.30
Step response of the dualstage system.
FIGURE 7.31
3σ value of PES NRRO versus frequency.
8
Generalized KYP LemmaBased Loop
Shaping Control Design
8.1 Introduction
To shape frequency responses of closedloop transfer functions such as sensitiv
ity/complementary sensitivity functions, H
∞
optimization together with frequency
weighting is a commonly used method. The additional weight functions, however,
increase the system and controller complexity since the weighting functions usually
have to be of high orders in order to capture the desired speciﬁcations accurately.
This is especially so when a controller is to be designed, aiming at achieving a wider
bandwidth while simultaneously suppressing disturbances of particular frequencies
within or beyond the servo bandwidth. Further, the process of choosing appropriate
weights is tedious and timeconsuming.
The KYP Lemma [61], being one of the most fundamental results in systems the
ory and control, establishes the equivalence between a frequency domain inequality
(FDI) for a transfer function and a linear matrix inequality associated with its state
space realization. It allows us to characterize various properties of dynamic systems
in the frequency domain in terms of linear matrix inequalities. The standard KYP
Lemma is only applicable for the inﬁnite frequency range, while the generalized
KYP Lemma [62] establishes the equivalence between a frequency domain prop
erty and a linear matrix inequality over a ﬁnite frequency range, allowing designers
to impose performance requirements over chosen ﬁnite or inﬁnite frequency ranges.
Hence, it is very suitable for analysis and synthesis problems in practical applications
where different speciﬁcations over different frequency ranges are usually required.
In this chapter, the generalized KYP Lemma is applied to design a feedback con
trol such that the speciﬁcations of the sensitivity function, required to suppress some
speciﬁc frequency disturbances, are satisﬁed. Unlike the standard KYP Lemma, the
matrix inequality in the generalized KYP Lemma involves a matrix variable which is
not necessarily positive deﬁnite and thus the Schur complement cannot be applied to
convexify the controller design. To overcome this difﬁculty, the Youla parametriza
tion approach is used to parameterize the closedloop transfer function. The search
for the coefﬁcients of the parameter Q(z) is then converted to a linear matrix in
equality problem within the generalized KYP Lemma framework. An application of
the method in the rejection of narrowband highfrequency and midfrequency distur
161
162 Modeling and Control of Vibration in Mechanical Systems
bances is presented to demonstrate the simplicity of the design and the improvement
of the positioning accuracy by the resultant controller.
8.2 Problem description
It is known that the power spectrum of the error e in Figure 2.20 is given by
S
e
= P(z)S(z)
2
d
1

2
+ S(z)
2
d
2

2
+S(z)
2
n
2
(8.1)
which implies that the sensitivity function S(z) is important in determining the dis
turbance rejection of a control loop. Thus, the design of a controller that gives
rise to a sensitivity function which can reject speciﬁc disturbances with known fre
quencies becomes rather signiﬁcant. The purpose here is then stated as: to de
sign a dynamic feedback controller C(z) for plant P(z) such that the closedloop
system is stable and for some prescribed positive scalars r
i
and frequency ranges
(f
i1
, f
i2
), i = 1, ..., N,
S(f) < r
i
, f
i1
≤ f ≤ f
i2
(8.2)
where S(f) = 1/(1 + C(f)P(f)). Smaller r
i
means the less contribution of the
disturbance in frequency range (f
i1
, f
i2
) to the error.
In view of the constraint stated in the Bode integral theorem, it is impossible to
achieve disturbance rejection at all frequencies higher than the bandwidth of the con
trol loop for actuators or microactuators used in mechanical motion systems such as
hard disk drives. The speciﬁcation (8.2) is considered for a speciﬁc frequency range,
which is however possible to be achieved through shaping the sensitivity function.
The above design problem may be approached by selecting a proper frequency
weighting function and carrying out an H
∞
optimization. However, the problem
of how to select a proper frequency weighting function that can give an accurate
shaping of the sensitivity function is generally difﬁcult and timeconsuming. Further,
the resultant controller order will depend on the order of the weighting function and
the plant. Note that a more accurate frequency shaping usually requires a higher
order weighting function.
The generalized KYP Lemma [62] gives a necessary and sufﬁcient condition for a
given transfer function to satisfy a required frequency domain property over a ﬁnite
frequency range in terms of a matrix inequality condition. Thus, it may be applied to
address the above design problem. In what follows, we employ the generalized KYP
Lemma to design a feedback control such that the sensitivity function satisﬁes the
required speciﬁcations so as to reject disturbances at speciﬁc frequencies. In order
to convexify the matrix inequality, the Youla parametrization approach as shown in
Figure 8.1 is used to design a controller with the generalized KYP Lemma.
Generalized KYP LemmaBased Loop Shaping Control Design 163
FIGURE 8.1
Q parameterization for control design.
8.3 Generalized KYP lemmabased control design method
The previous analysis indicates that the sensitivity function plays a key role in distur
bance rejection. To reject disturbance of frequency within a certain frequency range,
a proper shaping of the sensitivity function can be carried out. In this section, we
shall present a frequency shaping method based on the generalized KYP Lemma.
First, it is easy to see that the sensitivity function S(z) is equal to the transfer
function from w to z in Figure 8.1. The statespace model of the plant under consid
eration is denoted as (A
p
, B
p
, C
p
, D
p
). Then, a statespace representation of the
system in Figure 8.1 is given by
x(k + 1) = A
p
x(k) +B
p
u(k), (8.3)
z(k) = −C
p
x(k) + w(k) − D
p
u(k), (8.4)
where x ∈ R
n
x
is the state.
Let a statespace representation of the controller C(z) be given by (A
c
, B
c
, C
c
,
D
c
). Assuming that D = 1 + D
c
D
p
is invertible, then a statespace representation
of the sensitivity function can be given by (
˜
A,
˜
B,
˜
C,
˜
D), where
˜
A =
A
p
−B
p
D
−1
D
c
C
p
B
p
D
−1
C
c
−B
c
C
p
+B
c
D
p
D
−1
D
c
D
p
A
c
− B
c
D
p
D
−1
C
c
, (8.5)
˜
B =
B
p
D
−1
D
c
B
c
−B
c
D
p
D
−1
D
c
,
˜
C =
−C
p
+D
p
D
−1
D
c
C
p
−D
p
D
−1
C
c
,
˜
D = 1 −D
p
D
−1
D
c
. (8.6)
A special case of the generalized KYP Lemma that relates the bounded realness of
the sensitivity function over ﬁnite frequency ranges to its state space representation
is given below.
164 Modeling and Control of Vibration in Mechanical Systems
LEMMA 8.1
[62] Consider the sensitivity function S(z) =
˜
C(zI −
˜
A)
−1
˜
B+
˜
D with
˜
A being
stable. Then, for a given scalar r > 0, S(e
jθ
) ≤ r over a ﬁnite frequency
range if and only if there exist Hermitian matrices U and V ≥ 0 such that
˜
A
˜
B
I 0
∗
Σ
˜
A
˜
B
I 0
+
0 0
0 −r
2
˜
C
˜
D
∗
˜
C
˜
D
−I
≤ 0, (8.7)
where
(i) for low frequency range θ ≤ θ
l
,
Σ =
−U V
V U − (2cosθ
l
)V
; (8.8)
(ii) for middle frequency range θ
1
≤ θ ≤ θ
2
,
Σ =
−U e
jθ
c
V
e
−jθ
c
V U −(2cosθ
d
)V
,
θ
c
= (θ
1
+θ
2
)/2, θ
d
= (θ
2
−θ
1
)/2; (8.9)
(iii) for high frequency range θ ≥ θ
h
,
Σ =
−U −V
−V U + (2cosθ
h
)V
. (8.10)
For a given controller C(z), the above gives a necessary and sufﬁcient condition
for evaluating if S(z) ≤ r over some given frequency range in terms of LMI.
However, (8.7) is no longer an LMI when the controller C(z) = (A
c
, B
c
, C
c
, D
c
)
is to be designed since
˜
A and
˜
B involve the unknown parameters A
c
, B
c
, C
c
and
D
c
. Further, it is not possible to be converted to an LMI by Schur complement since
the matrix Σ is not deﬁnite. To overcome this difﬁculty, in the following we apply a
Youla parametrization approach where the controller C(z) is now of the structure as
shown in Figure 8.1.
Let K(z) be a given observer based controller that can be designed using ap
proaches such as the LQG control:
ˆ x(k + 1) = A
p
ˆ x(k) + B
p
u(k) +L(z(k) + C
p
ˆ x(k)), (8.11)
u(k) = −Mˆ x(k). (8.12)
Then, as stated in Chapter 5, the set of sensitivity functions can be parameterized as
S(z) = T
11
(z) +T
12
(z)Q(z)T
21
(z) (8.13)
Generalized KYP LemmaBased Loop Shaping Control Design 165
where Q(z) is a stable transfer function to be designed and
T
11
(z) T
12
(z)
T
21
(z) 0
= C
T
(zI − A
T
)
−1
B
T
+D
T
,
A
T
B
T
C
T
D
T
=
A
p
−B
p
M 0 B
p
−LC
p
A
p
− B
p
M + LC
p
L B
p
−C
p
D
p
M 1 −D
p
−C
p
C
p
1 0
. (8.14)
If Q(z) has the statespace realization (A
q
, B
q
, C
q
, D
q
), then the designed con
troller C(z) is given by
A
c
=
A
p
−B
p
M +LC
p
+B
p
D
q
C
p
B
p
C
q
B
q
C
p
A
q
, (8.15)
B
c
=
L +B
p
D
q
B
q
, C
c
=
−M + D
q
C
p
C
q
,
D
c
= D
q
. (8.16)
Denote the state space representations of T
11
(z) and T
12
(z)T
21
(z) by (A
t11
, B
t11
,
C
t11
, D
t11
) and (A
t
, B
t
, C
t
, D
t
), respectively, then from (8.13) a state space model
of S(z) can be written as
˜
A =
A
t11
0 0
0 A
t
0
0 B
q
C
t
A
q
,
˜
B =
B
t11
B
t
B
q
D
t
, (8.17)
˜
C =
C
t11
D
q
C
t
C
q
,
˜
D = D
t11
+ D
q
D
t
. (8.18)
Let Q(z) be an FIR ﬁlter:
Q(z) = q
0
+ q
1
z
−1
+q
2
z
−2
+ ... +q
τ
z
−τ
, (8.19)
q = [q
0
q
1
q
2
, ... , q
τ
] (8.20)
which is to be designed so that the required bounded realness of the sensitivity func
tion is satisﬁed. It is known that a state space realization for Q(z) can be given
by
A
q
=
0 I
τ−1
0 0
, B
q
=
0
(τ−1)×1
1
,
C
q
= [q
τ
q
τ−1
· · · q
1
] , D
q
= q
0
,
where I
τ−1
is the identity matrix of dimension (τ −1) ×(τ −1) and 0
(τ−1)×1
is the
zero matrix of dimension (τ −1) ×1. Note that the ﬁlter parameter q to be designed
only appears in C
q
and D
q
. Therefore, from (8.17)−(8.18), we know that q exists
in
˜
C and
˜
D only. In this case, (8.7) deﬁnes an LMI in terms of the variables U,
V , and q. Hence, U, V , and the design parameter q can be computed via a convex
optimization.
166 Modeling and Control of Vibration in Mechanical Systems
The KYP Lemmabased control design can be carried out following the steps:
Procedure 8.1
Step 1. Compute M and Lusing MATLAB commands M = dlqr(A
p
, B
p
, C
T
p
C
p
,
R) and L = A
p
· dlqe(A
p
, B
p
, − C
p
, W
d
, W
v
), where R is the weighting for the
control input in the cost function
J =
(ˆ x
T
C
T
p
C
p
ˆ x +u
T
Ru)
for linear quadratic regulator design, and W
d
and W
v
are the variance matrices of
process noise and measurement noise for the Kalman estimator design. Here in
the KYP Lemmabased control design, R, W
d
, and W
v
can be chosen as identity
matrices.
Step 2. Compute T
11
(z), T
12
(z) and T
21
(z) from (8.14).
Step 3. Obtain the state space model (
˜
A,
˜
B,
˜
C,
˜
D) in (8.17)−(8.18).
Step 4. Based on disturbance spectrum, specify the positive scalars r
i
and fre
quency points f
i
(i = 1, ..., N) for the sensitivity function
S(f
i
) < r
i
, f
i1
≤ f
i
≤ f
i2
(8.21)
where f
i1
and f
i2
deﬁne the frequency range. For each speciﬁcation on the resultant
sensitivity function in the frequency range, construct the LMI (8.7) in terms of the
variables U, V , and q with r = r
i
.
Step 5. Obtain q, U and V by solving these LMIs using MATLAB LMI toolbox.
If the LMIs are not solvable, the speciﬁcations given in Step 4 are to be adjusted.
Step 6. Obtain the controller parameters from (8.15)−(8.16).
8.4 Peak ﬁlter
Many other control methods are available to reject narrowband disturbance, such as
linear time invariant feedback methods using classical design and modern frequency
shaping and ﬁlter shaping, and adaptive feedforward methods using higher harmonic
control and LMS algorithm[57]. It has been shown that most of these methods would
result in a compensator with lightly damped complex poles at the center frequency
of the disturbance. Such a compensator has high gains at speciﬁc frequency. It is
named as peak ﬁlter according to its shape of frequency response. The peak ﬁlter
F(s) works in the control loop as in Figure 8.2 with its discretized form F(z).
8.4.1 Conventional peak ﬁlter
Conventionally, the peak ﬁlter is described as the following model
F(s) =
s
2
+ 2ξ
1
ω
p
s + ω
2
p
s
2
+ 2ξ
2
ω
p
s + ω
2
p
, (8.22)
Generalized KYP LemmaBased Loop Shaping Control Design 167
FIGURE 8.2
Peak ﬁlter F in the nominal feedback loop.
FIGURE 8.3
Peak ﬁlter in the frequency domain.
168 Modeling and Control of Vibration in Mechanical Systems
where ω
p
= 2πf
p
is the center frequency in radian/sec and ξ
1
and ξ
2
are the damping
ratios with ξ
1
> ξ
2
. The peak ﬁlter F(s) can be designed according to its Bode plot,
as illustrated in Figure 8.3. M and N denote the peak height, ∆ stands for the peak
width corresponding to N. Let
m = 10
M/20
, n = 10
N/20
. (8.23)
When M ≫N, ξ
1
and ξ
2
are determined approximately by
ξ
1
=
∆
2
+ 2∆
2∆ + 2
√
n −1, (8.24)
ξ
2
=
ξ
1
m
. (8.25)
The maximum phase loss θ caused by the pair of the lightly damped complex poles
is given by
θ = arctan
m−1
2
√
m
. (8.26)
It is noticed that the conventional version of the peak ﬁlter induces additional
phase loss. The phase loss negatively impacts the phase margin and distorts the
gain of the sensitivity function around the disturbance frequency, particularly for
the disturbance near the 0 dB crossover frequency. To overcome the drawback, a
phaselead peak ﬁlter [118] was proposed.
8.4.2 Phase lead peak ﬁlter
The phase ﬁlter is improved by adding a differentiator to provide additional phase
lead such that the phase margin is preserved and the sensitivity function curve is
smoothly shaped.
Let
T
0
(s) =
P(s)C(s)
1 + P(s)C(s)
. (8.27)
The peak ﬁlter adopts the following form
F(s) = K
−sin(φ)s
2
+ω
0
cos(φ)s
s
2
+ 2ξω
0
s +ω
2
0
, (8.28)
where ω
0
is the disturbance frequency at which high disturbance rejection is required,
ξ ∈ (0, 1) is the damping ratio, φ is the phase angle determined by
φ = arg[T
0
(jω
0
)] ∈ [−π, π], (8.29)
and 0 < K < γ is the positive ﬁlter gain. Let
G(ω, K) = 1 +T
0
(jω)F(jω), (8.30)
Generalized KYP LemmaBased Loop Shaping Control Design 169
and γ is the minimal positive real solution of the following two equations:
Re[G(ω, γ)] = 0, (8.31)
Im[G(ω, γ)] = 0. (8.32)
An estimate of ξ and K in the ﬁlter (8.28) is given by
ξ =
∆
d
(ω
0
+ 0.5∆
d
)
4ω
2
0
, (8.33)
K = (10
M/20
− 1)
2ξ
T
0
(jω
0
)
, (8.34)
where the disturbance bandwidth ∆
d
is deﬁned as the frequency difference between
the two points whose magnitudes are 1/
√
2 times of the peak value, and M in unit
dB is the desired reduction ratio of the narrowband disturbance.
The disturbance ﬁlter in (8.28) is a general highgain controller to reject narrow
band disturbances in a speciﬁc frequency range because the ﬁlter zero location can
be automatically shifted according to the disturbance frequency associated with the
baseline servo system with C(s) and P(s).
8.4.3 Group peak ﬁlter
The group ﬁltering scheme is used to compensate multiple frequency narrowband
disturbance. A parallel structure of the group peak ﬁlter is described as
F(s) =
n
i=1
F
i
(s), (8.35)
where the subﬁlter F
i
(s) is given by (8.28) with the corresponding ﬁlter parameters
(ω
i
, φ
i
, ξ
i
, K
i
).
An illustration example is shown in Figure 8.4 with the sensitivity functions plot
ted before and after activating the group ﬁlter with two subﬁlters at ω
1
= 2π700
and ω
2
= 2π2000. The ﬁlters associated with solid and dashed curves have different
parameters.
8.5 Application in high frequency vibration rejection
In this section, we shall provide an example to demonstrate the design and effective
ness of the proposed KYP based control to reject narrowband disturbances at high
frequencies. The previously introduced peak ﬁltering method ﬁnds it difﬁcult to deal
with disturbances in high frequency range, especially near the actuator resonance
modes.
170 Modeling and Control of Vibration in Mechanical Systems
10
2
10
3
10
4
−35
−30
−25
−20
−15
−10
−5
0
5
10
Frequency (Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
with filter
with filter
without filter
FIGURE 8.4
Sensitivity functions before and after group peak ﬁltering activated.
FIGURE 8.5
PZT microactuator attached to VCM actuator arm.
Generalized KYP LemmaBased Loop Shaping Control Design 171
0 2000 4000 6000 8000 10000 12000
0
0.5
1
1.5
2
2.5
3
x 10
−3
Frequency(Hz)
P
E
S
N
R
R
O
p
o
w
e
r
s
e
p
c
t
r
u
m
(
µ
m
)
8 kHz
10 kHz
FIGURE 8.6
Power spectrum of the position error before servo control.
A. System models
Figure 8.5 shows one kind of microactuators for HDDs, which is a PZT actuated
suspension attached to a VCM actuator arm. Figure 8.6 shows an example of the
main disturbances that exist in the hard disk drive servo system. Besides the distur
bances of low frequencies, the system suffers from the high frequency disturbances
around 8 kHz and 10 kHz, which are induced by air turbulence to suspensions or
sliders in HDDs. For higher rotational speed HDDs, these disturbances appear more
prominent and create a signiﬁcant impact on the positioning accuracy of the R/W
head. Here we employ the generalized KYP Lemma to design a feedback control
C(z) for the microactuator such that the sensitivity function S(z) satisﬁes the re
quired speciﬁcations so as to reject disturbances around 8 kHz and 10 kHz. The
desired speciﬁcations for the sensitivity function S(z) are set as
Spec.(a) S(f) < 0 dB, f ≤ 2 kHz,
Spec.(b) S(f) < −0.35 dB, 9950 Hz ≤ f ≤ 10050 Hz,
Spec.(c) S(f) < −1.41 dB, 7950 Hz ≤ f ≤ 8050 Hz.
Spec. (a) means to guarantee a 2 kHz bandwidth at least.
The transfer function of the PZT microactuator is given by
Pzt(s) =
180828605599509(s
2
+ 3079s + 1.934 × 10
9
)
(s + 1.257 × 10
5
)(s
2
+ 791.7s + 1.567 ×10
9
)(s
2
+ 5089s + 7.195 × 10
9
)
,
(8.36)
172 Modeling and Control of Vibration in Mechanical Systems
which is obtained by curveﬁtting to its frequency response measured via DSA. A
comparison between the measured and the modeled frequency responses is shown in
Figure 8.7, where it is noticed that there are two dominant resonance modes at 6.3
kHz and 13.5 kHz. Pzt(s) is discretized using the zeroorderhold method to obtain
Pzt(z). The sampling frequency being used is 40 kHz.
10
3
10
4
−40
−30
−20
−10
0
10
20
M
a
g
n
i
t
u
d
e
(
d
B
)
measured
modeled
10
3
10
4
−300
−200
−100
0
100
200
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
FIGURE 8.7
PZT micro actuator frequency response.
B. Controller design and LDV based experiment results
The plant Pzt(s) is precompensated by the integrator:
Int(z) =
6.3z
z − 1
, (8.37)
and two notch ﬁlters for suppressing the two main resonances at 6.3 kHz and 13.5
kHz before the KYP Lemma approach is applied to design a controller so that the
speciﬁcations (a), (b) and (c) can be satisﬁed. The notch ﬁlter is designed as the
following form:
Notch =
ω
2
2
ω
2
1
·
s
2
+ 2ξ
1
ω
1
s + ω
2
1
s
2
+ 2ξ
2
ω
2
s + ω
2
2
, (8.38)
where ω
1
is the frequency of the resonance to be suppressed , ξ
1
< ξ
2
, and ω
2
and
ω
1
should be chosen to be close to each other so that the resultant notch ﬁlter will
Generalized KYP LemmaBased Loop Shaping Control Design 173
not inﬂuence the system stability. Here for the model (8.36) the notch ﬁlters after
discretization are as follows.
Notch1(z) =
1.177z
2
−1.279z + 1.154
z
2
−0.8751z + 0.9259
,
Notch2(z) =
0.4011z
2
+ 0.3972z + 0.3532
z
2
+ 0.1391z + 0.007922
.
(A
p
, B
p
, C
p
, D
p
) in (8.3)−(8.4) is a state space description of the combined
system P(z) = Int(z) · Notch1(z) · Notch2(z) · Pzt(z). Q(z) is chosen as the
1
st
order FIR ﬁlter (8.19), and q
0
and q
1
are to be solved via the KYP Lemma.
As mentioned previously in Step 1 of the control design procedure, M and L are
obtained using MATLAB commands, i.e. M = dlqr(A, B
2
, C
T
1
C
1
, R) and L =
A· dlqe(A, B
2
, C
1
, W
d
, W
v
).
Next we use the KYP Lemma to search for the coefﬁcients of Q(z). Three LMIs of
the form(8.7) with Σin (8.8) and (8.9) need to be solved in order to achieve Spec. (a),
(b) and (c), respectively. The obtained q
0
and q
1
are q
0
= −0.4117, q
1
= 0.6371.
As such, C(z) can be obtained via (8.15)−(8.16). The resulting controller for the
PZT microactuator is C(z) · Int(z) · Notch1(z) · Notch2(z), and the sensitivity
function is shown in Figure 8.8. It can be observed in Figure 8.8 that with Q(z), the
sensitivity function is less than −4 dB and −2 dB at 8 kHz and 10 kHz respectively,
which means that speciﬁcations (a)−(c) have been satisﬁed by the searched Q(z).
The sensitivity function before the KYP Lemma based design is also shown in Figure
8.8 where it can be seen that the required speciﬁcations are not met. Moreover,
as seen from Figure 8.9, the gain margin and the phase margin are 12 dB and 67
deg, higher than 7.7 dB and 63 deg before the KYP design. Observed from Figures
8.8 and 8.9, the KYP design increases the loop gain around 9 kHz, which results
in the reduction of the sensitivity function from 8 to 10 kHz. The price for this
compensation is a bit lower loop gain at lower frequencies, which is consistent with
the Bode Integral constraint for sensitivity function.
In the experiment, the dSpace 1103 on TMS320C240 DSP board was used to im
plement the controller, and an LDV was used to measure the actuator displacement,
as shown in Figure 8.10. Channel 2 over Channel 1 of DSA is the measured fre
quency response of the sensitivity function with a swept sine signal as the reference.
Figure 8.11 shows the experimental sensitivity functions, which agree with the sim
ulation results in Figure 8.8. From (8.1), the σ values of the position errors before
and after the KYP Lemmabased control versus frequencies are obtained and shown
in Figure 8.12. It can be seen that the performance with the KYP Lemmabased con
trol is much better from 5 kHz onwards, and slightly worse before 5 kHz, which is
consistent with the sensitivity functions in Figure 8.8.
174 Modeling and Control of Vibration in Mechanical Systems
10
1
10
2
10
3
10
4
−50
−40
−30
−20
−10
0
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
Before KYP design
After KYP design
FIGURE 8.8
Sensitivity functions before and after the KYP lemmabased design: simulation re
sult.
Generalized KYP LemmaBased Loop Shaping Control Design 175
10
1
10
2
10
3
10
4
−20
−10
0
10
20
30
40
50
M
a
g
n
i
t
u
d
e
(
d
B
)
After KYP optimization
Before KYP optimization
10
1
10
2
10
3
10
4
−800
−600
−400
−200
0
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
FIGURE 8.9
Openloop Bode plot before and after the KYP lemmabased design.
FIGURE 8.10
Structure of experimental setup.
176 Modeling and Control of Vibration in Mechanical Systems
10
2
10
3
10
4
−25
−20
−15
−10
−5
0
5
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
Before KYP design
After KYP design
< 0dB
FIGURE 8.11
Sensitivity functions before and after the KYP lemmabased design: experimental
results.
0 2000 4000 6000 8000 10000 12000
0
0.002
0.004
0.006
0.008
0.01
0.012
Frequency(Hz)
P
E
S
σ
(
µ
m
)
Before KYP design
After KYP design
FIGURE 8.12
σ value of PES versus frequency.
Generalized KYP LemmaBased Loop Shaping Control Design 177
8.6 Application in midfrequency vibration rejection
The frequency responses of the microactuator are shown in Figure 8.13. Six reso
nance modes at 3.7, 4.9, 6.9, 9, 12.7 and 15 kHz are included in the model.
The disturbance distribution is reﬂected in the nonrepeatable runout power spec
trum of the measured PES in Figure 8.14. It is noticed that there is a vibration mode
at 650 Hz due to disk vibration. The objective here is to use the above KYP method
to design a linear dynamic output feedback controller C(z) for the microactuator in
Figure 8.13 such that its closedloop system is stable and the disturbance centering at
650 Hz is suppressed sufﬁciently. 45 kHz sampling rate is used in the servo control
design. The control algorithm is implemented with the digital position error signal
generated from DSP TMS320C6711. Currently, due to the limitation by the DSP
speed, with this sampling rate the platform can support up to 10
th
order controller.
Because 650 Hz is at a relative low frequency range, we just involve the static
part of the microactuator represented by a pade delay in the control design with the
KYP Lemma. After that, notch ﬁlters for the resonance modes at 3.7, 9 and 15
kHz will be used to compensate the dynamic part, which will not change a lot the
obtained performance of the low frequency part. The 4.9 and 6.9 kHz resonance
modes, seen in Figure 8.13, have relatively small magnitudes and can be ignored as
long as they are not excited in the control loop. The resonance mode at 12.7 kHz is
not considered in the control design as it is not excited easily and does not affect the
whole loop stability when the 15 kHz mode is compensated.
The pade delay model is given by
P
pade−delay
= −5.6234
s −2 · π · 17000
s + 2 · π · 17000
, (8.39)
which is precompensated by the proportionalintegral (PI) controller:
Int(z) = 0.027(−
z
z − 0.999
+ 0.5). (8.40)
Due to the ﬁrst order padedelay model used in the computation of LMIs, the
computation of controller can be very efﬁcient.
The desired speciﬁcations for the sensitivity function S(z) are set as:
Spec.(a) S(f) < 0 dB, f ≤ 500 Hz,
Spec.(b) S(f) < −10 dB, 610 Hz ≤ f ≤ 670 Hz,
Spec.(c) S(f) < 9.54 dB, f ≥ 19 kHz.
Spec. (b) means to attenuate the disturbances centering at 650 Hz by 10 dB at least.
The parameters of Q(z) in (8.19) with τ = 1 are attained by solving three LMIs of
the form (8.7) corresponding to Spec. (a), (b) and (c). The resultant C(z) is a 10
th
order controller.
For the sake of comparison, the phaselead peak ﬁlter (PLPF) of the form in (8.28)
with values K = 0.4, φ = −0.584, ω
0
= 2π650, and ξ = 0.0632, is also applied to
178 Modeling and Control of Vibration in Mechanical Systems
suppress the low frequency disturbances around 650 Hz, and the sensitivity function
comparison is shown in Figure 8.15. It can be seen that the KYP method achieves
better disturbance rejection from 60 Hz to 1 kHz, although they have almost the
same rejection capability in the very narrow band around 650 Hz. However, the
KYP method gives a poorer disturbance rejection performance for frequency below
60 Hz than the PLPF method.
In the open loop comparison in Figure 8.16, the phase margin (PM) with the PLPF
method is much higher, while the bandwidth is lower and the gain margin (GM) is
comparable with the KYP Lemma method. Consistent with the sensitivity functions
in Figure 8.15, the PES NRRO power spectrum comparison is shown in Figure 8.17
which clearly shows that the KYP based design gives a better disturbance rejection
around 650 Hz than the PLPF although at 650 Hz they offer a similar performance.
From Figure 2.20, it is known that the spectrum of the true PES y is given by
S
y
= P(z)S(z)
2
×d
1

2
+S(z)
2
d
2

2
+ T(z)
2
× n
2
(8.41)
= S
e
− S(z)
2
×n
2
+ T(z)
2
×n
2
, (8.42)
where S
e
is in (8.1), and T(z) = 1 −S(z) is the closed loop transfer function. Thus
the 3σ value of the true PES can be assessed from the power spectrum S
e
in Figure
8.17 with the known level of noise n. As a result it is improved from 6.4 nm with
the PLPF method to 6 nm with the KYP Lemma method.
In the above application, only the ﬁrst order Q(z) is used. A higher order Q(z)
offers more design freedom and has the potential of achieving better results. How
ever, whatever Q(z) is used, the resultant sensitivity function has to comply with the
Bode integral theorem, meaning that it is not possible to achieve disturbance rejec
tion across the entire frequency range. To further improve the disturbance rejection
at low frequency for the KYP Lemmabased design, we shall incorporate a nonlinear
compensation in Chapter 13.
8.7 Conclusion
This chapter has applied the generalized KYP Lemma in the microactuator closed
loop design to suppress the narrow band disturbances. The system design problem
with multiple speciﬁcations on the gain properties of the sensitivity function over
several frequency ranges has been solved by the LMI optimization based on the
KYP Lemma. The Youla parametrization approach has been used in the feedback
controller design. Practical applications have been demonstrated for narrowband
high frequency and midfrequency disturbance rejection. The resultant controller
veriﬁes that the desired speciﬁcations to reject the disturbances have been satisﬁed
via the search for the coefﬁcients of Q(z) in the Youla parametrization approach.
Generalized KYP LemmaBased Loop Shaping Control Design 179
10
2
10
3
10
4
0
10
20
30
40
M
a
g
n
i
t
u
d
e
(
d
B
)
10
2
10
3
10
4
−600
−500
−400
−300
−200
−100
0
100
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
Measured
Modeled
FIGURE 8.13
Frequency response of the PZT microactuator.
180 Modeling and Control of Vibration in Mechanical Systems
0 1000 2000 3000 4000 5000 6000
0
0.5
1
1.5
2
2.5
3
3.5
4
x 10
−3
Frequency(Hz)
N
R
R
O
m
a
g
n
i
t
u
d
e
(
µ
m
)
FIGURE 8.14
PES NRRO power spectrum calculated from measured PES signal without servo
control, reﬂecting the vibration distribution of the system (3σ = 21 nm including the
noise 3σ = 15.2 nm).
Generalized KYP LemmaBased Loop Shaping Control Design 181
10
1
10
2
10
3
10
4
−60
−50
−40
−30
−20
−10
0
10
M
a
g
n
i
t
u
d
e
(
d
B
)
Frequency(Hz)
KYP
PLPF
FIGURE 8.15
Comparison of sensitivity functions.
FIGURE 8.16
Open loop frequency responses (PLPF (GM: 6 dB, PM: 50 deg., Bandwidth
1.4kHz)); KYP(GM: 6 dB, PM: 34 deg., Bandwidth: 1.7 kHz))).
182 Modeling and Control of Vibration in Mechanical Systems
0 1000 2000 3000 4000 5000 6000
0
0.2
0.4
0.6
0.8
1
x 10
−3
Frequency(Hz)
N
R
R
O
m
a
g
n
i
t
u
d
e
(
µ
m
)
KYP
PLPF
650Hz
FIGURE 8.17
NRRO power spectrum with PLPF and KYP (50% reduction before 1 kHz).
9
Combined H
2
and KYP LemmaBased
Control Design
9.1 Introduction
As a closedloop shaping method, the KYP Lemmabased approach allows designers
to impose performance requirements over selected ﬁnite frequency ranges so as to
have the desired sensitivity function that is able to reject the disturbances in these
speciﬁc frequency ranges. Subsequently, the positioning accuracy can be improved
to some extent. However, the KYP Lemmabased loop shaping method does not
count for overall positioning error minimization which can be translated into the
H
2
optimal control problem by taking into consideration the disturbance and noise
models. On the other hand, the H
2
control design which incorporates all disturbance
and noise models can result in an average performance across the entire frequency
range and a high order controller. Thus it usually does not have the ﬂexibility to
speciﬁcally reject disturbances at certain frequency ranges, which however may be
dominant factors that inﬂuence the overall performance. Therefore there is a need to
suppress disturbances of speciﬁc frequencies when minimizing the positioning error.
This motivates us to incorporate the KYP Lemmabased method with the H
2
control
method in this chapter. With the selected speciﬁc disturbances handled by the KYP
Lemmabased design, the H
2
control is formulated with a lower order disturbance
model, excluding the disturbances covered in the above design. This will not only
release the computation burden in the H
2
control design but also result in a lower
order controller.
In this chapter, we will apply the combined control design method to a PZT mi
croactuator such that a disturbance at 650 Hz is rejected with the KYP Lemmabased
design and at the same time overall positioning error is minimized via the H
2
control
design. Then one more disturbance at 2 kHz near the servo bandwidth 1kHz is also
considered as a speciﬁc disturbance to be rejected via the the KYP Lemmabased
design. The design procedure will be illustrated and the resultant controller will be
veriﬁed via an experiment. A series of simulation and experimental results will show
the effectiveness of the control design method in terms of enhancing positioning ac
curacy.
183
184 Modeling and Control of Vibration in Mechanical Systems
9.2 Problem formulation
FIGURE 9.1
H
2
control scheme with Q parametrization for controller design.
In the previous chapter, speciﬁcations on sensitivity function S(z) are described
as
S(f
i
) < r
i
, f
i1
< f
i
< f
i2
, i = 1, 2, · · · , m (9.1)
where r
i
< 1 is a positive scalar, and f
i1
and f
i2
deﬁne the frequency range.
Such an upperbound speciﬁcation as in (9.1) will lead to a problem when the fre
quency f
i
is larger than and especially near the desired bandwidth or 0dB crossover
frequency of S(z). The 0dB crossover frequency of S(z) will be pushed away to
wards a higher frequency, as seen in Figure 9.9, which tends to damage the system
stability and deteriorate the system highfrequency performance. In view of this, a
lowerbound speciﬁcation, i.e.,
S(f
i
) ≥ 1, f
i1
≤ f
i
≤ f
i2
(9.2)
is required. This speciﬁcation helps to ﬁx the bandwidth or 0dBcrossover frequency
of S(z), which will be seen later in the application results.
The problem of the speciﬁc disturbance rejection can be solved by imposing such
performance speciﬁcations in (9.1) and then using the KYP Lemmabased control
design method in Chapter 8. However, as shown in Figure 9.1 associated with Fig
ure 8.14, the servo mechanical system suffers from various kinds of disturbances
and sensing noise. The KYP Lemmabased control design cannot include all dis
turbances and noises which contribute to the position error. In view of this, we also
Combined H
2
and KYP LemmaBased Control Design 185
take into account the overall performance of the servo control system, which is repre
sented as the socalled track misregistration (TMR) induced by w =
w
1
w
2
w
3
T
passing through D
1
(s), D
2
(s), and N(s). It is expressed by the standard deviation
σ
z
of z, and
T
zw
2
= σ
z
, (9.3)
when w is a white noise with zero mean and identity covariance matrix, where T
zw
is the transfer function from w to z.
In the next section, we proceed to the controller design to achieve the speciﬁca
tions in (9.1) and (9.2), and meanwhile to optimize (9.3).
9.3 Controller design for speciﬁc disturbance rejection and over
all error minimization
The generalized KYP Lemmabased design method in Chapter 8 is used to design a
controller for speciﬁc disturbance attenuation.
Let (A
p
, B
p
, C
p
, D
p
) and (A
c
, B
c
, C
c
, D
c
) respectively be the statespace model
of plant P(z) and controller C(z). In order to convexify matrix inequalities, the
Youla parametrization approach with the Q(z) in a FIR ﬁlter form is applied and the
controller structure is shown in Figure 9.1. K(z) is an observer based controller that
can be designed using the LQG method as in (8.11)−(8.12).
For the presentation of the KYP Lemma, we denote
σ(S, Π) :=
S
I
∗
Π
S
I
(9.4)
where S(z) = S(e
jθ
), I stands for an identity matrix and Π a Hermitian matrix of
the form
Π =
Π
11
Π
12
Π
∗
12
Π
22
, (9.5)
which speciﬁes the frequency domain property to be investigated.
9.3.1 Qparametrization to meet speciﬁc speciﬁcations
A. Speciﬁcation (9.1)
Recall from (8.13)−(8.14) that a set of sensitivity functions S(z): (
˜
A,
˜
B,
˜
C,
˜
D)
can be Qparameterized.
According to the denotation (9.4)−(9.5), the speciﬁcation S(z) ≤ r is written as
σ(S, Π) ≤ 0 with
Π =
Π
11
Π
12
Π
∗
12
Π
22
=
1 0
0 − r
2
. (9.6)
186 Modeling and Control of Vibration in Mechanical Systems
Thus based on the KYP Lemma in Chapter 8, achieving
S
e
jθ
≤ r for the
frequency range θ
1
≤ θ ≤ θ
2
can be obtained by solving the following matrix
inequality
˜
A
˜
B
I 0
∗
Σ
˜
A
˜
B
I 0
+
˜
C
˜
D
0 I
∗
Π
˜
C
˜
D
0 I
≤ 0, (9.7)
which is, since Π
11
> 0, equivalent to
˜
A
˜
B
I 0
∗
Σ
˜
A
˜
B
I 0
+
0 0
0 −r
2
˜
C
˜
D
∗
Π
11
Π
11
˜
C
˜
D
−Π
11
≤ 0, (9.8)
where
Σ =
−U e
jθ
c
V
e
−jθ
c
V U −(2 cos θ
d
) V
, (9.9)
θ
c
=
(θ
1
+ θ
2
)
/
2
, θ
d
=
(θ
2
−θ
1
)
/
2
, (9.10)
U and V are Hermitian matrices and V ≥ 0.
To convexify the matrix inequality (9.8), we shall give a state space realization
of S(z) = T
11
(z) + T
12
(z) Q(z) T
21
(z). Denote the statespace representation
of T
11
(z) and T
12
(z)T
21
(z) by (A
t11
, B
t11
, C
t11
, D
t11
) and (A
t
, B
t
, C
t
, D
t
),
respectively. Then a statespace model of S(z) can be written as (8.17)−(8.18).
B. Speciﬁcation (9.2)
Again, according to the denotation (9.4)−(9.5) the speciﬁcation S(z) ≥ r is
equivalent to σ(S, Π) ≤ 0 with
Π =
Π
11
Π
12
Π
∗
12
Π
22
=
−1 0
0 r
2
. (9.11)
However, because Π
11
< 0, (9.7) can not be converted equivalently to (9.8), which
means (9.7) is not possibly convexiﬁed according to the method in Section 9.3.1.
Hence we resort to the following speciﬁcation
σ (S, Π) = aR(S) +bI (S) + c, Π :=
0 a +jb
a − jb 2c
(9.12)
where R and I denote the real and the imaginary parts of S(e
jθ
). When a, b and
c are properly selected, S(z) ≥ r can be achieved. A simple selection is a = 0,
b = −1, c = r, and
σ (S, Π) = −I (S) + r. (9.13)
Thus σ(S, Π) ≤ 0 means I (S) ≥ r, and subsequently S(z) ≥ r. In this situation,
Π =
Π
11
Π
12
Π
∗
12
Π
22
=
0 −j
j 2r
, (9.14)
Combined H
2
and KYP LemmaBased Control Design 187
where Π
11
= 0 and (9.7) is equivalent to
˜
A
˜
B
I 0
∗
Σ
˜
A
˜
B
I 0
+
0
˜
C
∗
Π
12
Π
∗
12
˜
C
˜
D
∗
Π
12
+ Π
∗
12
˜
D + Π
22
≤ 0, (9.15)
which is a linear matrix inequality with unknown variables in
˜
C and
˜
D only, and can
be solved using the same method as in Section 9.3.1A.
It should be mentioned that R(S) ≥ r can also be used to achieve S(z) ≥ r, if
it is suitable for a speciﬁc application. In this case,
Π =
Π
11
Π
12
Π
∗
12
Π
22
=
0 −1
−1 2r
, (9.16)
and the linear matrix inequality (9.15) remains applicable.
9.3.2 Qparametrization to minimize H
2
performance
Next we focus on the design of Q(z) to minimize the H
2
normT
zw
2
. From Figure
9.1 we have
−z = N (z) w
3
+S (z) [P (z) D
1
(z) w
1
+ D
2
(z) w
2
−N (z) w
3
] . (9.17)
Denote a statespace realization of P(z)D
1
(z), D
2
(z) and N(z) by (A
1
, B
1
, C
1
, D
1
),
(A
2
, B
2
, C
2
, D
2
), and (A
3
, B
3
, C
3
, D
3
), respectively. It follows from (8.13) and
(8.17)−(8.18) that
x (k + 1) =
¯
Ax (k) +
¯
Bw(k), (9.18)
−z (k) =
¯
Cx (k) +
¯
Dw(k), (9.19)
where,
¯
A =
A
1
0 0 0
0 A
2
0 0
0 0 A
3
0
˜
BC
1
˜
BC
2
−
˜
BC
3
˜
A
,
¯
B =
B
1
0 0
0 B
2
0
0 0 B
3
˜
BD
1
˜
BD
2
−
˜
BD
3
, (9.20)
¯
C =
˜
DC
1
˜
DC
2
−
˜
DC
3
+C
3
˜
C
,
¯
D =
˜
DD
1
˜
DD
2
−
˜
DD
3
+ D
3
.
The H
2
norm T
zw
2
can be minimized as follows:
min
(Ξ=Ξ
T
>0, Ω=Ω
T
>0)
Trace (Ω) (9.21)
subject to
¯
A
T
Ξ
¯
A−Ξ +
¯
C
T
¯
C < 0 (9.22)
¯
B
T
Ξ
¯
B +
¯
D
T
¯
D < Ω (9.23)
188 Modeling and Control of Vibration in Mechanical Systems
or equivalently,
¯
A
T
Ξ
¯
A− Ξ
¯
C
T
¯
C −I
< 0 (9.24)
−Ω +
¯
B
T
Ξ
¯
B
¯
D
T
¯
D −I
< 0 (9.25)
where
¯
A =
A
1
0 0 0 0 0
0 A
2
0 0 0 0
0 0 A
3
0 0 0
B
t11
C
1
B
t11
C
2
−B
t11
C
3
A
t11
0 0
B
t
C
1
B
t
C
2
−B
t
C
3
0 A
t
0
B
q
D
t
C
1
B
q
D
t
C −B
q
D
t
C
3
0 B
q
C
t
A
q
,
¯
B =
B
1
0 0
0 B
2
0
0 0 B
3
B
t11
D
1
B
t11
D
2
−B
t11
D
3
B
t
D
1
B
t
D
2
−B
t
D
3
B
q
D
t
D
1
B
q
D
t
D
2
−B
q
D
t
D
3
,
¯
C =
(D
t11
+ D
q
D
t
) C
1
(D
t11
+D
q
D
t
) C
2
(D
t11
+ D
q
D
t
) C
3
C
t11
D
q
C
t
C
q
,
¯
D =
(D
t11
+ D
q
D
t
) D
1
(D
t11
+ D
q
D
t
) D
2
−(D
t11
+ D
q
D
t
) D
3
+D
3
] . (9.26)
Note that the Q(z) coefﬁcients q
i
(i = 0, 1, . . . , τ) only appear in C
q
and D
q
.
Therefore, from (8.17)−(8.18) and (9.26), we know that q
i
exists only in
˜
C,
˜
D,
¯
C
and
¯
D. In this case, (9.8), (9.15) and (9.24)−(9.25) deﬁne LMIs in terms of the
variables U, V , Ξ, Ω and q
i
. Hence, the Q(z) coefﬁcients q
i
can be computed via a
convex optimization.
With the solved Q(z): (A
q
, B
q
, C
q
, D
q
), the controller C(z) is then given by
A
c
=
A
p
−B
p
M + LC
p
+B
p
D
q
C
p
B
p
C
q
B
q
C
p
A
q
,
B
c
=
L + B
p
D
q
B
q
,
C
c
=
−M +D
q
C
p
C
q
,
D
c
= D
q
.
(9.27)
9.3.3 Design steps
To summarize, a design procedure for controller C(z) is given as follows.
Combined H
2
and KYP LemmaBased Control Design 189
Step 1. Design K(z) from (8.11)−(8.12).
Step 2. Compute T
11
(z), T
12
(z) and T
21
(z) from(8.14), and obtain the state space
model (
˜
A,
˜
B,
˜
C,
˜
D) in (8.17)−(8.18).
Step 3. Based on disturbance spectrum and bandwidth requirement, specify the
positive scalars r
i
and r
j
, and the frequency points f
i
(i = 1, . . . , m) and f
j
(j = 1,
. . . , n) for the sensitivity function S(z), i.e.,
S (f
i
) < r
i
, f
i1
≤ f
i
≤ f
i2
, (9.28)
and
S (f
j
) > r
j
, f
j1
≤ f
j
≤ f
j2
. (9.29)
For each speciﬁcation, construct the LMIs (9.8) and (9.15) in terms of the variables
U, V , C
q
and D
q
.
Step 4. Construct the LMIs (9.24)−(9.25) in terms of the variables Ξ, Ω, C
q
and
D
q
.
Step 5. Obtain Q(z) : (A
q
, B
q
, C
q
, D
q
) by solving the above LMIs using the
MATLAB LMI toolbox.
Step 6. Obtain the controller C(z) from (9.27).
9.4 Simulation and implementation results
This section will apply the control design method in Section 9.3 for a PZT microac
tuator to separately reject one or two speciﬁc disturbances and meanwhile minimize
the H
2
norm of the PES.
9.4.1 System models
The frequency response of the PZT microactuator, shown in Figure 9.2, was obtained
using a LDV and a DSA. The main resonance modes of the plant are at frequencies
6.5 kHz, 9.5 kHz, 11.3 kHz, and 20 kHz. The identiﬁed plant model of the micro
actuator P(s) has the following parameters:
Zeros = 10
5
× [−0.0296 ± 0.4927j, − 0.0110 ± 1.0964j,
− 0.7116 ± 0.6275j, −0.0093 ±0.6198j, 0.8168],
Poles = 10
5
×[−0.0255 ±1.2733j, −0.0050 ±0.7100j,
− 0.0048 ± 0.5981j, −0.0245 ±0.4071j, −0.8168],
Gain = −0.4819.
The frequency response of the plant model is plotted against the measured data in
Figure 9.2 for comparison and it is subsequently discretized in MATLAB using the
“zoh” method with a sampling rate of 40 kHz.
190 Modeling and Control of Vibration in Mechanical Systems
10
2
10
3
10
4
−40
−20
0
20
40
Frequency (Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
10
2
10
3
10
4
−250
−200
−150
−100
−50
0
50
Frequency (Hz)
P
h
a
s
e
(
d
e
g
)
Measured
Modeled
FIGURE 9.2
Frequency response of a PZT microactuator.
9.4.2 Rejection of speciﬁc disturbance and H
2
performance minimiza
tion
Consider the disturbances in Figure 8.14; the disturbance around 650 Hz is due to
the disk vibration. A suitable feedback controller, C(z) has to be designed for the
system so that the overall system is stable and the disturbance around 650 Hz is
suppressed sufﬁciently, while ensuring that the H
2
norm of the position error signal
is minimized. Hence, the desired speciﬁcation of the sensitivity function S(z) is
S (f) < −10 dB for 610 Hz ≤ f ≤ 670 Hz and at the same time the position error
signal is to be minimized.
The parameters of a ﬁrstorder FIR Q(z) are obtained by solving the three LMIs
(9.8), (9.24) and (9.25). For comparison, another controller is designed without H
2
minimization and just to suppress the vibration around 650 Hz. With each of the
two designed controllers, the frequency response of the openloop C(z) ×P(z) and
the sensitivity function S(z) are depicted in Figure 9.3 and Figure 9.4. It is seen
that the hump of S(z) is reduced to about 3 dB with the controller designed by
the combined method, i.e., the combined H
2
optimization and speciﬁc disturbance
rejection method. On the other hand, the performance speciﬁcations listed in Table
9.1 show that the proposed method offers better stability margins although the open
loop crossover frequency is a bit lower.
Experiments are carried out for the KYP Lemmabased controller and the KYP+H
2
controller to verify the simulation results. The controllers are implemented using
dSPACE 1103 on a TMS320C240 DSP board and the structure of the experimental
setup is the same as in Figure 8.10. The sensitivity function of the system is obtained
by the ratio of the measurements in Channel 2 and Channel 1 of the DSA, where a
swept sine signal is the reference input. The sensitivity function obtained from the
Combined H
2
and KYP LemmaBased Control Design 191
TABLE 9.1
Comparison of performance speciﬁcations
Method KYP KYP+H
2
Crossover frequency (kHz) 1.85 1.78
Gain margin (dB) 7.9 12.6
Phase margin (deg) 35.4 45.9
experiment, as shown in Figure 9.5, demonstrates the effectiveness of the proposed
method. As a result, as seen in Figure 9.6, the position error is reduced and its σ
value has a 4% improvement.
10
1
10
2
10
3
10
4
−50
0
50
M
a
g
n
i
t
u
d
e
(
d
B
)
10
1
10
2
10
3
10
4
−250
−200
−150
−100
−50
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
KYP+H
2
KYP
Frequency(Hz)
FIGURE 9.3
Openloop frequency responses.
192 Modeling and Control of Vibration in Mechanical Systems
10
1
10
2
10
3
10
4
−50
−40
−30
−20
−10
0
10
M
a
g
n
i
t
u
d
e
(
d
B
)
Frequency(Hz)
KYP+H
2
KYP
FIGURE 9.4
Designed sensitivity functions.
10
2
10
3
10
4
−35
−30
−25
−20
−15
−10
−5
0
5
10
Frequency (Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
KYP+H
2
KYP
FIGURE 9.5
Comparison of sensitivity functions obtained from experiment.
Combined H
2
and KYP LemmaBased Control Design 193
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
x 10
4
0
0.05
0.10
0.15
0.20
0.25
0.30
0.35
Frequency (Hz)
N
R
R
O
m
a
g
n
i
t
u
d
e
(
n
m
)
KYP+H
2
KYP
FIGURE 9.6
NRRO power spectrum with KYP Lemmabased controller with and without H
2
minimization.
9.4.3 Rejection of two disturbances with H
2
performance minimiza
tion
In what follows, we include one more speciﬁc disturbance rejection at 2 kHz (see
Figure 9.7), which is caused by air ﬂow and is near the desired servo bandwidth 1
kHz. In this case, the speciﬁcations on S(z) are set as:
i. S (f) < −10 dB for 610 Hz ≤ f ≤ 670 Hz;
ii. S (f) < −5 dB for 1950 Hz ≤ f ≤ 2050 Hz;
iii. S (f) > 0 dB for 990 Hz ≤ f ≤ 1010 Hz,
where in addition to the rejection of disturbance at 650 Hz, the disturbance around
2 kHz is also to be suppressed. Note that the third speciﬁcation aims to ﬁx the
bandwidth.
By solving the LMIs (9.8), (9.15) and (9.24)−(9.25) in terms of the variables
P, V , Ξ, Ω, C
q
and D
q
, a controller C(z) which leads to the S(z) satisfying the
speciﬁcations can be obtained with a fourthorder FIR Q(z). With the resultant
controller, the openloop frequency response and the sensitivity function are shown
in Figures 9.8 and 9.9, where the speciﬁcations (i), (ii) and (iii) have been achieved
and the closedloop system remains stable. The openloop gain and phase margins
are 14.6 dB and 49.5 degrees, respectively.
We have also carried out experiments to verify the designed controller. The ex
perimental results of the openloop frequency response and sensitivity function are
194 Modeling and Control of Vibration in Mechanical Systems
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
0
0.5
1
1.5
2
2.5
3
x 10
−3
Frequency (Hz)
N
R
R
O
m
a
g
n
i
t
u
d
e
(
µ
m
)
FIGURE 9.7
PES NRRO spectrum without servo control.
plotted in Figure 9.8 and Figure 9.10, respectively. The NRRO spectrum of the sys
tem with the designed controller is shown in Figure 9.11. Based on Figure 9.11, the
attenuation of disturbance centering at 2 kHz has improved by 35%.
9.5 Conclusion
This chapter has addressed a combined control design method that incorporates the
generalized KYP Lemmabased design and the H
2
optimization. With the incorpo
rated control design method, speciﬁc narrowband disturbances have been attenuated
and simultaneously the positioning error of the control system has been minimized.
The method has been applied to design a controller for a PZT microactuator to at
tenuate disturbances at 650 Hz and 2 kHz where disk vibrations are dominant and
minimize the 3σ value of the overall PES. Simulation and experimental results have
demonstrated the effectiveness of the proposed method and veriﬁed that the position
ing accuracy has been improved.
Combined H
2
and KYP LemmaBased Control Design 195
10
2
10
3
10
4
−60
−40
−20
0
20
40
M
a
g
n
i
t
u
d
e
(
d
B
)
10
2
10
3
10
4
−800
−600
−400
−200
0
200
Frequency (Hz)
P
h
a
s
e
(
d
e
g
)
Simulation
Experiment
FIGURE 9.8
Openloop frequency response.
10
1
10
2
10
3
10
4
−50
−40
−30
−20
−10
0
10
M
a
g
n
i
t
u
d
e
(
d
B
)
Frequency (Hz)
Bandwidth is
increased
FIGURE 9.9
Resultant sensitivity function (Solid line: with Spec. (i), (ii) and (iii); Dashed line:
with Spec. (i) and (ii)).
196 Modeling and Control of Vibration in Mechanical Systems
10
2
10
3
10
4
−30
−25
−20
−15
−10
−5
0
5
10
Frequency (Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
Simulation
Experiment
FIGURE 9.10
Resultant sensitivity function with all the three requirements fulﬁlled.
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x 10
−3
Frequency (Hz)
N
R
R
O
m
a
g
n
i
t
u
d
e
(
µ
m
)
Without control
Closed loop
FIGURE 9.11
NRROpower spectrum with rejection of two speciﬁc disturbances at 0.65 and 2 kHz.
10
Blending Control for MultiFrequency
Disturbance Rejection
10.1 Introduction
The blending control technique [120] aims to design a controller capable of simul
taneously coping with different optimality criteria deﬁned for various inputoutput
channels. If several targeted transfer functions are achieved respectively by their
designed controllers, it is shown that under some mild assumptions there exists a
unique controller capable of replicating these transfer functions and thus simultane
ously achieving the performance attained by each individual controller. Each indi
vidual controller can be a static or a dynamic feedback controller derived by means
of standard optimal and robust control methods such as the LQG/LTR control, H
2
or
H
∞
control. The blended controller can be easily computed by a procedure based on
simple linear algebra. In this chapter, we shall apply the blending control idea to deal
with the problem of rejecting several narrowband disturbances. The multifrequency
disturbance rejection is formulated as a control blending problem. Control design for
each disturbance rejection is carried out by using the H
2
optimal control method. The
ultimate controller is obtained by blending all these H
2
controllers and is expected
to be able to reject all these disturbances simultaneously. In this chapter, the con
trol blending is respectively applied in the control design for a 1.8inch HDD VCM
actuator in three cases, (1) reject two disturbances of frequencies higher than band
width, (2) reject three disturbances with frequencies higher than bandwidth, and, (3)
reject one disturbance with frequency lower than bandwidth and another one higher
than bandwidth. Note that this chapter presents a systematic approach using dis
turbance models only, which leads to the achievement of a satisfactory rejection of
multifrequency narrow band disturbances.
10.2 Control blending
Control blending accounts for the problem of simultaneous performance achieve
ments. It involves designing an individual controller for each performance speciﬁca
197
198 Modeling and Control of Vibration in Mechanical Systems
tion and blending all the controllers to obtain an ultimate controller.
Consider the system in state space:
˙ x(t) = Ax(t) + Bu(t) +
r
i=1
E
i
w
i
(t), (10.1)
y
1
(t) = C
1
x(t) +D
1
u(t), (10.2)
y
2
(t) = C
2
x(t) +D
2
u(t), (10.3)
· · ·
y
r
(t) = C
r
x(t) + D
r
u(t), (10.4)
e(k) = C
e
x(k) +D
e
u(k) +v(k), (10.5)
where x(t) ∈ R
n
x
, u(t) ∈ R
n
u
, y
i
(t) ∈ R
p
i
, w
i
(t) ∈ R
q
i
, e(t) ∈ R, and A, B, C
i
,
D
i
, E
i
, C
e
, and D
e
are matrices of appropriate dimensions.
The following two assumptions are necessary in the control blending scheme.
Assumption 10.1: (A, B) is stabilizable.
Assumption 10.2: The n
x
×s matrix,
E =
E
1
E
2
· · · E
r
, (10.6)
where s =
r
i=1
q
i
, has full column rank.
FIGURE 10.1
Blending control scheme.
Blending Control for MultiFrequency Disturbance Rejection 199
10.2.1 State feedback control blending
The feedback control scheme of the above system is shown in Figure 10.1, where the
controller is expressed as
˙ x
s
(t) = A
s
x
s
(t) +B
s
x(t) (10.7)
u(t) = C
s
x
s
(t) + D
s
x(t). (10.8)
The closedloop system from w
i
to y
i
, i = 1, 2, · · · , r is then described by
A
cl
E
cl
C
cl
0
=
A +BD
s
BC
s
E
1
E
2
· · · E
r
B
s
A
s
0 0 · · · 0
C
1
+ D
1
D
s
D
1
C
s
0 0 · · · 0
C
2
+ D
2
D
s
D
2
C
s
0 0 · · · 0
.
.
.
.
.
.
.
.
.
.
.
. · · ·
.
.
.
C
r
+ D
r
D
s
D
r
C
s
0 0 · · · 0
. (10.9)
For the system (10.1)−(10.5), we consider the static state feedback control
u(t) = K
i
x(t), i = 1, 2, · · · , r. (10.10)
The individual closedloop transfer function is given by
¯
T
i
(s) =
¯
C
i
(sI −
¯
A
i
)
−1
E
i
, (10.11)
with
¯
A
i
= A +BK
i
,
¯
C
i
= C
i
+D
i
K
i
. (10.12)
The result in [120] reads that there exists a single dynamic compensator of the
form (10.7)−(10.8) of order µ = n
x
(r − 1), such that 1). the closedloop system is
stable with spectrum
σ(A
cl
) = ∪
i
σ(
¯
A
i
); (10.13)
and 2). for each i, the transfer function T
i
(s) from the i
th
input to the i
th
output
coincides with the one given by the i
th
gain K
i
, i.e.,
T
i
(s) =
¯
T
i
(s). (10.14)
The signiﬁcance of the result lies in the fact that if K
i
, i = 1, 2, · · · , r, are the
static compensators each of which optimizes some performance criteria, then there
exists a single dynamic compensator which simultaneously achieves the same per
formances. The single dynamic compensator of the form in (10.7)−(10.8) can be
found following the procedure below.
Procedure 10.1
200 Modeling and Control of Vibration in Mechanical Systems
• Let K
i
, i = 1, 2, · · · , r be the static state feedback controllers each of which
asymptotically stabilizes the plant, i.e.,
¯
A
i
= A + BK
i
is asymptotically
stable.
• For each i, choose a matrix
˜
E
i
of dimension n
x
×(n
x
−q
i
) such that [E
i
˜
E
i
]
is square and invertible. This is possible under Assumption 10.2.
• Choose matrices
˜
Z
i
of dimension µ×(n
x
−q
i
) (µ = n
x
(r −1)) such that the
matrix
E
1
E
2
· · · E
r
˜
E
1
˜
E
2
· · ·
˜
E
r
0 0 · · · 0
˜
Z
1
˜
Z
2
· · ·
˜
Z
r
(10.15)
is invertible.
• Deﬁne
Z
i
= [0
˜
Z
i
][E
i
˜
E
i
]
−1
, (10.16)
and
V
i
= Z
i
¯
A
i
. (10.17)
• Form the square matrix Ξ as
Ξ =
I I · · · I
Z
1
Z
2
· · · Z
r
. (10.18)
• Compute the controller matrices A
s
, B
s
, C
s
, D
s
as
D
s
C
s
B
s
A
s
=
K
1
K
2
· · · K
r
V
1
V
2
· · · V
r
Ξ
−1
. (10.19)
The order of the blended controller is n
x
(r −1) which is very high. Theoretically
to match r different transfer functions T
i
, n
x
(r − 1) + n
x
= n
x
r different poles
have to be allocated and a full order controller is required. In practical application,
however, reducing the order is needed as long as it is possible.
10.2.2 Output feedback control blending
When the full state variables of the plant are not available, the feedback control has
to rely on some measurements given by
y
p
(t) = C
y
x(t), (10.20)
where y
p
∈ R
p
. The output feedback controller takes the form
˙ x
c
(t) = A
c
x
c
(t) +B
c
y(t) (10.21)
u(t) = C
c
x
c
(t) + D
c
y(t). (10.22)
Blending Control for MultiFrequency Disturbance Rejection 201
The closedloop system from w
i
to y
i
, i = 1, 2, · · · , r is then described by
A
cl
B
cl
C
cl
0
=
A+ BD
c
C
y
BC
c
E
1
E
2
· · · E
r
B
c
C
y
A
c
0 0 · · · 0
C
1
+D
1
D
c
C
y
D
1
C
c
0 0 · · · 0
C
2
+D
2
D
c
C
y
D
2
C
c
0 0 · · · 0
.
.
.
.
.
.
.
.
.
.
.
. · · ·
.
.
.
C
r
+D
r
D
c
C
y
D
r
C
c
0 0 · · · 0
. (10.23)
Assumption 10.3: The composed p × s matrix,
E =
C
y
E
1
C
y
E
2
· · · C
y
E
r
, (10.24)
where s =
r
i=1
q
i
, has full column rank.
Consider the static state feedback control as stabilizing compensators
u(t) = K
i
x(t), i = 1, 2, · · · , r. (10.25)
The corresponding closedloop transfer function from w
i
to y
i
is given by
¯
T
i
(s) =
¯
C
i
(sI −
¯
A
i
)
−1
E
i
, (10.26)
with
¯
A
i
= A+BK
i
,
¯
C
i
= C
i
+ D
i
K
i
. (10.27)
The theorem on output blending control in [120] says that there exists a single
dynamic compensator of the form (10.21)−(10.22) of order µ = n
x
r − p, such that
1). the spectrum of the closedloop system satisﬁes
∪
i
σ(
¯
A
i
) ⊆ σ(A
cl
); (10.28)
and 2). for each i, the transfer function T
i
(s) from the i
th
input w
i
to the i
th
output
y
i
coincides with the one given by the i
th
gain K
i
, i.e.,
T
i
(s) =
¯
T
i
(s). (10.29)
The single dynamic output controller of the form in (10.21)−(10.22) can be found
following the procedure below.
Procedure 10.2
• Let K
i
, i = 1, 2, · · · , r be the static state feedback controllers each of which
asymptotically stabilizes the plant, i.e., the matrix
¯
A
i
= A + BK
i
is asymp
totically stable.
• For each i, choose a matrix
˜
E
i
of dimension n ×(n −q
i
) such that [E
i
˜
E
i
] is
square and invertible. This is possible under Assumption 10.3.
202 Modeling and Control of Vibration in Mechanical Systems
• Choose matrices
˜
Z
i
of dimension µ × (n
x
− q
i
) (µ = n
x
r − p) such that the
matrix
C
y
E
1
C
y
E
2
· · · C
y
E
r
C
y
˜
E
1
C
y
˜
E
2
· · · C
y
˜
E
r
0 0 · · · 0
˜
Z
1
˜
Z
2
· · ·
˜
Z
r
(10.30)
is invertible.
• Deﬁne
Z
i
= [0
˜
Z
i
][E
i
˜
E
i
]
−1
, (10.31)
and
V
i
= Z
i
¯
A
i
. (10.32)
• Form the square matrix Ξ as
Ξ =
C
y
C
y
· · · C
y
Z
1
Z
2
· · · Z
r
. (10.33)
• Compute the controller matrices A
c
, B
c
, C
c
, D
c
as
D
c
C
c
B
c
A
c
=
K
1
K
2
· · · K
r
V
1
V
2
· · · V
r
Ξ
−1
. (10.34)
The procedure is quite simple to implement, and it is proved in [120] that the
matrix Ξ is indeed an invertible matrix.
Consider the system
¯
A
¯
E
¯
N 0
=
¯
A
1
0 · · · 0 E
1
0 · · · 0
0
¯
A
2
· · · 0 0 E
2
· · · 0
.
.
.
.
.
. · · ·
.
.
.
.
.
.
.
.
. · · ·
.
.
.
0 0 · · ·
¯
A
r
0 0 · · · E
r
C
y
C
y
· · · C
y
0 0 · · · 0
. (10.35)
The following Proposition advises on the stability of the controller matrix A
c
.
Proposition 10.1 [120]
i). The controller matrix A
c
is not stable if system (10.35) has unstable invariant
zeros.
ii). The controller matrix can be generically chosen as a stability matrix if system
(10.35) is of minimum phase.
iii). If E is square then the controller matrix is stable if and only if system (10.35)
is of minimum phase.
Blending Control for MultiFrequency Disturbance Rejection 203
10.3 Control blending applicationin multifrequency disturbance
rejection
In this section, we ﬁrst formulate the problem of multifrequency disturbance rejec
tion into the above control blending framework. Then, Procedure 10.1 is used to
design a blended controller associated with the static state feedback controller de
signed via the H
2
optimal control method. To facilitate the single dynamic state
feedback controller, a full state observer is required.
10.3.1 Problem formulation
The closed control loop of a mechanical actuation system is shown in Figure 10.2,
where P(s) is the actuator model, C(z) is the feedback controller, e is the measured
error signal. D
i
(s), i = 1, 2, · · · , r are disturbance models and w
i
are white noises
with unity variances. v is the measurement noise with σ
v
as its standard deviation.
Here, we focus on the rejection of output disturbance D
i
(s), considering that an
input disturbance can be converted to an output disturbance.
From Figure 10.2, with Ref = 0 we have
y = S(z)(D
1
(z)w
1
+D
2
(z)w
2
+ · · · + D
r
(z)w
r
) + (1 −S(z))v. (10.36)
The multifrequency disturbance rejection problem is stated as follows: For nar
rowband disturbances d
i
, i = 1, 2, · · · , r, design a feedback controller C(z) such
that the closedloop system is stable, and the disturbances d
i
in different frequency
ranges [f
iL
, f
iH
] can be suppressed simultaneously.
In Figure 10.2, denote (A
p
, B
p
, C
p
, D
p
) as the statespace realizations of P(z)
with state vector x
p
∈R
n
p
, and (A
d
i
, B
d
i
, C
d
i
, D
d
i
) as the statespace realizations
of D
i
(z) with state vector x
d
i
∈R
n
d
i
. D
i
(s) is proposed to have the form:
D
i
(s) =
k
di
(s + 2ζ
i
ω
i
)
s
2
+ 2ξ
i
ω
i
s +ω
2
i
, i = 1, 2, · · · , r, (10.37)
with frequency ω
i
, damping ratio ξ
i
, ζ
i
used to adjust the damping of D
i
, and gain
k
di
. Clearly, D
d
i
= 0.
To ensure necessary rejection of lowfrequency disturbances, a precompensation
integrator Int(z) is introduced which is shown in Figure 10.3 and given by
˜ u(k + 1) =
˜
ke(k) + ˜ u(k). (10.38)
Denote the output due to d
i
by y
i
. Then the combined system is given as follows.
x(k + 1) = Ax(k) +Bu(k) +
r
i=1
E
i
w
i
(k), (10.39)
y
1
(k) = C
1
x(k) +D
1
u(k), (10.40)
204 Modeling and Control of Vibration in Mechanical Systems
FIGURE 10.2
Control loop with injected disturbances at different frequencies.
FIGURE 10.3
Control structure.
Blending Control for MultiFrequency Disturbance Rejection 205
y
2
(k) = C
2
x(k) +D
2
u(k), (10.41)
· · ·
y
r
(k) = C
r
x(k) + D
r
u(k), (10.42)
e(k) = C
e
x(k) + D
e
u(k) + v(k), (10.43)
where x =
x
T
p
˜ u
T
x
T
d
1
x
T
d
2
· · · x
T
d
r
T
∈ R
n
x
, u(k) ∈ R
n
u
, y
i
(k) ∈ R
p
i
, w
i
(k) ∈
R
q
i
, n
x
= n
p
+ 1 +
r
i=1
n
d
i
, n
u
= 1, p
i
= 1, q
i
= 1,
A =
A
p
B
p
0 0 · · · 0
−
˜
kC
p
1 −
˜
kD
p
˜
kC
d
1
˜
kC
d
2
· · ·
˜
kC
d
r
0 0 A
d
1
0 · · · 0
0 0 0 A
d
2
· · · 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 0 · · · A
d
r
, B =
B
p
−
˜
kD
p
0
0
.
.
.
0
, (10.44)
E
1
=
0
−
˜
kD
d
1
B
d
1
0
.
.
.
0
, E
2
=
0
−
˜
kD
d
2
0
B
d
2
.
.
.
0
, · · · , E
r
=
0
−
˜
kD
d
r
0
0
.
.
.
B
d
r
, (10.45)
C
1
=
−C
p
−D
p
−C
d
1
0 · · · 0
, D
1
= −D
p
, (10.46)
C
2
=
−C
p
−D
p
0 −C
d
2
· · · 0
, D
2
= −D
p
, (10.47)
C
r
=
−C
p
−D
p
0 0 · · · −C
d
r
, D
r
= −D
p
, (10.48)
C
e
=
−C
p
−D
p
−C
d
1
−C
d
2
· · · −C
d
r
, D
e
= −D
p
. (10.49)
Note that the measurement error signal e in (10.43) includes all disturbance out
puts y
i
, i = 1, 2, · · · , r.
10.3.2 Controller design via the control blending technique
In this section, we present how to design the controller C(z) by using the control
blending technique. Before the blending, the H
2
optimal control method will be
used to design a static state feedback controller to minimize the error caused by each
disturbance. Subsequently, the blending technique will be utilized to yield one single
dynamic state feedback controller that is able to reject all disturbances. Next a state
observer will be designed to facilitate the state feedback controller with the mea
surement signal. Finally, the SISO (single input and single output) output feedback
controller C
b
(z) will be given based on the dynamic state feedback controller and
the state observer.
Denote the dynamic state feedback controller by
Σ(z) : x
s
(k + 1) = A
s
x
s
(k) + B
s
x(k), (10.50)
206 Modeling and Control of Vibration in Mechanical Systems
u(k) = C
s
x
s
(k) +D
s
x(k). (10.51)
This dynamic state feedback controller is designed following Procedure 10.1,
where K
i
in step 1 is designed based on the H
2
optimal control method, which
is described as follows.
Let K
i
, i = 1, 2, · · · , r be the static state feedback controllers each of which
asymptotically stabilizes the plant and minimizes the H
2
norm of the transfer func
tion from w
i
to y
i
. K
i
can therefore be obtained by the following optimization:
min Trace(S
i
)
subject to the LMIs
−P
i
∗ ∗
P
i
A
T
+ W
T
i
B
T
−P
i
∗
0 C
i
P + D
i
W
i
−1
< 0, (10.52)
−S
i
∗
E
i
−P
i
< 0. (10.53)
With the solved W
i
and P
i
, K
i
is obtained as
K
i
= W
i
P
−1
i
. (10.54)
Next, a dynamic state feedback controller of the form(10.50)−(10.51) is designed
based on Procedure 10.1. Since only the measurement is available for feedback, a
state observer, as in Figure 10.3, is needed to facilitate the controller Σ(z). The state
observer O(z) is designed as follows.
ˆ x(k + 1) = A
o
ˆ x(k) + [B
o1
B
o2
]
u(k)
e(k)
, (10.55)
y
o
(k) = C
o
ˆ x(k) +D
o
u(k)
e(k)
, (10.56)
where
A
o
= A− ALC
e
, B
o1
= B, B
o2
= AL, (10.57)
C
o
= I, D
o
= 0, (10.58)
L = dlqe(A, E, C
e
, Q, R), E is as in (10.6), Q = I, and R = σ
2
v
.
The overall controller C(z) is then given by
A
c
=
A
s
B
s
C
o
B
o1
C
s
A
o
+B
o1
D
s
C
o
, B
c
=
0
B
o2
, (10.59)
C
c
=
C
s
D
s
C
o
, D
c
= 0. (10.60)
Blending Control for MultiFrequency Disturbance Rejection 207
10.4 Simulation and experimental results
The disk drive under consideration is a 1.8inch small hard disk drive with a spindle
motor rotational speed of 4200 RPM. Figure 10.4 shows the frequency response of
the VCM actuator. The VCM actuator model P(s) is described by
P(s) =
8.326628 × 10
17
(s
2
+ 1081s + 7.3 × 10
8
)(s
2
+ 6635s + 2.852 ×10
9
)
(s
2
+ 552.9s + 4.777 × 10
5
)(s
2
+ 1056s + 6.964 × 10
8
)
×
1
(s
2
+ 2815s + 2.527 × 10
9
)(s
2
+ 1.131 ×10
4
s + 3.948 × 10
9
)
. (10.61)
10
1
10
2
10
3
10
4
−60
−40
−20
0
20
40
60
G
a
i
n
[
d
B
]
Measured
Modeled
10
1
10
2
10
3
10
4
−600
−500
−400
−300
−200
−100
0
Frequency [Hz]
P
h
a
s
e
[
d
e
g
r
e
e
]
FIGURE 10.4
Frequency response of the VCM actuator.
10.4.1 Rejecting highfrequency disturbances
In this section, controller C(z) will be designed to reject a few disturbances with
frequencies higher than bandwidth. First, two disturbances will be rejected, and
next, three disturbances are to be attenuated simultaneously.
208 Modeling and Control of Vibration in Mechanical Systems
We consider two disturbances at frequencies 4 kHz and 8 kHz which are the res
onance frequencies of the plant as shown in Figure 10.4. By ﬁtting the model in
(10.37), we have ω
1
= 2π4000, ξ
1
= 0.085, ζ
1
= 0.6, k
d1
= 20 for D
1
(s), and
ω
2
= 2π8000, ξ
2
= 0.01, ζ
2
= 0.6, k
d2
= 43 for D
2
(s).
Take k
i
= 100 · T
s
. The parameters in (10.19) of Σ(z) are obtained by the design
procedure outlined in the last section. With Σ(z), the observer O(z) (10.55)−(10.56),
we have controller C
b
(z) by (10.59)−(10.60). Incorporated with an integrator, a 27th
order C(z) is obtained. After simple order reduction by canceling close zeros and
poles, C(z) can be reduced to 14th order. The ﬁnal controller C(z) leads to the fre
quency response of the openloop P(z)C(z) as shown in Figure 10.5. The openloop
bandwidth is 1 kHz. The simulated sensitivity function S(z) = 1/(1 + P(z)C(z))
is shown in Figure 10.6, where there are two obvious rejection (i.e., S(z) < 1)
frequency ranges around 4 kHz and 8 kHz.
Rejection capability in terms of S(z) for D
i
(i = 1, 2) can be changed by
adjusting k
di
and ξ
i
. More simulations show that the change of S(z) for D
1
at 4
kHz leads to slight variation of S(z) for D
2
at 8 kHz, while the change of S(z)
for D
2
at 8 kHz affects more on S(z) around 4 kHz. Moreover, changing k
d2
will
affect rejection at 4 kHz more than changing damping ξ
2
.
The controller is implemented via dSpace 1103 with the utilization of LDV to
measure the displacement of the actuator. The sensitivity function S(z) of the closed
control loop is measured via DSA and plotted in Figure 10.7, which agrees well with
the simulated one (the dashed curve).
10
1
10
2
10
3
10
4
−20
0
20
40
60
80
M
a
g
n
i
t
u
d
e
(
d
B
)
10
1
10
2
10
3
10
4
−200
−100
0
100
200
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
FIGURE 10.5
Openloop frequency response with disturbance rejection at 4 and 8 kHz.
Blending Control for MultiFrequency Disturbance Rejection 209
10
1
10
2
10
3
10
4
−60
−50
−40
−30
−20
−10
0
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
4 kHz
8 kHz
FIGURE 10.6
Simulated sensitivity function with disturbance rejection at 4 and 8 kHz.
Next, assume that disturbances at three frequencies are to be rejected simultane
ously, and the frequencies are not the plant resonance frequencies. The disturbances
are modeled as (10.37) with ω
1
= 2π3000, ξ
1
= 0.02, ζ
1
= 0.7, k
d1
= 40 for D
1
(s),
ω
2
= 2π6500, ξ
2
= 0.085, ζ
2
= 0.61, k
d2
= 50 for D
2
(s), and ω
3
= 2π10000,
ξ
3
= 0.3, ζ
3
= 0.6, k
d3
= 52 for D
3
(s).
The order of the designed C(z) is 46. After order reduction, a 31st order C(z) is
obtained. With the lower order C(z), the compensated openloop frequency response
is shown in Figure 10.8. The simulated and the measured sensitivity functions are
shown in Figure 10.9 where three rejection frequency bands around 3 kHz, 6.5 kHz,
and 10 kHz are clearly seen.
The rejection capabilities for these three disturbances are further examined via
more simulations. It is shown that changing rejection capability for 3 kHz distur
bance by adjusting k
d1
and/or ξ
1
will change that for 6.5 kHz, while will not affect
much on that for 10 kHz. The rejection change for 6.5 kHz disturbance by chang
ing k
d2
, ξ
2
or ζ
2
will not lead to much change to other two disturbance rejections.
Changing parameters ξ
3
, ζ
3
and k
d3
of D
3
for 10 kHz disturbance will affect the
rejection for the other two disturbances more than itself.
210 Modeling and Control of Vibration in Mechanical Systems
10
1
10
2
10
3
10
4
−70
−60
−50
−40
−30
−20
−10
0
10
20
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
FIGURE 10.7
Measured (solid curve) sensitivity function with disturbance rejection at 4 and 8 kHz.
10
1
10
2
10
3
10
4
−20
0
20
40
60
80
M
a
g
n
i
t
u
d
e
(
d
B
)
10
1
10
2
10
3
10
4
−200
−100
0
100
200
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
FIGURE 10.8
Openloop with disturbance rejection at 3, 6.5, and 10 kHz.
Blending Control for MultiFrequency Disturbance Rejection 211
10
1
10
2
10
3
10
4
−60
−50
−40
−30
−20
−10
0
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
FIGURE 10.9
Sensitivity function with disturbance rejection at 3, 6.5, and 10 kHz.
10.4.2 Rejecting a combined mid and high frequency disturbance
Previously, the frequencies of the disturbances that need to be rejected are higher than
the bandwidth. In this section, we consider two disturbances near the bandwidth:
one at 2 kHz, higher than the bandwidth, and the other at 650 Hz, lower than the
bandwidth. The disturbance models are as in (10.37) where ω
1
= 2π650, ξ
1
= 0.05,
ζ
1
= 0.8, k
d1
= 77 for D
1
(s) at 650 Hz, and ω
2
= 2π2000, ξ
2
= 0.0046, ζ
2
= 0.85,
k
d2
= 0.1 for D
2
(s) at 2 kHz.
The resultant controller C(z) is of 27th order. By close zeropole cancelations, a
21st order controller can be obtained, and is able to accomplish the task of rejecting
D
1
and D
2
, which is illustrated in Figures 10.10 and 10.11. In Figure 10.10, a
signiﬁcant point is that corresponding to the peaks at 650 Hz and 2 kHz, the phases
are both lifted, which is desired in order not to lose the phase margin.
As observed in Figure 10.11, S(z) < 0 at 2 kHz. The experimentally measured
S(z) is plotted; see the rough curve in Figure 10.11. Additional simulations show
that decreasing damping ξ
1
will not only lower S(z) at 650 Hz, but also lower it
at 2 kHz. Increasing k
d1
will lower S(z) at 650 Hz, but will increase the hump of
S(z) and cause slight increase of S(z) at 2 kHz. On the other hand, increasing k
d2
in D
2
(s) for 2 kHz disturbance will not affect much on S(z) at 650 Hz. Changing
ζ
2
will easily cause phase loss at 2 kHz. It is thus suggested to change k
d2
and ξ
2
to
have a satisfactory sensitivity function S(z) at 2 kHz without causing phase loss.
Using the method of phase lead peak ﬁlter in Chapter 8, the rejection capability
212 Modeling and Control of Vibration in Mechanical Systems
for vibrations at 650 Hz and 2 kHz can be achieved with two peak ﬁlters, say pf
1
and pf
2
, which are individually designed to deal with the two vibrations, and then
the two ﬁlters are connected in parallel, i.e., 1 + pf
1
+ pf
2
. In this design method,
a baseline controller needs to be designed ﬁrst, then the two peak ﬁlters are obtained
based on the predesigned baseline controller. Thus the rejection performance of the
system for vibration in other frequency ranges depends on the predesigned base
line controller. Moreover, the phase lead peak ﬁlter method is difﬁcult to deal with
multifrequency vibrations in higher frequency range and at frequencies near plant
resonance modes, that however have been treated in this chapter.
10
2
10
3
10
4
−20
0
20
40
60
M
a
g
n
i
t
u
d
e
(
d
B
)
10
2
10
3
10
4
−200
−100
0
100
200
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
FIGURE 10.10
Openloop with disturbance rejections at 0.65 and 2 kHz.
REMARK 10.1 From the above application results and analysis, it is
noted that to deal with lower (say ≤ 2 kHz) frequency disturbances, the pro
posed control design method based on the blending technique works like the
phase lead peak ﬁltering method. For higher (say > 2 kHz) frequency distur
bances, it works like phasestabilized control [86]. This feature is beneﬁcial to
the stability of the closedloop system.
Blending Control for MultiFrequency Disturbance Rejection 213
10
1
10
2
10
3
10
4
−70
−60
−50
−40
−30
−20
−10
0
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
FIGURE 10.11
Sensitivity function with disturbance rejections at 0.65 and 2 kHz.
10.5 Conclusion
The rejection problem of several disturbances around different frequencies has been
formulated as a control blending problem. A controller for each disturbance rejec
tion has been designed individually by using the H
2
optimal control method. The
ultimate controller has been obtained by blending all these H
2
controllers so that all
these disturbances can be rejected simultaneously. The control blending method has
been respectively applied to design a controller for a 1.8inch HDD VCM actuator
in three cases, (1) rejecting two disturbances of frequencies higher than bandwidth
and close to actuator resonance frequencies, (2) rejecting three disturbances with fre
quencies higher than bandwidth and different from resonance frequencies, and, (3)
rejecting one disturbance with frequency lower than bandwidth and another higher
than bandwidth. Simulation and experimental results have shown that the control
blending technique results in a simultaneous attenuation for these disturbances. In
addition, it is worth noting that the method is able to prevent phase loss when it is
used to deal with disturbances near the bandwidth.
11
H
∞
Based Design for Disturbance Observer
11.1 Introduction
The idea of observing disturbance to improve the performance of a servomechanism
was ﬁrst introduced in [123]. It was suggested that if the disturbances were suppos
edly generated by a linear dynamic system and the model of the system was known,
they could be estimated from the system output measurement by an asymptotic esti
mator (Luenberger observer) and the effect of the disturbances could be neutralized
by feeding the disturbance estimates back into the system [124]. Over the years, the
method has been modiﬁed and applied. However, it is not always easy to identify the
disturbance model. Further, it is not always true that the disturbance model is linear
timeinvariant. Subsequently, a new type of disturbance observer (DOB) has been
introduced [128]. This new method does not require control designers to have the
full information of the disturbance model and does not need the assumption that the
disturbance model is linear timeinvariant. However, it requires the model of the con
trolled plant to be accurately known and invertible, at least within the bandwidth of
interest [129]. Recently, it has been proven that under certain assumptions imposed
on the plant and disturbance models, the two different methods are equivalent in that
the original disturbance observer introduced in [123] is actually a generalization of
the latter method [126]. In this chapter, we study the latter method where a general
form of disturbance observers, which does not need to solve the plant model inverse,
will be introduced.
If majority of the disturbances are of relatively lower frequency, when a DOB
is added on to attenuate the effect of disturbances, the standard and conventional
way of designing a Qﬁlter is to design it to be a lowpass ﬁlter with unity DC gain
[128]. In this chapter, we introduce the conventional disturbance observer ﬁrst, and
then present a general form of disturbance observer. An H
∞
control based method is
applied to the design of a Qﬁlter. The designed disturbance observer is applied to an
HDD servo system and its effectiveness in disturbance attenuation is demonstrated
by simulations and experiments.
215
216 Modeling and Control of Vibration in Mechanical Systems
11.2 Conventional disturbance observer
Figure 11.1 shows the block diagramof the conventional disturbance observer (DOB)
structure, where C is the feedback controller, P
n
is the nominal model of the plant
P, d
1
is the input disturbance, n is the noise,
ˆ
d
1
is an estimate of d
1
, and P
−1
n
is the
inverse of P
n
. τ represents a delay in the plant P. Theoretically, P(z) = z
−τ
P
n
(z).
Ignoring the nominal feedback loop with C, the transfer functions from the distur
bance d
1
and noise n to the output y are given by
T
yd
1
=
PP
n
(1 − Qz
−τ
)
P
n
+ Q(P − P
n
z
−τ
)
, (11.1)
T
yn
=
−PQ
P
n
+ Q(P − P
n
z
−τ
)
. (11.2)
To reject the disturbance d
1
, Q can be set as unity because T
yd
1
≈ 0 when the
delay is negligible. However, when Q = 1, T
yn
≈ −1 which means the measure
ment noise n is not attenuated. Thus, to eliminate the noise effect, it is known from
(11.2) that an ideal solution of Q is zero, but this will mean T
yd
1
≈ P and hence the
disturbance will be ampliﬁed.
FIGURE 11.1
Block diagram of the control loop with a conventional disturbance observer.
H
∞
Based Design for Disturbance Observer 217
Considering the overall system in Figure 11.1, the sensitivity function is given by
S(z) =
1 −Qz
−τ
1 − Qz
−τ
+PC + PQP
−1
n
. (11.3)
Theoretically, the plant model is P = P
n
z
−τ
, thus (11.3) becomes
S(z) =
1 −Qz
−τ
1 + PC
, (11.4)
which is stable as long as Q is stable since the loop is stable before the disturbance
observer is added on. Moreover, it is deduced from (11.3) that when Qz
−τ
= 1 with
zero phase around the disturbance frequency, the disturbance can be rejected because
S(z) → 0. Hence Q is designed such that the phase of Qz
−τ
is almost zero degree
and the magnitude is close to one in the frequency range where the disturbance d
1
dominates. Note that in Figure 11.1, it is needed to solve the inverse P
−1
n
of the
nominal plant model. In what follows, a general form of the disturbance observer is
proposed and the plant model inverse is not needed.
11.3 A general form of disturbance observer
Let
P
n
(z) = B(z)/A(z), (11.5)
B(z) = b
m
z
m
+ b
m−1
z
m−1
+· · · + b
0
,
A(z) = z
n
+ a
n−1
z
n−1
+ · · · +a
0
be the nominal model of the plant P. With the same notations as in Figure 11.1,
Figure 11.2 displays a general form of disturbance observer, where
M(z) =
B(z)
z
d
m
, N(z) =
A(z)
z
d
n
, (11.6)
and d
2
is the output disturbance.
In the conventional disturbance observer in Figure 11.1,
M(z) = z
−τ
, N(z) = P
−1
n
(z). (11.7)
Denote the statespace descriptions P(z) : (A
p
, B
p
, C
p
, D
p
); C(z) : (A
c
, B
c
,
C
c
, D
c
); M(z) : (A
M
, B
M
, C
M
, D
M
); N(z) : (A
N
, B
N
, C
N
, D
N
). From
Figure 11.2,
x(k + 1) = Ax(k) +B
1
w(k) +B
2
u
q
(k), (11.8)
y
q
(k) = C
y
x(k) +D
yw
w(k) +D
yu
u
q
(k), (11.9)
y(k) = C
z
x(k), (11.10)
218 Modeling and Control of Vibration in Mechanical Systems
FIGURE 11.2
Block diagram of the control loop with a general disturbance observer.
where
A =
A
p
−B
p
D
c
C
p
B
p
C
c
0 0
−B
c
C
p
A
c
0 0
−B
M
D
c
C
p
B
M
C
c
A
M
0
−B
N
C
p
0 0 A
N
, (11.11)
B
1
=
B
p
−B
p
D
c
B
p
D
c
0 −B
c
B
c
0 −B
M
D
c
B
M
D
c
0 −B
N
B
N
, B
2
=
−B
p
0
−B
M
0
,
C
y
=
−D
N
C
p
−D
M
D
c
C
p
D
M
C
c
C
M
C
N
, (11.12)
D
yw
=
0 −(D
N
+ D
m
D
c
) D
N
+ D
M
D
c
, D
yu
= −D
M
,
C
z
=
−C
p
0 0 0
, D
zw
= [0 − 1 0], D
zu
= 0, (11.13)
and x
T
(k) = [x
T
p
(k) x
T
c
(k) x
T
M
(k) x
T
N
(k)] is the augmented state of P(z), C(z),
M(z), and N(z), and w
T
(k) = [d
1
(k) d
2
(k) n(k)].
Denote the transfer function from w to y as T
yw
= [T
yd
1
T
yd
2
T
yn
]. The H
∞
optimization method will be applied to design Q(z) to minimize the H
∞
norm
W
d
1
T
yd
1
W
d
2
T
yd
2
W
n
T
yn
∞
. (11.14)
W
d
1
, W
d
2
and W
n
are weightings. Here the models for disturbances d
1
, d
2
and
noise n are not needed. However, in order to have a desired suppression of the
H
∞
Based Design for Disturbance Observer 219
disturbances d
1
, d
2
and the noise n, appropriately chosen weightings W
d
1
, W
d
2
and
W
n
are required which relies on our knowledge of the disturbances.
Different from the conventional disturbance observer, the general disturbance ob
server does not need the inverse of the nominal plant model. As shown later, it is
able to suppress the disturbances in a low frequency range without much perfor
mance degradation to higher frequency disturbances and noise.
The objective of the general disturbance observer design is then stated as: Given
a positive scalar γ and appropriate weightings W
d
1
, W
d
2
and W
n
, design a stable
Q(z) : (A
Q
, B
Q
, C
Q
, D
Q
) such that
[W
d
1
T
yd
1
W
d
2
T
yd
2
W
n
T
yn
]
T
∞
< γ. (11.15)
The H
∞
control design problem can be solved via Theorem 5.4.
REMARK 11.1 The sensitivity function of the control system in Figure
11.2 with the general disturbance observer is given by
S(z) =
1 + QM
1 + PC + QM − PQN
. (11.16)
With the conventional disturbance observer in (11.7), (11.16) is equal to
S(z) =
1 + Qz
−d
1 +PC +Qz
−d
−PQP
−1
n
. (11.17)
Assume that in (11.16)
Q(z) = Q
n
(z) ×
z
d
n
B(z)
, P = P
n
, (11.18)
then
S(z) =
1 + Q
n
z
d
n
/z
d
m
1 + PC + Q
n
z
d
n
/z
d
m
−Q
n
, (11.19)
which recovers the form in (11.17) when d
m
≥ d
n
, meaning that the con
ventional disturbance observer is a special case of the general disturbance
observer.
REMARK 11.2
In Figure 11.2, if the loop is cut at u, the transfer function from u to ˜ e or
y is considered as a new plant, and derived as
P
equ
=
y
u
=
P
1 − (M − PN)Q
. (11.20)
220 Modeling and Control of Vibration in Mechanical Systems
An equivalent open loop is denoted by T
EQ−OL
(z), and
T
EQ−OL
(z) = P
equ
C =
PC
1 − (M −PN)Q
. (11.21)
Assume ideally that P
n
= P and d
m
= d
n
. Equation (11.20) becomes
P
equ
= P, which implies that the proposed general disturbance observer does
not inﬂuence the characteristics of openloop P(z)C(z) greatly and thus the
stability and performance achieved by the nominal control loop with controller
C(z) are maintained. This is similar to the conventional disturbance observer.
11.4 Application results
It has been shown that a disturbance observer is capable of estimating disturbances
and modeling error [128]. Hence, disturbance observers can be used to increase the
R/W headpositioning accuracy in hard disk drives (HDDs) by using its estimation
results to cancel the effect of the disturbances and modeling error. Further, due to its
cost effectiveness and easy “addon” implementation with minimal change required
to the existing feedback controller, the disturbance observer without using additional
sensors is frequently used to enhance the tracking performance of a hard disk drive
servo system, such as the attenuation of disturbances [127], and the compensation of
VCM pivot friction [130].
Consider the VCM plant with model P(s) given in Chapter 10. The discretized
model with the sampling time T
s
= 1/30000 sec is given by (11.5) with
B(z) = 0.02399z
5
+ 0.1868z
4
− 0.03483z
3
−0.02165z
2
+ 0.1728z + 0.02025,
(11.22)
A(z) = z
6
− 3.052z
5
+ 4.657z
4
− 4.979z
3
+ 3.997z
2
− 2.399z + 0.7765.
(11.23)
The controller C(z) is the combination of a PID controller and two notch ﬁlters, and
given by
C(z) =
1.474 ×10
−5
z
6
−4.416 ×10
−5
z
5
+ 6.517 ×10
−5
z
4
− 6.621 ×10
−5
z
3
+5.015 × 10
−5
z
2
−2.872 × 10
−5
z + 9.053 × 10
−6
3.333 × 10
−5
z
6
−7.59 × 10
−5
z
5
+ 8.081 × 10
−5
z
4
−4.846 × 10
−5
z
3
+1.545 × 10
−5
z
2
−5.234 × 10
−6
z
.
(11.24)
Let
M(z) =
B(z)
z
6
, N(z) =
A(z)
z
6
, (11.25)
H
∞
Based Design for Disturbance Observer 221
W
d
1
= 0.3, W
d
2
= 1 and W
n
= 1.5. A stable Q(z) is then obtained via the H
∞
optimization in (11.14) and its frequency response is shown in Figure 11.3. Note
that the plant model inverse is not required in the general disturbance observer. This
beneﬁt is of great signiﬁcance, especially for nonminimum phase plant.
The sensitivity function S(z) is plotted in Figure 11.4, from which it can be
seen that the designed disturbance observer is able to suppress disturbance with fre
quency lower than 1 kHz without causing much degradation for rejection of higher
frequency disturbance. The servo performance, such as bandwidth, will change with
different weightings W
d
1
, W
d
2
and W
n
. Thus by adjusting the weightings accord
ing to the weights of d
1
, d
2
, and noise n in the position error signal, the designed
disturbance observer will result in a desired reduction rate of the error. To demon
strate the effectiveness of the disturbance estimation in the time domain, we assume
that the disturbances d
1
and d
2
and the noise n are generated by d
1
= D
1
(s)w
1
,
d
2
= D
2
(s)w
2
and n = N
n
(s)w
3
, where
D
1
(s) =
0.0004(s
2
−83.39s + 9.741 ×10
5
)(s
2
+ 1616s + 9.626 ×10
6
)
(s
2
+ 125.7s + 3.948 ×10
5
)(s
2
+ 10.05s + 1.011 ×10
6
)
,
D
2
(s) and N
n
(s) are in (2.60) and (2.61), and w
i
(i = 1, 2, 3) are independent white
noises with unity variance.
With the designed general disturbance observer, the estimate
ˆ
d
1
of d
1
is shown in
Figure 11.6. It follows the original d
1
approximately. As a result, the error signal is
shown in Figure 11.7. 50% reduction is achieved. The general disturbance observer
is more effective to compensate for the input disturbance d
1
than d
2
and n. W
d
2
and
W
n
are selected as 1 and 1.5 is to keep the attenuation to d
2
and n achieved by the
nominal feedback controller C(z). With lower W
d
2
and W
n
and higher W
d
1
, the
attenuation to d
2
and n will be degraded, although more suppression to d
1
will be
attained by using the disturbance observer.
Moreover, the conventional disturbance observer is designed for comparison. M(z)
and N(z) are given by
M(z) = z
−1
, (11.26)
N(z) = P
−1
n
(z)
=
5.2494(z
2
− 1.983z + 0.9834)(z
2
−1.253z + 0.9654)(z
2
+ 0.1836z + 0.8179)
z(z + 0.9501)(z + 0.1259)(z + 0.116)(z
2
− 1.22z + 0.9646)
.
(11.27)
A stable Q(z) for the conventional disturbance observer is designed with the H
∞
control method. Figure 11.5 shows the resultant sensitivity function, which is similar
to the one from the general disturbance observer. However, the plant model inverse
needs to be calculated.
Experiment has been done with a LDV and a dSpace 1103. The measured sensi
tivity functions are shown in Figure 11.8, which agree with the simulation results in
Figure 11.4. To evaluate the effect of the disturbance observer on the stability and
performance achieved by the nominal controller C(z), T
EQ−OL
is measured with the
222 Modeling and Control of Vibration in Mechanical Systems
general disturbance observer and the conventional disturbance observer, and shown
in Figure 11.9 and Figure 11.10. As expected, performance measures such as gain
margin and phase margin are not affected.
−10
−5
0
5
10
15
20
25
30
35
40
M
a
g
n
i
t
u
d
e
(
d
B
)
10
1
10
2
10
3
10
4
45
90
135
180
225
270
315
P
h
a
s
e
(
d
e
g
)
Bode Diagram
Frequency (Hz)
FIGURE 11.3
Frequency response of the designed Q(z).
11.5 Conclusion
A general form of disturbance observer has been presented and designed based on
the H
∞
control method to achieve desired disturbance/noise rejection. The distur
bance observer does not need to solve the plant model inverse, and thus its design
is simpliﬁed and has great advantages over the conventional disturbance observer,
especially for nonminimum phase plant. The simulation and implementation results
show that the general disturbance observer designed using the method employed in
this chapter is able to effectively improve the attenuation of disturbance in low fre
quency, and will not sacriﬁce the stability and performance of the nominal feedback
control loop.
H
∞
Based Design for Disturbance Observer 223
10
2
10
3
10
4
−50
−40
−30
−20
−10
0
10
M
a
g
n
i
t
u
d
e
(
d
B
)
Frequency(Hz)
No DOB
With DOB
FIGURE 11.4
The sensitivity functions without and with the general disturbance observer.
10
1
10
2
10
3
10
4
−50
−40
−30
−20
−10
0
10
M
a
g
n
i
t
u
d
e
(
d
B
)
Frequency(Hz)
Conventional DOB
General DOB
FIGURE 11.5
The sensitivity function comparison with the general and the conventional distur
bance observers.
224 Modeling and Control of Vibration in Mechanical Systems
0 0.05 0.1 0.15 0.2 0.25 0.3
−8
−6
−4
−2
0
2
4
6
8
x 10
−3
Time(sec)
A
m
p
l
i
t
u
d
e
o
f
d
i
s
t
u
r
b
a
n
c
e
d
1
a
n
d
i
t
s
e
s
t
i
m
a
t
e
(
µ
m
)
d
1
estimate of d
1
FIGURE 11.6
Disturbance d
1
.
0 0.05 0.1 0.15 0.2 0.25 0.3
−0.03
−0.02
−0.01
0
0.01
0.02
0.03
Time(sec)
A
m
p
l
i
t
u
d
e
o
f
e
r
r
o
r
e
(
µ
m
)
No DOB
With DOB
FIGURE 11.7
Error signal e.
H
∞
Based Design for Disturbance Observer 225
10
2
10
3
10
4
−50
−40
−30
−20
−10
0
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
No DOB
With general DOB
FIGURE 11.8
Measured sensitivity functions without and with the general disturbance observer.
10
1
10
2
10
3
10
4
−40
−20
0
20
40
60
M
a
g
n
i
t
u
d
e
(
d
B
)
10
1
10
2
10
3
10
4
−500
−400
−300
−200
−100
0
P
h
a
s
e
(
d
e
g
)
Freuqency(Hz)
Nominal
with the general DOB
FIGURE 11.9
Comparison of T
EQ−OL
about the general disturbance observer.
226 Modeling and Control of Vibration in Mechanical Systems
10
1
10
2
10
3
10
4
−40
−20
0
20
40
60
M
a
g
n
i
t
u
d
e
(
d
B
)
10
1
10
2
10
3
10
4
−500
−400
−300
−200
−100
0
P
h
a
s
e
(
d
e
g
)
Freuqency(Hz)
Nominal
with the conventional DOB
FIGURE 11.10
Comparison of T
EQ−OL
about conventional disturbance observer.
12
TwoDimensional H
2
Control for Error
Minimization
12.1 Introduction
The H
2
optimal control for 1D systems is a classical problem in linear systems
theory. Its objective is to minimize the error energy of the system when the system is
subject to a unit impulse input or, equivalently, a white noise input of unit variance.
Because of this analytically and practically meaningful speciﬁcation, the H
2
problem
and solution has been well studied and applied for several decades. Recently, the
H
2
control problem has been studied for 2D systems and a sufﬁcient condition for
the evaluation of 2D system H
2
performance in terms of LMIs is derived [139].
Using the condition, a systematic method for the design of the H
2
controller for 2
D systems in terms of LMIs has been developed. The developed 2D H
2
control
design method is of great importance to those systems that have 2D behavior and
can be modeled using 2D linear system models. Selfservo track writer (SSTW)
for data storage devices is one of these systems [17]. In the selfservo track writing,
due to vibrations and noise the servo controller causes the actuator to follow the
resulting noncircular trajectory in the next burst writing step, so that the new bursts
are written at locations reﬂecting the errors present in the preceding track via the
closedloop response of the servo loop, as well as in the present track. Consequently,
each step in the process carries a “memory” of all preceding track shape errors.
This “memory” depends on the particular closedloop response of the servo loop.
Because of the interdependency of propagation tracks, track shape errors may be
ampliﬁed from one track to the next through the closedloop response when writing
the propagation tracks. Thus selfservo writing systems must provide a means of
accurately writing servopatterns while controlling the propagation of track shape
errors. Therefore, error propagation containment is critically important. The target
of preventing error propagation is to reject the track shape error due to track non
circularity recorded in propagated tracks so that the circular concentric tracks are
achieved in every propagation trace.
In this chapter we describe the SSTW process with a twodimensional (2D) model.
Then the error propagation containment problem of the SSTW process is formulated
as a 2D stabilization problem. Instead of the conventional feedforward control, the
2D stabilizing control is able to prevent the error propagation. Furthermore, the
227
228 Modeling and Control of Vibration in Mechanical Systems
TMR minimization problem of the SSTW process is formulated as a 2D H
2
con
trol problem. A 2D H
2
controller is designed which is able to prevent the error
propagation and minimize the TMR.
12.2 2D stabilization control
We consider the following 2D system model [138]:
_
x
h
(i + 1, j)
x
v
(i, j + 1)
_
= A
_
x
h
(i, j)
x
v
(i, j)
_
+B
1
w(i, j) +B
2
u(i, j), (12.1)
y(i, j) = C
1
_
x
h
(i, j)
x
v
(i, j)
_
+D
11
w(i, j) + D
12
u(i, j), (12.2)
e(i, j) = C
2
_
x
h
(i, j)
x
v
(i, j)
_
+D
21
w(i, j) + D
22
u(i, j), (12.3)
where x
h
∈ R
n
1
, x
v
∈ R
n
2
, w(i, j) ∈ R
q
, u(i, j) ∈ R
m
, y(i, j) ∈ R
p
and
e(i, j) ∈ R
l
are, respectively, the horizontal state, the vertical state, the disturbance
input, the control input, the controlled output, and the measurement of the plant.
Let
x(i, j) =
_
x
h
(i, j)
x
v
(i, j)
_
, (12.4)
the above system is equivalent to
x(i + 1, j + 1) = A
1
x(i, j + 1) + A
2
x(i + 1, j)
+B
11
w(i, j + 1) +B
12
w(i + 1, j)
+B
21
u(i, j + 1) +B
22
u(i + 1, j), (12.5)
y(i, j) = C
1
x(i, j) + D
11
w(i, j) +D
12
u(i, j), (12.6)
e(i, j) = C
2
x(i, j) + D
22
u(i, j) (12.7)
where
A
1
=
_
I
n
1
0
0 0
_
A, A
2
=
_
0 0
0 I
n
2
_
A,
B
k1
=
_
I
n
1
0
0 0
_
B
k
, B
k2
=
_
0 0
0 I
n
2
_
B
k
, k = 1, 2. (12.8)
Introduce the following 2D output feedback controller C(z
1
, z
2
):
x
c
(i + 1, j + 1) = A
c1
x
c
(i, j + 1) + A
c2
x
c
(i + 1, j)
+B
c1
e(i, j + 1) +B
c2
e(i + 1, j), (12.9)
u(i, j) = C
c
x
c
(i, j) +D
c
e(i, j). (12.10)
TwoDimensional H
2
Control for Error Minimization 229
Associated with the 2D controller (12.9)−(12.10), the 2D stabilization prob
lem is stated as follows: for the 2D system (12.1)−(12.3) or (12.5)−(12.7) with
w(i, j) = 0, design a dynamic output feedback controller of the formin (12.9)−(12.10)
such that the resulting closedloop SSTW servo system is asymptotically stable.
Deﬁne
Z = D
c
CR +C
c
Ξ
T
, V
k
= SB
2k
D
c
+ ΛB
ck
, (12.11)
U
k
= S(A
k
+ B
2k
D
c
C)R+ SB
2k
C
c
Ξ
T
+ ΛB
ck
CR + ΛA
ck
Ξ
T
, k = 1, 2,
(12.12)
where R > 0, S > 0 and Ξ and Λ are invertible matrices satisfying ΞΛ
T
= I −RS.
THEOREM 12.1
[136] Consider the 2D system (12.5)−(12.7). Then, there exists a full order
output feedback controller of the form in (12.9)−(12.10) that asymptotically
stabilizes the system (12.5)−(12.7) if there exist matrices R > 0, S > 0,
˜
Ω
X
> 0, D
c
, U
k
, V
k
(k = 1, 2) and Z such that the following LMI holds:
_
_
−
˜
Ω
X
0
˜
Ω
T
A1
0 −(
˜
Ω
F
−
˜
Ω
X
)
˜
Ω
T
A2
˜
Ω
A1
˜
Ω
A2
−
˜
Ω
F
_
_
< 0 (12.13)
where
˜
Ω
F
=
_
S I
I R
_
,
˜
Ω
A1
=
_
SA
1
+ V
1
C U
1
A
1
+B
21
D
c
C A
1
R +B
21
Z
_
,
˜
Ω
A2
=
_
SA
2
+V
2
C U
2
A
2
+ B
22
D
c
C A
2
R + B
22
Z
_
. (12.14)
In this situation, the controller parameters of (12.9)−(12.10) can be given by
C
c
= (Z − D
c
CR)Ξ
−T
, B
ck
= Λ
−1
(V
k
− SB
2k
D
c
), (12.15)
A
ck
= Λ
−1
[U
k
−S(A
k
+B
2k
D
c
C
2
)R −SB
2k
C
c
Ξ
T
− ΛB
ck
C
2
R]Ξ
−T
, k = 1, 2.
(12.16)
12.3 2D H
2
control
Let T
yw
: w → y denote the closedloop system subject to the white noise w. The
H
2
norm of T
yw
is approximately given by
T
yw
2
=
¸
¸
¸
_
1
L
1
K − 1
L,K−1
i, j=1
y(i, j)
2
(12.17)
230 Modeling and Control of Vibration in Mechanical Systems
where L and K are large enough. The control design problem to minimize the 2
D H
2
norm is stated as follows: ﬁnd a 2D output feedback controller of the form
in (12.9)−(12.10) for the 2D system (12.1)−(12.3) or (12.5)−(12.7) such that the
closedloop system is stable and the H
2
performance T
yw
2
is minimized.
An LMI approach will be given as follows to design a 2D H
2
controller for the
2D system (12.1)−(12.3) such that the closedloop system is stable and the error is
minimized.
THEOREM 12.2
[139] The 2D H
2
control problem for the plant (12.1)−(12.3) is solvable if
there exist matrices S > 0, Θ, Λ, Γ, D
c
and blockdiagonal matrices X =
diag{X
h
, X
v
}, Y = diag{Y
h
, Y
v
}, M = diag{M
h
, M
v
}, H
11
= diag{H
h
11
,
H
v
11
} > 0, H
22
= diag{H
h
22
, H
v
22
} > 0, and H
12
= diag{H
h
12
, H
v
12
}, of appro
priate dimensions, such that
_
¸
¸
¸
¸
_
H
11
∗ ∗ ∗ ∗
H
T
12
H
22
∗ ∗ ∗
XA + ΓC
2
Θ X +X
T
− H
11
∗ ∗
A +B
2
D
c
C
2
AY
T
+B
2
Λ I + M
T
−H
T
12
Y + Y
T
−H
22
∗
C
1
+ D
12
D
c
C
2
C
1
Y
T
+ D
12
Λ 0 0 I
_
¸
¸
¸
¸
_
> 0,
(12.18)
_
¸
¸
_
S ∗ ∗ ∗
XB
1
+ ΓD
21
X +X
T
− H
11
∗ ∗
B
1
+B
2
D
c
D
21
I + M
T
−H
T
12
Y +Y
T
−H
22
∗
D
11
+D
12
D
c
D
21
0 0 I
_
¸
¸
_
> 0, (12.19)
Trace(S) < λ
2
, (12.20)
are satisﬁed. If the above stated conditions are satisﬁed, a feasible H
2
con
troller is given by
_
x
h
c
(i + 1, j)
x
v
c
(i, j + 1)
_
= A
c
_
x
h
c
(i, j)
x
v
c
(i, j)
_
+ B
c
e(i, j), x
h
c
∈ R
n
h
, x
v
c
∈ R
n
v
, (12.21)
u(i, j) = C
c
_
x
h
c
(i, j)
x
v
c
(i, j)
_
+ D
c
e(i, j), (12.22)
with
A
c
= U
−1
[Θ−X(A + B
2
D
c
C
2
)Y
T
− XB
2
C
c
V
T
−UB
c
C
2
Y
T
]V
−T
,(12.23)
C
c
= (Λ −D
c
C
2
Y
T
)V
−T
, B
c
= U
−1
(Γ −XB
2
D
c
), (12.24)
where U = diag{U
h
, U
v
} and V = diag{V
h
, V
v
} satisfy XY
T
+UV
T
= M.
REMARK 12.1 The solution of the 2D controller in the above theorem
is in terms of LMIs which can be eﬃciently solved by convex optimization just
TwoDimensional H
2
Control for Error Minimization 231
like other 1D control problems involving LMIs. In [139], it is proved that if
there exist solutions X, Y , M, D
c
, S > 0, Θ, Λ, Γ, H
11
, H
22
, and H
12
for
the LMIs (12.18)(12.20), M − XY
T
is invertible. From UV
T
= M − XY
T
,
nonsingular matrices U and V can be computed. Then, the computation of
the controller parameters C
c
, A
c
, B
c
can be carried out by solving (12.23)
and (12.24).
12.4 SSTW process and modeling
FIGURE 12.1
SSTW process.
The process of selfservo writing is shown in Figure 12.1 and is generally known
to involve the following distinct steps [135]: writing some tracks or at least one track
called seed tracks; reader reads back the seed track and writer writes actual product
servopattern for the next track based on the readback signal; writing servopattern for
232 Modeling and Control of Vibration in Mechanical Systems
the next track based on the readback signal from the previous written track till the
whole process is completed. During the process, SSTW generates radial informa
tion progressively to deploy servopattern. The radial error in track N is inevitably
compounded to the following tracks: tracks N + 1, N + 2, · · ·. This leads to error
propagation in SSTW.
In selfservo track writing, track shape errors such as noncircularity are intro
duced by mechanical disturbances, spindle motor vibration and other factors when
writing the propagation tracks. The servo controller causes the actuator to follow
the resulting noncircular trajectory in the next burst writing step, so that the new
bursts are written at locations reﬂecting the errors present in the preceding step via
the closedloop response of the servo loop, as well as in the present step. Conse
quently, each step in the process carries a “memory” of all preceding track shape
errors. This “memory” depends on the particular closedloop response of the servo
loop. Selfservo writing systems must provide a means of accurately writing ser
vopatterns while controlling the propagation of track shape errors.
12.4.1 SSTW servo loop
FIGURE 12.2
SSTW servo loop with disturbances and noise models.
Figure 12.2 shows the SSTW servo loop with disturbances and noises. D
1
(s),
D
2
(s), and N(s) are respectively the models of disturbances d
1
, d
2
and noise n.
y(k) is the position of the write head with respect to a perfectly circular track on the
disk and PES(k) is the position error signal. Let T
p
be the rotational period of the
disk, and T
s
be the sampling rate of the position error signal. Then the sector number
K = T
p
/T
s
. y(k − K) represents the track proﬁle of the previous track. Similarly,
PES(k−K) represents the position error when writing the previous track. The read
head follows on the track y(k −K) which is the reference input for the SSTW servo
system, i.e., one revolution of y(k) becomes the reference of the next written track
TwoDimensional H
2
Control for Error Minimization 233
or a few subsequent tracks due to the action of selfservo writing.
It follows from Figure 12.2 that
y(k) = T(z)y(k −K) +P(z)S(z)d
1
(k) + S(z)d
2
(k) +T(z)n(k), (12.25)
where S(z) = 1/(1 +P(z)C(z)) is the sensitivity function, T(z) = 1 −S(z) is the
closedloop transfer function. The typical closedloop transfer function will amplify
the error at the frequency where its magnitude is more than 0 dB. In other words,
these frequency components in disturbances or noise will be ampliﬁed during prop
agation, while others will compound to following tracks and decay gradually. With
the same disturbance models D
1
(s) and D
2
(s) and noise model N(s) as in Chapter
11, Figure 12.3 shows PES NRRO in the time domain and its σ value versus track
number when a 1D feedback controller C(z) is used in the closed loop. The error
propagation mentioned previously is clearly observed. Hence the error propagation
problem must be addressed in the servo control design for SSTW.
0 1 2 3 4 5 6 7 8 9 10
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
Track number
P
E
S
a
n
d
σ
o
f
P
E
S
Position error signal
Sigma value
FIGURE 12.3
PES NRRO and its σ values versus track number during propagation. (The time
sequence and the σ value increase with the track number.)
12.4.2 Twodimensional model
For the 2D modeling, the following notations are used:
234 Modeling and Control of Vibration in Mechanical Systems
FIGURE 12.4
SSTW servo loop modeling in two dimensions.
i = 0, 1, 2, · · · , L: the ith track.
j · T
s
: time with j = 0, 1, 2, · · · , K −1.
As shown in Figure 12.4, y(i, j) is the position of the write head at track i in the
radial dimension and time j · T
s
in the axial dimension and PES(i, j) the position
error. y(i − 1, j) (i.e., y(j − K)) represents the track proﬁle of the (i − 1)th track.
Similarly, PES(i − 1, j) represents the position error of the (i − 1)th track. The
read head follows the track y(i − 1, j) which is the reference input for the SSTW
servo system, i.e., one revolution of y(i, j) becomes the reference of the next written
track due to the action of selfservo track writing.
Based on Figure 12.4, we have
x
p
(i, j + 1) = A
p
x
p
(i, j) + B
p
(u(i, j) + d
1
(i, j)), (12.26)
y(i, j) = C
p
x
p
(i, j) +D
p
(u(i, j) +d
1
(i, j)) + d
2
(i, j), (12.27)
PES(i, j) = y(i −1, j) − y(i, j) + n(i, j). (12.28)
Denote x
p
, x
d
1
, x
d
2
and x
n
the corresponding state vectors of P(z), D
1
(z), D
2
(z)
and N(z), respectively. Let
x
h
(i + 1, j) = y(i, j), x
h
∈ R
n
h
, x
v
∈ R
n
v
,
x
v
(i, j + 1)
T
= [x
p
(i, j + 1)
T
x
d
1
(i, j + 1)
T
x
d
2
(i, j + 1)
T
x
n
(i, j + 1)
T
],
e(i, j) = PES(i, j), w(i, j)
T
= [w
1
(i, j)
T
w
2
(i, j)
T
w
3
(i, j)
T
], (12.29)
it follows from (12.26)−(12.28) that
_
x
h
(i + 1, j)
x
v
(i, j + 1)
_
= A
_
x
h
(i, j)
x
v
(i, j)
_
+B
1
w(i, j) +B
2
u(i, j), (12.30)
y(i, j) = C
1
_
x
h
(i, j)
x
v
(i, j)
_
+D
11
w(i, j) + D
12
u(i, j), (12.31)
TwoDimensional H
2
Control for Error Minimization 235
e(i, j) = C
2
_
x
h
(i, j)
x
v
(i, j)
_
+D
21
w(i, j) + D
22
u(i, j), (12.32)
where
A =
_
¸
¸
¸
¸
_
0 C
p
D
p
C
d
1
C
d
2
0
0 A
p
B
p
C
d
1
0 0
0 0 A
d
1
0 0
0 0 0 A
d
2
0
0 0 0 0 A
n
_
¸
¸
¸
¸
_
, B
1
=
_
¸
¸
¸
¸
_
D
p
D
d
1
D
d
2
0
B
p
D
d
1
0 0
B
d
1
0 0
0 B
d
2
0
0 0 B
n
_
¸
¸
¸
¸
_
, (12.33)
B
2
=
_
¸
¸
¸
¸
_
D
p
B
p
0
0
0
_
¸
¸
¸
¸
_
, C
1
=
_
0 C
p
D
p
C
d
1
C
d
2
0
¸
, (12.34)
D
11
=
_
D
p
D
d
1
D
d
2
0
¸
, D
12
= D
p
, (12.35)
C
2
=
_
1 −C
p
− D
p
C
d
1
− C
d
2
C
n
¸
, (12.36)
D
21
=
_
−D
p
D
d
1
−D
d
2
D
n
¸
, D
22
= −D
p
. (12.37)
As such, the SSTW servo loop is modeled as the 2DRoesser model (12.30)−(12.37)
[138] where disturbance and noise models are taken into consideration. The SSTW
error propagation problem can then be simpliﬁed as the stabilization problem of a
2D system. Unlike 1D feedback control plus feedforward compensation, it does
not need an additional feedforward controller to prevent the error propagation. In the
next section, feedforward compensation on the basis of 1D feedback control will be
presented, followed by 2D control in a later section.
12.5 Feedforward compensation method
To contain the error propagation, it is easy to come up with the idea of injecting a
correction signal f(k) to the PES by using the available signal PES(k−K) through
a feedforward compensator F(z). Thus a feedforward (FF) compensation method is
used as shown in Figure 12.5.
During the servo writing process, by tracking the previously written adjacent
tracks, adjusting servo reference and closing the existing VCM loop, the head can
be offset to a known radial position with reference to the adjacent tracks and gen
erates radial information progressively to deploy servo pattern. The radial error in
one track is inevitably compounded to the following tracks and propagates accord
ing to the closedloop transfer function from y(k − K) to y(k), as seen in (12.25).
The target of error propagation containment is thus to reject the writtenin error due
to track noncircularity recorded in propagated tracks so that the circular concentric
236 Modeling and Control of Vibration in Mechanical Systems
FIGURE 12.5
SSTW servo loop.
tracks are achieved in every propagation trace. Obviously, when there is no hump in
the closedloop transfer function T(z), the error propagation will be contained.
Next, we use three control schemes for the selfservo track writer: (1). PD con
trol since it can produce a ﬂat closedloop response such that the error propagation
is contained even without a feedforward compensation; (2). PID control plus feed
forward compensation; (3). To minimize TMR or equivalently the H
2
norm of the
transfer function from w = [w
1
w
2
w
3
]
T
to track proﬁle y, the H
2
control tech
nique is employed to design the feedback controller C(z) and then a feedforward
compensator F(z) is designed to contain the error propagation.
The frequency response of the VCM plant under consideration is shown in Figure
12.6, and its transfer function P(s) is described by the following zeros, poles and
gain:
zeros = 10
4
×[3.4558, − 0.6158 ± 8.7749j, 0.7540 ± 4.9697j,
−0.1960 ± 3.2614j, −0.9425 ± 1.6324j], (12.38)
poles = 10
5
×[−0.2513, − 0.3456, − 0.0726 ± 1.0342j, − 0.0090 ± 0.5968j,
−0.0377 ± 0.3751j, − 0.0283 ± 0.2813j,
−0.0011 ± 0.0036j], (12.39)
gain = 5.8987 ×10
12
. (12.40)
(1). PD feedback control
A typical form of PID controllers is given by
C(z) = K
p
+ K
i
z
z − 1
+K
d
z − 1
z
. (12.41)
Let K
i
= 0, a PD controller leading to a ﬂat closedloop transfer function can
be obtained by adjusting K
p
and K
d
. Denote the closedloop transfer function as
TwoDimensional H
2
Control for Error Minimization 237
10
2
10
3
10
4
−60
−40
−20
0
20
40
60
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
Measured
Modeled
10
2
10
3
10
4
−1000
−800
−600
−400
−200
0
P
h
a
s
e
(
d
e
g
)
FIGURE 12.6
Frequency response of a VCM actuator.
238 Modeling and Control of Vibration in Mechanical Systems
FIGURE 12.7
Frequency response of the closedloop transfer function with the PD controller, PID
controller, and H
2
controller.
T
PD
(z). Figure 12.7 shows a ﬂat magnitude response of T
PD
(z) resulting from the
PD controller given below. The sampling rate is 12.64 kHz.
C(z) =
4.79 ×10
−5
− 9.573 ×10
−5
+ 4.783 ×10
−5
z
2
− z
. (12.42)
Theoretically the error propagation would disappear as T
PD
∞
≤ 1. However,
the PD control is not able to deal with bias force due to lack of integrator. Simulation
shows that the propagation stops after several tracks when a small bias force of 0.001
is added. Therefore, PD control is not practically applicable.
(2). PID feedback control plus feedforward compensation
The closedloop transfer function with the PID controller
C(z) =
2.435 × 10
−5
z
2
−4.406 × 10
−5
z + 1.999 × 10
−5
z
2
−1.209 + 0.24
(12.43)
has a region greater than 0 dB as seen in Figure 12.7. As mentioned earlier, this will
cause the error propagation problem. Thus a feedforward compensator F(z) will be
designed to contain the error propagation. According to Figure 12.5,
y(k) =
_
PC
1 + PC
+
F
1 + PC
_
y(k − K) +
P
1 + PC
d
1
(k) +
1
1 +PC
d
2
(k)
TwoDimensional H
2
Control for Error Minimization 239
+
PC
1 + PC
n(k) −
PF
1 +PC
d
1
(k −K) −
F
1 + PC
d
2
(k − K), (12.44)
which means that if the feedforward compensator F(z) is designed such that the
magnitude of the transfer function
Φ =
PC
1 +PC
+
F
1 + PC
(12.45)
is less than one, i.e., Φ
∞
< 1, the error propagation can be contained. It is
straightforward from (12.45) that
F(z) = Φ(z)(1 +P(z)C(z)) − P(z)C(z). (12.46)
When Φ is selected as the closedloop transfer function with the previous PD con
trol, a 7
th
order F(z) is obtained from (12.46) after order reduction. The designed
F(z) can prevent the error propagation when the PID feedback control is applied in
the servo loop since Φ
∞
< 1 with the designed F(z), as observed in Figure 12.8.
Figure 12.9 shows the 3σ value of the PES versus frequency with error propagation
containment.
10
1
10
2
10
3
10
4
−9
−8
−7
−6
−5
−4
−3
−2
−1
0
1
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
FF with H2
FF with PID
FIGURE 12.8
Φ versus frequency.
240 Modeling and Control of Vibration in Mechanical Systems
0 1000 2000 3000 4000 5000 6000 7000
2
4
6
8
10
12
14
Frequency(Hz)
3
σ
(
p
e
r
c
e
n
t
o
f
t
r
a
c
k
)
PID
H2
FIGURE 12.9
3σ of PES NRRO.
(3). H
2
feedback control plus feedforward compensation
With the identiﬁed disturbance models in Chapter 2, the augmented system for the
optimal H
2
control design is described as follows.
x(k + 1) = Ax(k) + B
1
w(k) +B
2
u(k), (12.47)
PES(k) = C
1
x(k) + D
11
w(k), (12.48)
z(k) = C
2
x(k) + D
21
w(k) + D
22
u(k), (12.49)
where
A =
_
¸
¸
_
A
p
B
p
C
d
1
0 0
0 A
d
1
0 0
0 0 A
d
2
0
0 0 0 A
n
_
¸
¸
_
, B
1
=
_
¸
¸
_
B
p
D
d
1
0 0
B
d
1
0 0
0 B
d
2
0
0 0 B
n
_
¸
¸
_
, B
2
=
_
¸
¸
_
B
p
0
0
0
_
¸
¸
_
,(12.50)
C
1
=
_
C
p
0 C
d
2
C
n
¸
, D
11
=
_
0 D
d
2
D
n
¸
, (12.51)
C
2
=
_
C
p
0 C
d
2
0
¸
, D
21
= 0, D
22
= 0, (12.52)
x is the combined state variables from the VCM actuator model P(z), the input dis
turbance model D
1
(z), the output disturbance model D
2
(z), and the measurement
noise model N(z). (A
p
, B
p
, C
p
, D
p
), (A
d
1
, B
d
1
, C
d
1
, D
d
1
), (A
d
2
, B
d
2
, C
d
2
, D
d
2
),
and (A
n
, B
n
, C
n
, D
n
) are respectively the statespace models of P(z), D
1
(z),
D
2
(z) and N(z). PES is the measured position error signal and z stands for y in
Figure 12.5. The H
2
control problem can be solved using the method in Chapter 5.
TwoDimensional H
2
Control for Error Minimization 241
FIGURE 12.10
Frequency response of the H
2
controller.
FIGURE 12.11
Frequency response of the openloop system with the H
2
controller.
242 Modeling and Control of Vibration in Mechanical Systems
FIGURE 12.12
Comparison of sensitivity functions.
Figures 12.10 and 12.11 show the frequency responses of the designed H
2
feed
back controller and the compensated openloop system with bandwidth of 1.07 kHz,
gain margin 7.3 dB, and phase margin 45.9 deg. Figure 12.12 shows the comparison
of error rejection functions with the PID control and the H
2
control, where we can
see that the H
2
control outperforms the PID control in error rejection.
The frequency response of the closedloop transfer function with the H
2
control
is shown in Figure 12.7, which implies that a feedforward compensator is needed to
contain the error propagation when the servo loop uses the H
2
feedback controller.
For the design of the feedforward compensator F(z), the selected Φ depends on the
feedback controller and it is found that lower Φ may not give a better error propaga
tion containment. When Φ = 0.9T
PD
, a satisfactory error propagation containment
is guaranteed when the H
2
feedback control is applied in the SSTW servo loop. A
6
th
order F(z) is thus obtained after model reduction. The resultant Φ with the
designed F(z) is shown in Figure 12.8, and it is seen that Φ < 1. Figure 12.13
shows the σ value comparison of PES at different tracks with error propagation con
tainment. It can be observed that PES σ is up and down as track propagation is going
on. Compared with that of the PID control, the σ of PES versus track number is
improved by around 27% when the H
2
control is employed in the SSTW servo loop.
TwoDimensional H
2
Control for Error Minimization 243
FIGURE 12.13
σ value of PES NRRO versus track number.
12.6 2D control formulation for SSTW
With the 2D model (12.30)−(12.37), the SSTW error propagation problem is sim
pliﬁed as the stabilization problem of a 2D system. That is to design a dynamic out
put feedback controller of the form in (12.9)−(12.10) for the model (12.30)−(12.32)
with w(i, j) = 0 such that the resulting closedloop system is asymptotically sta
ble. Unlike the design of 1D feedback plus feedforward compensation previously, it
does not need an additional feedforward controller to prevent the error propagation.
It is not difﬁcult to see from (12.9)−(12.10) that (A
c2
, B
c2
, C
c
, D
c
) is acting
along the time direction only, and thus actually it works like the 1D feedback con
troller that we are concerned with conventionally.
The 2Dcontroller (12.21)−(12.22) can be written equivalently to the form(12.9)
(12.10). Let
A
c1
=
_
I
n
h
0
0 0
_
A
c
, A
c2
=
_
0 0
0 I
n
v
_
A
c
, (12.53)
B
c1
=
_
I
n
h
0
0 0
_
B
c
, B
c2
=
_
0 0
0 I
n
v
_
B
c
. (12.54)
The performance of (A
c2
, B
c2
, C
c
, D
c
) can thus be evaluated as a normal 1D
244 Modeling and Control of Vibration in Mechanical Systems
control in time dimension.
As is known that one of the most important performance measures for SSTW is
the track misregistration or TMR, the total amount of random ﬂuctuation about the
desired track location. TMR is used to judge the required accuracy of positioning. To
achieve a high positioning accuracy, one way in servo control is to minimize TMR,
which is expressed as the standard deviation of the true PES, i.e.
σ
y(i,j)
=
¸
¸
¸
_
1
L
1
K −1
L,K−1
i,j=0
y(i, j)
2
. (12.55)
Let T
yw
: w → y denote the closedloop system subject to a white noise w. When
L and K are large enough, the H
2
norm of T
yw
can be approximately given by
T
yw
2
=
¸
¸
¸
_
1
L
1
K −1
L,K−1
i, j=1
y(i, j)
2
. (12.56)
Thus, the control design problem to minimize TMR can be treated as a 2D H
2
optimal control problem, which is stated as follows: ﬁnd a 2D output feedback
controller of the form in (12.9)−(12.10) for the SSTW plant P(s) such that the
closedloop system is stable and the H
2
performance T
yw
2
is minimized.
The problem of minimizing the TMR of the SSTW process is thus formulated as
the 2D H
2
control problem. A 2D H
2
controller will subsequently be designed to
minimize the track misregistration. Note that since the 2D controller stabilizes the
system, it also contains the error propagation. Thus, the 2D H
2
control approach
simultaneously addresses the error propagation and TMR minimization problems,
which is different from the previous 1D method where the problems are addressed
separately by feedback and feedforward controls.
12.7 2D stabilization control for error propagation containment
12.7.1 Simulation results
The simulation block diagram on 2D control is shown in Figure 12.14. The sim
ulation is carried out in MATLAB/Simulink. In the simulation, the sector number
K = 270, the spindle rotational speed is 7200 RPM, and thus the sampling fre
quency is 270 × (7200/60) = 32400 Hz.
To show the capability of the designed 2D controller to prevent the error propa
gation, 100 tracks are propagated in the simulation. The results show that the SSTW
process is stabilized and the error propagation is contained. The σ value of PES
NRRO is plotted versus the track number in Figure 12.15, which implies that the
error amplitude is oscillating steadily.
TwoDimensional H
2
Control for Error Minimization 245
FIGURE 12.14
2D controller for SSTW servo loop.
From (12.9)−(12.10) and Figure 12.14, it is not difﬁcult to see that (A
c2
, B
c2
,
C
c
, D
c
) is acting along the time direction only, and thus actually it works like the
1D controller we usually take into account. Subsequently, the openloop and the
sensitivity function frequency responses are obtained and shown in Figures 12.16
and 12.17. The openloop crossover frequency is 1.6 kHz, the gain margin is 6 dB,
and the phase margin is 50 degrees. A bad high frequency part of the sensitivity
function has appeared, because here only the stabilization problemis considered and
no performance optimization is involved. Thus the 2D H
2
control will be presented
which gives a better performance than the stabilizing only controller.
12.8 2D H
2
control for error minimization
12.8.1 Simulation results
The σ value of PES NRRO is plotted versus the track number in Figure 12.15, where
the error amplitude is oscillating steadily, which implies that error propagation is
contained. Additionally, the position error has been minimized in the H
2
normsense.
The σ values of PES NRRO with the stabilization control and the H
2
control are
compared in Figure 12.15 and it is seen that the error is reduced by an average of
60% via the 2D H
2
control.
The conventional 1D H
2
feedback control has also been designed to minimize
the error signal, and based on the feedback control a feedforward control is further
designed to contain the error propagation. This method is also compared with the
above two methods in Figure 12.15. It is seen that the proposed 2D control method
is comparable to the previous 1D control method.
246 Modeling and Control of Vibration in Mechanical Systems
0 10 20 30 40 50 60 70 80 90 100
0
1
2
3
4
5
6
7
8
9
x 10
−3
Track number
σ
o
f
p
e
s
N
R
R
O
(
µ
m
)
2−D H2 control
2−D Stabiliation
1−D H2 feedback+feedforward control
FIGURE 12.15
σ of PES NRRO versus track number.
10
2
10
3
10
4
−20
−10
0
10
20
30
M
a
g
n
i
t
u
d
e
(
d
B
)
10
2
10
3
10
4
−250
−200
−150
−100
−50
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
FIGURE 12.16
Openloop frequency response with stabilization controller.
TwoDimensional H
2
Control for Error Minimization 247
10
2
10
3
10
4
−30
−25
−20
−15
−10
−5
0
5
10
15
20
M
a
g
n
i
t
u
d
e
(
d
B
)
Frequency(Hz)
FIGURE 12.17
Sensitivity function with stabilization controller (A
c2
, B
c2
, C
c
, D
c
).
12.8.2 Experimental results
It is noted that in the 2D controller (12.9)−(12.10), (A
c2
, B
c2
, C
c
, D
c
) acts in
the time dimension, and runs on the same track as a 1D controller C(z) we have
usually considered. Thus in this section we particularly take into account the 1D
controller C(z) : (A
c2
, B
c2
, C
c
, D
c
) for the plant P(s) (12.38)−(12.40) in the
time dimension.
Straightforwardly from the designed 2D controller C(z
1
, z
2
) : (A
c1
, A
c2
, B
c1
,
B
c2
, C
c
, D
c
), the 1D controller C(z) : (A
c2
, B
c2
, C
c
, D
c
) can be obtained
and its frequency response is shown in Figure 12.18, where the notches around 4, 6
and 9 kHz are to suppress the resonances as observed in Figure 12.6. With the 1D
controller C(z), the simulated frequency responses of the open loop C(z)P(z) and
the sensitivity function S(z) = 1/(1 + C(z)P(z)) are drawn in Figures 12.19 and
12.20.
To practically verify the 1Dcontroller C(z), experiment is carried out with dSpace
1103 on TMS320C240 DSP board and the LDV used to measure the displacement of
the actuator. The measured openloop frequency response compared with the simu
lated one is shown in Figure 12.19, and the crossover frequency is 1.2 kHz, the gain
margin is 9 dB, and the phase margin is 54 degrees. Figure 12.20 also shows the
measured sensitivity function with comparison to the simulated one. Additionally,
the step response and the control signal of the closed loop are taken from the oscillo
248 Modeling and Control of Vibration in Mechanical Systems
scope as a bmp picture shown in Figure 12.21, which testiﬁes that the 1D controller
drawn from the designed 2D controller works well.
The results in this section show that the 1D controller C(z) extracted from the 2
D controller can stabilize the plant P(s) in the time dimension. However, as shown
in Figure 12.3, one 1D feedback controller C(z) is not capable enough to contain
error propagation and stabilize the SSTW process modeled in (12.30)−(12.37) in
two dimensions. On the basis of the 2D model, the designed 2D controller can
stabilize the SSTW process and minimize the position error simultaneously.
Bode Diagram
Frequency (Hz)
P
h
a
s
e
(
d
e
g
)
M
a
g
n
i
t
u
d
e
(
d
B
)
−30
−20
−10
0
10
20
30
10
0
10
1
10
2
10
3
10
4
−90
−45
0
45
90
135
FIGURE 12.18
Frequency response of controller (A
c2
, B
c2
, C
c
, D
c
).
12.9 Conclusion
This chapter has employed 2D controllers in stabilization and error minimization
of systems that have 2D behavior and can be modeled as a 2D model. The ap
plication in a selfservo track writing process described by a 2D model has been
addressed in detail. By applying the twodimensional model, the error propagation
TwoDimensional H
2
Control for Error Minimization 249
10
2
10
3
10
4
−60
−40
−20
0
20
40
M
a
g
n
i
t
u
d
e
(
d
B
)
Measured
Simulated
10
2
10
3
10
4
−600
−500
−400
−300
−200
−100
0
100
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
FIGURE 12.19
Openloop frequency response with controller (A
c2
, B
c2
, C
c
, D
c
).
10
2
10
3
10
4
−40
−35
−30
−25
−20
−15
−10
−5
0
5
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
Measured
Simulated
FIGURE 12.20
Sensitivity function with controller (A
c2
, B
c2
, C
c
, D
c
).
250 Modeling and Control of Vibration in Mechanical Systems
FIGURE 12.21
Step response (Channel 1/2/3: Reference/Output/Control signal).
containment problem and the TMR minimization problem of the selfservo track
writing process has been formulated as a 2D H
2
control problem. With the stored
error information of the preceding track, the adopted 2D controller is applicable.
The 2D approach provides a systematic way of achieving both the error propagation
containment and the TMR minimization simultaneously. The simulation results have
demonstrated that the error propagation is prevented by the 2D control scheme, and
good positioning accuracy is achieved with the 2D H
2
control scheme. Also, the
time dimensional portion, as a 1D controller, of the designed 2D controller has
been implemented via LDV and dSpace and the implementation results have been
shown to verify that the 1D controller performs well.
13
Nonlinearity Compensation and Nonlinear
Control
13.1 Introduction
Nonlinearities such as friction in the actuator pivot are known to limit the low fre
quency gain of a control loop. Translated to the error rejection function or sensitivity
function, it lifts the magnitude of the sensitivity function at low frequencies, and
thus reduces the ability of the control loop to reject vibrations at lowfrequencies and
affects the system performance. Based on an identiﬁed friction model, the friction
can be compensated for by injecting an estimated friction force into the actuator. A
friction compensation method based on a nonlinear hysteresis model is thus studied.
Moreover, on the basis of a linear feedback control, to further improve the re
jection of lowfrequency disturbances such as nonlinear disturbances arising from
friction torque or bias or other unknown disturbances, an adaptive nonlinear com
pensation scheme will be adopted in this chapter to cancel their effects through a
proper estimation of the disturbances.
13.2 Nonlinearity compensation
As stated in Chapter 2, the nonlinear friction model f
e
= F(x(k)) is identiﬁed as
in the form of (2.20) using the operator based method. With the model, the friction
can be compensated for by injecting the friction force f
e
into the plant, as seen in
Figure 13.1. A sinusoidal signal of 50 Hz and 1 V amplitude is injected as the
reference. The input u versus the actuator displacement x is compared for the cases
of with and without the nonlinear compensation in Figure 13.2. It is evident that
with the nonlinear compensation the relationship between u and x is linearized very
well. In the frequency domain, the actuator frequency responses before and after the
compensation are measured with a swept sine wave via a DSA and shown in Figure
13.3. The compensated magnitude and phase responses approach those of the pure
double integrator much more closely than those before compensation.
With the friction compensation, the VCM actuator frequency responses are mea
251
252 Modeling and Control of Vibration in Mechanical Systems
sured with different sinusoidal reference amplitudes and plotted in Figures 13.4 and
13.5, where the straight smooth lines are from the pure double integrator. It is seen
that the linearization effect becomes better when the reference amplitude is higher
than 1 V (0.5 µm/V), and it is not so satisfactory for 0.25 V. Although the friction
model is obtained on the basis of the measurement for 0.5, 1, and 3 V displacement
amplitudes, the compensation based on the obtained model is able to achieve good
linearization effect for any displacement ranging from 0.5 V and above.
FIGURE 13.1
Friction compensation for the actuation system.
Figure 13.6 shows the corresponding simulated and measured sensitivity func
tions. With the compensation the magnitude at 10 Hz is reduced by around 20 dB
due to the increased openloop gain at low frequencies as seen in Figures 13.4 and
13.5. The slightly lower magnitude from 60 to 100 Hz before compensation, corre
sponding to the higher magnitude of the original actuator model from 60 to 100 Hz
in Figure 13.3, is caused by the nonlinearity of the original actuator.
In the above, a model based nonlinearity compensation has improved the ability of
the closed control loop to reject vibrations in low frequency range. In the rest of the
chapter, a nonmodel based compensation, which is an adaptive and a more ﬂexible
scheme, will be applied to compensate the nonlinear or unknown vibrations in low
frequency range.
13.3 Nonlinear control
While a KYP Lemmabased linear control can achieve disturbance rejection over
some chosen frequency ranges, it cannot run away from performance limitation,
Nonlinearity Compensation and Nonlinear Control 253
−1.5 −1 −0.5 0 0.5 1 1.5
−1.5
−1
−0.5
0
0.5
1
1.5
x 10
−3
Displacement (V, 0.5 µm/V)
I
n
p
u
t
u
(
V
)
No compensation
With compensation
FIGURE 13.2
Input u versus displacement x with and without compensation.
10
1
10
2
40
50
60
70
80
90
100
After comp.(measured)
After comp.(simulated)
1/s
2
Before comp.
10
1
10
2
−400
−300
−200
−100
0
Frequency(Hz)
P
h
a
s
e
(
d
e
g
)
M
a
g
n
i
t
u
d
e
(
d
B
)
FIGURE 13.3
Actuator frequency responses with and without friction compensation.
254 Modeling and Control of Vibration in Mechanical Systems
10
1
10
2
40
60
80
100
10
1
10
2
−300
−200
−100
0
Frequency(Hz)
P
h
a
s
e
(
d
e
g
)
M
a
g
n
i
t
u
d
e
(
d
B
)
0.25V
0.5V
FIGURE 13.4
Actuator frequency responses with friction compensation for different displacements
in voltage with 0.5µm/V (Straight smooth lines: the pure double integrator).
10
1
10
2
40
60
80
100
10
1
10
2
−300
−200
−100
0
Frequency(Hz)
P
h
a
s
e
(
d
e
g
)
M
a
g
n
i
t
u
d
e
(
d
B
)
1V
3V
FIGURE 13.5
Actuator frequency responses with friction compensation for different displacements
in voltage with 0.5 µm/V (Straight smooth lines: the pure double integrator).
Nonlinearity Compensation and Nonlinear Control 255
10
1
10
2
10
3
−90
−80
−70
−60
−50
−40
−30
−20
−10
0
10
After compensation
After compensation
Before compensation
M
a
g
n
i
t
u
d
e
(
d
B
)
Frequency(Hz)
FIGURE 13.6
Sensitivity functions with and without friction compensation.
as observed from the well known Bode integral constraint. For example, the KYP
Lemmabased control design in Chapter 8 results in an excellent attenuation of dis
turbance around 650 Hz. But its ability in rejecting low frequency disturbance is
not very desirable. On the other hand, in Chapter 2, the lowfrequency disturbance
is modeled as the output of an adaptive nonlinear scheme with the error signal as
the input. Here it is used to compensate for the lowfrequency disturbance, i.e., a
nonlinear controller is augmented with the KYP Lemmabased linear feedback con
troller, which remarkably improves the system disturbance rejection capability in
low frequency range without sacriﬁcing performance at other frequencies.
Considering Figure 13.7 with plant P(z): (A
p
, B
p
, C
p
, D
p
), we have the fol
lowing discrete time statespace realization:
x(k + 1) = A
p
x(k) + B
p
u(k) +B
p
d
1
(k), (13.1)
z(k) = −C
p
x(k) + w(k) −D
p
u(k) − D
p
d
1
(k) −d
2
(k), (13.2)
e(k) = −C
p
x(k) + w(k) −D
p
u(k) − D
p
d
1
(k) −d
2
(k) + n(k). (13.3)
The design of control law u = u
L
+ u
N
includes two parts:
1. The ﬁrst is to design a linear dynamic output feedback controller u
L
= C(z)e
for the plant P(z) such that the closedloop system is stable and satisﬁes the
performance in (8.2).
256 Modeling and Control of Vibration in Mechanical Systems
FIGURE 13.7
Control structure of a plant P(s) with Youla parametrization approach and adaptive
nonlinear compensation.
2. The second is to design a nonlinear control law u
N
such that the contribution
of the lowfrequency disturbance in d
1
to the error can be compensated.
13.3.1 Design of a composite control law
The linear control law design based on the KYP Lemma in Chapter 8 can give good
rejection of disturbance of particular frequencies such as that around 650 Hz. How
ever, the performance at lowfrequency needs to be improved. Thus we shall design a
nonlinear compensation based on the nonlinear modeling discussed in Section 2.4.2.
Note that the modeling in Section 2.4.2 is based on the time history of the mea
surement e. We consider the modeled disturbance
ˆ
d
1
given by
ˆ
d
1
= ˜ ω
T
(k)s(Φ
k
) (13.4)
with the update law
˜ ω(k + 1) = (1 − δ)˜ ω(k) −Γs(Φ
k
)e(k), (13.5)
where s(Φ
k
) is given in (2.56).
As demonstrated later, by injecting
u
N
(k) = −
ˆ
d
1
(13.6)
to the plant combined with u
L
as shown in Figure 13.7, we are able to compensate
the lowfrequency disturbance in d
1
, which is modeled as in (2.55).
Nonlinearity Compensation and Nonlinear Control 257
13.3.2 Experimental results in hard disk drives
The plant under consideration and the linear controller are the same as in Section
8.6 of Chapter 8. The nonlinear control signal u
N
in Figure 13.7 is calculated from
(13.5)−(13.6) and (2.56) in Section 2.4.2 of Chapter 2. The center positions c
e
i
and
c
∆e
i
for the measurement e and velocity ˙ e are chosen as zero. The variances are
σ
2
e
i
= σ
2
∆e
i
= 10, i = 1, ..., p. The forgetting factor δ = 0.5. Γ affects the learning
speed and should be selected to be as large as possible.
To evaluate the disturbance rejection performance of the combined linear control
C(z) designed in Section 8.6 in Chapter 8 and the nonlinear control (13.6), a sinu
soidal signal with the logarithmically spaced frequency from 10 Hz to 22.5 kHz is
respectively injected as w in Figure 13.7. For each frequency sinusoidal input, the
error signal e will involve multiple frequency components due to the nonlinear con
trol. In this situation, it is reasonable that the error rejection capability is directly
measured as the amplitude ratio e/w in time domain. At each frequency point, the
error rejection e/w is then plotted and shown in Figure 13.8.
The error rejection capability is evaluated for each value of p = 1, 5, 9. As in
Figure 13.8, the two cases with p = 1 and p = 5 give similar results, and both are
better than that given by p = 9. This implies that a higher p may not necessarily
lead to a better result. This phenomenon is consistent with the observation in mod
eling. Overall, from the simulation result, the nonlinear control produces a better
rejection of disturbances of low frequencies, while not affecting the high frequency
disturbance rejection performance.
Figure 13.8 also shows the effect of Γ on the error rejection. 10% of the Γ value
in (2.58) is used in the calculation. It is observed that the larger Γ yields a better
accuracy. This also agrees with the modeling result in Section 2.4.2 in Chapter 2.
Consequently, corresponding to the error rejection in Figure 13.8 with p = 1, the
power spectrum of PES NRRO is shown in Fig 13.9. It is observed that the error
is much lowered by 80% before 400 Hz, and no cost in higher frequency range is
paid. Overall it is evaluated from calculation that the 3σ of the true PES NRRO
is improved from 6.0 nm with the KYP Lemma method to 5.5 nm with the KYP
Lemmabased linear control augmented with the nonlinear control.
REMARK 13.1 It should be mentioned that the nonlinear compensation
scheme can be combined with any linear control to improve low frequency
vibration rejection without sacriﬁcing disturbance rejection capability in other
frequency ranges.
258 Modeling and Control of Vibration in Mechanical Systems
10
1
10
2
10
3
10
4
−90
−80
−70
−60
−50
−40
−30
−20
−10
0
10
p=1
p=5
p=9
No u
N
M
a
g
n
i
t
u
d
e
(
d
B
)
Frequency(Hz)
0.1Γ
FIGURE 13.8
Comparison of error rejection frequency response without and with u
N
of different
p and Γ.
Nonlinearity Compensation and Nonlinear Control 259
0 1000 2000 3000 4000 5000 6000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x 10
−3
KYP
KYP+nonlinear control
N
R
R
O
m
a
g
n
i
t
u
d
e
(
µ
m
)
Frequency(Hz)
400 Hz
FIGURE 13.9
NRRO power spectrum with KYP lemmabased linear control and nonlinear com
pensation (80% reduction before 400 Hz).
13.4 Conclusion
The nonlinear friction has been compensated for by injecting the modeled friction
force in Chapter 2 into the actuator. With the model based compensation, the lin
earization effect for the VCM actuator has been veriﬁed via the measurement of the
hysteresis in time domain and the frequency response in frequency domain. The
measured error rejection function showed an increased error rejection capability in
low frequency range, as a result of the compensation.
To further improve the disturbance rejection in low frequency range, the linear
control is combined with an adaptive nonlinear compensation. The simulation results
have demonstrated that the proposed controller can effectively reject disturbances at
low frequencies, resulting in a marked improvement for the 3σ value of the PES
NRRO in the data storage system.
14
Quantization Effect on Vibration Rejection
and Its Compensation
14.1 Introduction
A/D and D/A converters are inevitable in digital control systems. Chapter 5 men
tioned quantization performed by the A/D converter. This chapter investigates the
quantization effect on vibration rejection capability of the closedloop control sys
tem. With a proposed quantizer model, the inﬂuence of different quantizer bits on
the error rejection ability in low frequency range is analyzed and evaluated based
on frequency response measurement. On the other hand, it is known that nonlinear
behaviors such as actuator pivot friction limit the low frequency gain of the open
loop. Translated to the error rejection function or sensitivity function, it lifts the
magnitude at low frequencies and thus reduces the ability of the closedloop system
to reject vibrations in the low frequency range. Therefore, quantization and friction
induced problems should be treated differently for more effective control.
A simple and low cost scaling scheme is developed in this chapter to compen
sate for the effect of the quantizer. It is demonstrated that the compensation scheme
is effective in improving the sensitivity function in the low frequency range with
out deteriorating performances at other frequencies. With the compensation for the
quantization effect, the impact on the rejection ability is mostly due to the friction.
As such, the effects from quantization and friction on the error rejection function can
be differentiated, and the quantization and friction induced problems can be tack
led separately in the control loop. Additionally, through the proposed quantization
model and measurement methodology, suitable bit resolution for the quantizer can
be identiﬁed with and without the compensation.
14.2 Description of control system with quantizer
The system under investigation is the VCM actuator in a commercial 1.8−inch disk
drive with a spindle motor rotational speed of 4200 RPM. The experiment setup is
shown in Figure 14.1, where the spindle driver is used to spin the spindle motor, the
261
262 Modeling and Control of Vibration in Mechanical Systems
LDV is to measure the position of the read/write head, and a VCM driver to drive
the VCM actuator. The LDV displacement range is set as 2 µm/V. The frequency
response of the VCM actuator is measured via DSA by injecting a swept sine wave
of 5 mV amplitude. Due to experimental limitations, measurements in this chapter
are taken when the head is positioned in the middlediameter region of the disk. The
measured frequency response of the VCM actuator is shown in Figure 14.2, and its
model P(s) is obtained as
zeros = 10
5
× [1.6965; −0.0077 ±0.2575j];
poles = 10
5
× [−1.6965; −0.0251 ± 0.5020j;
−0.0101 ± 0.2511j; −0.0019 ± 0.0073j];
gain = −5.3167 × 10
17
.
A PID controller combined with notch ﬁlters is designed for the VCM actuator and
shown in Figure 14.4. The closed servo loop of the VCM actuator is seen in Figure
14.3, where dSpace DS1103 with TMS320C240 DSP on board are used to imple
ment the controller with the sampling rate of 30 kHz. The quantizer model Q(·) is
given by
Q(e) =
max(e)
2
n
× floor
e ×
2
n
max(e)
+ 0.5
, (14.1)
where n is the quantizer bit, max(e) means the maximum amplitude of the error
signal e, and floor(e) means rounding e to the nearest integer towards minus inﬁnity.
The sensitivity function of the control loop without Q(e) is given by
S(z) =
1
1 + P(z)C(z)
. (14.2)
A swept sine wave with 50 mV amplitude is injected as the reference signal to the
closed control loop and the sensitivity function S(z) is measured and plotted in Fig
ure 14.5. It is known that as the excitation level to the VCM actuator decreases, the
VCM actuator gain in the lowfrequency range is lowered due to friction nonlinearity
effect [40], which correspondingly leads to the increased S(z). This is illustrated
also in Figure 14.5, where S(z) changes for different reference amplitudes. In this
chapter, we focus on the effect of quantizer (14.1) on the sensitivity function and
seek to differentiate the quantization and friction nonlinearity effect.
With the quantizer Q(e), the sensitivity function is given by
S
Q
(z) =
e
reference
=
1
1 + P(z)Q
e
(z)C(z)
(14.3)
where
Q
e
(z) =
e
3
(z)
e(z)
(14.4)
stands for an approximation of the quantizer (14.1).
Quantization Effect on Vibration Rejection and Its Compensation 263
FIGURE 14.1
Experimental setup.
10
1
10
2
10
3
10
4
−40
−20
0
20
40
60
M
a
g
n
i
t
u
d
e
(
d
B
)
Measured
Modeled
10
1
10
2
10
3
10
4
−600
−400
−200
0
200
Frequency(Hz)
P
h
a
s
e
(
d
e
g
)
FIGURE 14.2
Frequency response of the VCM actuator measured by injecting a swept sine wave
of 5mV amplitude.
264 Modeling and Control of Vibration in Mechanical Systems
FIGURE 14.3
The servo loop in experiment.
10
1
10
2
10
3
10
4
−20
−15
−10
−5
0
5
10
15
M
a
g
n
i
t
u
d
e
(
d
B
)
10
1
10
2
10
3
10
4
−100
−50
0
50
100
P
h
a
s
e
(
d
e
g
)
Frequency(Hz)
FIGURE 14.4
Frequency response of the controller C(z).
Quantization Effect on Vibration Rejection and Its Compensation 265
10
1
10
2
10
3
10
4
−80
−70
−60
−50
−40
−30
−20
−10
0
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
0.05v
0.1v
0.5v
1v
FIGURE 14.5
Frequency response of the sensitivity function S(z) with different reference levels
(i.e., actuator moving ranges are different).
266 Modeling and Control of Vibration in Mechanical Systems
14.3 Quantization effect on error rejection
In this section, we investigate how the quantizer (14.1) with different bits affects the
error rejection function, and showthat the lower lowfrequency disturbance rejection
reﬂected in the sensitivity function may also be caused by quantization in addition to
friction.
14.3.1 Quantizer frequency response measurement
Before we proceed to investigate the quantization effect on error rejection, we exam
ine the transfer function (14.4) for the quantizer (14.1) with different bits n through
the measurement of its frequency response.
The frequency response of e
3
over e in Figure 14.3 was measured via DSA by
injecting the swept sinusoidal signal as the reference signal. The quantizer model
of the form in (14.1) with max(e) = 0.2 V and different resolution bits of n =
6, 8, 10 are investigated respectively. The dashed curves in Figures 14.6−14.8
are the corresponding measured frequency responses of Q(e). It is observed that the
magnitude difference at frequencies less than 100 Hz between bits 6 and 8 are almost
10 dBand the phase difference is about 250 deg. From200 Hz upwards, the quantizer
gains in the three cases are exactly unity. Moreover, as the bit number increases, the
quantizer approaches unity gain. When bit n = 10, it can be approximated as 1. The
frequency response when n = 12 is almost the same as that when n = 10, and thus
is omitted here.
14.3.2 Quantization effect on error rejection
Figures 14.6−14.8 show that the bit number n affects Q(e) mainly in low frequency
range. In addition, we shall see that Q(e) with lower bit number n will deteriorate
the error rejection capability of the servo system in low frequency range.
The measured sensitivity functions with different quantizer bits are shown in Fig
ure 14.9, where the effect of the bit n on the lowfrequency part can be seen. The
averaged difference of S(z) at the lowfrequency range when the bit changes from
10 to 6 is about 10 dB. The trend is that the lower number of bits leads to a lower
effective magnitude of the compensated open loop, and thus higher S(z), which
means poorer error rejection ability.
Note that a lower bit number means that the known part due to the quantization
is less. When the error signal is too low for the A/D converter to differentiate, a
high level of error rejection can not be reﬂected in the sensitivity transfer function.
In Figure 14.5, the lowest sensitivity function level at low frequencies is below −60
dB. This implies that a gain of at least 1000 requires more than 10 bit resolution.
Hence, as shown in Figure 14.9, for the two cases of n = 10 and n = 12, no obvious
impact on the sensitivity functions is seen.
Quantization Effect on Vibration Rejection and Its Compensation 267
10
1
10
2
10
3
−50
−40
−30
−20
−10
0
10
20
M
a
g
n
i
t
u
d
e
(
d
B
)
After compensation
Before compensation
10
1
10
2
10
3
−400
−300
−200
−100
0
100
200
Frequency(Hz)
P
h
a
s
e
(
d
e
g
)
FIGURE 14.6
Frequency response of the quantizer before and after compensation (bit number n =
6).
10
1
10
2
10
3
−20
−10
0
10
20
M
a
g
n
i
t
u
d
e
(
d
B
)
After compensation
Before compensation
10
1
10
2
10
3
−150
−100
−50
0
50
100
150
Frequency(Hz)
P
h
a
s
e
(
d
e
g
)
FIGURE 14.7
Frequency response of the quantizer with compensation (bit number n = 8).
268 Modeling and Control of Vibration in Mechanical Systems
10
1
10
2
10
3
−10
−5
0
5
10
M
a
g
n
i
t
u
d
e
(
d
B
)
After compensation
Before compensation
10
1
10
2
10
3
−200
−100
0
100
200
Frequency(Hz)
P
h
a
s
e
(
d
e
g
)
FIGURE 14.8
Frequency response of the quantizer with compensation (bit number n = 10).
10
1
10
2
10
3
10
4
−80
−70
−60
−50
−40
−30
−20
−10
0
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
n=12
n=10
n=8
n=6
FIGURE 14.9
Measured sensitivity function S
Q
(z) with different bits n.
Quantization Effect on Vibration Rejection and Its Compensation 269
We take S
Q
(f) at f = 10 Hz as the representative value of S
Q
(z) at low
frequencies. Figure 14.10 shows the relation of S
Q
(f) versus bit number n at
f = 10 Hz, which is decreasing dramatically until n = 10. This means that bit
number n = 10 is necessary for satisfactory performance. Approximately S
Q
(z) =
0.4604n
3
−12.1375n
2
+99.2833n−303.2000. Similar to Figure 14.5, S
Q
(z) with
a ﬁxed bit number n also changes with plant excitation levels. The trend of S
Q
(z)
with the bit number n is almost the same for each excitation level.
6 7 8 9 10 11 12
−64
−62
−60
−58
−56
−54
−52
−50
−48
−46
−44
Bit n

S
(
f
)

f
=
1
0
FIGURE 14.10
Sensitivity function S
Q
(f) with f = 10 Hz versus bit n.
14.4 Compensation of quantization effect on error rejection
In this section, a scaling method is used to compensate for the effect of the quantizer
on the sensitivity function at lowfrequencies. The compensation scheme is shown in
Figure 14.11. When the error signal amplitude is less than a threshold δ, it is scaled
up by a factor a > 1, and then scaled down by the factor a
−1
after undergoing quan
tization. The method of choosing the threshold δ and the scaling factor a is illustrated
in Figure 14.12. M is the value at the beginning point where the quantization effect
270 Modeling and Control of Vibration in Mechanical Systems
FIGURE 14.11
Compensation scheme of quantization effect.
can be seen obviously. D is the biggest difference between S
Q
and S. δ and a can
be generally chosen as follows.
δ = 10
M/20
· reference, a = 10
D/20
. (14.5)
10
1
10
2
10
3
10
4
−80
−70
−60
−50
−40
−30
−20
−10
0
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
Target S
Measured S
Q
M
D
FIGURE 14.12
Choosing the threshold δ and the scaling factor a.
The compensation is carried out for three cases with n = 6, 8, and 10 respectively,
and the improvement of the error rejection after compensation will be discussed.
A. n = 6
Quantization Effect on Vibration Rejection and Its Compensation 271
In this case, from the sensitivity function in Figure 14.9, δ and a are obtained as
δ = 10
−40/20
· 50 mV = 0.5 mV, a = 10
18/20
= 8.
Figure 14.6 shows the improved frequency response of the quantizer after com
pensation. The sensitivity function improvement after compensation in the low fre
quency range is shown in Figures 14.9 and 14.13. There is no further improvement
when a is increased. When a = 5, the result is as good as when a = 8, which means
that the optimal value of a is roughly between 5 and 8.
10
1
10
2
10
3
10
4
−80
−70
−60
−50
−40
−30
−20
−10
0
10
Frequency(Hz)
M
a
g
n
i
t
u
d
e
(
d
B
)
n=10
n=8
n=6
FIGURE 14.13
Sensitivity function with quantization compensation.
B. n = 8
When n = 8,
δ = 10
−48/20
· 50 mv = 0.2 mv, a = 10
15/20
= 5.6.
Figures 14.7 and 14.13 show the obvious improvement of the quantizer and the
sensitivity function after compensation. When a is increased to 10 and above, no
further improvement is observed. Thus the best value of a can be chosen around 5.6.
C. n = 10
In the case with bit n = 10,
δ = 10
−48/20
· 50 mV = 0.2 mV, a = 10
10/20
= 3.
272 Modeling and Control of Vibration in Mechanical Systems
TABLE 14.1
Quantization and friction effect
Quantizer bit n 6 8 10
Actually measured S
f=10Hz
(dB) −45 −50 −65
Compensated quantization effect (dB) 13 18 4
Estimated S
f=10Hz
with friction effect (dB) −58 −68 −69
No obvious difference can be seen in Figure 14.8. When compared with Figure
14.9, Figure 14.13 shows the obvious improvement of the sensitivity function after
compensation. When a = 6, the result is almost the same. As a is increased to 10,
the result becomes worse. Thus in this case a = 3 can be chosen as one of the best
scaling factors.
It can be seen fromFigure 14.13 that after compensation, S
Q
(z) is much closer to
S(z) in the dotted line in Figure 14.5, which is measured without the quantizer and
thus considered as subjected to nonlinear friction effect only, e.g., S(z)
f=10Hz
=
−69 dB with friction effect alone. From Figure 14.13, we can also see that instead
of bit number n = 10 previously shown in Section 14.3.2 without compensation, bit
number n = 8 is adequate with the compensation for a similar performance.
It has been shown that the low frequency range of the sensitivity function is also
affected by quantization in addition to actuator pivot friction. When the quantiza
tion effect is compensated for sufﬁciently, the low frequency portion of the sensi
tivity function can be regarded as the estimated effect from friction alone. Thus, by
comparing the compensated and the uncompensated S
Q
(z) with Figure 14.13, the
quantization and the friction effect on S
Q
(z) can then be differentiated and shown in
Table 14.1. We can see that the estimated friction effect for n = 6 is not as accurate
as the other two cases. This is because the compensation for n = 6 is insufﬁcient,
as seen in Figure 14.13. In this situation, a multistage scaling, i.e., a = a
i
for
δ
i1
< e < δ
i2
, i = 1, 2, ..., may be necessary.
To this end, the scaling compensation scheme is conducted for quantization com
pensation in the disk drive. The series of results demonstrates that the compensation
scheme is effective in improving the sensitivity function in the low frequency range
without inﬂuencing other frequencies. Additionally, it is noted that the implementa
tion of the scheme is simple and therefore incurs low cost.
14.5 Conclusion
The quantization effect on the closedloop system performance has been investi
gated. The frequency responses of the quantizers with different bits have been mea
sured and analyzed. Its effect on the error rejection function or sensitivity function
at low frequencies has shown that the rejection ability of the servo loop for low
Quantization Effect on Vibration Rejection and Its Compensation 273
frequency disturbances is relevant to the quantizer in addition to the pivot friction.
Moreover, a simple and lowcost scaling scheme has been used to compensate for the
effect of the quantizer. With sufﬁcient compensation, the effects due to quantization
and friction can be differentiated, and thus the friction and the quantization impact on
the system performance such as disturbance rejection capability can be treated sepa
rately. Through the proposed quantization model and measurement methodology, a
suitable bit resolution for the quantizer can be easily identiﬁed while the pivot fric
tion nonlinearity effect can be decoupled and tackled for more effective closedloop
system control.
15
Adaptive Filtering Algorithms for Active
Vibration Control
15.1 Introduction
Many advanced control techniques have been studied for active noise and vibration
control, including optimal and robust control and adaptive control strategies. The ad
vance of computer technologies has made digital signal processing techniques much
more useful in modern control systems. Adaptive ﬁlters have been widely used in
the implementation of adaptive algorithms for active vibration control (AVC). Sam
pling time constraint is the major drawback of a digital control system that requires
very high processing speed for real time control. Adaptive ﬁltering offers signiﬁcant
advantages over passive silencers at low frequencies where lower sampling rates are
adequate.
There are two main approaches to adaptive ﬁltering in AVC, feedforward and feed
back adaptive ﬁltering. The feedforward algorithm requires the source information
in order to attenuate the vibration. In many applications, the source of vibration is
impractical or expensive to measure, and a feedforward algorithm becomes impos
sible or difﬁcult to implement in those applications. A feedback algorithm has an
advantage of utilizing only the vibration signal to be controlled.
15.2 Adaptive feedforward algorithm
Figure 15.1 shows the socalled ﬁlteredX LMS (FXLMS) algorithmwhich was ﬁrst
introduced by Bernard Widrow in 1981 [37].
The ﬁlteredX LMS algorithm is derived as follows with the notation deﬁned in
Chapter 3. The error signal e(k) is given by
e(k) = d(k) + y(k) = d(k) + P(z)u(k)
= d(k) + P(z)(W
T
(z)x(k)). (15.1)
275
276 Modeling and Control of Vibration in Mechanical Systems
FIGURE 15.1
Block diagram of FXLMS algorithm.
Considering that
∇e(k) =
∂e(k)
∂W(k)
= p(k) ∗ x(k), (15.2)
the gradient estimate of the meansquared error will be
∇
ˆ
ξ(k) = ∇e
2
(k) = 2[∇e(k)]e(k)
= 2[p(k) ∗ x(k)]e(k), (15.3)
where p(k) is the impulse response of P(z). In a practical situation, an exact model
P(z) is not available, and therefore its estimated model
ˆ
P(z) is used to represent
P(z) in the algorithm. Either an FIR or IIR ﬁlter can be used to model the AVC
system. Then the reference signal for the LMS algorithm will be
ˆ x(k) =
ˆ
P(z)x(k). (15.4)
ˆ x(k) is called the ﬁltered reference signal because the reference input signal is passed
through the estimated model of the AVC system. Equation (15.3) becomes
∇
ˆ
ξ(k) = 2[ ˆ p(k) ∗ x(k)]e(k) = 2ˆ x(k)e(k). (15.5)
The coefﬁcients of the weight vector will be updated by the following equation.
Adaptive Filtering Algorithms for Active Vibration Control 277
W(k + 1) = W(k) −
µ
2
∇
ˆ
ξ(k)
= W(k) −µˆ x(k)e(k). (15.6)
The resulting adaptive algorithm is known as the FilteredX LMS algorithm.
In the normalized case, the algorithmis known as the FilteredXNLMS (FXNLMS)
algorithm whereby the weight vector will be updated by
W(k + 1) = W(k) − µ(k)ˆ x(k)e(k), (15.7)
µ(k) =
α
ε + ˆ x
T
(k)ˆ x(k)
. (15.8)
So far, the feedforward adaptive algorithm has been discussed. The feedforward
algorithmrequires the reference signal that cannot be available in many applications.
To overcome this problem an adaptive feedback algorithm is considered below.
15.3 Adaptive feedback algorithm
There is another algorithm that combines the traditional feedback control and adap
tive ﬁltering approach, and is therefore referred to as the Adaptive Feedback Algo
rithm. The algorithmutilizes only the feedback error signal to cancel the disturbance
vibration. Since there is no direct reference information available for the vibration
control, the disturbance signal is regenerated (extracted) fromthe error signal and the
approximated input reference signal is then fed back to the adaptive ﬁlter, as shown
in Figure 15.2.
An estimate of the reference signal x(k) is obtained by subtracting the estimated
cancellation signal ˆ y(k) of y(k) from the error signal e(k).
ˆ x(k) =
ˆ
d(k) = e(k) − ˆ y(k). (15.9)
The weight updating algorithm is the same as the feedforward control scheme,
FXLMS or FXNLMS. The feedback loop W(z)
ˆ
P(z) will introduce poles to the
system. The characteristic equation of the system will be:
α(z) = 1 + W(z)
ˆ
P(z). (15.10)
Adaptive feedback control can be seen as adaptive inverse control where an adap
tive ﬁlter is used to track the inverse model of P(z). The analogy is illustrated in
Figure 15.3.
In this control system, a compensator C(z) will adapt to track the inverse model
P
−1
(z) [38]. The system is analyzed as follows.
X
′
(z) = R(z) −X(z) = R(z) − E(z) +
ˆ
P(z)C(z)X
′
(z), (15.11)
278 Modeling and Control of Vibration in Mechanical Systems
i.e.,
X
′
(z) =
R(z) − E(z)
1 −
ˆ
P(z)C(z)
. (15.12)
E(z) = D(z) +
P(z)C(z)R(z) − P(z)C(z)E(z)
1 −
ˆ
P(z)C(z)
, (15.13)
from which,
E(z) =
D(z)(1 −
ˆ
P(z)C(z))
1 −
ˆ
P(z)C(z) + P(z)C(z)
+
R(z)
1 −
ˆ
P(z)C(z) + P(z)C(z)
. (15.14)
With the reference signal R(z) set to zero, the transfer function from D(z) to E(z)
will be:
E(z)
D(z)
=
1 −
ˆ
P(z)C(z)
1 −
ˆ
P(z)C(z) + P(z)C(z)
. (15.15)
If the estimated model
ˆ
P(z) is exactly the same as P(z) , (15.15) will be reduced to:
E(z)
D(z)
= 1 −P(z)C(z). (15.16)
An adaptive ﬁlter will be applied in the place of C(z) and the ﬁlter will adapt
its weight vector to approach the inverse model of P(z), P
−1
(z) [38]. When C(z)
becomes the inverse model of P(z), the effect of D(z) over E(z) will be totally
eliminated.
Therefore correct modeling of P(z) is crucial to the adaptive feedback control
system. In a practical situation, it is difﬁcult to obtain the exact model of the system.
Modeling error may lead the system to an unstable situation.
Adaptive Filtering Algorithms for Active Vibration Control 279
FIGURE 15.2
FilteredX LMS adaptive feedback algorithm.
FIGURE 15.3
Adaptive inverse control scheme.
280 Modeling and Control of Vibration in Mechanical Systems
15.4 Comparison between feedforward and feedback controls
In spite of the advantage of not requiring the information over feedforward control,
adaptive feedback control can only be applicable for narrow band disturbance rejec
tion. The traditional feedback controller cannot react to the disturbance before the
control error has already occurred. But the adaptive ﬁlter has the ability to capture
the statistics of the disturbance signal. For periodic disturbance signals, it is possible
for the adaptive feedback control to track the frequency of the signal and progres
sively attenuate the disturbance signal. Therefore adaptive feedback control is still
efﬁcient for pure sinusoid signals.
On the other hand, feedforward control has an ability to reject the wide band dis
turbances, because the controller receives the disturbance signal before it reaches the
point to be controlled and takes a control action in advance to eliminate the distur
bance impact.
An approximated model of the system appears in the feedback loop of the adaptive
feedback system, which introduces poles to the system. Therefore robustness of sta
bility of the adaptive feedback system can be improved by minimizing the modeling
error of the system.
15.5 Application in Stewart platform
We have chosen the adaptive feedback algorithm for hexapod smart structure intro
duced in Chapter 3. Since the structure has six actuators, a multiplechannel adaptive
feedback control system is studied.
15.5.1 Multichannel adaptive feedback AVC system
In a multiplechannel AVC system, let the number of the secondary sources be K
and the number of error sensors be M. The reference signal synthesizer uses K
secondary signals, M error signals, and K×M secondary path estimates to generate
M reference signals for K × M adaptive ﬁlters. The synthesized reference signals
are expressed as:
x
m
(k) = e
m
(k) +
K
n=1
ˆ p
mn
(k) ∗ u
n
(k), m = 1, 2, · · · , M, (15.17)
Adaptive Filtering Algorithms for Active Vibration Control 281
where ˆ p
mn
(k) is the impulse response of the secondarypath estimate
ˆ
P
mn
(z) and
u
n
(k) is the nth secondary signal expressed as
u
n
(k) =
M
m=1
w
nm
(k) ∗ x
n
(k), n = 1, 2, · · · , K, (15.18)
where w
nm
(k) is the impulse response of the adaptive ﬁlter W
nm
(z) .
A 2 × 2 adaptive feedback AVC system shown in Figure 15.4 is presented as an
example for multiple channel AVC systems. Two secondary signals, u
1
(k) and u
2
(k)
are generated as
u
1
(k) = w
11
(k) ∗ x
1
(k) + w
12
(k) ∗ x
2
(k), (15.19)
u
2
(k) = w
21
(k) ∗ x
1
(k) + w
22
(k) ∗ x
2
(k). (15.20)
The reference signals are synthesized as:
x
1
(k) = e
1
(k) + ˆ p
11
(k) ∗ u
1
(k) + ˆ p
12
(k) ∗ u
2
(k), (15.21)
x
2
(k) = e
2
(k) + ˆ p
21
(k) ∗ u
1
(k) + ˆ p
22
(k) ∗ u
2
(k). (15.22)
The coefﬁcients of the four adaptive ﬁlters are adjusted using the FXLMS algo
rithm expressed as:
W
11
(k + 1) = W
11
(k) + µ{[ ˆ p
11
(k) ∗ X
1
(k)]e
1
(k) + [ ˆ p
21
(k) ∗ X
1
(k)]e
2
(k)},
W
21
(k + 1) = W
21
(k) + µ{[ ˆ p
12
(k) ∗ X
1
(k)]e
1
(k) + [ ˆ p
22
(k) ∗ X
1
(k)]e
2
(k)},
W
12
(k + 1) = W
12
(k) + µ{[ ˆ p
11
(k) ∗ X
2
(k)]e
1
(k) + [ ˆ p
21
(k) ∗ X
2
(k)]e
2
(k)},
W
22
(k + 1) = W
22
(k) + µ{[ ˆ p
12
(k) ∗ X
2
(k)]e
1
(k) + [ ˆ p
22
(k) ∗ X
2
(k)]e
2
(k)}.
15.5.2 Multichannel adaptive feedback algorithm for hexapod plat
form
The hexapod smart structure has six secondary sources. If we use six error sensors,
there will be 72 adaptive ﬁlters including 36 ﬁlters for reference signal synthesizer
and 36 ﬁlters for FilteredXpurpose. There will be a very high computational burden
in real time implementation. Therefore a (6 × 1) adaptive control scheme, as shown
in Figure 15.5, has been chosen for the smart structure with only one sensor placed
at the center of the upper plate surface.
In this system, there are six secondary path actuators and one error sensor. The
error signal will be:
e(k) = d(k) +
6
n=1
p
n
(k) ∗ u
n
(k), (15.23)
where p
n
(k), n = 1, 2, · · · , 6 is the impulse response of the secondary path P
n
(z),
and u
n
(k), n = 1, 2, · · · , 6 is the secondary signal of the adaptive ﬁlter W
n
(z).
282 Modeling and Control of Vibration in Mechanical Systems
FIGURE 15.4
Block diagram of 2 × 2 adaptive feedback algorithm.
Adaptive Filtering Algorithms for Active Vibration Control 283
The reference signal x(k) is synthesized as an estimate of the primary disturbance.
x(k) =
ˆ
d(k) = e(k) −
6
n=1
ˆ p
n
(k) ∗ u
n
(k), (15.24)
where ˆ p
n
(k) is the impulse response of
ˆ
P
n
(z).
The FXLMS algorithm is used to minimize the error signal e(k) by adjusting the
weight vector for each adaptive ﬁlter W
n
(z) according to:
W
n
(k + 1) = W
n
(k) + µX
′
n
(k)e(k), n = 1, 2, · · · , 6, (15.25)
where X
′
n
(k) = ˆ p
n
(k) ∗X(k) is the reference signal vector ﬁltered by the secondary
path estimate,
ˆ
P
n
(z).
FIGURE 15.5
Block diagram of 6 × 1 FXLMS adaptive feedback control system.
284 Modeling and Control of Vibration in Mechanical Systems
15.5.3 Simulation and implementation
Identiﬁcation of the actuators in the platform can be found in Chapter 3. The 6 × 1
adaptive feedback controller is developed using the MATLAB Simulink platform.
Improvement and modiﬁcation on the controller structure was performed during an
experiment. There are six FilteredXLMS adaptive ﬁlters which form the main parts
of the controller. The outputs of the adaptive ﬁlters are fed into the six DA converters
of the dSPACE real time interface board (DS1104). Another major portion is the
primary disturbance synthesizer with the six estimated ﬁlters
ˆ
P
n
(z) of the secondary
paths P
n
(z). The secondary signals from the adaptive ﬁlters are also inputed into
these ﬁlters in order to regenerate the primary disturbance signal.
Normally a feedback controller can get into oscillation due to internally generated
noise. To prevent this unstable situation, small nonlinear dead zones are placed at the
output of the primary signal synthesizer and the receiving point of the error signal.
The dead zone levels are small enough to ensure that only the noise signal is prohib
ited from passing through. An automatic gain control, which will be introduced in
Section 15.5.3.2, is placed at the reference signal input to the six adaptive ﬁlters to
improve the stability of the controller.
Sampling frequency is set at 1 kHz. Therefore, the entire control process is carried
out within 1 ms for each sampled signal and the secondary signals are sent out with
1 kHz sampling rate. We may obtain better performance if the sampling frequency is
increased. But the computing demand of the six adaptive ﬁlters limits the sampling
frequency. The adaptation step size µ is set between 0.0005 and 0.002 depending on
the vibration frequency. During the experiment, the controller is able to achieve 20
dB to 30 dB attenuation for vibration of frequency from 60 Hz to 220 Hz. The result
can be further improved by minimizing the modeling error of the secondary path.
Figure 15.6 shows the experiment setup and the connections between various de
vices. The Stewart platform is mounted on a shaker (Labworks ET139) through
a custom−made mounting interface. The shaker is powered by a power ampliﬁer,
which in turn is driven by a signal generator. A sine wave is used as the reference
signal for the shaker. The piezoelectric accelerometer (357B21 from PCB Piezoelec
tronics) is placed at the center of the top plate to detect the error signal of the system
as a whole. A charge ampliﬁer (Sinocera YE5852) will amplify the detected error
signal from the accelerometer and ﬁlter the high frequency noise before the signal
is sent to the controller via the ADC unit. An interface unit of DS1104 DSP board,
which includes peripheral outlets for the DS1104 card, serves as a junction point
among the DSP board, charge ampliﬁer and DMU (Drive and Monitoring Unit).
Adaptive Filtering Algorithms for Active Vibration Control 285
15.5.3.1 Experimental results
Experiments are conducted for different frequencies of vibration. Depending on the
frequency of the disturbance (vibration) signal, an adaptation step size µ is chosen.
Recall that the step size is inversely proportional to the reference input power; a
smaller step size µ is chosen for higher frequency since the signal power increases
as the frequency is higher.
The PZT actuator has a maximum stroke length of 50 µm. In order not to exceed
the maximum stroke length of the actuator, the vibration signal is chosen with 200
mV peaktopeak at the ADC input. Error signal at the input of the algorithmwill be
20 mV peaktopeak (−40 dB) due to the scaling factor of the A/D converter.
During the experiments, the controller is observed to be able to attenuate distur
bances with frequency up to 220 Hz. The actual resonance frequency of the system
is approximately at 230 Hz. When the vibration frequency is around 230 Hz, the
system runs into an unstable state. This may be due to a large modeling error around
the resonance frequency. Theoretically, there is a phase shift of close to 180
◦
around
the resonance frequency. The phase response of the secondary path also shows that
there is a major phase shift around the resonance frequency. But the phase response
in the identiﬁcation result may not have sufﬁcient changes to represent the resonance
region.
Therefore, the controller is modiﬁed to compensate the phase error. Inverted gain
(−1) is inserted in the error signal path and the controller is tested with frequency
starting from 240 Hz. With this modiﬁcation, disturbance frequency up to 280 Hz is
observed to be attenuated.
Figures (15.7)−(15.10) are the experimental results captured by dSPACE control
desk with frequency 60 Hz, 210 Hz, 240 Hz, and 270 Hz, respectively. The straight
line is the point where AVC is switched on. The error signal is inverted when the
disturbance frequency is set at 240 Hz and 270 Hz.
286 Modeling and Control of Vibration in Mechanical Systems
FIGURE 15.6
General layout of the experimental setup.
Adaptive Filtering Algorithms for Active Vibration Control 287
FIGURE 15.7
60 Hz error signal in dB unit.
FIGURE 15.8
210 Hz error signal in dB unit.
288 Modeling and Control of Vibration in Mechanical Systems
FIGURE 15.9
240 Hz error signal in dB unit.
FIGURE 15.10
270 Hz error signal in dB unit.
Adaptive Filtering Algorithms for Active Vibration Control 289
15.5.3.2 Controller modiﬁcation and discussion
Observe from Figure 15.7−Figure 15.10 that the level of vibration attenuation varies
with frequency. It may be because of the error of magnitude in the frequency re
sponse of the secondary path. An accurate modeling is required not only at the
fundamental frequency but also at the harmonics of the disturbance signal. Because
once the controller starts to suppress the disturbance signal, harmonics will be intro
duced in the reference signal path of the controller.
Therefore, modeling is a big challenge. Nonlinearity of the secondary path ac
tuator introduces modeling error and instability. Modeling errors include magnitude
error and phase error. Phase error is a crucial factor since the feedback control system
must be able to predict the disturbance before the control signal is sent out. Phase
error in the secondary path model of the synthesizer section will introduce the wrong
phase in the reference signal. Then the phase error in the reference signal makes the
adaptive controller output the wrong phase of cancellation signal to the PZT actuator
as well as to the synthesizer again.
The magnitude response of the secondary path model is altered by ± 6 dB, which
means that the FIR models of the secondary path are multiplied by 2 or divided by
2, to introduce the magnitude error. Then the experiments are conducted again. The
stability of the controller is not affected and there is only a slight decrease in vibration
attenuation level. Furthermore, there is a magnitude peak at 240 Hz in the frequency
response of the secondary path model. Based on the experiments, the actual peak
response is estimated to be at 230 Hz. But the controller is able to attenuate the
disturbance frequency of 240 Hz with the phase compensating gain (−1) in the error
signal path. Therefore, we can say that the adaptive feedback controller is able to
tolerate some magnitude error in the secondary path model.
During the experiment, the controller is observed to have a faster convergence
rate for a higher frequency vibration. It veriﬁes that a larger signal power at higher
frequency can drive the adaptive ﬁlter to converge at a faster rate. But the faster
convergence rate due to the higher signal power can lead the system to instability if
the adaptation step size “µ” is not sufﬁciently small. Minimizing the adaptation step
size can improve the stability of the controller but a slow convergence rate has to be
borne.
Another strategy used to compromise between the stability and convergence rate is
to use the normalized LMS algorithm. The NFXLMS algorithmis able to work well
in the feedforward algorithm. But in the practical implementation of adaptive feed
back system, surged convergence of the normalized adaptive ﬁlter is again observed
to cause some instability to the feedback system.
Therefore, instead of applying the normalized algorithm, another strategy is devel
oped and applied to the system to ensure the stability and to improve the convergence
of the adaptive algorithm. That is an automatic gain control (AGC), introduced in
the reference signal path as shown in Figure 15.11.
When the adaptive controllers start to adapt their ﬁlter weights, it is observed that
the reference signal becomes larger than the normal level before the adaptation is
triggered. It may be because of the error of magnitude of the secondary path model.
290 Modeling and Control of Vibration in Mechanical Systems
In order to prevent the actuator from being overdriven and the control system from
being driven into an unstable situation, an automatic gain control, AGC is inserted
in the reference signal path. AGC is a simple nonlinear attenuator. The reference
signal generated by the synthesizer will be transformed into a 16order vector form
by a delay line. “Max” block will ﬁlter out the maximum value of the vector signal
and this signal will be transformed back to a scalar value by “To Sample” block.
This scalar value is ampliﬁed with “Slider Gain” which can be adjusted during the
experiment to obtain the best performance. Constant value 1 is added to avoid the
division by zero. The summed value will be reciprocated by the math function “1/µ”.
This results in a variable attenuating factor for the reference signal by inputting it to
the “Cross Product” block together with the reference signal. Therefore, when the
reference signal becomes larger, AGC will attenuate more and maintain the reference
signal within a safe level. The mathematical description of the automatic gain control
is
X
o
(k + 1) =
X
i
(k + 1)
1 + gX
o
(k)
, (15.26)
where g is the AGC gain (slider gain).
The different results between with and without automatic gain control are shown
Figures 15.12 and 15.13. It can be seen that automatic gain control can greatly
improve the performance of the adaptive feedback controller. Another way of min
imizing the overgrowing of reference signal is to restrict the adaptive ﬁlter weights
by putting in a leakage factor. The leakage factor of the adaptive ﬁlter prevents the
weight vector from growing without bound after the convergence of the error signal.
A leakage factor of about 0.9999 is chosen so that 1 − 0.9999 = 0.0001 is still less
than the smallest adaptation size 0.0005 in the experiments. But experiments with
the leakage factor show that the attenuation level is decreased from 30 dB to 20 dB.
Although the leakage factor can improve the stability, it can degrade the performance
of the AVC.
15.6 Conclusion
In this chapter, with the adaptive ﬁltering algorithmbeing conceptualized, a multiple
reference adaptive feedback controller has been developed. A6×1 adaptive feedback
system has been implemented using Simulink and Control Desk. Adaptive identiﬁ
cation has been used to model the six PZT actuators in the secondary path.
Experiments have been performed with different adaptation step µ. Lower step
size has to be chosen for a high frequency vibration signal in order to take into
account its high signal power. Improperly chosen µ has been observed to affect the
stability and performance of the control system.
The regenerated reference signal has been observed to become larger than the nor
mal level before the adaptive ﬁlter has started to adapt. The leaky FXLMS algorithm
Adaptive Filtering Algorithms for Active Vibration Control 291
has been applied to restrict the overﬂowof the weight vector of the adaptive ﬁlter. But
the controller performance is degraded with introducing the leakage factor. Finally
the automatic gain controller has been considered and developed in the reference sig
nal path. The experimental results are greatly improved by ﬁnetuning the slider gain
of the automatic gain controller. Actually, the nonlinear property of the automatic
gain controller partially compensates for the modeling error so that the performance
and stability of the adaptive feedback controller are comparatively better than those
without using the automatic gain controller. For vibrations with frequencies from 60
to 220 Hz, an attenuation level of up to 30 dB has been achieved. Therefore, the
adaptive feedback controller can provide satisfactory performance among various
constraints.
FIGURE 15.11
Simulink diagram of automatic gain control.
292 Modeling and Control of Vibration in Mechanical Systems
FIGURE 15.12
180 Hz error signal without automatic gain control.
FIGURE 15.13
180 Hz error signal with automatic gain control.
References
[1] S. Deeyiengyang, and K. Ono, “Suppression of resonance amplitude
of disk vibrations by squeeze air bearing plate,” IEEE Trans. Magnet
ics, vol. 37, no. 2, pp. 820−825, 2001.
[2] B. A. Francis, and J. C. Doyle, “Linear control theory with an H
∞
optimality criteron,” SIAM J. Control and Optimization, vol. 25, pp.
815−844, 1987.
[3] S. Boyd, and C. A. Desoer, “Subharmonic functions and performance
bounds in linear timeinvariant feedback systems,” IMA J. Mathemat
ical Control and Information, vol. 2, pp. 153−170, 1985.
[4] D. E. Baxa and R. A. Dykstra, “Pneumatic isolation systems control
forging hammer vibration,” Sound and Vibration, vol. 14, pp. 22−25,
1980.
[5] A. Soon and M. S. Lee, “Optimal design of linear and nonlinear ab
sorbers for damped systems,” Journal of Vibration, Acoustics, Stress,
and Reliability in Design, vol. 105, pp. 112−119, 1983.
[6] K. Zhou, J. C. Doyle, Essentials of Robust Control, Prentice Hall Inc.,
1998.
[7] Ali Saberi, P. Sannuti, B. M. Chen, H
2
Optimal Control, Prentice Hall
Inc., 1995.
[8] T. Krysinski, F. Malburet, Mechanical Vibrations: Active and Passive
Control, ISTE Ltd, 2007.
[9] S. S. Rao, Mechanical Vibrations, AddisonWesley Publishing Com
pany, 1990.
[10] Guzik Company Website [Online], Spin Stand Products Information,
http://www.guzik.com.
[11] B. Heo, I. Y. Shen, “Taming disk and spindle rocking by damped lam
inated disks − an experimental study,” IEEE Trans. Magnetics, 35(5),
pp. 2304−2306, 1999.
[12] C. Du, G. Guo and D. Wu, “Lowhump sensitivity function design for
dualstage systems with different micro actuators,” IEE Proc. Control
Theory and Application, 152(6), pp. 655−661, 2005.
293
294 Modeling and Control of Vibration in Mechanical Systems
[13] C. Du, L. Xie, J. N. Teoh, and G. Guo, “An improved mixed H
2
/H
∞
control design for hard disk drives,” IEEE Trans. Control Systems
Technology, 13(5), Sept. 2005.
[14] C. Du, W. Wong, and G. Guo, “Experimental study of disk vibration
reduction via stacked disks,” Journal of Sound and Vibration, vol. 301,
pp. 226−235, 2007.
[15] C. Du, L Xie, and G. Guo, “A generalized KYP Lemmabased ap
proach for disturbance rejection in data storage systems,” Automatica,
43, pp. 2112−2118, 2007.
[16] C. Du, G. Guo, and W. E. Wong, “Evaluation and compensation of
quantization effect on servo performance for Hard Disk Drives,” IEEE
Trans. Magnetics, 34(11), November, 2007.
[17] C. Du, L Xie, G. Guo, “H
2
control for head positioning in axial and
radial dimensions for selfservo track writing,” IEEE Trans. Control
Systems and Technology, vol. 16, no. 1, pp. 177−181, Jan., 2008.
[18] C. Du, L.Xie, and J. Zhang, “An operator based modeling and com
pensation for VCM actuator pivot friction in a 1.8inch HDD,” Proc.
the 17th IFAC (International Federation of Automatic Control) World
Congress, Seoul, Korea, 7−11 July, 2008.
[19] C. Du, L. Xie, and F. L. Lewis, “Using blending control to suppress
multifrequency disturbances,” Proc. the 10th International Confer
ence on Control, Automation, Robotics and Vision, pp. 1708−1713,
Hanoi, Vietnam, 17−20, Dec. 2008.
[20] C. Du, L. Xie, J. Zhang, and G. Guo, “Disturbance rejection for a
data storage systemvia sensitivity loop shaping and adaptive nonlinear
compensation,” IEEE/ASME Transactions on Mechatronics, 13(5), pp.
493−501, 2008.
[21] C. K. Thum, C. Du, B. M. Chen, E. H. Ong, and K. P. Tan,
“Midfrequency runout compensation in hard disk drives via a time
varying group ﬁltering scheme,” IEEE Trans. on Magnetics, 44(12),
pp. 4769−4779, 2008.
[22] Jr. W. E. West, U. Ummethala, and F. Aguirre, “Disk vibration induced
TMR: Thicker is better,” IDEMA Insight, 1998.
[23] J. S. McAllister, “Characterization of disk vibrations on aluminum
and alternate substrates,” IEEE Trans. Magnetics, 33(1), pp. 968−973,
1997.
[24] S. Deeyiengyang and K. Ono, “Suppression of resonance amplitude of
disk vibrations by squeeze air bearing plate,” IEEE Trans. Magnetics,
37(2), pp. 820−825, 2001.
References 295
[25] K. Takashi, Y. Uematsu, T. Yamada, M. Kamimura, M. Fukushi, and
Y. Kuroba, “HDD servo technology for media level servo track writ
ing,” IEEE Trans. Magnetics, 39(2), pp. 851−856, 2003.
[26] S. Futami, A. Furutani, and S. Yoshida, Nanometer positioning and its
microdynamics, Nanotechnology, 1, pp. 3137, 1990.
[27] S. Yu, S. Mishra, and M. Tomizuka, “Online force ripple identiﬁ
cation and compensation in precision positioning of a wafer stage,”
ASME International Mechanical Engineering Congress and Exposi
tion, Seattle WA, USA, 2007.
[28] D. Y. Abramovitch, S. B. Anderson, L. Y. Pao and G. Schitter, “A
tutorial on the mechanisms, dynamics and control of atomic force mi
croscopes,” Proc. of American Control Conference, pp. 3488−3502,
Marriott Marquis Hotel at Times Square, New York City, USA, 11−13
July, 2007.
[29] H. Kataoka, K. Ohishi, T. Miyazaki, S. Katsura, and H. Takuma,
“Motion control strategy of industrial robot for vibration suppression
and little positioning phase error,” Proc. the 10th International Work
shop on Advanced Motion Control, pp. 661−666, Trento, Italy, 26−28
March 2008.
[30] D. Stewart, “A Platform with six degrees of freedom,” Proceeding In
stitute of Mechanical Engineering, 180(15), Part I, pp. 371−386, Lon
don, 1965−1966.
[31] Kai Liu, J. M. Fitzgerald, F. L. Lewis, “Kinematic analysis of a Stewart
Platform manipulator,” IEEE Transaction on Industrial Electronics,
vol. 40, no. 2, pp. 282−293, 1993.
[32] J. Spanos. Z. Rahman, and G. Blackwood, “A soft 6aixs active vibra
tion isolator,” Proceedings of the American Control Conference, The
Westin Hotel, Seattle, Washington, USA, 21−23 June, 1995.
[33] D. Thayer, M. Kampell, and J. Vagners, “Six axis vibration isolation
using modern control technique,” Proc. the 21st Annual AAS Guidance
and Control Conference, February, 1998.
[34] F. M. Raimondi, M. Melluso, “Trajectory decentralized fuzzy con
trol of multiple UAVs,” Proc. the 10th International Workshop on Ad
vanced Motion Control, pp. 455−461, Trento, Italy, 26−28 March,
2008.
[35] S. Haykin, Adaptive Filter Theory, Upper Saddle River, NJ: Prentice
Hall, 2002.
[36] S. M. Kuo and D. R. Morgan, Active Noise Control Systems Algo
rithms and DSP Implementations, John Wiley & Sons, Inc., 1996.
296 Modeling and Control of Vibration in Mechanical Systems
[37] B. Widrowand S. D. Stearns, Adaptive Signal Processing, Englewood
Cliffs, NJ: Prentice Hall, 1985.
[38] B. Widrow and E. Walach, Adaptive Inverse Control, PrenticeHall In
formation and System Science Series, Upper Saddle River, NJ: Pren
tice Hall, 1996.
[39] J. Juang and J. Chen, “On combining adaptive estimation and robust
control for friction compensation,” Proc. of the 5th World Congress
on Intelligent Control and Automation, pp. 4396−4400, Hangzhou,
China, 15−19 June, 2004.
[40] D. Abramovitch, F. Wang, and G. Franklin, “Disk drive pivot nonlin
earity modeling Part 1: Frequency domain,” Proc. of American Con
trol Conf., pp. 2600−2603, Baltimore, Maryland, USA, June, 1994.
[41] T. Yan and R. Lin, “Experimental modeling and compensation of pivot
nonlinearity in Hard Disk Drives,” IEEE Transactions on Magnetics,
vol. 39, no. 2, pp. 1064−1069, March, 2003.
[42] F. Wang, T. Hurst, D. Abramovitch, G. Franklin, “Disk drive pivot
nonlinearity modeling part II: Time domain,” Proc. of American Con
trol Conference, pp. 2604−2607, Baltimore, Maryland, USA, June,
1994.
[43] C. Canudas de Wit, H. Olsson, K. J. Astrom, and P. Lischinsky, “Anew
model for control of systems with friction,” IEEE Trans. Automatic
Control, vol. 40, no. 3, pp. 419−425, 1995.
[44] J. Swevers, F. AlBender, C. G. Ganseman, and T. Prajogo, “An inte
grated friction model structure with improved presliding Behavior for
accurate friction compensation,” IEEE Trans. Automatic Control, vol.
45, no. 4, pp. 675−686, 2000.
[45] J. Ishikawa, M. Tomizuka, “A novel addOn compensator for cancel
lation of pivot nonlinearities in Hard Disk Drives,” IEEE Transactions
on Magnetics, pp. 1895−1897, vol. 34, no. 4, 1998.
[46] M. Kemal Ciliz and M. Tomizuka, “Modeling and compensation
of frictional uncertainties in motion control: A Neural network
Based Approach,” Proc. of the American Control Conference, pp.
3269−3273, The Westin Hotel, Seattle, Washington USA, 21−23
June, 1995.
[47] T. Huang, Y. Ding, S. Weerasooriya, T. S. Low, “Disk drive pivot
nonlinearity modeling and compensation through fuzzy logic,” IEEE
Transactions on Magnetics, pp. 30−35, vol. 34, no. 1, 1998.
[48] R. Diversi, R. Guidorzi, and U. Soverini, “Identiﬁcation of autoregres
sive models in the presence of additive noise,” International Journal of
Adaptive Control and Signal Processing, vol. 22, pp. 465−481, 2008.
References 297
[49] P. Krejci and K. Kuhnen, “Inverse control of systems with hysteresis
and creep,” IEE Proc. Control Theory and Application, vol. 148, no.
3, pp. 185−192, 2001.
[50] M. Brokate, and J. Sprekels, Hysteresis and Phase Transitions,
SpringerVerlag, BerlinHeidelbergNew York, 1996.
[51] R. W. Wood, and H. Takano, “Prospects for magnetic recording over
the next 10 years,” IEEE International Magnetics Conf., pp. 98, San
Diego, California, USA, 8−12 May, 2006.
[52] S. Jayantha, H. Dang, et al., “TrueTrack servo technology for high TPI
disk drives,” IEEE Trans. Magnetics, pp. 871−876, 2001.
[53] C. Lee, “Servo writer: a critical tool in hard disk manufacturing,” Solid
State Technology, vol. 34, no. 5, pp. 207−211, 1991.
[54] K. Noda and M. Naoe, “Novel NRRO minimization algorithm for an
ultrahighdensity servo track writer,” IEEE Trans. on Magnetics, vol.
33, Issue: 5, Part: 1, pp. 2626−2628, 1997.
[55] Y. B. Chang, D. K. Park, N. C. Park, and Y. P. Park, “Prediction
of track misregistration due to disk ﬂutter in hard disk drive,” IEEE
Trans. Magnetics, 38(2), pp. 1441−1446, 2002.
[56] H. S. Shimizu, T. Shimizu, M. Tokuyama, H. Masuda, and S. Naka
mura, “Numerical simulation of positioning error caused by airﬂow
induced vibration of head gimbals assembly in hard disk drive,” IEEE
Trans. Magnetics, 39(2), pp. 806−811, 2003.
[57] L. A. Sievers and A. Flotow, “Comparison and extensions of control
methods for narrowband disturbance rejection,” IEEE Trans. Signal
Processing, vol. 40, no. 10, pp. 2377−2391, 1992.
[58] D. Wu, G. Guo, and T. C. Chong, “Midfrequency disturbance sup
pression via microactuator in dualstage HDDs,” IEEE Trans. on
Magnetics, 38(5), pp. 2189−2191, 2002.
[59] R. Ehrlich, J. Adler, and H. Hindi, “Rejecting oscillatory, non
synchronous mechanical disturbances in hard disk drives,” IEEE
Trans. Magnetics, vol. 37, no. 2, pp. 646−650, 2001.
[60] S. M. SriJayantha, H. Dang, A. Sharma, I. Yoneda, N. Kitazaki, and
S. Yamamoto, “TrueTrack
TM
servo technology for high TPI disk
drives,” IEEE Trans. Magnetics, vol. 37, pp. 871−876, 2001.
[61] A. Rantzer, “On the KalmanYakubovicPopov lemma,” Systems and
Control Letter, 28(1), pp. 7−10, 1996.
[62] T. Iwasaki, G. Meinsma and M. Fu, “Generalized Sprocedure and ﬁ
nite frequency KYP lemma,” Mathematical Problems in Engineering,
6, pp. 305−320, 2000.
298 Modeling and Control of Vibration in Mechanical Systems
[63] G. Herrmann, S. S. Ge, and G. Guo, “A neural network controller
augmented to a high performance linear controller and its application
to a HDDtrack following servo system,” 16th IFAC World Congress,
MoE16TO/6, Prague, 4−8 July, 2005.
[64] C. Du, J. Zhang and G. Guo, “Disturbance modeling and control de
sign for selfservo track writing,” IEEE/ASME Trans. Mechatronics,
vol. 10, no. 1, pp. 122−127, 2005.
[65] C. Du, L. Xie, G. Guo, J. Zhang, Q. Li, B. Hredzak, J. Teoh, “A gen
eralized KYP Lemmabased control design and application for 425
kTPI servo track writing,” Proc. American Control Conference, pp.
1303−1308, Minneapolis, Minnesota, USA, 14−16 June, 2006.
[66] J. Zheng, C. Du, G. Guo, Y. Wang, J. Zhang, Q. Li, and B. Hredzak,
“Phase lead peak ﬁlter method to high TPI servo track writer with mi
croactuators,” Proc. American Control Conference, pp. 1309−1314,
Minneapolis, Minnesota, USA, 14−16 June, 2006.
[67] H. W. Bode, Network Analysis and Feedback Ampliﬁer Design, New
York: Van Nostrand, 1945.
[68] S. Boyd and C. H. Barratt, Linear Controller Design: Limits of Per
formance, Prentice Hall, Eaglewood Cliffs, 1991.
[69] P. Gahinet, A. Nemirovski, A. J. Laub, and M. Chilali, LMI Control
Toolbox, Natick, MA: The Math Works, Inc., 1995.
[70] K. Zhou, J. C. Doyle and K. Glover, Robust and Optimal Control,
PrenticeHall, Inc., New Jersey, 1996.
[71] M. Green, J. N. Limebeer, Linear Robust Control, PrenticeHall, Inc.,
1995.
[72] S. S. Ge, T. H. Lee, G. Y. Li, and J. Zhang, “Adaptive NNcontrol for
a class of discrete time nonlinear systems,” International Journal of
Control, 76(4), pp. 334−354, 2003.
[73] M. Brown and C. J. Harris, Neurofuzzy Adaptive Modeling and Con
trol, Prentice Hall, New York, 1994.
[74] S. S. Ge, T. H. Lee, and C. J. Harris, Adaptive Neural Network Control
of Robotic Manipulators, World Scientiﬁc, Singapore.
[75] J. R. Rice, The Approximation of Functions, AddisonWesley, Read
ing, MA, 1964.
[76] MicroE PA100 Positioner Installation Guide and User’s Manual, Mi
croE, Colorado.
[77] T. Coleman, M. Branch, A. Grace, Optimization Toolbox For Use with
MATLAB, The MathWorks Inc. Natick, Mass, 1999.
References 299
[78] X. Chen and K. Zhou, “Multiobjective H
2
/H
∞
design,” SIAM J.
Control and Optimization, vol. 40, no. 2, pp. 628−660, 2001.
[79] S. M. Djouadi, C. D. Charalambous, D. W. Repperger, “A convex pro
gramming approach to the multiobjective H
2
/H
∞
problem,” Proceed
ings of 2002 American Control Conference, pp. 4315−4320, Anchor
age, Alaska, USA, 8−10 May, 2002.
[80] R. Ehrlich and D. Curran, “Major HDD TMR sources and projected
scaling with TPI,” IEEE Trans. on Magnetics, 35(2), pp. 885−891,
1999.
[81] G. F. Franklin, J. D. Powell, and A. EmamiNaeini, Feedback Control
of Dynamic Systems, AddisonWesley Publishing Company, 1994.
[82] L. Guo and Y. Chen, “Disk ﬂutter and its impact on HDD servo per
formance,” IEEE Trans. on Magnetics, vol. 37, pp. 866−870, 2001.
[83] H. T. Ho, “Noise impact on servo TMR,” Proc. 1997 American Control
Conference, pp. 2906−2909, Albuquerque, New Mexico, USA, 4−6
June 1997.
[84] M. Karaman, and W. C. Messner, “Robust dual stage HDD track
follow control systems design for handoff shaping,” Digest of Asia
Paciﬁc Magnetic Recording Conference 2002, Singapore, 27−29 Au
gust, 2002.
[85] M. Kobayashi, S. Nakagawa, and S. Nakamura, “A phasestabilized
servo controller for dualstage actuators in hard disk drives,” IEEE
Trans. on Magnetics, 39(2), pp. 844−850, 2003.
[86] M. Kobayashi, S. Nakagawa, T. Atsumi, and T. Yamaguchi,
“Highbandwidth servo control designs for magnetic disk drives,”
Proc. IEEE/ASME International Conference on Advanced Intelligent
Mechatronics 2001, vol. 2, pp. 1124−1129, 2001.
[87] S. H. Lee, C. C. Chung, and S. M. Suh, “Multirate digital control for
high track density magnetic disk drives,” IEEE Trans. on Magnetics,
39(2), pp. 832−837, 2003.
[88] Z. Li, G. Guo, B. M. Chen, and T. H. Lee, “An optimal track following
design towards the highest trackperinch in hard disk drives,” Journal
of Information Storage and Processing Systems, vol. 3, pp. 27−41,
2001.
[89] K. Glover, D. J. N. Limebeer, J. C. Doyle, E. M. Kasenally and M. G.
Safonov, “A characterization of all Solutions to the four block general
distance problem,” SIAM J. Control and Optimization, vol. 27, pp.
283−324, 1991.
300 Modeling and Control of Vibration in Mechanical Systems
[90] M. C. de Oliviera, J.C. Geromel and J. Bernussou, “An LMI op
timization approach to multiobjective controller design for discrete
time systems,” Proc. 38th IEEE Conference Decision and Control, pp.
3611−3616, Phoenix, AZ, December 1999.
[91] D. Peaucelle, D. Arzelier, O. Bachelier and J. Bernussou, “A new ro
bust Dstability condition for real convex polytopic uncertainty,” Sys
tems & Control Letters, vol. 40, pp. 21−30, 2000.
[92] M. A. Peters and A. A. Stoorvogel, “Mixed H
2
/H
∞
control in a
stochastic framework,” Linear Algebra and its Applications, vol. 205,
971−996, 1992.
[93] T. Shimomura, T. Fujii, “An iterative method for mixed H
2
/H
∞
con
trol design with uncommon LMI solutions,” Proceedings of the 1999
American Control Conference, pp. 3292−3296, San Diego, CA, USA,
2−4 June, 1999.
[94] M. Sznaier and H, Rotsein, “An exact solution to general 4blocks
discretetime mixed H
2
/H
∞
problems via convex optimization,”
Proc. American Control Conference, pp. 2251−2256, Baltimore, MD,
USA, 29 June − 1 July, 1994.
[95] K. Zhou, J. Doyle, K. Glover and B. Bodenheimer, “Mixed H
2
and
H
∞
control,” Proc. of American Control Conference, pp. 2502−2507,
San Diego, CA, USA, 23−25 May, 1990.
[96] S. K. Aggarwal, D. A. Horsley, R. Horowitz, and A. P. Pisano, “Micro
actuators for high density disk drives,” Proc. 1997 American Control
Conference, pp. 3979−3984, Albuquerque, New Mexico, USA, 4−6
June, 1997.
[97] C. Du, J. Zhang, and G. Guo, “Vibration analysis and control design
comparison of HDDs using ﬂuid bearing and ball bearing spindles,”
Proc. 2002 American Control Conference, pp. 1378−1383, Anchor
age, Alaska, USA, 8−10 May, 2002.
[98] G. Guo, D. Wu, and T. C. Chong, “Modiﬁed dualstage controller for
dealing with secondary stage actuator saturation,” IEEE Trans. Mag
netics, 39(6), pp. 3587−3592, 2003.
[99] G. Guo and J. Zhang, “Control strategies for writing servo tracks nar
rower than 5 micro inches,” The 2003 JSMEIIP/ASMEISPS Joint
Conference on Micromechatronics for Information and Precision
Equipment (IIP/ISPS Joint MIPE03), Yokahama, Japan, 16−18 June,
2003.
[100] C. Chung, C. W. Seo, and S. H. Lee, “Two degreeoffreedom dual
stage actuator controller design for hard disk drives,” IEEE Trans.
Magnetics, vol. 36, no. 5, pp. 2255−2257, 2000.
References 301
[101] Y. Li and R. Horowitz, “Mechatronics of electrostatic microactuators
for computer disk drive dualstage servo systems,” IEEE/ASME Trans.
Mechatronics, vol. 6, Issue: 2, pp. 111−121, 2001.
[102] M. C. de Oliveira, J. C. Geromel, and J. Bernussou, “An LMI opti
mization approach to multiobjective and robust H
∞
controller design
for discretetime systems,” Proc. of the 38th IEEE Conference on De
cision and Control, vol. 4, pp. 3611−3616, 1999.
[103] D. Wu, G. Guo, and T. C. Chong, “Adaptive compensation of micro
actuator resonances in hard disk drives,” IEEE Trans. on Magnetics,
vol. 36, no. 5, pp. 2247−2250, 2000.
[104] B. Wu and E. Jonckheere, “A simpliﬁed approach to Bode’s theorem
for continuoustime and discretetime systems,” IEEE Trans. on Auto
matic Control, vol. 37, no. 11, pp. 1797−1802, 1992.
[105] J. Zhang, R. Chen, G. Guo, and T. S. Low, “Modiﬁed adaptive feedfor
ward runout compensation for dual stage servo system,” IEEE Trans.
on Magnetics, vol. 36, Issue: 5, Part: 1, pp. 3581−3584, 2000.
[106] T. Semba, T. Hirano, J. Hong, and L. S. Fan, “Dualstage servo con
troller for HDD using MEMS microactuator,” IEEE Trans. Magnetics,
vol. 35, no. 5, pp. 2271−2273, 1999.
[107] K. Okada, Y. Low, P. Gao, B. Qin, E. H. Ong and G. Guo, “A novel
inslider piezoelectric micro actuator for hard disc drives,” Digest of
AsiaPaciﬁc Magnetic Recording Conference 2002, Singapore, 27−29
August, 2002.
[108] C. Mohtadi, “Bode’s integral theorem for discretetime systems,” Pro
ceedings of the IEE, vol. 137, pp. 57−66, March, 1990.
[109] C. Scherer, P. Gahinet, and M. Chilali, “Multiobjective output
feedback control via LMI optimization,” IEEE Trans. Automatic Con
trol, vol. 42, no. 7, pp. 896−911, 1997.
[110] C. Du and G. Guo, “Lowering the hump of sensitivity functions for
discretetime dualstage systems,” IEEE Trans. Control Systems Tech
nology, vol. 13, no. 5, pp. 791−797, 2005.
[111] L. S. Fan, H. H. Ottesen, T. C. Reiley, and R. W. Wood, “Mag
netic recording head positioning at very high track densities using a
microactuatorbased twostage servo system,” IEEE Trans. Industrial
Electronics, vol. 42, no. 3, pp. 222−233, 1995.
[112] A. Beghi, R. Oboe, P. Capretta, and F. C. Soldavini, “Loop shaping
issues in hard disk drive servo system design,” IEEE/ASME Conf. on
Advanced Intelligent Mechatronics Proc., pp. 828−832, Como, Italy,
8−12 July, 2001.
302 Modeling and Control of Vibration in Mechanical Systems
[113] H. Fujimoto, F. Kawakami, and S. Kondo, “Multirate repetitive con
trol and applications,” Proc. of the American Control, pp. 2875−2880,
Denver, Colorado, USA, 4−6 June, 2003.
[114] A. H. Sacks, M. Bodson, and W. Messner, “Advanced Methods for
Repeatable Runout Compensation,” IEEE Trans. on Magnetics, vol.
31, pp. 1031−1036, 1995.
[115] M. Kawafuku, M. Iwasaki, H. Hirai, M. Kobayashi, and A. Okuyama,
“Rejection of repeatable runout for HDDs using adaptive ﬁlter,” The
8th IEEE Intl. Workshop on Advanced Motion Control, pp. 305−310,
Kawasaki, Japan, 25−28 March, 2004.
[116] Y. H. Kim, C. I. Kang, and M. Tomizuka,“Adaptive and optimal re
jection of nonrepeatable disturbance in Hard Disk Drives,” Proc.
IEEE/ASME Conference on Advanced Intelligent Mechatronics, pp.
1−6, Monterey, California, USA, 24−28, July, 2005.
[117] J. Zheng, G. Guo, and Y. Wang, “Identiﬁcation and decentralized con
trol of a dualactuator hard disk drive system,” IEEE Trans. Magnetics,
vol. 41, pp. 2515−2521, 2005.
[118] J. Zheng, G. Guo, Y. Wang, and W. E. Wong, “Optimal narrowband
disturbance ﬁlter for PZTactuated head positioning control on a spin
stand,” IEEE Trans. Magnetics, vol. 42, no. 11, pp. 3745−3751, 2006.
[119] T. Iwasaki, and S. Hara, “Generalized KYP lemma: Uniﬁed frequency
domain inequalities with design applications,” IEEE Trans. Automatic
Control, vol. 55, pp. 41−59, 2005.
[120] F. Blanchini, P. Colaneri, F. A. Pellegrino, “Simultaneous performance
achievement via compensator blending,” Automatica, 44(1), pp. 1−14,
2008.
[121] H. S. Lee, “Controller optimization for minimumposition error signals
of hard disk drives,” IEEE Trans. on Industrial Electronics, vol. 48, no.
5, pp. 945−950, 2001.
[122] H. Takahashi, and D. B. Bogy, “Vibration of head suspensions for
proximity recording,” IEEE Trans. Magnetics, 34(4), pp. 1756−1758,
1998.
[123] C. Johnson, “Accommodation of external disturbances in linear regu
lator and servomechanism problems,” IEEE Trans. Automatic Control,
vol. 163, no. 6, pp. 635 − 644, 1971.
[124] E. Davison, “The robust decentralized control of a general servomech
anism problem,” IEEE Trans. Automatic Control, vol. 21, no. 1, pp.
14−24, 1976.
References 303
[125] E. Davison, “The robust control of a servomechanism problemfor lin
ear timeinvariant multivariable systems,” IEEE Transactions on Au
tomatic Control, vol. 21, no. 1, pp. 25−34, 1976.
[126] E. Schrijver, and J. van Dijk, “Disturbance observers for rigid mechan
ical systems: Equivalence, stability, and design,” ASME J. Dynamic
Systems, Measurement, and Control, vol. 124, pp. 539−548, Decem
ber 2002.
[127] C. C. Wang, and M. Tomizuka, “Design of robustly stable disturbance
observers based on closed loop consideration using H
∞
optimization
and its applications to motion control systems,” Proc. of American
Control Conference, pp. 3764−3769, Boston, USA, 30 June − 2 July,
2004.
[128] K. Ohnishi, “A new servo method in mechatronics,” Trans. Jpn. Soc.
Elect. Eng., vol. 107D, pp. 83−86, 1987.
[129] M. T. White, M. Tomizuka, and C. Smith, “Improved track following
in magnetic disk drives using a disturbance observer,” IEEE/ASME
Trans. Mechatronics, vol. 5, no. 1, pp. 3−11, 2000.
[130] J. Ishikawa, and M. Tomizuka, “Pivot friction compensation using
an accelerometer and a disturbance observer for hard disk drives,”
IEEE/ASME Trans. Mechatronics, vol. 3, no. 3, pp. 194−201, 1998.
[131] M. Tomizuka, “Zero phase error tracking algorithm for digital con
trol,” ASME Journal of Dynamic Systems, Measurement, and Control,
vol. 116, pp. 33−38, March, 1994.
[132] G. C. Goodwin, “Predicting the performance of soft sensors as a route
to low cost automation,” Annual Reviews in Control, vol. 24, pp.
55−56, 2000.
[133] J. R. Ryoo, T.Y. Doh, and M. J. Chung, “Robust disturbance observer
for the trackfollowing control system of an optical disk drive,” Con
trol Engineering Practice, vol. 12, pp. 577−585, 2004.
[134] K. Yang, Y. Choi, and W. K. Chung, “On the tracking performance im
provement of optical disk drive servo systems using errorbased dis
turbance observer,” IEEE Trans. Industrial Electronics, vol. 52, pp.
270−279, 2005.
[135] D. F. Cribbs, M. L. Ellenberger, and J. W. Hassler Jr., “Self servo
writing disk drive and method,” Patent US5448429, 1995.
[136] C. Du, L. Xie, H
∞
Control and Filtering of Twodimensional Systems,
SpringerVerlag, 2002.
[137] E. Fornasini and G. Marchesini, “Doubly indexed dynamical systems;
statespace models and structural properties,” Mathematical Systems
Theory, vol. 1, no. 1, 1978.
304 Modeling and Control of Vibration in Mechanical Systems
[138] R. P. Roesser, “A discrete statespace model for linear image process
ing,” IEEE Trans. on Automatic Control, vol. 20, pp. 1−10, 1975.
[139] R. Yang, L. Xie, and C. Zhang, “H
2
and robust H
2
/H
∞
control of
2D systems in Roesser model,” Automatica, 42(9), pp. 1507−1514,
September, 2006.
[140] H. Ye, V. Sng, C. Du, J. Zhang, and G. Guo, “Radial error propagation
issues in self servo track writing technology,” IEEE Trans. on Magnet
ics, vol. 38, pp. 2180−2182, September, 2002.
[141] R. K. Miller, M. S. Mousa, and A. N. Michel, “Quantization and over
ﬂow effects in digital implementation of linear dynamic controllers,”
IEEE Trans. Automatic Control, vol. 33, pp. 698−704, July, 1988.
Modeling and Control of Vibration in Mechanical Systems
AUTOMATION AND CONTROL ENGINEERING
A Series of Reference Books and Textbooks Series Editors FRANK L. LEWIS, PH.D., FELLOW IEEE, FELLOW IFAC
Professor Automation and Robotics Research Institute The University of Texas at Arlington
SHUZHI SAM GE, PH.D., FELLOW IEEE
Professor Interactive Digital Media Institute The National University of Singapore
Modeling and Control of Vibration in Mechanical Systems, Chungling Du and Lihua Xie Analysis and Synthesis of Fuzzy Control Systems: A ModelBased Approach, Gang Feng LyapunovBased Control of Robotic Systems, Aman Behal, Warren Dixon, Darren M. Dawson, and Bin Xian System Modeling and Control with ResourceOriented Petri Nets, Naiqi Wu and MengChu Zhou Sliding Mode Control in ElectroMechanical Systems, Second Edition, Vadim Utkin, Jürgen Guldner, and Jingxin Shi Optimal Control: Weakly Coupled Systems and Applications, Zoran Gajic MyoTaeg Lim, Dobrila Skataric WuChung Su, ´, ´, and Vojislav Kecman Intelligent Systems: Modeling, Optimization, and Control, Yung C. Shin and Chengying Xu Optimal and Robust Estimation: With an Introduction to Stochastic Control Theory, Second Edition, Frank L. Lewis; Lihua Xie and Dan Popa Feedback Control of Dynamic Bipedal Robot Locomotion, Eric R. Westervelt, Jessy W. Grizzle, Christine Chevallereau, Jun Ho Choi, and Benjamin Morris Intelligent Freight Transportation, edited by Petros A. Ioannou Modeling and Control of Complex Systems, edited by Petros A. Ioannou and Andreas Pitsillides Wireless Ad Hoc and Sensor Networks: Protocols, Performance, and Control, Jagannathan Sarangapani Stochastic Hybrid Systems, edited by Christos G. Cassandras and John Lygeros Hard Disk Drive: Mechatronics and Control, Abdullah Al Mamun, Guo Xiao Guo, and Chao Bi Autonomous Mobile Robots: Sensing, Control, Decision Making and Applications, edited by Shuzhi Sam Ge and Frank L. Lewis Neural Network Control of Nonlinear DiscreteTime Systems, Jagannathan Sarangapani
Modeling and Control of Vibration in Mechanical Systems
Chunling Du
Data Storage Institute Singapore
Lihua Xie
Nanyang Technological University Singapore
Boca Raton London New York
CRC Press is an imprint of the Taylor & Francis Group, an informa business
MATLAB® and Simulink® are trademarks of The MathWorks, Inc. and are used with permission. The MathWorks does not warrant the accuracy of the text of exercises in this book. This book’s use or discussion of MATLAB® and Simulink® software or related products does not constitute endorsement or sponsorship by The MathWorks of a particular pedagogical approach or particular use of the MATLAB® and Simulink® software.
CRC Press Taylor & Francis Group 6000 Broken Sound Parkway NW, Suite 300 Boca Raton, FL 334872742 © 2010 by Taylor and Francis Group, LLC CRC Press is an imprint of Taylor & Francis Group, an Informa business No claim to original U.S. Government works Printed in the United States of America on acidfree paper 10 9 8 7 6 5 4 3 2 1 International Standard Book Number: 9781439817988 (Hardback) This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint. Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, please access www.copyright.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 9787508400. CCC is a notforprofit organization that provides licenses and registration for a variety of users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe. Visit the Taylor & Francis Web site at http://www.taylorandfrancis.com and the CRC Press Web site at http://www.crcpress.com
Contents
Preface List of Tables List of Figures Symbols and Acronyms 1 Mechanical Systems and Vibration 1.1 Magnetic recording system . . . . . . . . . . . . . . . . . . . . 1.2 Stewart platform . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Vibration sources and descriptions . . . . . . . . . . . . . . . . 1.4 Types of vibration . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.1 Free and forced vibration . . . . . . . . . . . . . . . . . 1.4.2 Damped and undamped vibration . . . . . . . . . . . . 1.4.3 Linear and nonlinear vibration . . . . . . . . . . . . . . 1.4.4 Deterministic and random vibration . . . . . . . . . . . 1.4.5 Periodic and nonperiodic vibration . . . . . . . . . . . . 1.4.6 Broadband and narrowband vibration . . . . . . . . . 1.5 Random vibration . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.1 Random process . . . . . . . . . . . . . . . . . . . . . 1.5.2 Stationary random process . . . . . . . . . . . . . . . . 1.5.3 Gaussian random process . . . . . . . . . . . . . . . . . 1.6 Vibration analysis . . . . . . . . . . . . . . . . . . . . . . . . . 1.6.1 Fourier transform and spectrum analysis . . . . . . . . . 1.6.2 Relationship between the Fourier and Laplace transforms 1.6.3 Spectral analysis . . . . . . . . . . . . . . . . . . . . . 2 Modeling of Disk Drive System and Its Vibration 2.1 Introduction . . . . . . . . . . . . . . . . . 2.2 System description . . . . . . . . . . . . . 2.3 System modeling . . . . . . . . . . . . . . 2.3.1 Modeling of a VCM actuator . . . . 2.3.2 Modeling of friction . . . . . . . . 2.3.3 Modeling of a PZT microactuator . 2.3.4 An example . . . . . . . . . . . . . 2.4 Vibration modeling . . . . . . . . . . . . .
xi xiii xv xxiii 1 1 2 4 5 6 6 6 6 7 8 11 11 12 12 13 13 14 14 17 17 17 19 19 23 29 30 39
. . . . . . . . . . . . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
v
. . . . . . . . .5 3 Modeling of Stewart Platform 3. . . . . . .2 Discretetime case . . 2. . . . . 4. . . . . . .2 Adaptive modeling of disturbance . .4. . . . .2 Passive control . . . . . . . . . . . . 5. . . . . . . . . . . 3. . . . . . . . .1 Isolators . . . . . . . . . . .2. . . . . . . . . . . . . . . . . . . 4. . . . . . . .1 Introduction . . . . . 4. . . . . . . . .3 Sampling . . . .2 H2 and H∞ norms .3. .4. . . . . . . . . . . . . . . . .2 Active systems . . .1 Bode integral constraint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Spectrumbased vibration modeling . . . . . . . . 5.2. . . . . . .4. . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . . . . . . .4. . . .3.vi Modeling and Control of Vibration in Mechanical Systems 2. . . . . . . . 3. . 5. . . . .1 H2 norm . . . . .6 Controller parametrization . . . . . 5. .4 Active vibration control . . reduction via . . . . .2 Absorbers . . 4. . . . . 5. . . . . . . . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . . .1 Continuoustime case . . . . . . . . . . . . . . . . . . . . . . 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Modeling using adaptive ﬁltering approach . .3 Selfadapting systems . . . . . . . . . . . . . . . . . . . .3 Control strategy .3. . . . . . . . . . .2. 4. . . . . . . . 4. . . . . . . . . . . . . . . .1 Actuators . . . .7 Performance limitation . . . . . . . . . . . . . . . . . 5. . . . . . . . . . . . . . . . . . . . . .4 H∞ control . . . . .4. . . . . . . . . . . . 3. . . . . . . . . . . . . . . . . . .2 Relationship between system gain and phase 5. . . . . .7. . . . . . . . . . . .2 System description and governing equations 3. . . . 5. . 5. . . . . . . . . .1 Introduction . . . . . . 5 Introduction to Optimal and Robust Control 5. . . . . . . .4 Conclusion . . . . 4. . . . . . . . . 5. . . . . . . . . . .2 Discretetime case . .5 Conclusion .2. . . . . . . . . 5. . . . . 4. . . . . . . 4. .5 Robust control . . . . . . . . . . . . . . . . . . . . .5 An application example − Disk vibration stacked disks . . . . . . . . . .7. 5. . . . . . . . . . . . . . . . .1 Introduction .3 Resonators . . . . 4 Classical Vibration Control 4. . . . . . . . . . . . . . . . . . . . . . 5. . . . . . . . . .4 Suspension . . . . . . . . . . . . . . . . . . . . 4. .2 Modeling of a Stewart platform . . .2 H∞ norm . . . . . . . . . . . . .7. . . . .3. . . . 4. . . . . 3.1 Adaptive ﬁltering theory . . . . . . . . 4. . .1 Continuoustime case . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 H2 optimal control .4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 43 48 53 53 53 55 55 58 62 63 63 63 63 64 64 65 66 82 83 84 84 86 87 89 89 89 89 91 92 92 94 96 96 99 101 104 108 108 111 111 2. . . . . . . . . . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2. . Conclusion . . . . . . . . . . . . . . . . . . . . . . . .
. . . .4 Method 2: an improved slack variable approach . . . 7. . . . . . . . . . . . . . . . . . . .2 An application example . .2 Mixed H2 /H∞ control problem . 6. .5. 8. . . . . . . . . . . . . . . . . . . . . . . . . .3 Design in continuoustime domain . . 7. . 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8. . . . . . . . . . . . 9. . . . . . . . . .4. . . . . . . . . . . . . . . . . . . . 8. . . . . . . . . . . . . . . . . . . . . . . . .3. . 7. .1 Synthesis method for lowhump sensitivity function . . . . . . . .2 Problem formulation . . . 7. . . . . . . . .2 Phase lead peak ﬁlter . . . 8. . . . . . . . 6. . . . . . . . 9 Combined H2 and KYP LemmaBased Control Design 9. . . . . . . . . . .4. . . . . . . . . .8 Conclusion vii .1 Introduction . . . . .7 Conclusion . . . . . .6 Application in midfrequency vibration rejection . . . . .3 Generalized KYP lemmabased control design method . . . . .1 Introduction .3 Implementation on a hard disk drive . . . . . . . .6 Conclusion . . . . . . . . . . . . . . . . . . . .1 H∞ loop shaping for lowhump sensitivity functions . . . . 9.3 Design steps . . . 8 Generalized KYP LemmaBased Loop Shaping Control Design 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . 8. . . . . 7. . . . . . . . . . . . . . . . . . . .5 Conclusion . . . .4. . . . . . .4 Design in discretetime domain . . . . . . . . . 7 LowHump Sensitivity Control Design for Hard Disk Drive Systems 7. . . . .3 Controller design for speciﬁc disturbance rejection and overall error minimization . . . . 6. . . . . . 7. . . . . . . . . . . . . . . . .4. . . .4 Peak ﬁlter . . . . . . .3. . . .4. . . .3 Implementation on a hard disk drive . . . . . . . . . . . . 8. . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Q parametrization to meet speciﬁc speciﬁcations . . . . . . 8. . . . .5 Application in servo loop design for hard disk drives 6. . . . .3 Group peak ﬁlter . . . . . . . . . . . . . . . . . . . .5 Application in high frequency vibration rejection . .1 Introduction .2 Design results . . . . . . . .3 Method 1: slack variable approach . . . . . . . . . . 8. . . . . 6. .2 Problem statement . . . . . . . .2 Problem description . . 8. . . . 9. . . . . . . . . . . . . . . . . . . .1 Problem formulation . . . . . . . . . . .5.4. . . .1 Introduction . . . . . . . 9.3. . . 6. . . 9. . 7. . . 7. . . . . 7. . . . . . . . . . . . . . . . . . . . . . . . .2 Q parametrization to minimize H2 performance . . . . . . . . 7. . . . . . . .2 Application examples . . . . . . . . . . . . . 112 115 115 115 116 117 123 123 128 131 133 133 133 137 137 141 148 152 152 153 158 158 161 161 162 163 166 166 168 169 169 177 178 183 183 184 185 185 187 188 6 Mixed H2 /H∞ Control Design for Vibration Rejection 6. . . . .1 Conventional peak ﬁlter . . . . . . . . . . . . . . . . . . . .Table of Contents 5. . . . . .3. . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 State feedback control blending . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Introduction . . . 12. . . . . . . . . . . . . . . . .2 Twodimensional model . . . . . . . . 11. . . . . . . . . . . . . . . .1 Rejecting highfrequency disturbances . . . . . . . . . . . . .7 2D stabilization control for error propagation containment 12. . . . . .2 Output feedback control blending . . . . . . . . . . . . . . . . . . . . . . . . . .8 2D H2 control for error minimization . . . . . . . . . . . . . . . . . .4. . . 12. . . . . . . . . . . . . . . . . 11. . . . . . . . .4. . . 12. . . . . . . . . . . . . . .2 2D stabilization control .3 2D H2 control . . . . . . . . .4. .3 A general form of disturbance observer . . . . . . . . . 12. . . . . . . .1 Simulation results . .2 Controller design via the control blending technique . . . . 12. . . . 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . 10. . . . . . . .3. . . . . . 9. . . . . . . . . . . . . . 9. . . . . .2 Experimental results . . . . . . . . . . . . 12. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11. .1 SSTW servo loop . . . . . 10. . . . . . . . 10.4 SSTW process and modeling . . . . . . . . . . . 12. . . . 12. . . . 9. .2. . . . . .1 Introduction . . .3 Rejection of two disturbances with H2 performance minimization . .5 Feedforward compensation method . . . . . . . . . .4. .3 Control blending application in multifrequency disturbance rejection .6 2D control formulation for SSTW .3. . . . . . . . . . . . . .1 System models . . . . . . . .7.4. . . 189 189 190 193 194 197 197 197 199 200 203 203 205 207 207 211 213 215 215 216 217 220 222 227 227 228 229 231 232 233 235 243 244 244 245 245 247 248 9. . . .4 Simulation and experimental results . . . . . . . . . . . . . . . . . . . . . . . . .2 Rejecting a combined mid and high frequency disturbance 10. . . . . 10. . . . . . . . . . 11. .2 Conventional disturbance observer .9 Conclusion . . . . .1 Introduction . . . . . . . . . . . . 12 TwoDimensional H2 Control for Error Minimization 12. .1 Simulation results . . . . . . . . . . . . . .8.5 10 Blending Control for MultiFrequency Disturbance Rejection 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12. 10. . .viii 9. . . . . . . . . . . . . . . . 10. . . . . . . . . . . . . . . . .5 Conclusion . . . 11 H∞ Based Design for Disturbance Observer 11. . . . . .5 Conclusion . . . . . . . . . . . . . .4. . . 10. . . . . . . . . . . . . . .4 Application results . . . .2 Rejection of speciﬁc disturbance and H2 performance minimization . . . . . . . .1 Problem formulation . . . .4. . . . . . . . 12. . . . . . . . . . .8. . . . 12. . . . . . . . . .4 Modeling and Control of Vibration in Mechanical Systems Simulation and implementation results . . . 10. .2. . . . . .2 Control blending . . . . . 12. . .
. . 14. . 13. . .1 Multichannel adaptive feedback AVC system . .3. . . . . .4 Compensation of quantization effect on error rejection . . .5. . . . . . .2 Quantization effect on error rejection . . . . . . . . . . . . 15. . . . . . . . . . . . . 15 Adaptive Filtering Algorithms for Active Vibration Control 15. . . . . . . . . . . 13. . . . . . .5. .3 Adaptive feedback algorithm . . . . . 14. . . . . . . . . . .5. . . . . . . . . . . . . . . . . .2 Description of control system with quantizer . . . 15. . . . . . . . . . . . . . . . .3 Quantization effect on error rejection . . . . . . . . . . . .2 Adaptive feedforward algorithm . . . . . . . . . 14. . . . . . . . . .1 Introduction . 15. . .1 Quantizer frequency response measurement . .1 Introduction . . . . . . . . . . .4 Conclusion . . . . . . . . . 14. . . . .1 Introduction . . 15. . . . . . . . . .3 Nonlinear control . . . . . . . . . . ix 251 251 251 252 256 257 259 261 261 261 266 266 266 269 272 275 275 275 277 280 280 280 281 284 290 293 305 . . . . . . . . . . . . . . . . 15. . . . . . . . . . . . . . . . . . . . . . . . . . References Index . . . .1 Design of a composite control law . 14.2 Nonlinearity compensation . . . .3. . . . . . .5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15. . . . . . . . . . . . . .3 Simulation and implementation . . . . . .2 Multichannel adaptive feedback algorithm for hexapod platform . . . . . . . .3. . . . . . . .2 Experimental results in hard disk drives 13. 15. . . . 14. . . . . . . . . . . . . . . . . . . . . . . . 13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . 14 Quantization Effect on Vibration Rejection and Its Compensation 14. . . .6 Conclusion . 15. 13. . . . . .4 Comparison between feedforward and feedback controls . . .5 Application in Stewart platform . . . . . .Table of Contents 13 Nonlinearity Compensation and Nonlinear Control 13. . . . . . . .
.
and 5. Jianliang Zhang and Jul Nee Teoh from Data Storage Institute (DSI) of Singapore. automotives. 4. Many realworld examples are given to demonstrate the modeling and control techniques. their effects on system stability and performance. lowering system performance. and techniques for rejecting vibrations of different frequency ranges and their limitations. aerospace and aeronautic systems. etc. It is our intention in this book to present to readers some of the recent developments in this ﬁeld. The book summarizes a collective research effort which we have had the pleasure to contribute to. Special attention is given to recently developed vibration modeling and control techniques in high precision systems. vibration control has drawn more intensive efforts from researchers and engineering practitioners in recent years. wasting energy and creating unwanted noise. The book aims at empowering readers with a clear understanding of characteristics of various vibrations. Examples are presented throughout the book so that the subject can be better understood. The theoretical developments and principles of control design are elaborated in detail so that the reader can apply the techniques developed to obtain solutions with the help of MATLAB . xi . and Frank Lewis from the University of Texas at Arlington. it remains and indeed becomes more challenging in many applications such as precision engineering and hard disk drives. Youyi Wang from Nanyang Technological University (NTU). Singapore.Preface This book is primarily intended for researchers and engineering practitioners in systems and control. The research work contained in this book was mainly performed at DSI and the School of Electrical and Electronic Engineering (EEE) of NTU. Vibration exists in a wide spectra of engineering systems such as hard disk drives. Therefore. manufacturing systems. especially those engaged in the area of modeling and control of vibrations in mechanical structures and systems. It focuses on vibration and disturbance rejections using recently developed control techniques for high precision positioning. Although the problem of vibration control has been studied for a long time. A number of simulation and experimental results with comprehensive evaluations are provided in each chapter. and demonstration of the beneﬁts gained from the applications of these techniques. Vibration is undesirable in most engineering applications. The book presents the latest results in vibration modeling and advanced control design for vibration attenuation in mechanical actuation systems to achieve high precision positioning performance. where an extremely high positioning accuracy is required. Many results reported in the book are due to the collaboration with Guoxiao Guo from Western Digital Corporation. which are dedicated to the review of related background knowledge. except Chapters 1.
For product information. Fax: 5086477001.xii Modeling and Control of Vibration in Mechanical Systems Algorithms applied in magnetic recording systems were implemented at DSI and those in the Stewart platform at the School of EEE. MA 017602098 USA. 3 Apple Hill Drive. Inc.. . Natick. and EEE. Ong Eng Hong and the colleagues in Mechatronics and Recording Channel Division of DSI. please contact: The MathWorks. Inc. We would like to express our sincere appreciation to DSI for its supportive environment and vibrant research atmosphere.mathworks. We are also sincerely grateful to Dr. com. NTU. NTU for their support. Web: www. Tel: 508 647 7000. Lihua Xie Chunling Du MATLAB is a registered trademark of The MathWorks. Email: info@mathworks.com.
. . . . . . . . . . . . . . . 146 Comparison of performance speciﬁcations . RRO and NRRO and disk vibration amplitude with stacked disks compared with single disk . . . . 49 81 Control performance comparison . . . .1 9. . .1 Quantization and friction effect . . . . . . . . . . . 191 14. . . . . . . . . . % reduction of σ values of PES. . .1 6. . 131 Control performance comparison. .1 7. . . . . . . .List of Tables 2. . . . . . . . . . . . . . . . .1 4. . . . .1 ˆ σ values of the modeling error (d1 − d1 ) for different p and Γ . . . 272 xiii .
.
. . . . . . . . . . . . .11 2. . . . . . . . . . . . Measured VCM Bode plots (straight lines: pure double integrator k/s2 ). . . . . . .10 2. .3 2. . . . . . . . Histogram of signal x. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .21 2. . respectively. . . . . . . A PZT actuated suspension. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .22 2. . . . . . .13 2. . . . .6 2. . . . . . Closedloop control system disturbances d1 . . .4 1. . . . . . . . . . . . . . . . .20 2. . . . . . Closed control loop of a disk drive with a VCM actuator for friction measurement via LDV. . . . Multiplicative uncertainty of a VCM. Measured and modeled friction and error.8 2. . .1 2. . . . . . . . . . . . . .2 1. . . . . . . . . . . . . .4 2. . . .19 2. . . . . . . . . . . .8inch hard disk drive. . . . d2 and noise n. Equivalent spring mass system of PZT microactuator. . xv . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .14 2. . An example of random vibration. . . . . . . . . . .7 2. . . . . . . . . . . . . . . . . . An opened 1. . . A typical frequency response of the PZT microactuator. .List of Figures 1. . Spectrum density of broadband vibration. . . . .1 1. . . . . . . . . . . . . . . . . Actuator frequency response for sinusoidal reference with amplitude of 0. . . Control signal u versus displacement x. . . . . . . . . . . . The operator zr versus x. . . Frequency responses of a second order transfer function. . . . . . . . . . . 2 3 4 7 9 10 13 20 21 22 24 26 28 28 29 30 31 32 33 34 34 36 37 37 38 38 39 40 41 42 A read back signal of embedded servo. . .17 2. . . . Actuator frequency response for sinusoidal reference with amplitude of 1 and 3 V. . . . Closedloop control system with disturbance and noise models. . . . . . . . .6 1. . . Preload and twoslope model for friction modeling. . . . . . . Spectrum density of narrowband vibration. . . . . . .18 2.9 2. . Hexapod from Micromega Dynamics. . . .7 2. . . .5 2. . . . . . .16 2. . . . .5 1. . . . . . . . . . . . . .5 µm/V). . . . .15 2. . . . . . . . . . . . . .12 2. . . . . . . Sensitivity function S(z). . . . . . . .23 The servo control loop of a hard disk drive. . . . . . . . . . VCM actuator modeling with friction nonlinearity model F (x). . . . . . . . .5 V. Zoomedin view of the hexapod. . . . . .2 2. . . . . . . Friction f versus actuator displacement x. . . . . . . . . PES NRRO spectrum. . . . . . . . . . . . . . . . . Measured and modeled frequency responses of the VCM actuation system (LDV range 0. . . . . . . . . . Plant input voltage u versus displacement x.3 1. . . . . . . . . . .
. . . . . Comparison of singledisk and dualdisk RRO power spectrum at 7200 RPM (23% improvement of σ value). . . . . System identiﬁcation using LMS adaptive ﬁlter. .4 4. . . . . . . . . .8 4. .4 3. .28 Modeling error (d1 − d 3. . . . . .3 3. . . . . . . . Comparison of singledisk and dualdisk axial vibrations measured via LDV at 8400 RPM. . . . . .1 4. . Generic feedback control system. Comparison of singledisk and dualdisk NRRO power spectrum at 7200 RPM (18% reduction of σ value). Linear discrete time adaptive ﬁlter. ˆ ˆ i. . . . . . . . . . . . . .6 4. . Comparison of singledisk and dualdisk PES in time domain at 7200 RPM (21% reduction of σ value). . . . . . . . . . . Comparison of singledisk and dualdisk axial vibrations measured via LDV at 10200 RPM. . . . . . . . . . .6 4. . . . . . . Comparison of singledisk and dualdisk PES in time domain at 10200 RPM (32% reduction of σ value). . . . . Block diagram of the LMS adaptive ﬁlter. . . .25 2. . . . . . . . . 102 . . . . . . . Comparison of singledisk and dualdisk NRRO power spectrum at 10200 RPM (27% reduction of σ value). . . . . . . . 33% improvement of σ value). . . . . . .2 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . Comparison of singledisk and dualdisk axial vibrations measured via LDV at 7200 RPM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7 4. . . . . ˆ Power spectrum of d1 . . 78 79 80 86 Conﬁguration of standard optimal control. . . . . . . . .1 3. . . ˆ ˆ1 ) for different Γ with p = 1. . . . . . . . . . . . . . . . . . Estimated and experimental frequency responses. . . . . .12 4. . . . . . . . . . . . .14 5. . . . . . . . . . . . . . . . . . . . . . . . . . . Frequency responses of a PZT actuator. . . . . . e = −P (s) · d1 − d2 + n. . . . . . . . . . . Comparison of singledisk and dualdisk RRO power spectrum at 10200 RPM (the 3rd and 5th harmonics reduced signiﬁcantly. . . . . . . . . . . . . . . . . . . . . . . . . NRRO power spectrum from measurement and disturbance models. . . . . . . . . . . . . .9 4. . . .1 5. . . . .26 2. . . 44 49 50 51 52 54 55 57 59 60 61 67 69 70 71 72 73 74 75 76 77 4. . . . . . . . . . 2. . .10 4. . Comparison of singledisk and dualdisk RRO power spectrum at 8400 RPM (41% improvement of σ value). . . .5 3. . Comparison of singledisk and dualdisk NRRO power spectrum at 8400 RPM (28% reduction of σ value). . . . . .2 Disk and spindle motor assembly of the spin stand. . .27 Modeling and Control of Vibration in Mechanical Systems Threelayer RBF neural network. . . . . . . . . . . . . . . . . . . . .xvi 2. . . . . . . . . . . . . . 94 A closedloop system with uncertainty. . . . . . . . . .11 Singleaxis system using piezoelectric stiff actuator. . . . . . . . . . . . . . . . . . . . . . .13 4. . . . . . . . . . .5 4. .24 2. .. . . . .e. . . . . .3 4. . ˆ Original disturbance d1 and the modeled d1 . . . . . Comparison of singledisk and dualdisk PES in time domain at 8400 RPM (38% reduction of σ value). . . . . . . . . .2 3. . . . . . . . .
.20 7. . 146 Frequency response of microactuator controller Cm (s). . Mixed H2 /H∞ control scheme for HDD servo bance models. . . . . . . . . . . . . . .7 5. . . . . . . . . . . . . . .4 7. . . . . . . . .4 5. . . . .5 5. . . . . . . . . . . . . . . . Frequency responses of the VCM actuator. . . . . . . . . . . . . . . 136 Frequency responses of Sv (s) and Sm (s). . . . . . . . . . . . . . . . . . . . . 147 Open loop frequency response of the dualstage system. . . . . . .1 6. . . . . . .) . . . . . xvii 103 104 107 110 110 113 123 125 126 129 130 Parallel structure of a dualstage actuation system with disturbances and noise injected.17 7. . . . . . . . . . . . . . . . . . . . . . . .5 7. . . . 141 Frequency response of VCM controller Cv (s). . . . . . . . . . Sensitivity function for discretetime system. . . . . . .24 A closedloop system with additive uncertainty for robust stability analysis. . . . .1 7. .6 5. . . . . . . . . . . . 151 3σ of PES NRRO versus frequencies. . . . . . . . . . . . . . . . . . . Multiplicative uncertainty of the VCM actuator. .2 6. . . .2 7. . . . . . . . . . 136 Structure of H∞ loop shaping. . . . . . 142 Frequency response of microactuator controller Cm (s). . . . . . . . . . . . . . . .15 7. 150 Step response of the dualstage system. . . . . . . . . . . . . . . . . . 139 Frequency responses of Pv (s)Cv (s) (solid line) and Pm (s)Cm (s)(dotted line). .7 7. . . . . . . . . . . . . . . 150 Open loop frequency response of the dualstage system (smooth line: simulation result. . . . .9 7. . . . . . dotted line: PID design). rough line: testing result. . . . . .List of Figures 5. 135 Decoupled structure of dualstage actuation systems. . . . . . . .3 7. . .6 7.10 7. . . . . . . . 140 Frequency response of VCM actuator P (s). . . . . 143 Sensitivity function Sm (s) (solid) and its weighting function inverse (dashed). . . . . . . . . .18 7. rough line: testing result. . .16 7. . . . . . . .4 6. . . . . . . . . . . . . . . . Control system structure for Youla parametrization. . . . . . . . . . . . . . . . . . . . . . . . . . . . .12 7. . . . . . . . . . . . . . . . . . . . . . . .13 7. . . . . . . . . .21 7. . . . . . . . . 144 Sensitivity and complementary sensitivity functions. . . . . . . . .19 7. . 156 . . . . Sensitivity function in discretetime domain. .11 7. . . . . .3 5. . . . . . . . . . . .8 7. . . . . . . . Sensitivity function for continuoustime system. .3 6. . . . . . . . . 147 Sensitivity and complementary sensitivity functions. . . . . . . . . . . 145 Microactuator frequency response. . . . . . . . 144 Open loop frequency response of the dualstage system. . . 148 Experimental structure. . . . . . 151 VCM controller Cv (z). . . . Frequency response of sensitivity functions. . . . . . . . 149 Sensitivity function of the dualstage system (smooth line: simulation result. . . . . . . . loop with distur. .5 7. . . . . . . 155 Sensitivity function of the dualstage system. . .8 6. Frequency response of sensitivity functions. . . . 155 Microactuator controller Cm (z). . . . . . . . . . . . . . . . . . . . . . . . . .23 7. . . . .22 7. . . . . . . . . . . . . . . . A closedloop system with multiplicative uncertainty for robust stability analysis. . 134 Power spectrum of PES nonrepeatable runout in open loop. . . . . . . . .14 7. . .
. . σ value of PES versus frequency. . . . . . . . NRRO power spectrum with PLPF and KYP (50% reduction before 1 kHz). . . . . . . . . 7. . . . . . . . . . . . . . Dashed line: with Spec. . . . . . . . . . . . . . . . .6 9. . . . . . . . . .xviii Modeling and Control of Vibration in Mechanical Systems 156 157 157 159 159 160 160 163 167 167 170 170 171 172 174 175 175 176 176 179 7. .4kHz)). . . . .8 8. . . . . . .7 9. . .. . . . . . reﬂecting the vibration distribution of the system (3σ = 21 nm including the noise 3σ = 15. . . . . . . . . Designed sensitivity functions. . . . . . . 8. . . . . . (i). .11 8. . .25 Sensitivity function. . . . . . . . . . . . . . . . . . . .9 . . . .28 Sensitivity function of the dualstage system. . . . . . . . . . . . . . . . . . . Comparison of sensitivity functions. . . . . . . . . . . . . KYP(GM: 6 dB. . . . . . . .3 8.10 8. 7. . . . . . . . . .9 8. . . Sensitivity functions before and after the KYP lemmabased design: simulation result. . . . . . . . Bandwidth 1.17 9. . . . . .29 Open loop frequency responses of the dualstage system. Openloop Bode plot before and after the KYP lemmabased design. . . . . . . . . . . . . . NRRO power spectrum with KYP Lemmabased controller with and without H2 minimization. . . . . PM: 50 deg. . . . . . . . . . . . . . . . . . . . .14 Q parameterization for control design.5 9. . . 7. . (i) and (ii)). . . . . . . .15 8. . . . . . . . . . . . Resultant sensitivity function (Solid line: with Spec. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 nm). . 7. . . . . . . . . . 8. . . Bandwidth: 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Openloop frequency responses. . . . . . . . . . PZT microactuator attached to VCM actuator arm. . . Power spectrum of the position error before servo control. .6 8. . H2 control scheme with Q parametrization for controller design. . . . . . . . .8 9. . . . . . . . . . . .1 9. . . . PM: 34 deg. . . Structure of experimental setup. . . PES NRRO spectrum without servo control. . . . . .16 180 181 181 182 184 190 191 192 192 193 194 195 195 8. . . . . Peak ﬁlter in the frequency domain. . . . . . Open loop frequency responses (PLPF (GM: 6 dB.5 8. . Comparison of sensitivity functions obtained from experiment. . . . . . . . . . . . . . Openloop frequency response..7 kHz))). . . . . . . . . . .12 8. . . . . 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . .26 Sensitivity function of the dualstage system. . . . . . . . . . PZT micro actuator frequency response. .4 9. . Sensitivity functions before and after the KYP lemmabased design: experimental results. . . .2 8. . . (ii) and (iii). . Frequency response of a PZT microactuator. . . . . . . . . . . 7. . Sensitivity functions before and after group peak ﬁltering activated. .3 9. . . . . . . . PES NRRO power spectrum calculated from measured PES signal without servo control. . . .7 8. .30 Step response of the dualstage system. . . .31 3σ value of PES NRRO versus frequency.4 8. . . .27 Frequency responses of Pv (z)Cv (z) (solid curve) and Pm (z)Cm (z) (dashed curve). . . . Frequency response of the PZT microactuator. . . . . . .2 9. . . . . . . . . . . . . . . .1 8. . . Peak ﬁlter F in the nominal feedback loop. . . . . . .13 8. . . .
. . . . . . . . . . . . . . . . . . . . 11. . . . . . . 11. . . . . . . . . . . . . . . .8 Openloop with disturbance rejection at 3. . Frequency response of the VCM actuator. . 11. . . . . . . . .65 and 2 kHz. . . . . . .3 10. . . .4 10. . . . . . . . . . .9 Comparison of TEQ−OL about the general disturbance observer. .5 SSTW servo loop. . . .1 SSTW process. . . . . . . . . . . . . .10 Comparison of TEQ−OL about conventional disturbance observer. . . . . . . . . . . . . . . . . . . . 11. . . . . . .10 Resultant sensitivity function with all the three requirements fulﬁlled. . .) 12. . . 6. . 11. . . . . . . and H2 controller. . .11 Sensitivity function with disturbance rejections at 0. . . 10. . . . .6 Frequency response of a VCM actuator. . . .4 SSTW servo loop modeling in two dimensions. . . . . . . . . . . . and 10 kHz. . . . . . . . . . . . 12. . . . . . .2 SSTW servo loop with disturbances and noise models. . . . . . . . . . . . 10. . . .65 and 2 kHz. . . . . . . . . . . . . . . . . . . . . . . 11. . . . . . . . . . . . . . . . (The time sequence and the σ value increase with the track number.7 Frequency response of the closedloop transfer function with the PD controller. . . . .1 10. . . . . . . . . . . . . 11. . . . .6 Simulated sensitivity function with disturbance rejection at 4 and 8 kHz. . . . . . . . 12. . . . . . . . . . . . . . . . . . . . . .5. . 12. . .7 Measured (solid curve) sensitivity function with disturbance rejection at 4 and 8 kHz. . . . . Control loop with injected disturbances at different frequencies. . . . . . . . . . . . . .5. . . . . . . . .5 198 204 204 207 208 209 210 210 211 212 213 216 218 222 223 223 224 224 225 225 226 231 232 233 234 236 237 238 . . . . . . . . . 10. . . . . . . . . . 12. . . . .8 Measured sensitivity functions without and with the general disturbance observer. . . . . . . . . . . . . . . . .7 Error signal e. .List of Figures xix 9. . . .10 Openloop with disturbance rejections at 0. . . . .5 The sensitivity function comparison with the general and the conventional disturbance observers. . . . . . . . . . . . . . . . PID controller. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Openloop frequency response with disturbance rejection at 4 and 8 kHz. . . 6. . . . . . . 12. . . . 196 9. 12. . . . . . . . . .6 Disturbance d1 . . . . . . . . . . . . . . . . . . . . . 11. . .4 The sensitivity functions without and with the general disturbance observer. . . . . . . . . . . . . . . .2 Block diagram of the control loop with a general disturbance observer. . . . . . . . . .9 Sensitivity function with disturbance rejection at 3. . . . . . .65 and 2 kHz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Frequency response of the designed Q(z). . . .3 PES NRRO and its σ values versus track number during propagation. . . 10. . . . . . . . . . and 10 kHz. 10. . . . . . . . . . 11. . . . . .11 NRRO power spectrum with rejection of two speciﬁc disturbances at 0. . . . . . . Control structure. . . 10. . . . . . . . . .2 10. . .1 Block diagram of the control loop with a conventional disturbance observer. . . . . . . . . . . 10. . . . . . . . 11. . 196 Blending control scheme. . . .
. . . . . . . . . . 14. 12. . . . . . .15 σ of PES NRRO versus track number. 13. . 14. . . . . . . 13. . . . . . . . . . . . .8 Comparison of error rejection frequency response without and with uN of different p and Γ. . . . . . 12. . . . . . . . . . . 12. . . . . . . 13. . . . . . . . . . . . .14 2D controller for SSTW servo loop. . . .9 3σ of PES NRRO. . . . . . . . . . . .5 Frequency response of the sensitivity function S(z) with different reference levels (i. . . . . . 13. . . 13. . . . . . . . Dc ). Bc2 . . . . . .3 The servo loop in experiment. . . 13. . . . . . . . . . . . . . .5µm/V (Straight smooth lines: the pure double integrator). . .3 Actuator frequency responses with and without friction compensation. . 14. . . . . . . . .12 Comparison of sensitivity functions. . .5 Actuator frequency responses with friction compensation for different displacements in voltage with 0. . . . . . .2 Frequency response of the VCM actuator measured by injecting a swept sine wave of 5mV amplitude. . . . . . . . .13 σ value of PES NRRO versus track number. . . Dc ). . . . . . . . . . . 14. . 239 240 241 241 242 243 245 246 246 247 248 249 249 250 252 253 253 254 254 255 256 258 259 263 263 264 264 265 267 267 .17 Sensitivity function with stabilization controller (Ac2 . . Cc. . . Bc2 . . . . . . . . Cc . .9 NRRO power spectrum with KYP lemmabased linear control and nonlinear compensation (80% reduction before 400 Hz). . . . 12. . . . . . .4 Frequency response of the controller C(z). . . 12. . . . 12. . . . . . . . 13. . . . . . . . . . . . . Dc ). . . . . . . . . . . . . . . .7 Control structure of a plant P (s) with Youla parametrization approach and adaptive nonlinear compensation. . . . . . . . . . . Bc2 . . . . . .xx Modeling and Control of Vibration in Mechanical Systems 12. . . . . . . . . . . . . .2 Input u versus displacement x with and without compensation. . . . . actuator moving ranges are different). . . . . . 12. . 14. . . . . . 12. . . . . . . . .. . . . 13. .4 Actuator frequency responses with friction compensation for different displacements in voltage with 0. . 12. 14. . . . . . . . . . . . Cc . . . . . . . . . . . 12. . . . . .e. . . . . . . . . .16 Openloop frequency response with stabilization controller. . .6 Sensitivity functions with and without friction compensation. . . . . . . . . . . . 13.6 Frequency response of the quantizer before and after compensation (bit number n = 6). . . . . . . . . . . . . . . . . . . . . 12. .19 Openloop frequency response with controller (Ac2 . . . 12. . . . . . . . 14. . . . . . . . . .7 Frequency response of the quantizer with compensation (bit number n = 8). . . . 12. . . . . .11 Frequency response of the openloop system with the H2 controller. . . . . . . . . . . . . . . . . . . . . . . . . Dc ). Cc.1 Friction compensation for the actuation system. .20 Sensitivity function with controller (Ac2 . . .8 Φ versus frequency. .21 Step response (Channel 1/2/3: Reference/Output/Control signal). . .1 Experimental setup.18 Frequency response of controller (Ac2 . . . . . . . . .5 µm/V (Straight smooth lines: the pure double integrator). . . . . . . . . . . . . . Bc2 . . .10 Frequency response of the H2 controller. . . . .
. .9 240 Hz error signal in dB unit.List of Figures 14. . . . . .9 Measured sensitivity function SQ (z) with different bits n.3 Adaptive inverse control scheme. . . . . .10 Sensitivity function SQ (f) with f = 10 Hz versus bit n. . . . . . . . . . . . . . . . . . . . . . .4 Block diagram of 2 × 2 adaptive feedback algorithm. . . . . . . . . . . . . . . . .7 60 Hz error signal in dB unit. . . . . . . . . 15.12 180 Hz error signal without automatic gain control. . . .8 210 Hz error signal in dB unit. . . . 15. . . . . . . . 14. . . . .13 180 Hz error signal with automatic gain control. . . . .8 Frequency response of the quantizer with compensation (bit number n = 10). . . . . . . . . . 14. . . . 15. . . . . . . . . . . . . 14. . . . . . 15. . . . . . . . . . . . . 15. . . . . . . . . . . . . . . . . . 15. . . . . . . . . .12 Choosing the threshold δ and the scaling factor a. . . . . . . . 14. . . . . . . 15. . . . .2 FilteredX LMS adaptive feedback algorithm. . . xxi 268 268 269 270 270 271 276 279 279 282 283 286 287 287 288 288 291 292 292 . . .6 General layout of the experimental setup. . . . . . . .1 Block diagram of FXLMS algorithm. . . . .10 270 Hz error signal in dB unit. .11 Compensation scheme of quantization effect. . . . . 15. . .13 Sensitivity function with quantization compensation. . . . . .5 Block diagram of 6 × 1 FXLMS adaptive feedback control system. . . . . . . . . . . . . . . . . . . 15. . 15. . . . . . . . . . . . . . . . 15. . . 15. . . . . . . . . . 15. . . . . .11 Simulink diagram of automatic gain control. . . . . . . . . 14. . . . .
.
n.e. 2.. · · · . j = 1. D): A B1 B2 C1 D11 D12 : C2 D21 D22 ndimensional real Euclidean space set of n × m real matrices n × n identity matrix statespace representation of a system compact representation of system: x(k + 1) = Ax(k) + B1 w(k) + B2 u(k) z(k) = C1 x(k) + D11 w(k) + D12 u(k) y(k) = C2 x(k) + D21 w(k) + D22 u(k) diag{A1 . Q ∈ Rn×n P − Q > 0 for symmetric P. 2 xxiii . B. i. · · · . on the diagonal XT : X∗: P ≥ 0: P > 0: P ≥ Q: P > Q: σ (X): ¯ Trace(X): · : w 2 transpose of matrix X complex conjugate transpose of matrix X symmetric positive semideﬁnite matrix P ∈ Rn×n symmetric positive deﬁnite matrix P ∈ Rn×n P − Q ≥ 0 for symmetric P. An }: block diagonal matrix with Aj ( not necessarily square). C. Q ∈ Rn×n largest singular value of X trace of X Euclidean vector norm : ℓ2 norm of a signal {w(k)}. ∞ k=0 w(k)) . A2 .Symbols and Acronyms Rn : Rn×m : In : (A.
∞)} or simply ℓ2 if w 2 < ∞. ∞)}. ∞)}: space of square summable sequences on {[0. The signal {w(k)} is said to be from ℓ2 {[0.xxiv Modeling and Control of Vibration in Mechanical Systems ℓ2 {[0. G 2: G ∞: H2 norm of transfer function G H∞ norm of transfer function G the real part of a complex number the imaginary part of a complex number spectral radius automatic gain control active vibration control degree determinant Dynamic Signal Analyzer fast Fourier transform ﬁlteredX LMS hard disk drive Laser Doppler Vibrometer Linear fractional transformation linear matrix inequality least mean square linear quadratic Gaussian loop transfer recovery micro electromechanical system mean square error nonrepeatable runout position error signal Re( ): Im( ): ρ( ): AGC : AV C : deg: det: DSA: FFT: F XLM S: HDD: LDV : LF T : LM I: LM S: LQG: LT R: M EM S: M SE: N RRO: P ES: .
Symbols and Acronyms P ID : P LP F : P ZT : RBF : RM S: RP M : RRO: SST W : ST W : T M R: V CM : proportionalintegralderivative phase lead peak ﬁlter lead zirconate titanate/piezoelectric radial basis function root mean square rotations per minute repeatable runout selfservo track writing servo track writing track misregistration voice coil motor xxv .
.
The magnetic recording system and Stewart platform are examples to be presented next to help in ﬁxing these ideas more ﬁrmly. industry robots [29]. This may require a design by using the value of the system response to generate additional forces according to certain rules or laws such that the modiﬁed response behaves according to desired performance and within certain bounds.1 Magnetic recording system Figure 1. positioning stages [27]. since forces and masses are taken into account in vibration analysis. Dynamics is a branch of mechanics that deals with motion and its effect on a body. called control forces which are functions of system response such as displacement. This results in a closedloop system that incorporates feedback controls. dynamics takes into account the effect of velocities and accelerations on the forces acting on bodies. we have to include the subject of dynamics and vibration.1 Mechanical Systems and Vibration When studying mechanical systems. we need to alter or control the vibration response and bring the response within acceptable levels by adding appropriate forces. the atomic force microscope (AFM) [28]. With such a grounding. Unlike statics. 1. If our analysis of system response to a vibration shows that it regularly affects the system performance in an unacceptable manner. Examples include magnetic recording systems. or temperature variation that affects the response of a system in an unacceptable manner. which deals with bodies at rest. Thus it is natural that both dynamics and vibrations of mechanical systems are studied in this book. To 1 . The purpose of designing a system with feedback force is to minimize unwanted behavior. A vibration may be a signal. as well as some automotive systems [34]. It consists of a stack of ﬂat rotating disks with positioning information or servo information embedded in their surfaces. Vibration is regarded as a branch of dynamics. force. Stewart platforms. The servo information is used to position the magnetic heads on the disk surfaces. Position measurement of the magnetic heads is achieved by means of analyzing the position error signal (PES) calculated from the read back signal.1 shows a servo control loop of a hard disk drive (HDD) with a voice coil motor (VCM) and a piezoelectric (PZT) actuated servo system. more advanced problems become accessible.
The closed kinematical linkage structure of a Stewart platform has major advantages over any serial link robots: great rigidity. (2) actuator pivot friction. FIGURE 1. (1) torque disturbances from spindle motor. and (5) head sensing and electronic noises. a force sensor and two ﬂexible joints. etc. the head positioning error with respect to the target track center needs to be as small as possible. 1. and subsequently the track density for a high capacity disk drive.2 Modeling and Control of Vibration in Mechanical Systems have a disk drive with a high storage capacity. suspension and slider vibrations. six degrees of freedom (DOF). consisting of two triangular parallel plates connected to each other by six active legs orthogonal to each other. [31] The Stewart platform is . media noise.2 Stewart platform The sixleg parallel linkage mechanism known as the “Stewart platform” was discovered as early as in 1965 [30]. It is designed according to a cubic conﬁguration. (3) airﬂowinduced nonrepeatable disk. high force to weight ratio. (4) mechanical resonance vibration. Each leg is equipped with a voice coil actuator.1 The servo control loop of a hard disk drive. and quantization noises. Hence how to deal with the variety of the disturbances is critical to the head positioning accuracy. The error is mainly due to.
The Stewart platform has been widely used in active vibration control.2. which comes with the use of spherical joints. This implies that if the axial forces can be measured and eliminated. The hexapod can be used to . a collocated force sensor. It has an important property for vibration control application: forces transmitted between the mobile plate and the base plate are totally axial forces of actuators. as well as a high precision pointing device and vibration isolator. FIGURE 1. each leg essentially acts as an axial springing parallel with a voice coil actuator.2 Hexapod from Micromega Dynamics. The Stewart platform (Hexapod) from MicromegaDynamics used as a vibration isolation device is shown in Figure 1. The plates are made of aluminum with a thickness of 20 mm and diameter of 250 mm. the vibration created by these forces can thus be eliminated. while the stiff design involves piezoelectric or magneto restrictive legs whose extensions can be controlled [32]. Stewart platforms can be divided into two main classes according to the stiffness of the legs: stiff type and soft type. Each leg of the active interface consists of a linear piezoelectric actuator. The hexapod has a cubic architecture and consists of two parallel plates connected to each other by six active legs. Flexible tips are used in order to avoid the problem of friction and backlash.Mechanical Systems and Vibration 3 widely used as space born structures. Thus the Stewart platform has become one of the most popular approaches for 6DOF active vibration control in precision systems due to its attractive properties. with the weight of the mobile plate at 1 kg. and ﬂexible tips to connect the two end plates. For the soft design.
motion induced in vehicles traveling over uneven surfaces. rotating unbalanced machinery. A collocated actuatorsensor pair thus enables the control of power that is supplied to the mechanical structure.3 Vibration sources and descriptions Any oscillatory motion of bodies that repeatedly appears is called vibration or oscillation. the associated signals for the actuator and sensor are power conjugated..3 shows the zoomedin view of the inside of the hexapod. respectively. 1. If an actuator and a sensor are collocated.3 Zoomedin view of the hexapod. Such situations can be complicated by the fact that the motion of the vibrat . actively increase the structural damping of ﬂexible systems attached to it. reciprocating machinery. Each of the legs in the Stewart platform consists of a PZT force sensor and an ampliﬁed PZT actuator. i. Flow induced vibration is generated by the forces exerted on an object by ﬂuid motion. and ground motion caused by earthquakes. structure ﬂexible modes. revealing the arrangement of the six collocated sensoractuator legs.4 Modeling and Control of Vibration in Mechanical Systems FIGURE 1. the modeling and vibration controls for the magnetic recording system and the Stewart platform will be detailed. Collocated actuatorsensor pairs are suitable in active vibration control applications in the sense that they guarantee damping and stability robustness if designed properly [33]. electrical torque.e. Some of them are: ﬂuid ﬂow. They form a collocated sensoractuator pair conﬁguration. In this book. There are usually forces associated with vibrations. Figure 1. the product of the actuated velocity and the measured force represents the power that is extracted from the mechanical structure. The wires shown in Figure 1. They can be induced by various types of excitation.2 are the outputs of the force sensors (each for one sensor) and inputs to the piezoelectric actuators (each for one actuator).
ground motion due to an earthquake. and grows indeﬁnitely when r > 0. and F0 is the force coefﬁcient. A model of the system undergoing such excitation is m¨ + cx + kx = F0 ω2 sinωt.3) (1.5) (1. and base motion that occurs when a vehicle travels over an uneven surface. and is equivalent to y(t) = Asin(ωt + φ). x is the system response. A = B2 + C 2 . Some oscillatory systems have simple harmonic motion of the form y(t) = Bsin(ωt) + Ccos(ωt). It is possible. in which Fourier analysis is used to identify the major frequency components in the vibration and they will be described in more detail in the later part of the chapter. Another complicating factor is the mass of the ﬂuid. however.4 Types of vibration There are several ways to categorize vibrations. vortexinduced vibration such as vibration of transmission cables. ω is the forcing frequency. to characterize them by means of statistical averages and spectrum plots. and cooling towers and chimneys. which increases the effective mass of the system. k and c are respectively stiffness and damping. the steadystate excitation force due to ﬂuid motion is sinusoidal with an amplitude proportional to the square of the forcing frequency. they can be calssiﬁed as follows. Further analysis of system response to ﬂow ﬂuid motion is quite complicated and requires detailed consideration of ﬂuid mechanics of the system.1) where m is the system mass. (1.4) (1. . In certain situations. 1. x ˙ (1. such as air ﬂow in hard disk drives. ﬂow through pipes and hoses having bends. C B cos(φ) = . sin(φ) = . Many types of motion cannot be easily represented by simple functions because they are essentially random.Mechanical Systems and Vibration 5 ing object can alter the ﬂuid ﬂow conditions. Examples include air turbulence to arm and suspension in hard disk drives. underwater cables used for towing and structural support. vibration caused by internal ﬂows. where y(t) is the displacement of a mass. thus changing the ﬂuid forces. Examples of vibration caused by ﬂuid motion include: wave action on structure.2) (1. structural vibration caused by ﬂuctuating aerodynamic forces such as turbulence [1][56].6) where oscillation amplitude decays exponentially when r < 0. Basically. A A There occur some oscillations having exponential amplitude as follows: y(t) = Aert sin(ωt + φ).
a large amount of excitation data collected may exhibit some statistical regularity.2 Damped and undamped vibration If during oscillation there is no energy lost or dissipated in friction or other resistance. mass and damper behave linearly. the principle of linear systems such as superposition holds. On the other hand. some knowledge of nonlinear vibration is desirable in dealing with practical vibratory systems. and the system undergoes oscillation. the ensuing vibration is known as free vibration. and techniques of analysis are more complicated and less well known. 1. In some cases.4 Deterministic and random vibration A vibration is known as deterministic vibration if it results from an excitation with value or amplitude known at any given time. Statistical methods can be used for analysis. if any of the basic components behaves nonlinearly. the superposition principle is not valid.4. bridges.6 Modeling and Control of Vibration in Mechanical Systems 1. Since all vibratory systems tend to behave nonlinearly with respect to amplitude level of oscillation.4. the excitation acting on a vibratory system is nondeterministic or random. if there is energy lost during oscillation. If a system vibration is due to an external force. If a vibration is linear. On the other hand.4. As for nonlinear vibration. the vibration is categorized as nonlinear vibration. and airplane wings. Linear and nonlinear differential equations are used to govern the behaviors of linear and nonlinear vibratory systems. as it is possible to estimate averages such as the mean and variance values of the random excitation. the resulting vibration is classiﬁed as linear vibration. the phenomenon known as resonance occurs. In these cases. and the value of the excitation at a given time cannot be predicted.1 Free and forced vibration If a system vibrates on its own after an initial disturbance and no external force acts on it. consideration of damping becomes extremely important. the vibration is known as undamped vibration. As is known. The occurrence of that resonance causing large oscillation may lead to failures of some structures such as buildings. A direct example of free vibration is the oscillation of a simple pendulum. If the frequency of the external force coincides with one of the natural frequencies of the system. turbines. 1. and there are well developed mathematical tools for analysis. the arising vibration is known as forced vibration. Examples of random excitations are air ﬂow inside hard . respectively. The oscillation in machines such as diesel engines that results from an external force is an example of forced vibration.3 Linear and nonlinear vibration If all the basic components in a vibratory system such as spring. a describing function is one approximation method used to analyze nonlinear vibratory systems. When analyzing vibration near resonance in physical systems.4. 1. it is called damped vibration.
8 0 0. If the excitation is random.4 0. 0 (1.08 0.4 An example of random vibration.8 0.04 0.7) 2 n=1 where ω = 2π/τ is the fundamental frequency. That is. the induced vibration is called random vibration..02 0.07 0.05 Time(sec) 0.01 0.2 −0.1 FIGURE 1.6 0. 2 ∞ a0 = + (an cos(nωt) + bn sin(nωt)). an . 1.4.4 −0. its Fourier series representation is given by a0 x(t) = + a1 cos(ωt) + a2 cos(2ωt) + · · · + b1 sin(ωt) + b2 sin(2ωt) + .4.09 0. bn are constant coefﬁcients given by ω π 2π/ω a0 = x(t)dt = 0 2 τ τ x(t)dt. if x(t) is a periodic function with period τ .. It can be described in terms of statistical quantities. such as that shown in Figure 1. (1.Mechanical Systems and Vibration 7 disk drives.03 0.6 −0.8) .06 0. and a0 .5 Periodic and nonperiodic vibration Periodic vibration can be represented by Fourier series as superposition of harmonic components of various frequencies.2 Vibration signal 0 −0. 0. road roughness and ground motion during earthquakes.
13) Thus the mean square value of x(t) is given by the sum of the squares of the absolute values of the Fourier coefﬁcients. 1. n n c0 = (1.4. A nonperiodic vibration x(t) can be represented by the integral Fourier transform pair: 1 2π ∞ −∞ x(t) = and X(ω)eiωt dω (1. (1.14) X(ω) = ∞ −∞ x(t)e−iωt dt.13) is known as Parseval’s formula for periodic functions. 0 τ (1.11) (1.6 Broadband and narrowband vibration A broadband vibration is a stationary random process whose spectral density function has signiﬁcant values over a range or band of frequencies which is approximately of the same order of magnitude as the center frequency of the band.10) x(t)sin(nωt)dt = 0 x(t)sin(nωt)dt. most periodic functions can be approximated by only a few harmonic functions.15) The RMS value of the nonperiodic function x(t) can be determined as RM S(x(t)) = ∞ −∞ X(ω)2 dω.8 an = bn = Let ω π ω π Modeling and Control of Vibration in Mechanical Systems 2π/ω x(t)cos(nωt)dt = 0 2π/ω 2 τ 2 τ τ x(t)cos(nωt)dt.9) (1.16) is known as Parseval’s formula for nonperiodic functions. The density in . Although the series in (1. 2πτ (1. Equation (1. 0 a0 .16) Equation (1.12) The root mean square (RMS) value of the periodic function x(t) can be determined as RM S(x(t)) = ∞ n=0 cn 2 .7) is an inﬁnite sum. 2 cn = (a2 + b2 )1/2 . (1.
A narrowband random vibration is a stationary vibration whose spectral density function has signiﬁcant values only in a range of frequency whose width is small compared to the magnitude of the center frequency.6 shows a vibration containing frequencies over a narrow band. 2S0 (ω2 − ω1 ). It is called bandlimited white noise if the band of frequencies has ﬁnite cutoff frequencies ω1 and ω2 . An ideal white noise is physically unrealizable.5 Spectrum density of broadband vibration. since the variance of such a random process would be inﬁnite because the area under the spectrum would be inﬁnite.5 describes a broadband vibration. A random process whose power spectral density is constant over a frequency range is called white noise. namely. The variance of a bandlimited white noise is given by the total area under the spectrum. Figure 1. which is composed of components containing frequencies over a wide or broad frequency range. . where S0 denotes the constant value of the spectral density. 16 14 12 10 Magnitude(dB) 8 6 4 2 0 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 1. It is called ideal white noise if the band of frequencies is inﬁnitely wide.Mechanical Systems and Vibration 9 Figure 1.
6 Spectrum density of narrowband vibration. .10 Modeling and Control of Vibration in Mechanical Systems 7 6 5 Magnitude(dB) 4 3 2 1 0 2 10 10 Frequency(Hz) 3 FIGURE 1.
So we also need a measure of the range of ﬂuctuation. variance and standard deviation. Td ) = 1 Td Td 0 [x(t) − µ]2 dt. in many applications the inputs are not well known and can be described only in terms of statistical measures such as mean value. For a signal continuous in time.5 1.5. such as those produced by digital data acquisition. The variance is calculated for a discrete signal as follows.Mechanical Systems and Vibration 11 1. (1. var(x) = σ 2 = 1 n n j=1 (xj − µ)2 .18) where the time duration of the data sample is Td . the mean is 1 Td Td E[x(Td )] = x(t)dt 0 (1. A measure that indicates the spread about the mean is the variance. which is deﬁned as the average value of the square of the difference between the signal and its mean.17) For a continuous function x(t). the mean is deﬁned as 1 n n E(x) = xj . The RMS value of x is E(x2 ). σ 2 = E(x2 ) − [E(x)]2. j=1 (1. Mean value The average is called the mean or expected value. and the mean is as follows. the variance is calculated from var(x. often represented by µ. The relation between the variance.19) where σ 2 is the variance and σ is called the standard deviation.21) . For discrete values. (1. (1. Variance Two signals may have the same mean value but one may ﬂuctuate with greater amplitude about the mean.20) The meansquare value of x is the expected value of x2 and is denoted E(x2 ). Simply specifying the minimum and maximum values is insufﬁcient because a single large ﬂuctuation (above or below the mean) can be misleading. the mean square.1 Random vibration Random process In contrast to deterministic excitations.
Random signals. like the example shown in Figure 1.23) 1.4. This means that if the excitation of a linear system is a Gaussian process. have no apparent pattern and never repeat. its mean and covariance will be independent of t: E[x(t)] = E[x(t + τ )] = µ. and is symmetric about the mean value. The graph of a Gaussian probability density function has a bellshaped envelope as seen in Figure 1. A very important property of a Gaussian process is that the form of its probability distribution is invariant with respect to linear operations. the standard deviation σx of x is calculated from σx = E(x2 ). which means that its statistical properties (such as its mean and variance) are timeindependent. (1. the steadystate response is generally a different random process.25) where x and σx denote the mean value and standard deviation of x. The mean x and ¯ ¯ standard deviation σx of x(t) vary with t for a nonstationary process but are constants for a stationary process.5.5. (1. and E[x(t)x(t + τ )] = σxx(τ ). but still a normal one.24) (1. even if we have the past values of the signal.7. Another characteristic of a random signal is that it is impossible to predict what the signal will be in the future. A random input will generate a random response. The probability density function of a Gaussian process x(t) is given by x 1 x−¯ 2 1 e− 2 ( σx ) .2 Stationary random process A special case of a random process is one that is stationary. 2πσx p(x) = √ (1. and the output signal given by its sensor will appear to be random. The only changes are that the magnitude of the mean and standard deviation of the response are different from those of the excitation.22) and is thus the same as the RMS value.12 Modeling and Control of Vibration in Mechanical Systems If the mean of x is zero. its spread is governed by the value of the standard deviation. 1. If x(t) is stationary.3 Gaussian random process A Gaussian or normal random process has a number of remarkable properties that permit the computation of random vibration characteristics in a simple manner. .
The Fourier series coefﬁcients.26) a(ω) = 1 2π 1 b(ω) = 2π ∞ −∞ ∞ −∞ x(t)cosωtdt. The spectrum graphically displays the frequency content of the signal.6.27) (1.7 Histogram of signal x.1 Vibration analysis Fourier transform and spectrum analysis We know that a periodic signal can be expressed as a Fourier series of harmonic functions. (1. gives a plot of the spectrum of the signal.6 1. 1. x(t)sinωtdt. (1.28) . A nonperiodic function is expressed with the following Fourier transform pair: ∞ −∞ ∞ −∞ x(t) = where a(ω)cos(ωt)dω + b(ω)sin(ωt)dω. when plotted versus frequency.Mechanical Systems and Vibration 180 13 160 140 120 Probability density 100 80 60 40 20 0 −4 −3 −2 −1 0 Signal x 1 2 3 4 FIGURE 1.
Sxx (ω). or power spectral density. Equation (1. 1. is one of the most useful functions in vibration testing.35) is the autocorrelation function and T is the time duration.33) 1. .34) T →∞ x(t)x(t + τ )dt −T /2 (1. while the time shift d affects only the phase angle of the transform. (1. (1. An equivalent form of the Fourier transform is given by ∞ −∞ (1.30) X(ω) = ∞ −∞ x(t)e−iωt dt.2 Relationship between the Fourier and Laplace transforms The Fourier transform X(ω) of x(t) is related to the Laplace transform X(s) as follows.3 Spectral analysis The spectral density. It is deﬁned as Sxx (ω) = where Rxx(τ ) = lim 1 T T /2 1 2π ∞ −∞ Rxx(τ )e−iωτ dτ.31) The spectrum of a nonperiodic signal is the magnitude of its Fourier transform.6.32) The transform of a timeshifted signal x(t − d) is X(ω)e−iωd and its spectrum is the same as the spectrum of x(t). 2π X(ω) = (1.29) x(t) = and 1 2π X(ω)eiωt dω.6.31) implies that the transform is symmetric about ω = 0. X(ω).14 Modeling and Control of Vibration in Mechanical Systems The Fourier transform of x(t) is X(ω) = a(ω) − ib(ω). 1 X(s)s=iω . that is X(−ω) = X(ω). that is. (1. (1.
(1.39) which means that the meansquare value can be computed from the spectral density. (1.38) Rxx(0) = Sxx (ω)dω = E(x2 ). . (1. A very useful relation is that ∞ −∞ 1 2π ∞ −∞ Rxy (τ )e−iωτ dτ.37) 1 T →∞ T T /2 x(t)y(t + τ )dt −T /2 (1.Mechanical Systems and Vibration Rxx(τ ) can be written as the inverse Fourier transform of Sxx (ω): ∞ −∞ 15 Rxx(τ ) = Sxx (ω)eiωτ dω.36) The crossspectral density is written as Sxy (ω) = where Rxy (τ ) = lim is the crosscorrelation function.
.
The system modeling in this chapter is discussed from the ﬁrst principle of actuators. A decoupling procedure is proposed which leads to approximate disturbance and noise models. 2.1 Introduction Modeling plays an important role in any control system analysis and design. To have a more realistic model. have to be involved inherently. nonlinearity induced by friction is measured under the condition that the actuator is controlled to avoid unsteady signal measurement. head/suspension assembly. An operator based modeling approach is adopted to model the hysteresis. 17 . especially due to high frequency unmodeled dynamics. lowfrequency disturbances are modeled as the output of an adaptive nonlinear mechanism with the error signal as the input. vibration modeling is based primarily on the spectral decomposition of the error signal measured in a closedloop system.2 Modeling of Disk Drive System and Its Vibration 2. and electronics card.2 System description A hard disk drive as shown in Figure 1. Particularly. actuator assembly. A physical system has to be modeled in order to design a control system. and an optimal model is obtained by minimizing the energy gain between the position and the modeling error. The hysteresis of friction versus actuator position is then obtained from the measurement in closedloop. Researchers have made many attempts to deal with a physical system: approximating it with a linear model at an operating point. Moreover. System dynamics measurement in the frequency domain is used to determine the parameters of a model in the form of transfer function. Physical systems are more or less nonlinear and may vary with time. disk. ﬁnding a control strategy that is robust and adaptable to the changes in the physical system. spindle and motor assembly. uncertainties.1 includes ﬁve major parts: baseplate and cover. In this chapter.
Servo bursts are conventionally written by costly dedicated servo writing equipment external to the disk drive. A secondary microactuator activated together with the VCM primary actuator is generally used to produce a higher bandwidth closedloop system. The electronics card involves drivers for the spindle motor and VCM. The major requirement in the seeking process is fast and smooth seeking with small or even no overshoot. The servo or position information is used to position the magnetic head on the disk surfaces. a high 0dB crossover frequency and a low hump of error rejection transfer function. and micro processors for servo control and control of interface to host computer. causing less offtrack. the servo performance is limited by mechanical factors in the actuator. spindle control to keep the spindle rotating at a speciﬁc speed with a minimum speed ﬂuctuation and a power driver to drive the spindle motor and the VCM actuator. A settling controller is used in between the above seeking and following modes. writing and seeking. Position measurement of the magnetic head is achieved by means of analyzing the position error signal (PES) calculated from the read back signal. the smaller the head position deviates from the desired track center. a servo controller to control the position of the R/W head during reading. acceleration current level and the control algorithm. disk platter. The headpositioning servomechanism is a control system that positions the R/W head from one track to another in minimum time. On the other hand. The seek time is limited by the actuator behavior. Once the actuator is regulating the position of the R/W head at the desired track. Dualstage actuation with a VCM as primary actuator and a microactuator as secondary actuator works as one way to achieve fast seeking and settling due to higher bandwidth. This typically requires a satisfactory servo loop based on the disturbance spectrum. The actuator servo channel consists of a demodulator producing position information from the servo burst read from the disk during seeking and following. ﬂex cable carrying head and VCM leads. Generally. it demands a high servo bandwidth. In the head/suspension assembly. the slider carrying heads ﬂies on top of the disk surface. a servo demodulator. the closer the tracks can be put together and the higher the track density becomes. an airbearing surface is created on surface next to the rotating disk. The defects such . etc. which uses a laserguided pushpin mechanism to position the write head on the disk surface until the servo burst information is written on the disk completely [53] [54]. An improved mechanical design is supposed to present less disturbance. and repositions the R/W head over a desired track with minimum statistical deviation from the track center. In a disk drive. The actuator assembly contains an actuator driven by voicecoil motor (VCM) and mounted via ballbearing at each end of a pivot shaft. the positioning information or servo information (“servo bursts”) is embedded in each disk surface. In this stage. The error transfer function must be well designed to yield a sufﬁciently small closedloop nonrepeatable runout. and a gimbal attaches the slider to the suspension.18 Modeling and Control of Vibration in Mechanical Systems The spindle and motor assembly includes disk clamps to clamp disks. The seek time is a measure of how fast the disk drive actuators can move the R/W head to a desired location. read/write (R/W) electronics. spindle motor. and arms to support suspension/head extension between the disks. a good closedloop servo system is expected to reject the disturbances.
The gray code track number provides coarse position information. Position error signal. B. It contains a coil which is rigidly attached to the structure to be moved and suspended in a magnetic ﬁeld created by permanent magnets. depending on the direction of the current. AGC ﬁeld. It determines the absolute position of the read/write head during seeking and tracking. the selfservo track writing creates the radial error propagation problem. More accurate placement of servo bursts is thus required accordingly. Figure 2. All tracks are divided into a ﬁxed number of radial sectors. a force is produced which accelerates the actuator radially inward or outward.3 2. A selfservo track writer regenerates timing and radial information from previously written tracks using the existing R/W head [135]. is the relative displacement of the R/W head from the track center.e. 2. i. The ﬁelds in an embedded servo pattern include Preamble. C. the position of the read/write head is obtained. sector or index mark. D bursts give ﬁne position error quantifying the amount of track misregistration. Approximately. or PES.Modeling of Disk Drive System and Its Vibration 19 as noncircularity caused by the spindle motor vibration in the servo bursts will make servo tracks more difﬁcult to follow in disk drives.3. the capability of writing and reading narrower tracks is improved. Conventional servo writers require a clean room environment because the disk and head will be exposed to the environment to allow access to the external head and actuator. data and servo information are written on all tracks in an interleaved manner. The demand of big storage capacity in hard disk drives requires high track density. When a current passes through the coil. The position feedback information for the disk drive servo mechanism consists of two components: gray code and position error signal. As technologies such as servo mechanism and head and media technology advance.1 shows one read back signal of the embedded servo. The external equipment is no longer needed in servo pattern writing and thus no open access is required for the head disk assembly and servo track writing does not have to be carried out in a clean room environment.1) . In an embedded servo. the deviation of the R/W head from the center of the track. (2.1 System modeling Modeling of a VCM actuator A linear VCM actuator moves in and out along a disk radius in one direction. Each sector starts with a servo pattern followed by user data information. resulting in an increased track density.. However. When PES and gray code are combined together. The pattern is repeated for each sector. fm = kt ic . which will hinder the whole process of servo track writing if not properly solved [140]. The produced force is a function of the current ic . track address in gray code and servo bursts. The A.
etc.6) . y and v are the displacement and the velocity of the read/write head. A second order transfer function can be used to represent the resonance. ˙ v = kv u.20 Modeling and Control of Vibration in Mechanical Systems FIGURE 2.. i.4) k u. and kv = kt /m with the actuator mass m.5) which includes the resonance model Pr (s). s2 (2. the actuator resonances have to be considered in the control design. With higher bandwidth.1 A read back signal of embedded servo. i. where kt is a linearized nominal value called torque constant. y= or y = ky v. and ξn be the associated damping coefﬁcient.. s2 (2. When the bandwidth of a control loop is very low and the resonance may not be a limiting factor to the control design. Then the actuator model becomes y= kv ky Pr (s)u. 2 + 2ξn ωn s + ωn Pr (s) = s2 (2. suspension.3) (2. ˙ (2. 2 ωn . the actuator model can be considered as the simpliﬁed and rigid one which is a double integrator with transfer function k/s2 .e.2) where u is the input to the actuator.e. arm. Let ωn = 2πfn correspond to a single resonance frequency fn . since the ﬂexible resonance modes will reduce the system stability and affect control performance if ignored. ky is the position measurement gain. The resonance of the actuator is mainly due to the ﬂexibility of the pivot bearing.
005 ≤ ξn ≤ 0. Other forms of Pr (s) include Pr (s) = and Pr (s) = 2 b2 s2 + b1 ωn s + b0 ωn 2 s2 + 2ξn ωn s + ωn 2 b1 ωn s + b0 ωn 2 s2 + 2ξn ωn s + ωn (2.05 is typical.7) (2. shown in Figure 2. 0. The peak of magnitude is higher when ξn decreases.2.2 Frequency responses of a second order transfer function.8) with zeros included to facilitate a phase lift which is usually associated with resonance modes.Modeling of Disk Drive System and Its Vibration 21 Different ξn gives different frequency responses. Phase(deg) Magnitude(dB) 0 0 Frequency(Hz) FIGURE 2. For lightly damped resonance. .
3 Measured VCM Bode plots (straight lines: pure double integrator k/s2 ).22 Modeling and Control of Vibration in Mechanical Systems FIGURE 2. .
v kv (2. we focus on friction modeling. Assume that nonlinearities such as friction and bias force and resonances are properly compensated. x 1 = ωn x 2 . we need to consider uncertainties inherent in the plant model. we have a statespace description y 0 0 ky 0 0 y ˙ v 0 0 k v b0 k v b1 v 0 ˙ + u. we need to properly handle the uncertain system so that it can perform under stricter margins. 2 + 2ξ ω s + ω 2 2 s s + 2ξn ωn s + ωn n n n (2. In this case.Modeling of Disk Drive System and Its Vibration 23 Figure 2. The deviation from the double integrator model in low frequencies in Figure 2. x1 = 0 0 0 ωn x 1 0 ˙ x2 ωn 0 0 −ωn −2ξn ωn x2 ˙ (2. Consider the resonance model in (2.10) Deﬁne x1 = i. the actuator model is a double integrator model.e. y = ky v.3 is due to the pivot friction and other nonlinearities. By curve ﬁtting each resonant mode. x2 = 2 u.14) On the other hand.11) (2.8) can be handled similarly). Friction models for the former can be . The nonlinearity modeling will be discussed in detail in the next section. 2. an uncertainty is introduced in the plant model.6) and (2.12) 2 sωn ωn u. Figure 2.2 Modeling of friction In this section. ˙ 2 b1 ωn s + b0 ωn v = kv 2 ˙ u..7) ((2.3.3 shows the measured frequency response of a VCM actuator system. To meet the robustness requirement against this unmodeled high frequency dynamics. There are basically two kinds of methodologies for friction control in the literature: Modelbased friction compensation and nonmodel based friction control. 2 s + 2ξn ωn s + ωn (2.9) (2.13) More resonance modes can be modeled similarly. For instance. It is written as y ˙ v ˙ = 0 ky 0 0 y 0 + u. a model with multiplicative uncertainty can be given by P = (1 +∆)Pn . Note that it is difﬁcult to involve all highfrequency dynamics in a modelbased servo control design. to make the mathematical models more realistic.4 shows the multiplicative uncertainty ∆ of a VCM. ˙ Then. one can obtain the parameters of the transfer function.
.24 Modeling and Control of Vibration in Mechanical Systems 20 10 0 −10 Magnitude(dB) −20 −30 −40 −50 −60 10 2 10 3 10 4 Frequency(Hz) FIGURE 2.4 Multiplicative uncertainty of a VCM.
a mathematical model will be developed to closely describe the friction hysteresis behavior. . the Coulomb model. where x(t) is the actuator response and π0 ∈ R is usually initialized to 0.e. and variations in breakaway forces. Therefore. x(t)). In order to have an accurate mathematical model for the hysteresis. Another dynamic model used in VCM pivot friction modeling is the preload and twoslope model. Hence we consider the operator model given by F (x(t)) = ax(t) + 0 ∞ wb (r)pr [π0 . However. 42] respectively in the frequency domain and the time domain. The elementary operator in the Prandtl hysteresis model is a rateindependent backlash or linear play operator. position dependence. a friction model involving dynamics is necessary to describe friction phenomena accurately. which is detailed in [40. In [45]. although the timedomain approach provides a good match between the time domain response of the model and the data collected. deﬁned by pr (π0 .5. but its implementation in real disk drives seems difﬁcult because of slow convergence [48]. it cannot guarantee a good match in the frequency domain. A nonlinear friction observer is then required for position control because the involved interim state is not measurable and has to be observed in order to estimate the friction force. x(t)]dλ. the creep model proposed in [49] is also incorporated. and the observer estimates the nonlinear disturbances due to pivot friction for disturbance cancelation. The accelerometer is employed to linearize the dynamics from the desired input signal to carriage angular acceleration. i.Modeling of Disk Drive System and Its Vibration 25 roughly classiﬁed into two categories: static models and dynamic models. a novel method for the cancelation of pivot nonlinearities is proposed and it consists of an accelerometer and a disturbance observer. a static friction model cannot capture observed friction phenomena like hysteresis. for example. Later. Fh (x(t)) = 0 ∞ wb (r)pr [π0 .15) where the weight wb deﬁnes the ratio of the backlash operator. x(t)]dr. There are various static models for friction. the Prandtl model [50] is less complex and more attractive in realtime applications. A relatively new dynamic friction model proposed in [43] combines the Dahl stiction behavior with arbitrary steady state friction characteristics. which is able to include the Stribeck effect. the viscous model. In this section. x(t)]dr + 0 ∞ wc (λ)lλ [ξ0 . The nonmodel based approaches include the neural network method [46][47] and the disturbance observer method [45]. Hysteresis nonlinearity can be modeled by a linearly weighted superposition of many backlash operators with different threshold r > 0 and weight values wb .. and vice versa. Among existing hysteresis models in the literature. as seen in Figure 2. which is used in [41] for VCM pivot friction modeling in HDDs. and friction models with the Stribeck effect. etc. an integrated model is proposed in [44]. However. to overcome the limitations of the above model. The resultant friction model needs to be iteratively improved and veriﬁed using the measured and the simulated responses. (2. The neural network method does not require full knowledge of the nonlinearity model.
26 Modeling and Control of Vibration in Mechanical Systems (2.18) . a.17) with the initial value zr (0) = max{x(0) − r. x(t)] = zλ (t) of the differential equation 1 zλ (t) + zλ (t) = x(t) ˙ λ with the initial value equation zλ (0) = ξ0 (λ). (2. if zr (t) − r < x(t) < zr (t) + r ˙ x(t). pr and lλ are the elementary hysteresis and linear creep operators. FIGURE 2. T ].5 The operator zr versus x. min{x(0) + r. wb (r) and wc (λ) are parameters to be determined. x(t)] = zr (t) of the rate independent hybrid differential equation ˙ x(t). The elementary hysteresis operator pr with threshold r is deﬁned as the solution operator pr [π0 .16) where t ∈ [0. π0 (r)}}. if x(t) = zr (t) − r zr (t) = 0. and are deﬁned as follows. if x(t) = zr (t) + r ˙ (2. Deﬁne the linear creep operator lλ with λ > 0 as the solution operator lλ [ξ0 .
and is denoted as Tex ∞ = ℓ (a. and thus a model (2.Modeling of Disk Drive System and Its Vibration The explicit integral formula for the linear creep operator lλ is as follows.6. (2. x(k)] + j=1 wcj lλj [ξ0 . wb2 . x(k)]. Note that Tex ∞ is a function of a. wb . if zri (k − 1) − ri < x(k) < zri (k − 1) + ri x(k) − ri . π0 (ri )}}. · · ·. Λ) . . wc = (wc1 . wc2 . L T k=1 x (k)x(k) (2. and Λ in (2. wc . wb . wb .22) and the initial value zλj (0) = ξ0 (λj ). then the approximation error e = f − fe . Denote wb = (wb1 . t 27 lλ [ξ0 . and Λ = (λ(1). 2) the discrete counterpart to the continuous elementary creep operator is given by (2.24) is minimized.20) where 1) the output sequence of the discrete hysteresis operator is calculated by pri [π0 . min{x(0) + ri . wbn ). x(k) + ri . and Λ.16) is developed as follows. · · · . Λ). wb .24) where L is the number of data points. wcm ).25) The MATLAB function fminsearch can be used to minimize ℓ (a. (2.20) can be obtained to approximate the friction f with the displacement x as the input. a discretetime model F (x(k)) of the operator F (x(t)) in (2. if zri (k − 1) + ri ≤ x(k) lλj [ξ0 .23) (2. wc . wb . n m F (x(k)) = ax(k) + i=1 wbi pri [π0 . x(k)] = zri (k). wc . Consider the hysteresis curve of the friction f versus the displacement x in Figure 2. x(t)] = e−λt ξ0 (λ) + λ 0 eλ(τ−t) x(τ )dτ. wc .19) For numerical implementation of the operatorbased modeling. We aim to ﬁnd optimal parameters a.20) so that (2. Let fe = F (x(k)) be the approximated friction. if zri (k − 1) − ri ≥ x(k) zri (k) = zri (k − 1). (2. x(k)] = zλj (k) with zλj (k + 1) = e−λj Ts · zλj (k) + (1 − e−λj Ts ) · x(k) (2. Λ) with respect to (a. wc . We deﬁne the energy gain between the actuator position and the error as Tex ∞ = L T k=1 e (k)e(k) . · · · . λ(2).21) with the initial value zri (0) = max{x(0) − ri . λ(m)).
28 Modeling and Control of Vibration in Mechanical Systems FIGURE 2. .7 A PZT actuated suspension.6 Friction f versus actuator displacement x. FIGURE 2.
An important point for PZT microactuator modeling is that the PZT element acts in series with the base plate springs. The compliance of the base plate is simpliﬁed as a single spring Kb . FIGURE 2. and the compliance of the ﬂex hinge elements is simpliﬁed as a single rotational spring Kr . Thus the displacement of the PZT element results in displacements of the springs. The free expansion of the piezo element is expressed as θf = Lm dexpV .Modeling of Disk Drive System and Its Vibration 29 2. cl1 (2.27) .3. (2.7 is considered in this section. The PZT and the base plate with spring constants Km and Kb can be equivalent to a single spring with constant KT = 1 Km 2 + 1 Kb .8 Equivalent spring mass system of PZT microactuator.3 Modeling of a PZT microactuator A piezoelectricbased microactuator located on the suspension as shown in Figure 2.26) The model is derived by applying forces at the interface of the piezo element and the base plate spring and by summing moments about the pivot point. The mechanical operation of the microactuator can be understood via an equivalent springmass system.
The displacement y is measured via . dexp is the piezo expansion coefﬁcient. c is the piezo thickness.3.1 An example Dynamics modeling The hard disk drive under consideration is shown in Figure 2. A typical frequency response of the PZT microactuator from voltage input to position output is shown in Figure 2. The following second order differential equation can be derived to capture the dynamic behavior of the microactutor: d2 θ dθ KT Lm dexpl1 2 +C + (Kr + KT l1 )θ = V.9.3. The actuator is driven by a driver which converts voltage differences into current differences linearly.28) where K is the torsional inertia. It includes the VCM actuator mounted with the arm and the suspension/head.9 A typical frequency response of the PZT microactuator. Kr is the torsional spring rate. V is the voltage. and l1 is the length as indicated in Figure 2.4 2. 2. 0 Magnitude(dB) −10 −20 −30 1 10 0 Phase(deg) −40 −80 −120 10 2 10 3 10 4 −160 −200 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 2.10.8.4. dt dt c K (2.30 Modeling and Control of Vibration in Mechanical Systems where Lm is the piezo length. The actuator dynamics measurement is taken using a Laser Doppler Vibrometer (LDV) and a Dynamic Signal Analyzer (DSA).
8209 × 1017 .527 × 109 ) (2. it is replaced with the pure double integrators model. i.964 × 108 )(s2 + 6032s + 2.. where the modeled one is plotted from the following transfer function P (s) obtained by curve ﬁtting to the measured frequency responses. P (s) = K Pres (s).3290 × 1017 . P (s) = KPf (s)Pres (s). K = 4.3 × 108 Pres (s) = 2 .10 An opened 1. The measured frequency response is shown in Figure 2. the LDV. K = 5. Translated to the error rejection function or sensitivity function. (s + 1056s + 6. However.30) which is plotted as the dotted curves in Figure 2.8inch hard disk drive. When the friction nonlinearity is neglected. 1 . it lifts the magnitude .11. and the frequency response is measured by using the DSA to generate a swept sine signal to excite the actuator. the friction in the actuator pivot [39] [40] is known to limit the low frequency gain of the control loop.25 × 2π120s + (2π120)2 s2 + 1081s + 7. s2 (2.Modeling of Disk Drive System and Its Vibration 31 FIGURE 2.e.29) The resonance of Pf (s) at 120 Hz is due to the nonlinearity of actuator pivot friction. Pf (s) = 2 s + 2 × 0.11.
11 Measured and modeled frequency responses of the VCM actuation system (LDV range 0.5 µm/V).32 Modeling and Control of Vibration in Mechanical Systems 100 80 60 Magnitude(dB) 40 20 0 −20 −40 1 10 10 2 Measured Modeled 10 3 10 4 200 0 Phase(deg) −200 −400 −600 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 2. .
7 dB.9467z + 0. and the phase margin is 49 deg.929z + 0. and shown in Figure 2. Thus the friction measurement is carried out under the closed loop control as shown in Figure 2.9023z 2 + 0. The VCM actuator model with consideration of nonlinearity F (x) is shown in Figure 2. which includes two pure integrators.3 ms.6442 kc = 0. The sampling time Ts is 83.7242 + ki )× .8. the resonance modes Pres (s) and the gain K given in (2.31) C(z) = kc(kp + kd FIGURE 2.12 Closed control loop of a disk drive with a VCM actuator for friction measurement via LDV. (2. and thus reduces the ability of the control loop to reject lowfrequency vibrations and affects the positioning accuracy. kd = 400 × 10−6 . and 3 V is respectively used as the reference signal in Figure 2.4. Ts z z−1 z 2 + 0. ki = 400. With the measured u and x.31). The relation between x and f can be obtained and shown as hysteresis curves in Figure 2. A 10 Hz sinusoidal signal with increasing amplitude of 0. it is necessary to compensate the friction impact. the openloop 0 dB crossover frequency is 945 Hz.12. kp = 0.14.2 Friction measurement and modeling Due to the ﬂuctuation of the head when the disk is rotating (rotational speed is 4200 RPM).12. The controller C(z) is a PID controller combined with a notch ﬁlter and is expressed in (2. The control signal u and displacement x are measured. Therefore.Modeling of Disk Drive System and Its Vibration 33 of the sensitivity function at low frequencies.5. 2.30). the gain margin is 8.13. . 1.0625. it is difﬁcult to have a steady displacement signal of the head.3. With the controller. z−1 Ts 0. ua can be obtained from x. and thus f = u − ua.6.
.5 µm/V) 2 3 4 FIGURE 2.5 −1 −1.34 Modeling and Control of Vibration in Mechanical Systems 2 1.5 0 −0. 0. FIGURE 2.14 VCM actuator modeling with friction nonlinearity model F (x).5 x 10 −3 −2 −4 −3 −2 −1 0 1 Displacement (V.5 1 Control signal u (V) 0.13 Control signal u versus displacement x.
08. r2 = 0.18. wb3 = −0. Since with the chosen peaktopeak values of x = 0.0024. With these parameters.33) . after 2000 iterations. the ri are respectively r1 = 0.5.20) can closely track f.. wcj = 0.6. The above mentioned operator based method to approximate hysteresis will be applied to model the hysteresis of f and x as shown in Figure 2. The preload model for velocity v and the twoslope model for position x are given by fv = kv v + ks sgn(v).08.4403. ξ0 = 0. Figures 2. It turns out that the model makes it possible to approximate hysteretic transfer characteristics without modeling the underlying physics.32) With m = 3.5783. r3 = 1.20). The modeled hysteresis from x to fe is drawn and compared with the measured one in Figure 2. m.6 as a mapping between the actuator position x and the friction force f. i. the modeling accuracy decreases.15.21) can be chosen as the amplitude of x. Note that the operator model fe = F (x(k)) describes the hysteretic characteristics in Figure 2. ri in (2. fe = F (x(k)) can be calculated from (2. Indeed. · · · . (2. (2.6.16 and 2. wb1 = −1. wb2 = 0. and fx = kax. For comparison. we shall ﬁnd F (x) to model the relationship between f and x.e.1293 > 0.5. It is observed that the time trace from the model (2. x > sx .20) is removed. and the error f −fe is small. in this case.1431. 1 and 3V. wc3 = 0. as shown in Figures 2. This is different from the friction models such as the preload and twoslope model in [40][42].0655.25. j = 1.17 . the minimal Tex ∞ = 0. wc1 = −0. which gives the minimal Tex ∞ = 0. an advantage of our operator based model is that the frequency response of the actuator can also ﬁt well to the measured one. as shown in Figure 2. Compared to the model in [40] [42]. If the creep term in the model (2.16 and 2.5. The time traces of fe and f are compared in Figure 2.1667.Modeling of Disk Drive System and Its Vibration 35 In what follows. n = 3 and the initial values π0 = 0. wc2 = −7.3341. It is seen that the modeled hysteresis and the measured one are close to each other. we also apply the preload and two slope model [40] for the hysteresis. x ≤ sx kb x + (ka − kb )sx .17 show that the quality of the agreement in the frequency domain between 70 Hz and 150 Hz decreases with lower excitation amplitude especially for the 0.5V case. a minimum error e is achieved and the optimal parameters are obtained as a = 9.
85e − 4. the magnitude response for the case of the amplitude of x(t) of 0.25 0.48.45 0.0145.25V. kb = 0.36 Modeling and Control of Vibration in Mechanical Systems Using the data with the amplitude of x(t) of 0. we obtain that ks = 1.0014. It can be observed that the plant input u versus position x ﬁts reasonably well to the measured results in the time domain.5 Time(sec) FIGURE 2.009.15 0. respectively. kv = 0.15 Measured and modeled friction and error.5 0 0.05 0.5 −1 −1. .4 0.5 x 10 −3 2 From measurement (f) From modeling(fe) Error (f−fe) 1.2 0. Using this model.1 0.5 1 Friction and error 0.17 deviates signiﬁcantly from the measurement results. sx = 0.3 0. However.5V. ka = 0. a comparison with the measured data in the time domain is shown in Figure 2.5 and 1 V. plotted as the dotted curve in Figure 2. in the frequency domain. 2.5 0 −0.5 −2 −2.19 for the amplitude of x(t) of 0.35 0.
. respectively.16 Actuator frequency response for sinusoidal reference with amplitude of 1 and 3 V. FIGURE 2.Modeling of Disk Drive System and Its Vibration 37 FIGURE 2.17 Actuator frequency response for sinusoidal reference with amplitude of 0.5 V.
2 0.38 Modeling and Control of Vibration in Mechanical Systems FIGURE 2.2 0.6 0.2 0 Displacement (µm) 0.6 FIGURE 2.8 −0. 1 0.8 −1 −0.18 Preload and twoslope model for friction modeling.4 −0.4 −0.2 −0.6 −0. .19 Plant input voltage u versus displacement x.6 −0.8 0.4 0.5V 0 −0.4 Input voltage u x 10 −3 Measured Modeled 1V 0.
1 Spectrumbased vibration modeling FIGURE 2. suspension and slider vibrations. y is the position of the R/W head and e is the position error signal. The noise signal n includes media and head sensing noises and also represents the effects of the PES demodulation noise which includes actual electrical noise and A/D quantization noise.20 Closedloop control system disturbances d1 . From Figure 2. (2. Since a controlled closedloop can provide steady signals. Figure 2. It is the combination of the repeatable runout which is synchronous with the spindle revolution and the nonrepeatable runout.Modeling of Disk Drive System and Its Vibration 39 2. The signal d1 represents all the torque disturbances to the system.20 shows a simpliﬁed blockdiagram of disk drive servo loop. The noise signal n is thus reasonably modeled as a broadband white noise. e(k) = −P (z)S(z)d1 (k) − S(z)d2 (k) + S(z)n(k).4 Vibration modeling Vibration in disk drives causes the deviation of the R/W head positioning from the desired track center.34) . which directly add to the relative position of the R/W head and the servo track.4. d2 and noise n. the suspension and the slider. the vibration source analysis is based on the signal that is collected from the closedloop system.20. 2. The signal d2 represents disturbances that are due to nonrepeatable motions of the disk and motor. Such disturbances include any torque due to airturbulence force to the actuator. The effects of the torque disturbances are dominant at frequencies that are relatively low when compared to the servo bandwidth.
and wi (i = 1. Two humps are obviously observed in the baseline curve. where P (z) is the transfer function of the discretized plant model P (s) and the sensitivity function or error rejection function is given by S(z) = 1 .36) Figure 2. One is in the frequency range lower than 100 Hz.23 shows the spectrum of NRRO component in the error signal e. d2 (s) and n(s) respectively. d2 .36) and Figure 2. Eq. models D1 (s). In Figure 2. Assume that d1 .22. Considering (2. The power spectrum denoted by Se of the error signal e is given by Se = P (z)S(z)2 d1 (k)2 + S(z)2 d2 (k)2 + S(z)2 n(k)2. Hence the disturbance and noise modeling can be carried out as follows. D2 (s) and N (s). and the ﬁrst hump is due to d1 through P (z)S(z) with a hump in a lower frequency range. D2 (z) and N (z) are the discrete forms of D1 (s). D2 (s) and N (s) are used to describe d1 (s). Hence the steps to obtain D1 . 2.37) where D1 (z). (2.35) Figure 2. (2. respectively.40 Modeling and Control of Vibration in Mechanical Systems FIGURE 2.21 Closedloop control system with disturbance and noise models. 1 + P (z)C(z) (2. (2.36) becomes Se = P (z)S(z)2 D1 (z)2 + S(z)2 D2 (z)2 + S(z)2 N (z)2 .22 shows the sensitivity function S(z) of a closedloop control system. N and D2 are . 3) are independent white noises with variance 1. and n are uncorrelated.21. D1 (z) and N (z) can be determined by ﬁtting weighted versions of P (z)S(z) and S(z) to the baseline curve of the spectrum and the spikes are considered as the effect of D2 (z). the other one is after 500 Hz. the second hump is caused by S(z) through S(z)2 n(k)2 .
.22 Sensitivity function S(z).Modeling of Disk Drive System and Its Vibration 41 FIGURE 2.
Till now.41) .39) where WL is a lowpass weighting function used to select Sb in low frequency range. jq Sb (j) = i=1+(j−1)q min Se (i).40) where WH is the highpass weighting function to select Sb in high frequency range. j = 1. D1 (z)2 = WL (z)Sb /P (z)S(z)2 . D(z)2 = {Se − [P (z)S(z)2 D1 (z)2 + S(z)2 N (z)2 ]}/S(z)2 . (2. Thus. Step 1). Compute D1 (z). 2. (2. N (z)2 = WH Sb /S(z)2 . Find Sb (j). (2. Step 4).23 PES NRRO spectrum. the baseline curve Sb can be ﬁt well by the identiﬁed D1 and N . Step 3). (2. The remaining part of the spectrum is then regarded as D2 . Compute N (z). · · · . L/q.38) where L is the length of Se and q is as small as possible. Step 2).42 Modeling and Control of Vibration in Mechanical Systems FIGURE 2.
and higher order neural networks. the output. The function approximation is deﬁned as follows.2. The RBF network is suitable for online nonlinear adaptive modeling and control. The different arrangement of the nodes and the interconnections deﬁnes various architectures of neural networks [73][74]. which are suitable to different kinds of applications.701588(s + 1. (2. the feedback controller C(z) can be optimized to minimize the error due to w = [w1 w2 w3 ]T by using the H2 optimal control method or other advanced control methods.1 [75] If f(x) : Rn → Rm is a continuous vector function deﬁned on a compact set Ω. and the activation function vectors. (s + 708.4s + 8.3916 × 10−5 (s + 575.42) (s2 + 315.24 is a RBF network. The nodes are interconnected by weighted links with weight parameters adjustable. because it is a linearly parameterized network.431 × 104 )(s + 766.0001317s + 4. radial basis function (RBF) networks.43) 1. Deﬁnition 2.45) for an acceptable small ε.178 × 104 ) 0. a multilayer neural network is usually used to generate the mapping from input to output since it can approximate any function under mild assumption with any desired accuracy.178 × 104 )(s2 + 315.Modeling of Disk Drive System and Its Vibration 43 D1 (s). x).4.6) .125 × 10−5 s + 4.373 × 108 ) D2 (s) = . The threelayer neural network shown in Figure 2. and wij is the second to the third layer interconnection .507 × 107 ) D1 (s) = With the disturbance and noise models.44) (s + 4603)(s + 1538)(s2 + 4451s + 1.04389s + 161.2 2. (2.1 Adaptive modeling of disturbance Neural network approximation A neural network usually consists of a large number of simple processing elements called nodes. the approximation problem is to determine the optimal W denoted by W⋆ . There are a number of neural networks studied for function approximation such as multilayer perceptron networks. and s ∈ Rp are respectively the input. for some index d such that d(y(W⋆ .6)(s2 + 0. 2.672 × 107 ) N (s) = .5s + 8. and exhibits a fast initial rate of learning convergence.2)(s2 + 8609s + 4. (2. 1.8)(s + 575.1695(s + 1. D2 (z) and N (z) using the bilinear approximation method [81]. x) : Rt × Rn → Rm is an approximating function that depends continuously on W and x. where x ∈ Rn .376 × 108 ) (2. In control engineering. then.4. y ∈ Rm . has spatially localized learning capability and thus has better memory during learning. and any y(W.271 × 104 )2 (s2 + 6. D2 (s) and N (s) are then obtained as follows from D1 (z). f(x)) ≤ ε.4)2(s2 + 0.
.44 Modeling and Control of Vibration in Mechanical Systems FIGURE 2.24 Threelayer RBF neural network.
They are separately written as sj (x) = exp[ −(x − cj )T (x − cj ) ]. The output yi is given by p 45 yi = j=1 wij sj ( x − cj 2 ). Several functions such as Gaussian. nonlinear or unknown vibration sources. cj ∈ Rn is the center vector.2.4. 2 σj (2. (2.48) where ε(x) ∈ Rm is the reconstruction error vector. (2. velocity and acceleration. past information such as x(k − 1) and x(k − 2) is used as input. (2. Hardy’s multiquadric and inverse Hardy’s multiquadric functions have been used as activation functions. a dynamic neural network is needed in order to fully emulate the dynamics. disturbances in frequency range lower than a few hundred Hz are quite dominant and they may be due to torque disturbances.50) 1 2 σj + (x − cj )T (x − cj ) . In discretetime neural networks. In a dynamic neural network.51) For a mechanical system with dynamics given as a function of position. and WT = [wij ]T . The approximation of a general function f(x) : Rn → Rm can then be expressed as f(x) = WT s(x) + ε(x) (2.47) where p is the number of nodes. Here we use d1 to represent these disturbances. x = [x1 x2 · · · xn ]T ∈ Rn is the network input vector. This section aims to model the lowfrequency disturbance d1 with an .Modeling of Disk Drive System and Its Vibration weights. (2. 2. m.2 Disturbance modeling In some mechanical motion systems. · · · . y(x) = WT s(x). dynamic variables such as velocity and acceleration are involved in the input x.46) or equivalently in a matrix form.49) sj (x) = and sj (x) = 2 σj + (x − cj )T (x − cj ). • denotes the Euclidean norm. 2. i = 1.
52). f(Φ(e(k))) is denoted as f(Φk ).. e = −P (s) · d1 − d2 + n. ˙ The weight vector W (k) in (2. (2. It is noted that the contribution of disturbances and noise to e is described as in (2. (2.46 Modeling and Control of Vibration in Mechanical Systems adaptive nonlinear strategy based on the measured error signal e. and (2) it possesses a localized response.53) is dependent on some time history of the measured signal e.. (2. the unknown function f can be approximated using neural networks. s(Φk ) ∈ Rp is a basis vector function with a suitable number of nodes p.54) where l is to be determined. the low frequency disturbances can be approximated as ˆ d1 = f(Φk ) = W T (k)s(Φk ) + ε. As stated in the previous section.52) where f(·) is an unknown function of the bounded vectorvalued function Φ(e(k)) with e(k) = [e(k) e(k − 1) · · · e(k − l)]T . The disturbance d1 is modeled as ˆ d1 = f(Φ(e(k))).55) where ε is the modeling error. sp (Φk )]T . Here we select a Gaussian RBF based neural network. For brevity. The properties are.55) can be calculated iteratively according to the following weight update law: W (k + 1) = (1 − δ)W (k) − Γs(Φk )e(k) (2. The speed of learning rate is related to the chosen Γ matrix. .57) with the adaptation gain matrix Γ ∈ Rp×p being diagonal and satisfying Γ > 0 and the forgetting factor 1 > δ > 0. Next. The forgetting factor δ is to ensure the boundedness of W (k) when the system is subjected to bounded disturbances.56) where σei and cei (σ∆ei and c∆ei ) are the variance and center position of the measurement e (velocity e). The Gaussian RBF network has some attractive properties and thus has been widely used in nonlinear control and signal approximation. (1) it is bounded and strictly positive. Based on the Gaussian RBF network. the power spectrum in Figure 2. (2.53) (2. Thus (2. − (e(k)−ce )2 i 2 σe i − (e(k)−e(k−1)−c∆e )2 i σ2 ∆yi . The radial basis function s(Φk ) is written as si (Φk ) = e ·e s(Φk ) = [s1 (Φk ) s2 (Φk ) . In the openloop without controller C(z).27 of a position error signal e will be used to verify the adaptive modeling scheme.
1. 38. the disturbance d2 and the noise n are represented approximately by ˆ d2 = D2 (s)w 2 . after some trials.60. two more cases with p = 5. 38.59) ˆ the time trace d1 calculated from (2.21.61) (2.55) gives the best approximation of the disturbance d1 .26.5.27 with the component of disturbance ˆ P (s) · d1 removed and considering that noise n is a white noise. It is observed that the time trace from the model (2.61 × 108 ) . σ increases from 4. the model from e to d1 is chosen to be nonlinear.55) can ˆ give a close tracking of the original one.60) The NRRO spectrum obtained by combining the nonlinear and linear models is compared with the measured one in Figure 2.5. (2. 33}. The effect of different p and Γ on the modeling accuracy will be evaluated ˆ later. and Γ = Ts · 106 · 220. from the spectrum in Figure 2.695 × 107 )(s2 + 3340s + 5.5. The disturbance d1 affects the error signal or the output e through the plant P (s). Γ = Ts · 106 · diag{38. 88. 88.62) (2. l = 1. 220. (s2 + 245s + 1. To evaluate the effects of different values of p on the modeling error. 220.Modeling of Disk Drive System and Its Vibration 47 A low frequency signal with the spectrum as in the low frequency range in Figure 2.43 to 4.0019(s2 + 3329s + 1. The spectrum of d1 is seen in Figure 2. ˆ Let the modeling error de = d1 − d1 .668 × 107 )(s2 + 477. p = 9. thus the nonlinear ˆ model from e to d1 aims to generate a cancelation signal of the disturbance d1 .58) (2. ˆ (2.5s + 5. 66. 88. it was found that 2 2 when σy1 = σ∆y1 = 10. Γ = Ts · 106 · diag{33. D2 (s) = 0. Moreover. Figure 2.5. 9.63) The σ value of the modeling error e can be seen in Table 2. 5.27. which means higher p may not give a better result . For p = 1 with zero center positions cy1 and c∆y1 . Since the low frequency disturbances due to torque and bias are generally nonlinear. The veriﬁcation is implemented with the sampling period Ts = 1/45000sec. δ = 0. 38. 9 and the following parameters for Γ are investigated.25 shows the simulated time trace comparison of d1 and d1 from the nonlinear model (2.005 w3 . 66. 88. (2.55).27 is generated and injected as disturbance d1 in Figure 2. It is found that the adaptive nonlinear modeling method can indeed be used to model the disturbance d1 that is dominant in low frequency range.5}. p = 5.701 × 108 ) and n = N (s)w3 = 0. p = 1. With p = 1.
The time sequence comparison ˆ of d1 − d1 is shown in Figure 2. Compared with p. The other two cases are similar.43 to 11.58). For p = 1. but also in the frequency domain. It turns out that the derived model matches well the measured model not only in the time domain. Γ inﬂuences the modeling accuracy more signiﬁcantly.5 Conclusion In this chapter the models of a VCM actuator and a PZT microactuator have been derived on the basis of their physical operations. 2. The developed vibration and noise modeling approaches are based on the error signal power spectrum of the closedloop system. The models have been veriﬁed with the measured frequency responses from voltage input to displacement output.28. A Prandtl operator based model has been used to model the friction nonlinearity. for low frequency disturbance modeling. and the optimal model parameters have been obtained by minimizing the energy gain between the actuator position and the modeling error. To have a complete model of an actuation system. an adaptive nonlinear scheme with system measurement as input has been applied to approximate the original disturbance for realtime compensation. with Γ being changed to 10% of the value in (2. where the difference is clearly seen. this chapter has also addressed the modeling of uncertainties in high frequency and nonlinearities such as actuator pivot friction. the σ value of the error e increases from 4. In particular. .48 Modeling and Control of Vibration in Mechanical Systems in this practical application.
55/11 4.Modeling of Disk Drive System and Its Vibration TABLE 2.1 49 ˆ σ values of the modeling error (d1 − d1 ) for different p and Γ p 1 5 9 ˆ1 − d1 ) for Γ and 0.43/11 4.05 0.15 0.60/10 σ values of (d −3 5 4 3 2 1 0 −1 −2 −3 −4 x 10 Original disturbance d1 Original disturbance and the modeled (µ m) d1 from nonlinear modeling 0 0.25 ˆ Original disturbance d1 and the modeled d1 .(×10−4) 4.25 FIGURE 2.1Γ. .2 0.1 Time(sec) 0.
.5 1 0.5 0 0 1000 2000 3000 Frequency(Hz) 4000 5000 6000 FIGURE 2.26 ˆ Power spectrum of d1 .50 Modeling and Control of Vibration in Mechanical Systems 4 x 10 −3 3.5 2 1.5 3 NRRO magnitude(µm) 2.
i.5 2 1. ˆ .5 0 Modeled Measured 0 1000 2000 Frequency(Hz) 3000 4000 5000 6000 FIGURE 2..5 NRRO magnitude(µm) 3 2.27 NRRO power spectrum from measurement and disturbance models.e. e = −P (s)· ˆ ˆ d 1 − d 2 + n.Modeling of Disk Drive System and Its Vibration 51 x 10 −3 4 3.5 1 0.
28 ˆ Modeling error (d1 − d1 ) for different Γ with p = 1.1 Time(sec) 0.2 0.52 Modeling and Control of Vibration in Mechanical Systems 4 3 2 Modeling error (µm) 1 0 −1 −2 −3 −4 x 10 −3 Γ 0.1Γ 0 0. .15 0.25 FIGURE 2.05 0.
adaptive ﬁltering is often adopted in control and signal processing ﬁelds. Indeed. an ampliﬁed piezoelectric actuator.3. This piezoelectric Stewart platform can be used as a precision pointing device.2 and Figure 1. Referring to the discussion in Chapter 1. adaptive identiﬁcation algorithms can be used. 3. a force sensor and two ﬂexible joints.2 System description and governing equations The study of active damping of ﬂexible structures has been induced by the requirement of structural stability. for the modeling of an unknown system. a vibration isolation mount. Each active leg consists of a force sensor. The Laplace form of the governing equation of the motion in this system is 53 . In this chapter. An adaptive ﬁltering approach will then be introduced and subsequently used to model the platform and verify the transfer function obtained. each leg of the hexapod Stewart platform consists of an ampliﬁed piezoelectric actuator. followed by the derivation of a transfer function from the actuator force to the sensor output based on frequency response data. In order to obtain more accurate models for physical systems. Shown in Figure 1. and two ﬂexible joints. Figure 3. the hexapod consists of two parallel plates connected to each other by six active legs. the governing motion equation of a piezoelectric Stewart platform will be obtained ﬁrst.3 Modeling of Stewart Platform 3. or an active damping mount. Most solutions to the active damping problem rely on the integration of smart actuators and sensors in the structure itself.1 Introduction An adaptive control system is able to adapt to the changes of a physical system.1 shows an equivalent system of each leg connecting two rigid bodies: the disturbance source m and the sensitive payload M that are connected by a force sensor and a piezoelectric actuator represented by its elongation δ and spring stiffness k. The legs are mounted in such a way to achieve the geometry of cubic conﬁguration.
(3.1 Singleaxis system using piezoelectric stiff actuator.3) Consider the piezoelectric hexapod integrated in a structure. A force sensor is located in each leg of the hexapod and collocated with the actuator. The corresponding sensor output is y = k(q − δ).6) .1) (3. δ M ms2 + k(M + m) (3.. (3.54 Modeling and Control of Vibration in Mechanical Systems FIGURE 3. x represents payload frame displacement. M s2 xp = −ms2 xd = k(xd − xa ) = y and δ = xp − xa . i. Let M and K be the mass and stiffness matrices of the global passive system.2) The transfer function between the extension δ of the piezo stack and the sensor output y is M ms2 y =k . and f = kδ represents the equivalent piezoelectric force in the leg. The dynamic equation governing the system is written as M s2 x + Kx = Bf. structure and hexapod.e. Taking into account the relationship between the leg extension and the payload frame displacement q = B T x.4) where B is the force Jacobian matrix.5) where y is the force sensor output and q is the leg extension from the equilibrium position. (3. (3.
e(k) that is then applied back to the ﬁlter to adjust the ﬁlter weights so as to minimize the estimation error in the statistical sense.Modeling of Stewart Platform the sensor output equation becomes y = k(B T x − δ). y(k) = W T (k)X(k). W 1.7) The sensor output y is to be used as the input to the controller. The difference between y(k) and d(k) is deﬁned as the estimation error. the statistically stationary time sequence of input signal. The ﬁlter output at time k. In most practical instances the adaptive process aims to minimize the meansquare value. Next an adaptive ﬁltering approach will be used to model the hexapod Stewart platform and verify the dynamics equation (3.3).2. 3. y(k). is to be as close as possible to the desired response. or average power of the error signal [37]. where W T (k) = [W 0 W 1 W 2 · · ·]. Optimization under this criterion is . x(1). d(k).3.2 Linear discrete time adaptive ﬁlter.3 3. · · ·. In Figure 3. is applied to the linear discrete time ﬁlter whose coefﬁcients are W 0.1 Modeling using adaptive ﬁltering approach Adaptive ﬁltering theory LMS algorithm FIGURE 3. · · ·. 55 (3. and X T (k) = [x(k) x(k − 1) x(k − 2) · · ·]. x(0).
11) The updating equation of the adaptive ﬁlter coefﬁcient: W (k + 1) = W (k) − becomes W (k + 1) = W (k) + µX(k)e(k).8) e(k) = d(k) − y(k) = d(k) − W T (k)X(k).10) where Rdx is the cross correlation function between d(k) and x(k). In the LMS algorithm. In many practical applications. ˆ ∇ξ(k) = ∇e2 (k) = 2[∇e(k)]e(k). ξ(k) = E[e2 (k)]. the statistics of d(k) and x(k) are unknown. 2 . e(k) = d(k) − W T (k)X(k). ˆ ξ(k) = e2 (k).9) ξ(k) = E[(d(k) − W T (k)X(k))2 ] = E[d2(k)] − 2W T (k)E[d(k)X(k)] + W T (k)E[X(k)X T (k)]W (k) = Rdd − 2W T (k)Rdx + W T (k)Rxx W (k). where E[ ] denotes the expected value.56 Modeling and Control of Vibration in Mechanical Systems most effective in statistical sense. ∂e(k) ∇e(k) = = −X(k). (3. (3. So the performance index of the adaptive ﬁlter is determined by the meansquared error (MSE) criterion in which the cost function is deﬁned as follows. Therefore another method called the Least Mean Square (LMS) algorithm was introduced by Bernard Widrow [37]. (3. the exact knowledge of gradient vector is not available. The mathematical derivation of the LMS algorithm is as follows. ∂W (k) ˆ ∇ξ(k) = −2X(k)e(k). (3. Thus the method of steepest descent [35] cannot be implemented in practical situations. (3.12) µ ˆ ∇ξ(k). Rdd and Rxx are respectively the autocorrelation functions of d(k) and x(k). instantaneous squared error is used as an estimation of the mean squared error.
The maximum stable step size is inversely proportional to the ﬁlter order L and the power of the reference signal x(k).12) is the wellknown LMS algorithm that is suitable for practical signal processing applications because of its simplicity and the availability of the instantaneous error in real time. Equation (3.Modeling of Stewart Platform 57 FIGURE 3. A technique used to optimize the convergence speed. The NLMS algorithm is given as follows. Px (k) = L (3. ˆx(k) LP (3. W (k + 1) = W (k) + µ(k)X(k)e(k).3 Block diagram of the LMS adaptive ﬁlter.14) where α is a normalized step size that satisﬁes 0 < α < 2. (3. is known as the normalized LMS algorithm (NLMS).15) . independent of the reference signal power. Normalized LMS algorithm The LMS adaptation process is very much dependant on the step size µ and the reference signal power. The step size determines the system convergence rate and stability [36][37].13) where each variable is identical to the one for the LMS algorithm except for µ(k) that is an adaptive step size computed as µ(k) = α . L is the ﬁlter length.3. X T (k)X(k) ˆ . The block diagram of the LMS algorithm is illustrated in Figure 3. ˆ The estimation of Px (k) can be done using the mean square value of the reference signal as follows. and ˆ Px (k) is the estimated power of the reference signal x(k).
3.58 Modeling and Control of Vibration in Mechanical Systems Then Equation (3. When the error signal reaches the minimum level. But the sampling frequency of the adaptive ﬁlter is set to 10 kHz to give enough time for adaptive ﬁlter to compute the ﬁlter weights within each sample of the training signal. P (z) is the transfer function of the system to be identiﬁed. W (z) is a digital ﬁlter used to model P (z) based on the LMS error minimization algorithm. convergence of the adaptive ﬁlter used in the identiﬁcation section of the system can be affected by a large amount of primary vibration signal [36].14) reduces to µ(k) = α .17) This constant ensures that the step size does not tend to inﬁnity in the case of a zero input signal. Sampling frequency of the white noise generator is set at 1 kHz so that the response of the PZT actuator will be conﬁned to the bandwidth of 500 Hz (half of sampling frequency). a salient feature of this algorithm is that it lowers the inﬂuence of the input signal on the noise ampliﬁcation effect.3. The LMS algorithm will adaptively adjust the coefﬁcients of ﬁlter W (z) to minimize e(k) based on the least MSE criterion. Filter tap values are recorded through the dSPACE manager . Therefore an ofﬂine identiﬁcation technique will be applied for the system modeling. A 16th order LMS adaptive ﬁlter with adaptation step size of 0. especially when the input signal x(k) is large. A Simulink program is developed for ofﬂine identiﬁcation of the platform. In an online identiﬁcation system. An adaptive identiﬁcation method can be applied either online along with the vibration control or ofﬂine prior to the vibration control. µ(k) = α . The block diagram of the adaptive identiﬁcation is illustrated in Figure 3. the ﬁlter W (z) represents a model of P (z). Therefore.4.2 Modeling of a Stewart platform Adaptive identiﬁcation is a technique that uses an adaptive ﬁlter to model an unknown system.16) This step size is the most widely used for the NLMS algorithm. where γ is the inducednorm from noise or signal inputs to the ﬁltering error. X T (k)X(k) (3. The difference between the two outputs d(k) and y(k) is fed back into the LMS algorithm as error signal e(k). Both systems P (z) and W (z) are excited by a band limited white noise. The Simulink program for the identiﬁcation is downloaded into a dSPACE real time interface board DS1104. The LMS adaptive ﬁlter approach is applied for the modeling of the Stewart platform. Furthermore. ε + X T (k)X(k) (3. A variant of the NLMS algorithm uses a small constant ε as follows. The NLMS algorithm guarantees an attenuation level of γ ≤ 1.01 is chosen. A bandlimited white noise is used as the training signal. the number of adaptive ﬁlters required for an adaptive control system will be increased. Identiﬁcation process for each PZT actuator is performed alternatively.
software during the identiﬁcation process.624 × 1011 .6 shows the frequency responses of an PZT actuator.6 agrees with that of the model in (3. The result shows that there is a resonance peak at about 240 Hz.9s4 − 1. Figure 3.943 × 1016 (3.4 System identiﬁcation using LMS adaptive ﬁlter.507 × 1010 s − 5. It is clearly shown that the estimated transfer function is well ﬁtted to the real one.5 is the phase and magnitude responses of one of the six PZT actuators. It is also noted that the shape of the frequency responses in Figure 3. Experimental results of the active vibration control system also indicate that the frequency region that can attenuate the vibration signal is between 60 Hz and 220 Hz. From the results of the identiﬁcation of six PZT actuators of the Stewart platform in Figure 1. The phase response between 60 Hz and 200 Hz is approximately linear.499 × 104 s3 − 5.378 × 106 s3 + 1.2.621 × 1010 s2 + 6. By applying a frequency domain modeling approach.82 × 1012 s + 1. s5 + 3232s4 + 5. Figure 3.3).Modeling of Stewart Platform 59 FIGURE 3. The transfer function of the PZT actuator is obtained as: P (s) = −190. one of the resonance frequencies of the smart structure is found to be around 240 Hz.703 × 108 s2 − 2.18) . an approximate model in terms of a transfer function for each leg can be obtained. including the estimated and the experimental ones.
5 Frequency responses of a PZT actuator.60 Modeling and Control of Vibration in Mechanical Systems FIGURE 3. .
Modeling of Stewart Platform 61 FIGURE 3.6 Estimated and experimental frequency responses. .
. The model obtained via the adaptive identiﬁcation or the experimental testing has shown consistency with the governing motion equation. The LMS adaptive ﬁltering approach has been adopted for the system identiﬁcation.4 Conclusion This chapter has studied the modeling of an active piezoelectric Stewart platform.62 Modeling and Control of Vibration in Mechanical Systems 3.
etc.4 Classical Vibration Control 4. and human pain and discomfort. (1) using hybrid mass dampers to apply a control force to a movable mass so as to reduce building sway caused by wind and seismic waves. 4.1 Passive control Isolators Vibration isolation methods are used to reduce the undesired effects of vibration. and absorbers.2. (3) damping out vibrations using piezoelectric devices installed on the trailing edge of helicopter blades. Vibrations can sometimes be eliminated theoretically. An active vibration control is required for a system if it needs an external power to perform its function.1 Introduction The presence of vibration often leads to undesirable effects such as structural or mechanical failure. isolators.2 4. control of natural frequency. It involves the insertion of a resilient member called an isolator between the vibrating mass and the source of vibration so that a reduction in the dynamic response of the system is achieved under speciﬁed conditions of vibration excitation [4]. worsening positioning performance. However. reduction of vibration is preferred so as to achieve a compromise between an acceptable amount of vibration and a reasonable manufacturing cost. such as balancing of rotating and reciprocating machine. An isolation system is said to be active or passive depending on whether or not external power is required for the isolator to perform its function. damping and stiffness modiﬁcation. Various classical techniques of vibration control for the purpose of vibration reduction have been presented. frequent and costly maintenance of machines. Examples of some active vibration control include. (2) reducing aircraft cabin noise by attenuating the vibration of the large panels of thin metal that form the cabin walls. A passive isolator consists 63 . Some of the techniques will be introduced brieﬂy in this chapter. due to the high manufacturing cost that may be involved in eliminating vibration.
cork.. 4.3) (4. a damped dynamic vibration absorber can be used [5]. the vibration of the system can be reduced by using a dynamic vibration absorber.2. and an energy dissipator. The motion equations of the masses m1 and m2 are written as m1 x1 + k1 x1 + k2 (x1 − x2 ) = F0 sinωt. The steadystate amplitudes of the mass m1 and m2 are given by (k2 − m2 ω2 )F0 2.4) the vibration amplitude of the machine. damping. ¨ (4. while operating at its original resonant frequency.2.. Typical examples of passive isolators include metal springs. We consider an auxiliary mass m2 attached to a machine of mass m1 through a spring of stiffness k2 . The main idea is to create another source of vibrations which will cancel out the original vibration if it is correctly tuned.64 Modeling and Control of Vibration in Mechanical Systems of a resilient member.5) (4. (k1 + k2 − m1 ω2 )(k2 − m2 ω2 ) − k2 k2 F0 X2 = 2. Consider the machine operating near its resonance ω2 ≃ k1 /m1 before the addition of the dynamic vibration absorber. where . and rubber springs. 4.2) where x1 and x2 represent displacements of m1 and m2 . The principle is to use the kinetic energy in the resonant mass system.3 Resonators Resonators are used in certain cases where it is easier and more efﬁcient to locate antivibration systems near the vibration source. pneumatic springs. m2 m1 (4. The above mentioned dynamic vibration absorber removes the original resonance peak in the response of the machine. If it is necessary to reduce the amplitude of vibration of the machine over a range of frequencies. (k1 + k2 − m1 ω2 )(k2 − m2 ω2 ) − k2 X1 = Equation (4. If the absorber is designed such that ω2 = ω2 = k2 k1 = . ¨ m2 x2 + k2 (x2 − x1 ) = 0. respectively.2 Absorbers If a system is acted upon by a force whose excitation frequency nearly coincides with its natural frequency. m2 the amplitude X1 of the machine m1 will be zero. stiffness. i.e. but introduces two new resonance peaks.3) implies that if k2 .1) (4. which is simply a springmass system. i.e. will be zero.
denotes the capacity of a system to store strain energy. For an oscillatory system. Dynamic behavior can be modeled in several ways. Damping modiﬁcation Damping deﬁnes the ability of a structure to dissipate energy. In the case of a helicopter. where the stiffness constant ks is expressed in the unit of force per unit length. The motion of the counterweight is mainly inplane motion so that the system will generate inplane loads to eliminate the inplane vibration. damping is a measure of how much energy is dissipated by the system during an oscillation cycle. More complicated laws exist. and the stiffness parameter k(x) is a function of how much the spring has been elongated. This is a concept for structure isolation that is achieved by canceling the variable loads transmitted to the support structure. Next we will brieﬂy introduce stiffness modiﬁcation and damping modiﬁcation. Dynamic forces transfer their energy to the structure. in a rotating structure the stator is often excited by the rotor via inertial effects such as imbalance or other effects such as the aerodynamic asymmetry of the blades or ﬂaps. Stiffness. Fs = ks ∆x. Stiffness modiﬁcation The dynamic behavior of structures results from the exchange and dissipation of energy. . For example. External dynamic forces are balanced in a more complex way with inertial and damping forces. For example. structural connections between components add damping to a structure.Classical Vibration Control 65 the mass is coupled to the vibrating source either by kinematic coupling or via a very ﬂexible link [8]. where the parameter a would depend on the particular material being modeled. This is a linear model where the spring displacement is measured from the rest length. for example. which then responds via several mechanisms.2. often schematically and conceptually represented by a spring. If the external force is static or quasistatic. Positioning a resonator on the rotor is to generate loads that oppose the excitation loads. In practice. The best known is Newton’s second law of motion. The stiffness force follows Hooke’s law.4 Suspension Suspension acts as the link between two structures that is designed in order to isolate one of the structures from the other. such as bending or extension. It is supported by three springs which allow it to vibrate in a plane perpendicular to the axis of rotation. An additional technique consists of introducing a ﬂapping mass whose inertial effects will neutralize the excitation inputs. the resonant mass is placed on the axis of the rotor hub. The characteristics of the link are determined by analyzing the dynamic behavior and vibrations without modifying its main function in static characteristics. structural stiffness forces develop to create an equilibrium. 4. the link characteristics can be modiﬁed by varying its stiffness or appropriate positioning of the system natural frequencies and its damping in terms of energy transfer. the nonlinear relation F = k(x)∆xa.
in this section we discuss an approach with minimal mechanical alternation. the positioning accuracy of test equipment such as spin stand and servo track writer has to be increased. Reference [24] investigated the effect of suppressing resonance amplitude of disk vibrations by applying a squeeze ﬁlm damping. are used on vibrating structures.66 Modeling and Control of Vibration in Mechanical Systems Most systems possess damping to some extent. spindle motor and disk vibrations [23] affect the positioning accuracy and limit how close adjacent tracks can be placed together. they are subjected to shear or direct strains.2. If a system undergoes a forced vibration. The disk vibrations in the axial . A servo track writer (STW) [25] is used to pattern the recording media with servo information for the HDD servo system. such as laminated or sandwiched materials. kη αEη (4. Another arrangement is that a viscoelastic layer is sandwiched between the elastic layers. by the use of structural materials having high internal damping. We know that the response amplitude of a system at resonance ω = ωn under harmonic excitation F (t) = F0 eiωt is given by F0 F0 = . 4.. damped laminated disks were used to reduce the amplitude of rocking modes by sandwiching a viscoelastic layer in between two aluminum layers to increase disk damping. Hence viscoelastic materials having large loss factor are used to provide internal damping for vibration control. and thus restrict the TPI number that can be achievable. which is helpful in vibration control.5 An application example − Disk vibration reduction via stacked disks To support higher track per inch (TPI) density hard disk drive. The Guzik spin stand shown in Figure 4. Damping can be modiﬁed in the system to control its response. When viscoelastic materials are used for vibration control.1 is used to evaluate the effectiveness of dualdisk stack in position accuracy improvement. In [11].6) where η is the energy loss factor. and the stiffness k is proportional to the Young’s modulus. The presence of damping always limits the amplitude of vibration. The material with the largest loss factor will be subjected to the smallest stress. The Spin stand is commonlyused equipment for magnetic recording component testing [10]. k = αE with a constant α. That is to stack more than one normal recording media together for reading and writing on the disk surfaces. its response or amplitude of vibration near resonance tends to become large if there is no damping. An example is the use of viscoelastic materials. Another application example is presented as follows with detailed discussion and evaluation [14].e. Damping tapes. A simplest arrangement is that a layer of the viscoelastic material is attached to an elastic one. consisting of thin metal foil covered with a viscoelastic adhesive. In addition to the laminated disks and the squeezed air bearing plate methods. while the stress is proportional to the displacement. i. In both cases.
5inch in diameter and 0. Step 1: Stack two disks together and mount on the spin stand. The position error signals are also collected for both the single disk and the dualdisk stack cases.Classical Vibration Control 67 direction are also measured at the outer diameter (OD) region of the disk via LDV. To write servo information on the disk surface in between the two disks.1 Disk and spindle motor assembly of the spin stand. steps 1.8 mm in thickness. The experimental process is described as follows. Two such disks are stacked together and mounted on the spindle motor of the spin stand. FIGURE 4. 8400 and 10200 RPM respectively. following which. Step 3: Read back the written information and obtain the PES. and the disk is an aluminum disk of 3. The disk vibration in the axial direction of the single disk and the dualdisk stack are measured via LDV. the spindle motor is stopped and the written surface of the disk is ﬂipped over. the spindle motor is spinning at 7200. 20 traces of the position error signal are collected and denoted by P ESi . In the experiment. 2 and 3 are repeated. Step 2: Spin the motor and write servo information on the top disk surface. To evaluate the improvement of position accuracy.
In Figure 4.2 are reﬂected in NRRO. other dominant harmonics before the 11th harmonic are all reduced. computed from the collected PES data. and PES by 38%. The power spectra of the RRO and NRRO are calculated from the tested PES and shown in Figure 4. It is found that the frequencies of the disk vibration modes T1T9 are shifted and their amplitudes are apparently reduced. as indicated by T1T6. N = 20. It can be seen that the amplitudes of the vibration modes are all reduced with shifts in frequencies. Positioning accuracy improvement at 7200 RPM Tested disk vibration results of the dualdisk and the single disk are shown in Figure 4.8. except for the 1st harmonic. and the PES σ value is reduced by 21%. N ). The vibration modes are all shifted by 40 Hz to higher frequencies. while T1 appears very close to S4. Positioning accuracy improvement at 8400 RPM The disk or the dualdisk stack is rotating at 8400 RPM. Figure 4.9 shows the comparison of singledisk and dualdisk PES in time domain.7 and Figure 4. N RROi = P ESi − RRO.5 shows the time sequences of PES in both the cases and the amplitude reduction is seen apparently.4. The resultant σ value of NRRO is improved by 28%. Notice that S1 is reduced signiﬁcantly. which may be due to better sliderdisk interaction. · · · . are shown in Figure 4. Besides the harmonics with respect to the spindle speed.3 and Figure 4. NRRO and PES are obtained. and the disk vibrations in the axial direction are shown in Figure 4. The repeatable runout (RRO) and the nonrepeatable runout (NRRO) are given by N i=1 RRO = 1 N P ESi . . It turns out that the σ value of RRO is reduced by 23%. Figure 4. (4.3. The other peaks indicated by S1S4 can be identiﬁed to be caused by slider vibrations. we can observe that almost all the reduced disk vibration modes T1T9 in Figure 4.6.6 can be found and they are all suppressed in the case of the dualdisk stack. This further veriﬁes that S1 is caused by the diskslider interaction.68 Modeling and Control of Vibration in Mechanical Systems (i = 1.6. As seen in Figure 4. Figure 4. the disk vibration modes can be seen clearly as denoted by T1T9. The power spectra of the RRO and NRRO. This phenomenon can be seen in other testing results which will be shown later. and the amplitude of PES in the case of the dualdisk stack approach is much lower as compared to conventional single disk approach. The standard deviations (or σ values) of RRO. They are both lowered signiﬁcantly.8.7 shows that almost all the ﬁrst seven harmonics are decreased and as a result. but can not be traced in the tested disk vibrations in Figure 4. the σ value of RRO is reduced by 41%. This randomly happens and the writing process at a different time may cause such a shifting. B. the corresponding T1T6 in Figure 4.7) A.4. In Figure 4. NRRO by 18%. The corresponding S4 cannot be found in this case. It is noted that the peak S2 for the single disk case is shifted to S3 in the dualdisk case.2.
5 2−disk 1−disk −5 −5.5 Dsplacement amplitude(log10(m)) −6 −6.Classical Vibration Control 69 −4.5 −7 T1 −7. .5 T2 T3 T4 T5 T6 T7 T8 T9 −8 −8.2 Comparison of singledisk and dualdisk axial vibrations measured via LDV at 7200 RPM.5 −9 0 500 1000 1500 Frequency(Hz) 2000 2500 3000 FIGURE 4.
70 Modeling and Control of Vibration in Mechanical Systems 1−disk 2−disk 0.3 Comparison of singledisk and dualdisk RRO power spectrum at 7200 RPM (23% improvement of σ value).1 FFT(RRO)(Vrms) 0. .05 0 0 500 1000 1500 2000 Frequency(Hz) 2500 3000 3500 4000 FIGURE 4.
05 S4 0.07 FFT(NRRO)(Vrms) 0.01 T2 T6 T4 T9 T8 0 0 500 1000 1500 2000 Frequency(Hz) 2500 3000 3500 4000 FIGURE 4.1 S1 0.04 S3 0.Classical Vibration Control 71 0.06 0.02 T1 0.4 Comparison of singledisk and dualdisk NRRO power spectrum at 7200 RPM (18% reduction of σ value).03 S2 T3 T7 0.08 0. .09 2−disk 1−disk 0.
.5 Comparison of singledisk and dualdisk PES in time domain at 7200 RPM (21% reduction of σ value).72 Modeling and Control of Vibration in Mechanical Systems FIGURE 4.
5 −9 0 500 1000 1500 Frequency(Hz) 2000 2500 3000 FIGURE 4.5 T1 T2 T3 T4 T5 T6 −8 −8.Classical Vibration Control 73 −5 2−disk 1−disk −5. .6 Comparison of singledisk and dualdisk axial vibrations measured via LDV at 8400 RPM.5 −7 −7.5 −6 Dsplacement amplitude(log10(m)) −6.
.05 0 0 500 1000 1500 2000 Frequency(Hz) 2500 3000 3500 4000 FIGURE 4.74 Modeling and Control of Vibration in Mechanical Systems 1−disk 2−disk 0.1 FFT(RRO)(Vrms) 0.7 Comparison of singledisk and dualdisk RRO power spectrum at 8400 RPM (41% improvement of σ value).
05 T1 0.03 T3 0.08 0.1 S1 0.06 0.09 1−disk 2−disk 0. .04 0.01 0 0 500 1000 1500 2000 Frequency(Hz) 2500 3000 3500 4000 FIGURE 4.07 FFT(NRRO)(Vrms) 0.02 T2 T4 S2(S3) T6 T5 0.8 Comparison of singledisk and dualdisk NRRO power spectrum at 8400 RPM (28% reduction of σ value).Classical Vibration Control 75 0.
.9 Comparison of singledisk and dualdisk PES in time domain at 8400 RPM (38% reduction of σ value).76 Modeling and Control of Vibration in Mechanical Systems FIGURE 4.
S1 is reduced signiﬁcantly again and S2 and S3 are shifted. The improvements of . and it is observed that the amplitude of PES in the case of the dualdisk stack is decreased.10 shows the obtained disk axial vibration by LDV. Figure 4.12 reﬂects the corresponding disk vibration modes T1T6 in Figure 4. 4. the σ value of RRO is improved by 33%. Figure 4. Positioning accuracy improvement at 10200 RPM A higher rotational speed of 10200 RPM is performed on both the dualdisk stack and single disk cases. −4.5 2−disk 1−disk −5 −5.5 T5 T4 −9 0 500 1000 1500 Frequency(Hz) 2000 2500 3000 FIGURE 4. As a result.10 and tabulated in Table 4.6 and 4.5 Dsplacement amplitude(log10(m)) −6 −6. NRRO by 27%. The 3rd and 5th harmonics are signiﬁcantly reduced as seen in Figure 4. The modes T1T6 are quite obvious with the amplitude reduced and the frequencies shifted higher. The reduction of disk vibration amplitude is evaluated from Figures 4. while S4 cannot be seen at all in Figure 4.11.5 −7 T1 −7. Figure 4.Classical Vibration Control 77 C.5 T2 −8 T3 T6 −8.12.1 for different rotational speeds.11 and Figure 4.2.13 shows the comparison of PES in time domain.10 Comparison of singledisk and dualdisk axial vibrations measured via LDV at 10200 RPM.10. Figure 4. compared with that of the single disk approach. and PES by 32%.12 show the power spectrum of RRO and NRRO. As for the slider related vibrations.
33% improvement of σ value).05 0 0 500 1000 1500 2000 Frequency(Hz) 2500 3000 3500 4000 FIGURE 4. .15 FFT(RRO)(Vrms) 0.78 Modeling and Control of Vibration in Mechanical Systems 1−disk 2−disk 0.2 0.11 Comparison of singledisk and dualdisk RRO power spectrum at 10200 RPM (the 3rd and 5th harmonics reduced signiﬁcantly.1 0.
05 0.06 0.07 FFT(NRRO)(Vrms) 0.01 0 0 500 1000 1500 2000 Frequency(Hz) 2500 3000 3500 4000 FIGURE 4.12 Comparison of singledisk and dualdisk NRRO power spectrum at 10200 RPM (27% reduction of σ value).03 T1 0. .Classical Vibration Control 79 0.09 0.08 0.02 T2 T3 T4 S2 S3 T5 T6 0.1 S1 1−disk 2−disk 0.04 0.
13 Comparison of singledisk and dualdisk PES in time domain at 10200 RPM (32% reduction of σ value).80 Modeling and Control of Vibration in Mechanical Systems FIGURE 4. .
˜ 8t3 E d (4. In the dualdisk case.8) indicates that the disk vibration is inversely proportional to t3 . we have also observed the frequency shifting of disk vibration modes. Equation (4. 4.6 and 4. The second best result obtained is performed at 10200 RPM. RRO and NRRO and disk vibration amplitude with stacked disks compared with single disk Speed(RPM) 7200 8400 10200 RRO NRRO PES 23 18 21 41 28 38 33 27 32 Disk vibration amplitude 24 45 35 the σ values of RRO. n) mode. where the dualdisk stack approach leads to an improvement of 32% in positioning accuracy. ρ is the gas density. (4. 4.1. t3 Ed V ∝ (4. the ratio of the .2. This indicates that the stacked disk is a simple way of reducing the disk vibrations in spin stand tests. particularly when a single disk surface is accessed. It can be seen clearly from the table that the best improvement is obtained when operating at 8400 RPM. the disk stack thickness ˜ = 2t.2. and λmn is the dimensionless frequency parameter which is generally a function of the boundary conditions on the plate. ν is the Poisson’s ratio. ω is the angular speed. In Figures 4. the amplitude of disk vibration ρR4 ω2 ˜ V ∝ .8) where R is the disk outer radius. it is known that the natural frequency of disk vibration mode is deﬁned by [23] fmn = λ2 Et3 mn 2 12γ(1 − ν 2 ) 2πR 1/2 . NRRO and PES are also summarized in Table 4. Hence t assuming that the change of E is negligible. γ is the mass per unit area of the disk.1 81 % reduction of σ values of PES. we have observed a remarkably reduced amplitude of the disk vibration and its reﬂection in positioning accuracy improvement via the dualdisk stack. Remarkable improvement is achieved for every rotational speed. The experimental results in Figures 4.10 veriﬁed the amplitude reduction of the disk vibration. On the other hand.Classical Vibration Control TABLE 4. For all the cases studied.10) where fmn is the natural frequency of the (m.6. t is the disk thickness. and d is the disk substrate damping. and 4. Theoretically the maximum amplitude of disk vibration is given by [22] ρR4 ω2 .10. E is the Young’s modulus of the disk substrate.9) ˜ 1 ˜ which means that V < V when the damping d > 8 d.
The tuning is usually performed by an actuator. the other is to add moving masses so that their inertial effects counteract the effects of the excitation variation. the tuning system will modify the feature of the system. 4.6 and 4. such as stiffness. Assuming that the change of ν is also negligible. Distinct ﬂight conditions such as ﬂight level. Thus such a disk vibration reduction approach can be used as an alternative or complementary technique to the air shroud design [24] and the active control approaches to further improve positioning accuracy. used in selftuning systems involves analyzing the system’s vibration frequency and its timewise variation.11) . 4. Further research for details on how the stacked disks change these properties would be signiﬁcant. One method. as can be seen in the experimental results in Figures 4. The selftuning system is tuned by a designed control algorithm.2. for example. and considering that ˜ = 2t. When the initial characteristics of the system no longer meet the requirements of the system’s working conditions. In addition to the doubled thickness. airplane or helicopter decreases as it consumes fuel. the frequency of the disk vibration mode is given by [23] f = fmn ± nΩ. the mass of a system like a car. the system characteristics therefore need to adapt to the parameter variations of the excitation and the system itself. or landing produce different types of excitation. damping and position of a mass. maneuvers. and Poisson’s ratio. system dynamic parameters vary with time. Some selfadapting systems in use are brieﬂy introduced as follows [8]. Selftuning suspension Selftuning suspension systems are equipped suitably for slow variation of the system characteristics. Denoting Ω as the rotational speed of the disk. Simulations using the control algorithm are required to identify the setting parameters of the . To summarize.3 Selfadapting systems In certain applications. (4. the mechanical properties such as d and λmn [22][23] are certain to change correspondingly after stacking two disks. such as multidisk servo track writers [25]. After the system status is measured. For example.10.12) which implies some shifting of the disk vibration modes. the proposed process is indeed an approach to reduce the disk vibration on both disk surfaces for reading and writing. t ˜ √ λ2 Et3 ˜ fmn = 2 2 mn2 2πR 12γ(1 − ν 2 ) 1/2 (4.82 Modeling and Control of Vibration in Mechanical Systems inner to outer diameter. Two ways are possible: one is to change the stiffness of the structure.
or acceleration. Active methods should be used to decrease the vibrations. depending on the application. It is able to adapt to various operating conﬁgurations depending on the type of driving. and as the mass vibrates. It is comprised of a servomechanism with an actuator. which is basically a massspring system. compared to the excitation frequency. a sensor. the vibratory level is decreased. The sensor measures the motion of the mass in terms of displacement. The mobile ﬂapping mass is supported by three ﬂexible elements indexed at 120◦. and it slides along the rotor axis. In certain structures. Selfadapting resonator Selfadapting resonators are capable of adjusting to changes in the excitation frequency. Selfadjusting absorbers The mechanical resonator. is used to control certain vibration sources. The z position of the mass makes the variation of the inertia of the assembly and thus the antiresonance frequency varies accordingly.Classical Vibration Control 83 algorithm and to verify the reliability of the system. The required position of the mobile mass is determined by control algorithms through an actuator and the mobile mass is moved by the actuator. The actuator applies a force to the mass whose vibration is to be reduced. The micropro . It consists of a ﬂapping mass that vibrates in a plane perpendicular to the rotation axis of the rotor. or the desired comfort. The car suspension is an example of selftuning suspensions. the excitation frequency evolves with time gradually. It is placed where the vibrations are to be reduced. The principle is that the resonance frequency of the massspring system as a resonator is adjusted with the excitation frequency in operation. this variation usually takes place between 1 and 3 seconds. velocity. These changes take place over a long period.4 Active vibration control The selfadapting systems presented previously will not be sufﬁcient in the cases where the source characteristics vary too fast for the involved algorithms or the required level of performance is too high. 4. The stiffness of the three elements is designed so that the resulting antiresonance corresponds to the excitation frequency. For a 20 Hz excitation frequency. One method to achieve this is to modify the suspension stiffness or its damping using various techniques. This kind of system actually works as the case of selfadjusting absorbers. The hub resonator used in a helicopter is an example. the driving conditions. and a microprocessorbased system. Subsequently. As a result. it is necessary to create a system whose resonance frequency would adjust automatically to the variations of the excitation frequency. the algorithm must be validated by experimental tests on the real structure. A vibration control system is called active if it uses external power to perform its function.
electric motors. and they generate an electric signal when they ﬂex. Active suspension Several active principles are used for suspensions to isolate one structure from another.2 Active systems An active system is applied to decrease vibrations by introducing dynamic loads locally in the structure [8]. Hydraulic or electrodynamic actuators suitably located on the structure are used to inject the loads. which are similar to piezoelectric materials. instead of dispersing the vibratory energy of the structure. The role of the actuators is to modify the distribution of vibratory energy for different modes to minimize the structure vibrations. One of them is internal load control to modify the distribution of internal loads in the structure. 4. The technology of generating loads is fundamental in control strategy. 4. These are ferromagnetic materials that expand or contract when subjected to an electric or a magnetic ﬁeld.4. . The loads depend on the vibratory condition of the structure. With a working hydraulic pressure of commonly 2000 psi. and vice versa. Its principle is to inject a set of dynamic loads into the structure in order to minimize its vibratory response. The previously mentioned voice coil motor (VCM) actuator in hard disk drives is a linear actuator moving in one direction. and as sensors to sense motion. called the control algorithm. It is driven when a force is produced to accelerate it radially as a current is passed through the coil. a cylinder containing a piston whose area is only 1 square inch will generate 1 ton of force. which is sensed with the help of a set of accelerometers or strain gauges. They change shape when an electric current passes through them. The control logic. The dynamic loads are controlled by a processor in order to minimize the vibratory level. programmed in the computer uses sensor measurement to decide how much force the actuator should apply. hydraulic cylinders are usually used. Active vehicle suspensions use hydraulic devices.1 Actuators Piezoelectric materials can convert electrical current into motion. and magnetorheological ﬂuid dampers. In building application. The materials used for highprecision actuation include electrostrictive and magnetostrictive materials. Because of its similarity to a loudspeaker.84 Modeling and Control of Vibration in Mechanical Systems cessor based system consists of analogtodigital converters to process sensor inputs and digitaltoanalog converters to convert the microprocessor’s output command into an input signal to the actuator. Thus they can be used as actuators to create force or motion. Some actuators cannot provide enough force for larger applications such as vehicle suspension and control of building motions.4. it is referred to as a voice coil motor. A VCM actuator has a coil of wire rigidly attached to the structure and suspended in a permanent magnetic ﬁeld.
A twostage electromagnetic resonator uses two masses called stages [8]. We take an electromagnetic actuator as an example.14) where k1 and c1 are respectively the stiffness and damping of the mass. It uses the dynamic ampliﬁcation of the mass to generate high loads using minimal energy. The control force FV is introduced between the two masses by an electromagnetic load. the force generated on the structure is ampliﬁed with respect to the force FV . ¨ ˙ (4. electromagnetic.17) ¨ ˙ ˙ ˙ with the primary stage M2 and the secondary stage M1 . ˙ (4. have been used to match the ﬁelds of application in terms of forces and frequencies. (4.Classical Vibration Control 85 Active resonator An active resonator works by displacing a mass with an actuator. The load transmitted to the structure is F T = c2 x 2 + k 2 x 2 . The function of control consists of tuning the system parameters such as mass and stiffness so that at the control frequency.18) . while maintaining a small relative displacement of the two stages. The natural frequency of the massspring system must be designed as close as possible to the frequency of the vibration that needs to be controlled. the Laplace transform of the displacement produced by the mass is 1 F0 . The motion equation of the mass M of a permanent magnet for small motion around its static position is given by M x1 + c1 x1 + k1 x1 = FV (t). The type of single stage resonator uses the principle of single stage mechanical systems with displacement operated by an electromagnetic force.13) With FV (t) = F0 cos(ωt). Different types of actuator technologies such as hydraulic. ˙ (4. The motion equations of the two masses are written as M1 x1 + c1 (x1 − x2 ) + k1 (x1 − x2 ) = FV (t). The role of the ampliﬁer is to ensure the law between the generated force FV and the control voltage V is linear.16) M2 x2 + c1 (x2 − x1 ) + c2 x2 + k1 (x2 − x1 ) + k2 x2 = −FV (t). The load transmitted to the structure equals to FT (t) = c1 x1 − k1 x1 − FV (t). ¨ ˙ ˙ (4. M s2 + c1 s + k1 X1 (s) = (4.15) The mechanical parameters such as mass and stiffness are designed so that the magnetic loads remain small and compatible with an acceptable energy consumption. or piezoelectric actuators. It is produced from a voltage V that is fed to a power ampliﬁer to generate a magnetic force via the current in the coil.
Feedback control provides a mechanism for tailoring system behavior to speciﬁc standards and needs.86 Modeling and Control of Vibration in Mechanical Systems 4. The block diagram depicting feedback control is in Figure 4. The ﬁrst objective implies that it is desirable for a control system to respond to a reference command rapidly. where the response is evaluated by a sensor and is fed back to an actuator that generates a force or motion.3 Control strategy When the system response is not acceptable. Standard performance measures are usually deﬁned in terms of the step response. In order to properly design a feedback control system. stability.14. The difference is an error that generates a control action.14 Generic feedback control system.4. The allowable error represents how close to the desired response the control force must bring the structure and how much the control system can reject vibrations. One can see that the signal from a block goes concurrently to other blocks or a summing point. The summing point indicates that the input and output signals are compared. The purpose of designing a system with feedback force is to minimize unwanted behaviors and eliminate the effect of vibration on system performance in a desired manner. The general design objectives are the speed of response. Stability is the prerequisite of any active control system design. the value of the response is used to generate additional forces according to rules or laws such that the modiﬁed response behaves according to the design and within certain bounds. and accuracy or allowable error. FIGURE 4. The disturbance or forcing is also shown being input directly to the plant or being injected to the output of the plant. This results in a closedloop system that incorporates feedback control. There are a number of control . The controller produces control action or signal by using the comparison signal of plant output with the desired reference value. performance must be deﬁned in terms of system speciﬁcations.
active vibration control was introduced with actuators. active systems. absorbers. There are thus fewer practical applications of adaptive controls. Subsequently.5 Conclusion Various classical techniques of vibration control were discussed in this chapter and some typical techniques such as using isolators.Classical Vibration Control 87 methods used by people to design control actions. In mathematics how the control gains or the control algorithm should adapt to changing conditions is difﬁcult. 4. and control strategies. An adaptive control system is one that can change or adapt either its gain values or even its control algorithm to accommodate changing conditions. and suspension were introduced individually. disk vibration reduction via stacked disks as an application example was presented with detailed discussion and evaluation. Classical control is represented by the wellknown proportionalintegralderivative (PID) control. Advanced control techniques include robust and optimal controls. which will be introduced brieﬂy in the next chapter. a resonator. . Speciﬁcally.
.
performance limitation of a feedback control system is discussed. will be investigated.5 Introduction to Optimal and Robust Control 5.2) Although G 2 can be computed from its deﬁnition. the socalled Youla parametrization. D).. 5. −∞ (5. What follow are the H2 and H∞ control problems and controller design via a linear matrix inequality approach. If 89 . H2 and H∞ norms and their calculation are ﬁrst introduced since they are two commonly used measures of performance speciﬁcations of a control system. there are some simple alternatives taking advantage of a statespace representation of G(s). and the parametrization of all stable controllers. which is necessary to help understand the vibration control schemes in the later chapters.1 [6] Consider a system G(s) with a statespace representation (A. −∞ (5.1 H2 and H∞ norms H2 norm The H2 norm of a matrix transfer function G(s) analytic in Re(s) > 0 (open righthalf plane) is deﬁned as G 2 := sup{ σ>0 1 2π +∞ T race[G∗ (σ + jω)G(σ + jω)]dω}.2 5.e.1 Introduction This chapter intends to brieﬂy review the optimal and robust control. B.1) or equivalently [2] G 2 = 1 2π +∞ T race[G∗ (jω)G(jω)]dω. LEMMA 5. i. Robust control of systems with multiplicative uncertainty and additive uncertainty. Moreover. C.2.
Tzw Tzw can equivalently be deﬁned as T race[ ∞ k=0 2 = g(k)gT (k)]. (5. w ∈ Rnw is the disturbance input.11) where Y2 and X2 are the reachability and observability Gramians that can be obtained from the following Lyapunov equations: AY2 AT − Y2 + BB T = 0. z(k) = Cx(k) + Dw(k).10) The H2 norm for the discretetime system Tzw can be computed as Tzw 2 = T race(DT D + B T Y2 B) = T race(DDT + CX2 C T ). (5. then we have G 2 2 = T race(B T Y2 B) = T race(CX2 C T ) (5. Let Tzw denote the transfer function from the input w to the output z.5) We also consider a discretetime linear timeinvariant system G(z) with the following statespace representation: x(k + 1) = Ax(k) + Bw(k). T T (5.3) where X2 and Y2 are the controllability and observability Gramians that can be obtained from the following Lyapunov equations: AX2 + X2 AT + BB T = 0. AT X2 A − X2 + C T C = 0. (5. That is Tzw 2 = E[zT z]. (5. Let the input w to the system be a wideband stationary stochastic process.6) (5.7) where x ∈ Rnx is the state. (5. A Y2 + Y2 A + C C = 0. (5.4) (5.12) (5.13) . Then the H2 norm is deﬁned as Tzw 2 = 1 T race[ 2π 2 π −π ∗ Tzw (ejω )Tzw (ejω )dω]. z ∈ Rnz is the controlled output. The H2 norm of Tzw can also be interpreted as the RMS value of the output z(k) when the system is subject to a white noise having zero mean and unit variance.9) where g(k) is the impulse response of Tzw .8) By Parseval’s theorem.90 Modeling and Control of Vibration in Mechanical Systems A is stable and D = 0.
17) . C.15) are linear in X2 and Y2 . C. 5.15) BT 0 I Observe that (5. (5.2 H∞ norm The H∞ norm of a matrix transfer function G(s) that is analytic and bounded in the open righthalf plane is deﬁned as [3] G := z sup σ [G(s)] = sup σ [G(jω)] = sup ¯ ¯ ω∈R w=0 w Re(s)>0 2 2 ∞ . Hence the H∞ norm of a transfer function can. and it also appears as the peak value on the magnitude plot of G(jω). in principle. LEMMA 5. The H2 norm of the system can be computed by minimizing µ using the function mincx in MATLAB Optimization Toolbox. be obtained graphically.2 Consider a discretetime transfer function Tzw of realization (A. B. Tzw 2 < µ if and only if there exist X2 = X2 and 2 T Y2 = Y2 such that T race(Π) < µ and Π CX2 D X2 C T X2 0 > 0. B.14) and (5. where H := A + BR−1 DT C −C T (I + DR−1 DT )C BR−1 B T .2. In general. D) with A stable. − (A + BR−1 DT C)T (5. T Given a scalar µ > 0.Introduction to Optimal and Robust Control 91 The following theorem presents an alternative LMI condition for bounding the H2 norm of the discretetime system Tzw . LEMMA 5. A control engineering interpretation of the inﬁnity norm of a scalar transfer function G(s) is the distance in the complex plane from the origin to the farthest point on the Nyquist plot of G. and hence can be solved by employing the LMI Tool [69] in MATLAB. D). the H∞ norm of a stable matrix transfer function can be read directly from its singular value plots.16) where w and z are respectively the input and output of G(s). Then G(s) ∞ < γ if and only if σ (D) < γ and the Hamiltonian matrix H has no eigenvalues on ¯ the imaginary axis. The H∞ norm can also be computed in statespace. (5.14) DT 0 I Y2 AY2 B Y2 AT Y2 0 > 0. (5.3 [6] Let γ > 0 and G(s) : (A.
B. C. Given a scalar λ > 0. Modeling and Control of Vibration in Mechanical Systems The following socalled Bounded Real Lemma gives a matrix inequality condition for the system G(s) to have a prespeciﬁed level of H∞ norm.4 Continuoustime Bounded Real Lemma Consider a continuoustime transfer function G(s) of realization G(s) = C(sI − A)−1 B + D. LEMMA 5. B.2π] 2 2 .20) C D −I 5. The continuoustime linear timeinvariant plant P (s) is described by the statespace equations: x(t) = Ax(t) + B1 w(t) + B2 u(t). for some given λ > 0.18) C D − λI In the discretetime case. the H∞ norm G ∞ < λ if and only if there exists X = X T > 0 such that T A X + XA XB CT BT X − λI DT < 0. Then. D).5 Discretetime Bounded Real Lemma Consider a discretetime and a stable transfer function matrix Tzw (z) with a statespace realization (A. Tzw ∞ < λ if and only if there exists X = X T > 0 such that T A XA − X AT XB CT B T XA B T XB − λ2 I DT < 0.21) . the H∞ norm of a stable transfer function matrix Tzw (z) with realization (A.1.3. (5. where A is a stable matrix. D) is deﬁned as Tzw ∞ = z sup σ [Tzw (ejω )] = sup ¯ w w=0 ω∈[0. (5. The following discretetime Bounded Real Lemma provides a linear matrix inequality condition for the system Tzw to have Tzw ∞ < γ. (5. C.1 H2 optimal control Continuoustime case We consider the closedloop system described by the block diagram in Figure 5. LEMMA 5.92 and R = γ 2 I − DT D. ˙ (5.3 5.19) where w and z are respectively the input and output of Tzw .
and A. Dc = 0. B= . The subsystem (A.26) (5. D = D12 Dc D21 + D11 . B2 . D12 is of full column rank. Cc = F. (5. C1 .24)−(5. B2 . Bc C2 Ac Bc D21 ¯ ¯ C = [C1 + D12 Dc C2 D12 Cc ] . real rational controller C(s) that stabilizes P internally and minimizes the H2 norm of the closedloop transfer function matrix Tzw from w to z in (5. (5. Bc = −K. The subsystem (A.25). Let X2 ≥ 0 and Y2 ≥ 0 be the solutions of the following Riccati equations: T T T AT X2 + X2 A − (X2 B2 + C1 D12 )(D12 D12 )−1 (X2 B2 + C1 D12 )T T +C1 C1 = 0.1 (1). (3). Introduce the following dynamic output feedback controller C(s): xc(t) = Ac xc(t) + Bc y(t).23) and (5. D21 ) has no invariant zeros on the imaginary axis.32) . u ∈ Rnu is the control input.25) Denote ξ = [xT xT ]T . D12 .23) where x ∈ Rnx is the state. C1 . 93 (5.Introduction to Optimal and Robust Control z(t) = C1 x(t) + D11 w(t) + D12 u(t). Assume that the system (5. D21 is of full row rank. (5. B1 . We assume D22 = 0 without loss of generality [89]. w ∈ Rnw is the disturbance input. B1 . the closedloop c system is given by ˙ ξ(t) = z(t) = where A + B2 Dc C2 B2 Cc B2 Dc D21 + B1 ¯ ¯ A= .26)−(5.24) (5. C2 . C2 .28) (5.30) T T T T T Y2 AT + AY2 − (Y2 C2 + B1 D21 )(D21 D21 )−1 (Y2 C2 + B1 D21 )T T +B1 B1 = 0. (5.31) According to the H2 optimal control theory. z ∈ Rnz is the controlled output. an H2 optimal controller can be obtained as [7] Ac = A + B2 F + KC2 .22) (5. y ∈ Rny is the measurement output. ˙ u(t) = Ccxc (t) + Dc y(t).27). and D21 are of appropriate dimensions.23) satisﬁes the following conditions: Assumption 5.21)−(5. (5. y(t) = C2 x(t) + D21 w(t) + D22 u(t). (2). ¯ ¯ Cξ(t) + Dw(t).29) ¯ ¯ Aξ(t) + Bw(t). (4).27) The continuoustime H2 control problem is to ﬁnd a proper. D12 ) has no invariant zeros on the imaginary axis.21)−(5. D11 . From (5. (5.
w ∈ Rnw is the disturbance input. T T T K = −(Y2 C2 + B1 D21 )(D21 D21 )−1 .(5. C1 . D22 = 0 is also assumed. D11 . y ∈ Rny is the measurement output.1 Conﬁguration of standard optimal control. where T T T F = −(D12 D12 )−1 (D12 C1 + B2 X2 ).37) where x ∈ Rnx is the state. (5. y(k) = C2 x(k) + D21 w(k) + D22 u(k). (5.3. the socalled perturbation method is applied [7] so that the above design method to ﬁnd an appropriate controller is still applicable. D12 . C2 .34) If the conditions (1)−(4) in Assumption 5.2 Discretetime case Consider the discretetime linear timeinvariant system P (z) with the following statespace representation: x(k + 1) = Ax(k) + B1 w(k) + B2 u(k). z(k) = C1 x(k) + D11 w(k) + D12 u(k).35) (5. z ∈ Rnz is the controlled output.36) (5. and D21 are of appropriate dimensions. and A. 5. u ∈ Rnu is the control input. B2 . B1 . .33) The minimal H2 norm of the transfer function Tzw is given by Tzw 2 = T T T race(B1 X2 B1 ) + T race[(AT X2 + X2 A + C1 C1 )Y2 ].1 are not satisﬁed.94 Modeling and Control of Vibration in Mechanical Systems FIGURE 5.
And the minimal H2 norm of the transfer function Tzw is given by Tzw 2 = T T T race(B1 X2 B1 ) + T race[(AT X2 A + C1 C1 )Y2 ]. (5.47) . From (5. real rational controller C(z) that stabilizes P (z) internally and minimizes the H2 norm of the transfer function matrix Tzw (z) from w to z of the closedloop system (5.14)−(5.30)−(5.41). D = D12 Dc D21 + D11 . T T T T AT X2 A − (AT X2 B2 + C1 D12 )(D12 D12 + B2 X2 B2 )−1 (AT X2 B2 + C1 D12 )T T +C1 C1 = 0. u(k) = Ccxc (k) + Dc y(k). (5. There exists a controller (5.44) T T T T T T AY2 AT − (AY2 C2 + B1 D21 )(D21 D21 + C2 Y2 C2 )−1 (AY2 C2 + B1 D21 )T T +B1 B1 = 0.40)−(5.35)−(5. where A + B2 Dc C2 B2 Cc B2 Dc D21 + B1 ¯ ¯ A= . THEOREM 5.15) for synthesis. (5. Bc C2 Ac Bc D21 ¯ ¯ C = [C1 + D12 Dc C2 D12 Cc ] .38)−(5. the closedloop c system is given by ¯ ¯ ξ(k + 1) = Aξ(k) + Bw(k).41) The discretetime H2 control problem is to ﬁnd a proper.39) Denote ξ = [xT xT ]T .39) such that Tzw 2 < µ if and only if the following linear matrix inequalities 2 and equality admit a solution: T race(Π) < µ.39). B= .46) A parametrization of all H2 controllers is developed in terms of LMIs as in the following theorem which linearizes the H2 norm conditions (5.1 [90] Consider system (5.37).31) for discretetime systems is as follows. 95 (5. The counterpart of the Riccati equations (5.43) (5. (5.42) (5.Introduction to Optimal and Robust Control Introduce the following dynamic output feedback controller C(z): xc (k + 1) = Ac xc(k) + Bc y(k). ¯ ¯ z(k) = Cξ(k) + Dw(k).45) A discretetime H2 optimal controller can then be obtained as (5.40) (5.35)−(5. (5.32).38)−(5.37) and (5.38) (5.
96 Π ∗ ∗ P2 ∗ ∗ ∗ ∗
and
Modeling and Control of Vibration in Mechanical Systems C1 X + D12 E C1 + D12 Dc C2 X + X T − P2 I + Z T − J > 0, (5.48) ∗ Y +YT −H J AX + B2 E A + B2 Dc C2 B1 + B2 Dc D21 H U Y A + W C2 Y B1 + W D21 ′ ′ > 0, ∗ X + X − P2 I +Z −J 0 (5.49) T ∗ ∗ Y +Y −H 0 ∗ ∗ ∗ I D11 + D12 Dc D21 = 0, (5.50)
where ∗ denotes an entry that can be deduced from the symmetry of the matrix, the matrices X, E, Y , W , U , Dc , Z, J, and the symmetric matrices P2 , H and Π are the variables. A feasible H2 controller is obtained by choosing N1 and M1 nonsingular such that N1 M1 = Z − Y X and calculating
−1 Cc = (E − Dc C2 X)M1 , −1 Bc = N1 (W − Y B2 Dc ),
Dc = Dc ,
(5.51) (5.52)
−1 −1 Ac = N1 [U − Y (A + B2 Dc C2 )X − N1 Bc C2 X − Y B2 Cc M1 ]M1 . (5.53)
5.4
5.4.1
H∞ control
Continuoustime case
We consider the continuoustime system P described by (5.21)−(5.23), and the class of causal, linear, timeinvariant and ﬁnitedimensional controllers that internally stabilize P , or namely, all admissible controllers for P . Our aim is to ﬁnd an admissible controller C such that the closedloop system Tzw satisﬁes Tzw
∞
< γ.
(5.54)
Assumption 5.2 (1). The pair (A, B2 ) is stabilizable and the pair (A, C2 ) is detectable. T T (2). The matrices D12 and D21 satisfy D12 D12 = Inu and D21 D21 = Iny . (3). rank (4). rank A − jωI B1 = nx + ny , for all real ω. C2 D21 (5.56) A − jωI B2 = nx + nu , for all real ω. C1 D12 (5.55)
Introduction to Optimal and Robust Control
97
The assumption that (A, B2 , C2 ) is stabilizable and detectable is necessary and sufﬁcient for the existence of admissible controllers. The full rank assumptions (3) and (4) are necessary for the existence of stabilizing solutions to the Riccati equations that are used to obtain the solution to the H∞ control problem. ˆ ˆ ˜ ˜ Deﬁne the matrices A, B, A and C as
T T T ˆ ˆˆ A = A − B1 D21 C2 , B B T = B1 (I − D21 D21 )B1 , T T ˜ T T ˜ ˜ A = A − B2 D12 C1 , C C = C1 (I − D12 D12 )C1 .
(5.57) (5.58)
Suppose that Uy maps y to u and has a minimal realization (Au , Bu , Cu , Du ) satisfying det(I − D22 Du ) = 0 for a wellposed closed loop. Then a realization in terms of LFT is given by LF T (P, Uy ) = A + B2 Du M C2 B2 (I + Du M D22 )Cu B1 + B2 Du M D21 , (5.59) Bu M C2 Au + Bu M D22 Cu Bu M D21 C1 + D12 Du M C2 D12 (I + Du M D22 )Cu D11 + D12 Du M D21
with M = (I − D22 Du )−1 .
THEOREM 5.2 [71] Consider the system (5.21)−(5.23) satisfying Assumption 5.2. There exists an admissible C such that the closedloop system (5.26)−(5.27) satisﬁes (5.54) if and only if 1. There is a solution X∞ ≥ 0 to the algebraic Riccati equation
T T ˜ ˜ ˜ ˜ X∞ A + AT X∞ − X∞ (B2 B2 − γ −2 B1 B1 )X∞ + C T C = 0, T T ˜ such that A − (B2 B2 − γ −2 B1 B1 )X∞ is asymptotically stable. 2. There is a solution Y∞ ≥ 0 to the algebraic Riccati equation T T ˆˆ ˆ ˆ AY∞ + Y∞ AT − Y∞ (C2 C2 − γ −2 C1 C1 )Y∞ + B B T = 0,
(5.60)
(5.61)
T T ˆ such that A − Y∞ (C2 C2 − γ −2 C1 C1 ) is asymptotically stable. 2 3. ρ(X∞ Y∞ ) < γ . In the case when these conditions hold, C is an admissible controller satisfying (5.54) if and only if C is given by the LFT
C = LF T (Ca , Uy ),
Uy
∞
< γ,
(5.62)
where Uy is a stable transfer function. The generator Ca is given by Ak Bk1 Bk2 Ca = Ck1 0 I , Ck2 I 0
(5.63)
98 where
Modeling and Control of Vibration in Mechanical Systems
T Ak = A + γ −2 B1 B1 X∞ − B2 F∞ − Bk1 C2z , T T T Bk1 Bk2 = B1 D21 + Z∞ C2z B2 + γ −2 Z∞ F∞ ,
(5.64) (5.65) (5.66) (5.67) (5.68)
T T T C2z = C2 + γ −2 D21 B1 X∞ , F∞ = D12 C1 + B2 X∞ , −2 −1 −2 Z∞ = Y∞ (I − γ X∞ Y∞ ) = (I − γ Y∞ X∞ )−1 Y∞ .
Ck1 −F∞ = , Ck2 −C2z
A point given by Theorem 5.2 is that a solution to the H∞ generalized regulator problem exists if and only if there exist stabilizing, nonnegative deﬁnite solutions X∞ and Y∞ to the algebraic Riccati equations associated with the full information H∞ control problem and the H∞ estimation of C1 x such that the coupling condition ρ(X∞ Y∞ ) < γ 2 is satisﬁed. The optimal H∞ control problem is to ﬁnd an internally stabilizing controller C(s) such that Tzw ∞ of the closedloop system (5.26)−(5.27) is minimized. However, in practice it is often not necessary to design an optimal controller, and it is usually appropriate to obtain a controller that gives rise to an H∞ norm of the closedloop system less than a prescribed value. More speciﬁcally, a suboptimal H∞ control problem is that given γ > 0, ﬁnd an admissible controller C, if there is any, such that Tzw ∞ < γ. The following theorem gives a design method for a suboptimal H∞ output feedback controller. THEOREM 5.3 [90] Consider system (5.21)−(5.23). Given a scalar γ > 0, there exists an output feedback controller (5.24)−(5.25) such that Tzw 2 < γ if the following ∞ LMI admits a solution (E, W, U, Dc, X, Y ): AX + XAT + B2 E + (B2 E)T U T + A + B2 Dc C2 T ∗ A Y + Y A + W C2 + (W C2 )T , ∗ ∗ ∗ ∗ B1 + B2 W Dc D21 (C1 X + D12 E)T Y B1 + W D21 (C1 + D12 Dc C2 )T < 0, (5.69) −γI (D11 + D12 Dc D21 )T ∗ −γI X I I Y
> 0. (5.70)
In this case, a feasible H∞ controller is obtained from (5.51)−(5.53), where N1 M1 = I − Y X.
Introduction to Optimal and Robust Control
99
5.4.2
Discretetime case
Assume that the timeinvariant discretetime system (5.35)−(5.37) satisﬁes: Assumption 5.3 (1). (A, B2 , C2 ) is stabilizable and detectable. T T (2). D12 D12 > 0 and D21 D21 > 0. (3). rank (4). rank A − ejθ I B1 = nx + ny , for all θ ∈ (−π, π]. C2 D21 (5.72) A − ejθ I B2 = nx + nu , for all θ ∈ (−π, π]. C1 D12 (5.71)
We seek a causal, linear, timeinvariant and ﬁnitedimensional controller C(z) such that the closedloop system (5.40)−(5.41) is stable and Tzw
∞
< γ,
(5.73)
or equivalently, under zero initial conditions, z
2 2
− γ2 w
2 2
≤ −ε w 2 , 2
(5.74)
for all w ∈ ℓ2 [0, ∞) and some ε > 0. Let B = B1 B2 , Cd = Dd = Js = Jt = D11 D12 , Inw 0 Inz 0 , 0 −γ 2 Inw C1 , 0 (5.75) (5.76) (5.77) (5.78)
Inw 0 . 0 −γ 2 Inu
With the assumptions (1)−(4), a causal, linear, ﬁnitedimensional stabilizing controller that leads to Tzw ∞ < γ exists if and only if the following two conditions hold [71]. 1. There exists a solution to the Riccati equation
T X∞ = Cd Js Cd + AT X∞ A − W T R−1 W,
(5.79)
with R= W = R1 RT 2 R2 R3
T = Dd Js Dd + B T X∞ B,
(5.80) (5.81)
W11 T = Dd Js Cd + B T X∞ A, W21 ∈ Rnu ×nx , W21
100
Modeling and Control of Vibration in Mechanical Systems
such that A − BR−1 W is asymptotically stable and X∞ ≥ 0, Denote ∇ = R1 − RT R−1 R2 , W∇ = W11 − RT R−1 W21 , 2 3 2 3 and let E1 be an nu × nu matrix such that
T E1 E1 = R 3 ,
(5.82) < 0, R1 ∈ R
nw ×nw
R1 − RT R−1 R2 2 3
, R3 ∈ R
nu ×nu
.
(5.83)
(5.84)
(5.85)
and E2 be an nw × nw matrix such that
T E2 E2 = −γ −2 (R1 − RT R−1 R2 ) = −γ −2 ∇. 2 3
(5.86)
Deﬁne the system At Bt Ct Dt −1 A − B1 ∇−1 W∇ B1 E2 0 −1 = E1 R−1 (W − R2 ∇−1 W∇ ) E1 R−1 R2 E2 I . 3 3 −1 −1 C2 − D21 ∇ W∇ D21 E2 0
(5.87)
2. There exists a solution to the Riccati equation
−1 T Y∞ = Bt Jt Bt + At Y∞ AT − Mt St MtT , t
(5.88)
where
T T St = Dt Jt Dt + Ct Y∞ Ct =
S1 S2 , T S2 S3
(5.89) (5.90)
T T Mt = Bt Jt Dt + At Y∞ Ct = Mt1 Mt2 , −1 such that At − Mt St Ct is asymptotically stable and
Y∞ ≥ 0, −1 T S1 − S2 S3 S2 < 0. Note that Ct in (5.87) is partitioned as Ct = Ct1 E1 R−1 (W − R2 ∇−1 W∇ ) 3 = . Ct2 C2 − D21 ∇−1 W∇
(5.91) (5.92)
(5.93)
A controller that achieves the objective (5.73) is given by
−1 xc (k + 1) = At xc (k) + B2 u(k) + Mt2 S3 (y(k) − Ct2xc (k)), −1 E1 u(k) = −Ct1 xc(k)1 − S2 S3 (y(k) − Ct2 xc (k)).
(5.94) (5.95)
Introduction to Optimal and Robust Control
101
All controllers that achieve the objective (5.73) are generated by the LFT C = LF T (Ca , Uc), where Uc is a linear causal system such that Uc ∞ < γ , and the generator Ca is given by T I − B2 − Mt2 (X2 )−1 xc (k + 1) T 0 E1 S2 (X2 )−1 u( k) 0 0 X2 η(k) −1 T xc (k) At 0 (Mt1 − Mt2 S3 S2 )(γ 2 X1 )−1 y(k) , = −Ct1 0 X1 (5.96) φ(k) −Ct2 I 0 with
T X2 X2 = S3 , −1 T T X1 X1 = −γ −2 (S1 − S2 S3 S2 ).
(5.97) (5.98)
The following theorem gives one parametrization approach of all suboptimal discretetime H∞ output feedback controllers. THEOREM 5.4 [90] Consider system (5.35)−(5.37). Given a scalar γ > 0, there exists an output feedback controller (5.38)−(5.39) such that Tzw 2 < γ if the following ∞ LMI admits a solution: P∞ J AX + B2 E A + B2 Dc C2 B1 + B2 Dc D21 0 ∗ H U Y A + V C2 Y B1 + V D21 0 ′ T T T T ∗ ∗ X + X ′ − P∞ I+Z −J 0 X C1 + E D12 T T T T ∗ ∗ ∗ Y +YT −H 0 C1 + C2 Dc D12 T T T T ∗ ∗ ∗ ∗ I D11 + D21 Dc D12 ∗ ∗ ∗ ∗ ∗ γI > 0, (5.99) where the matrices X, E, Y , V , U , Dc , Z, J, and the symmetric matrices P∞ and H are the variables. A feasible H∞ controller is obtained from (5.51)(5.53).
5.5
Robust control
The H∞ norm is used to test robust stability of a nominally stable system under unstructured perturbations. The following socalled small gain theorem is the basis for robust stability analysis.
102
Modeling and Control of Vibration in Mechanical Systems
THEOREM 5.5 Small Gain Theorem Consider a proper and stable transfer function matrix T (s). Suppose that a stable ∆(s) is connected from the output of T (s) to the input of T (s) as shown in Figure 5.2. Then the closedloop system given in Figure 5.2 is internally stable if σ [∆(jω)]¯ [T (jω)] < 1, ∀ω ∈ R ¯ σ ∞. (5.100)
FIGURE 5.2 A closedloop system with uncertainty. The small gain condition is sufﬁcient to guarantee internal stability of the closedloop system even if ∆ is nonlinear and timevarying. The small gain theorem tells us that an H∞ norm bound on T implies closedloop stability in the presence of certain H∞ norm bounded system uncertainties. The H∞ norm bound implies a certain stability robustness. Note from (5.100) that the size of tolerable uncertainty varies inversely proportional to the H∞ norm bound of T , which means that the robustness increases as the H∞ norm bound decreases. In what follows, the small gain theorem will be used to test robust stability under model uncertainties. The modeling error ∆ is assumed to be stable and suitably scaled with weighting functions W1 and W2 , i.e., the uncertainty can be represented as W1 ∆W2 . Additive uncertainty We assume that the model uncertainty can be represented by an additive perturbation: Π = P + W1 ∆W2 , (5.101)
Introduction to Optimal and Robust Control which is shown in Figure 5.3.
103
FIGURE 5.3 A closedloop system with additive uncertainty for robust stability analysis.
THEOREM 5.6 [70] Let Π = P + W1 ∆W2 , and C be a stabilizing controller for the nominal plant P . Then the closedloop system is wellposed and internally stable for all ∆ ∞ < 1 if and only if W2 CSW1 ∞ ≤ 1, where S= 1 . 1 + PC (5.102)
Similarly, the closedloop system is stable for all stable ∆ with ∆ ∞ ≤ 1 if and only if W2 CSW1 ∞ < 1. Multiplicative uncertainty The system model is described by the following multiplicative perturbation: Π = (I + W1 ∆W2 )P (5.103)
where W1 , W2 and ∆ are stable. Consider the feedback system shown in Figure 5.4. THEOREM 5.7 [70] Let Π = (I + W1 ∆W2 )P , C be a stabilizing controller for the nominal plant P , and T = 1 − S. Then (i) the closedloop system is wellposed and internally stable for all stable ∆ with ∆ ∞ < 1 if and only if W2 T W1 ∞ ≤ 1. (ii) the closedloop system is wellposed and internally stable for all stable ∆ with ∆ ∞ ≤ 1 if W2 T W1 ∞ < 1. (iii) the robust stability of the closedloop system for all stable ∆ with ∆ ∞ ≤ 1 does not necessarily imply W2 T W1 ∞ < 1.
104
Modeling and Control of Vibration in Mechanical Systems
FIGURE 5.4 A closedloop system with multiplicative uncertainty for robust stability analysis. (iv) the closedloop system is wellposed and internally stable for all stable ∆ with ∆ ∞ ≤ 1 only if W2 T W1 ∞ ≤ 1. (v) In addition, assume that neither P nor C has poles on the imaginary axis. Then the closedloop system is wellposed and internally stable for all stable ∆ with ∆ ∞ ≤ 1 if and only if W2 T W1 ∞ < 1.
5.6
Controller parametrization
Consider the standard system block diagram in Figure 5.1 with Ap B1 B2 P (s) = C1 D11 D12 . C2 D21 D22
(5.104)
Suppose (Ap , B2 ) is stabilizable, (C2 , Ap ) is detectable, and D22 = 0. The problem discussed here is that given a plant P , parameterize all controllers C that internally stabilize P . The parametrization for all stabilizing controllers is known as Youla parametrization [68], as shown in Figure 5.5. The Youla parametrization starts with a nominal controller that is an estimatedstate feedback. The estimated state feedback controller is given by u = −K x, ˆ (5.105)
where the state feedback gain K is some appropriate matrix and x is an estimate of ˆ the component of x, governed by the observer equation ˙ x = Ap x + B2 u + L(y − C2 x), ˆ ˆ ˆ (5.106)
Introduction to Optimal and Robust Control
105
where L is the estimator gain. The transfer function of the estimatedstate feedback controller is thus Kn (s) = −K(sI − Ap + B2 K + LC2 )−1 L. (5.107)
The nominal controller Kn (s) will stabilize P provided K and L are chosen such that Ap − B2 K and Ap − LC2 are stable. To augment the estimated state feedback controller, we inject v into u as shown in Figure 5.5, meaning that (5.105) is replaced by u = −K x + v, ˆ (5.108)
and therefore the signal v does not induce any observer error. For the signal e we take the output prediction error: e = y − C2 x. ˆ (5.109)
In Figure 5.5, the observer based controller applies output prediction error processed through a stable transfer function Q and added to the output of Kn . This augmentation is able to yield every controller that stabilizes the plant, which means every stabilizing controller can be realized as an observer based controller with some stable transfer function Q. Thus in the sequel we can form simple statespace equations for the parametrization of all controllers that stabilize the plant, and all closedloop transfer matrices achieved by controllers that stabilize the plant. The statespace equations of the augmented controller are given by ˙ x = (Ap − B2 K − LC2 )ˆ + Ly + B2 v, ˆ x (5.110) (5.111) (5.112)
u = −K x + v, ˆ e = y − C2 x. ˆ If Q has a statespace realization
xq = Aq xq + Bq e, ˙ v = Cq xq + Dq e,
(5.113) (5.114)
then a statespace realization of the observer based controller by eliminating e and v from the augmented controller equations (5.110)−(5.112) and the Q realization (5.113)−(5.114) is obtained as: ˙ x = (Ap − B2 K − LC2 − B2 Dq C2 )ˆ + B2 Cq xq + (L + B2 Dq )y, ˆ x xq = −Bq C2 x + Aq xq + Bq y, ˙ ˆ u = −(K + Dq C2 )ˆ + Cq xq + Dq y, x or equivalently, C(s) = Cc (sI − Ac )−1 Bc + Dc , (5.118) (5.115) (5.116) (5.117)
106 where
Modeling and Control of Vibration in Mechanical Systems
Ac = Bc =
Ap − B2 K − LC2 − B2 Dq C2 −Bq C2 L + B2 Dq , Bq
B2 Cq , Aq
(5.119) (5.120)
Cc = −K − Dq C2 Dc = Dq .
Cq ,
(5.121) (5.122)
On the other hand, the statespace equations for the closedloop system with only the augmented controller (5.110)−(5.112) are found as follows by eliminating u and y from (5.110)−(5.112) and the plant equations in (5.104). x = Ap x − B2 K x + B1 w + B2 v, ˙ ˆ ˙ = LC2 x + (Ap − B2 K − LC2 )ˆ + LD1 w + B2 v, x ˆ x z = C1 x − D12 K x + D11 w + D12 v, ˆ e = C2 x − C2 x + D21 w, ˆ (5.123) (5.124) (5.125) (5.126)
which are equivalently written as T11 (s) T12 (s) = CT (sI − AT )−1 BT + DT , T21 (s) 0 where Ap −B2 L , LC2 Ap − B2 K − LC2 B1 B2 , LD1 B2
(5.127)
AT = BT = CT = DT =
(5.128) (5.129) (5.130) (5.131)
C1 C2
D11 D12 . D21 0
− D12 K , − C2
It has been veriﬁed that by augmenting the stable transfer function Q, the closedloop transfer function from w to z is simply an afﬁne function of Q and equals T11 + T12 QT21 .
.5 Control system structure for Youla parametrization.Introduction to Optimal and Robust Control 107 FIGURE 5.
i. It is also important to understand how noise n will be ﬁltered through the closedloop system.136) becomes ∞ 0 K logSdω = 2π k=1 Re(pk ) > 0. If the plant P or compensator is not stable. hence T is also commonly called the complementary sensitivity function. because it explains how disturbance d2 goes through the closedloop system and shows up at the output y.135) (5.1 Performance limitation Bode integral constraint The block diagram in Figure 2.8 is that if the system is made less sensitive to disturbance at some frequencies. Denote Tyd2 the transfer function from d2 to y.. 1 + PC (5. (5. and given by S= Note that S + T ≡ 1.134) 1 .136) An implication of Theorem 5. note that S = Tyd2 = −Ted2 = Ten .20 shows a typical closedloop control system.20 .8 [67] Bode’s Integral Theorem for Continuoustime Systems For a stable. if P (s) and /or C(s) have a ﬁnite number of unstable poles pk . (5. THEOREM 5.6. then (5. In Figure 2. it will be more sensitive at other frequencies.133) The sensitivity function is thus important. ∞ 0 logSdω = 0. The closedloop transfer function from r to y is given by T = PC . 1 + PC (5.7 5.132) The sensitivity function is also known as the disturbance rejection function or error rejection function. The Bode plot of sensitivity function in continuoustime domain is shown in Figure 5.7. (5.e. It can be explained by the following Bode integral theorem.137) . rational P and C with P (s)C(s) having at least 2pole roll oﬀ. or the error signal e.108 Modeling and Control of Vibration in Mechanical Systems 5.
and φ = Ts ω with the sampling time Ts and the frequency ω in radians/sec.9 implies that if for some frequency S < 1.9 [108] Bode’s Integral Theorem for Discretetime Systems Given a stable closedloop discretetime feedback system. which is in general not practical for a physical system. and therefore in a ﬁnite frequency range.7 means that for a discretetime system. and the upper limit of the frequency spectrum is π/Ts . Unlike the continuous time result. if we want to attenuate disturbances at one frequency. K is the total number of unstable poles. Equation (5. Thus all S > 1 happen below the Nyquist frequency. Since the theorem is limited by frequencies up to the Nyquist frequency. In a word. better performance at low frequency may result in worse performance at high frequency.138) where βk are the openloop unstable poles of the system. Figure 5. if we have sufﬁcient knowledge of system disturbance and place the disturbance ampliﬁcation at places where the disturbance is negligible. with the linear feedback control whenever we improve the disturbance rejection at one frequency we pay for it at another. if the closedloop bandwidth is pushed up. its sensitivity function has to satisfy the following integral constraint: 1 π π 0 K logS(ejφ )dφ = k=1 lnβk .137) implies that any unstable poles in the system only make it worse. Further research beyond Nyquist frequency is needed to address vibrations at frequencies above the Nyquist frequency.Introduction to Optimal and Robust Control 109 where K is the number of unstable poles. Typical digital control systems assume P C is small and S ≈ 1 at or above the Nyquist frequency. Otherwise. most of the disturbances may be ampliﬁed. Note that Ts is the sampling period. THEOREM 5. the Nyquist frequency. In both cases. (5. there is no inﬁnite bandwidth to spread this over. the main difference is the Nyquist frequency limits the frequency range we have to work with. Nevertheless. then at some other frequency S > 1. then we succeed. in that more of the disturbance would have to be ampliﬁed. we must amplify some disturbance at another frequency. Theorem 5. which strives to overcome the limitation due to the integral theorem. .
6 Sensitivity function for continuoustime system. .7 Sensitivity function for discretetime system. Area of vibration amplification 0 Magnitude in dB (20log S) Area of vibration rejection Nyquist frequency 10 Frequency(Hz) FIGURE 5.110 Modeling and Control of Vibration in Mechanical Systems Area of vibration amplification 0 Magnitude in dB (20log S) Area of vibration rejection 10 Frequency(Hz) FIGURE 5.
An example is that if an A/D converter has a word length of 10 bits or more. and π + ∠L(jωc ) is the phase margin of the feedback system. Thus any measured continuous signal must be converted to a set of pulses by sampling. and converting to binary form is an analog to digital (A/D) converter. they are continuous in time and are not quantized in amplitude. The device that performs the sampling. Further 1 + L(jωc ) = 1 + L−1 (jωc ) = 2sin π + ∠L(jωc )  2 (5. The process is called quantization. The inﬁnite set of numbers represented by the smooth curve is replaced by a ﬁnite set of numbers. the resolution is 10/1024. the resulting sampled sequence will . which is an important characteristic related to the resolution of the converter. If the sampling frequency is not selected properly. The number of binary digits or bits generated by the device is its word length. Thus in order to produce a change in the output the input must change by at least 0. A nonminimum phase zero contributes an additional phase lag and imposes limitations upon the roll off rate of the openloop gain. quantization. If π + ∠L(jωc ) is forced to be very small by rapid gain attenuation.7. which is the process used to measure a continuoustime variable at separated instants of time. A typical design requires that the openloop transfer function has a high gain at low frequencies and a low gain at high frequencies while the transition should be well behaved [70]. where L(jωc ) = 1.3 Sampling Mathematical relations and operations can be handled by digital microprocessor only when they are expressed as a ﬁnite set of numbers rather than as functions having an inﬁnite number of possible values. In the classical feedback control theory. an input signal can be resolved to 1 in 210 or 1024.2 Relationship between system gain and phase In the classical feedback theory. In an analog device. Each pulse amplitude is then rounded off to one of a ﬁnite number of levels depending on the characteristics of the converter.01 V. signals are analog. The behavior of ∠L(jω) is particularly important near the crossover frequency ωc . A discretetime signal is extracted by sampling from a continuoustime signal. If the input signal has a range of 10 V.Introduction to Optimal and Robust Control 111 5. that is. the feedback system will amplify disturbances and exhibit little uncertainty tolerance at and near ωc . Thus the conﬂict between attenuation rate and loop quality near crossover is clearly evident. It is noted that ∠L(jω0 ) will be large if the gain L attenuates slowly near ω0 and small if it attenuates rapidly near ω0 .139) must not be too small for good stability robustness. The resolution measures the smallest change in the input signal that will produce a change in the output signal. or approximately 0. the Bode’s gainphase integral relation has been used as an important tool to express design constraints in feedback systems. Thus a digital device is one in which signals are quantized in both time and amplitude.7. Let L = P C denote the openloop system. 5.01 V. it has been common to express design goals in terms of the shape of the openloop transfer function.
140) where ω is the highest frequency contained in the signal. as seen in Figure 5. as well as performance limitation of linear feedback control systems. controller parametrization. as shown in Figure 5.7 implies that a certain rejection amount S < 1 must be accompanied by a certain area of S > 1. the ampliﬁcation S > 1 will essentially spread over a broader frequency band and the height of ampliﬁcation hump shrinks. Y (ω) = 0 ¯ for ω > ω. A proper sampling frequency is readily determined in many cases by means of the following sampling theorem. Figure 5.8 Conclusion Before presentating a series of advanced vibration control methodologies. It has reviewed H2 and H∞ performances. which helps understand problems to be addressed in the later chapters and possible solutions. robust control. which has to occur before the Nyquist frequency. ¯ If a system involves a sampling operation of continuoustime signals to generate discretetime signals. With a higher sampling rate and closedloop system bandwidth being kept constant. better performance in low frequency range may result in worse performance in high frequency range. 5.8. that is. which is half of the sampling rate. this chapter has been used to recall some standard advanced control techniques.10 Sampling Theorem A continuoustime signal y(t) can be reconstructed from its uniformly sampled values y(KTs ) if the sampling period Ts satisﬁes Ts ≤ π ω ¯ (5. H2 and H∞ controls. If we push the closedloop system bandwidth from f1 to f2 . . THEOREM 5. Any time delay added into a closedloop control system will decrease the stability of the system and in some cases may even cause system instability. a time delay may be induced. The Nyquist rate shown in Figure 5.8.7 signiﬁes that the freedom to spread the ampliﬁcation area around is limited by the Nyquist frequency.112 Modeling and Control of Vibration in Mechanical Systems not accurately represent the original continuous signal.
8 Sensitivity function in discretetime domain. .Introduction to Optimal and Robust Control 113 2f f 1 N1 0 f Magnitude in dB (20log S) 2 f N1 10 Frequency(Hz) FIGURE 5.
.
an H∞ constraint is to be satisﬁed when minimizing the H2 performance of the nominal system. z2 (k) = Cz2 x(k) + Dz21 w(k) + Dz22 u(k). (6.1) (6. 6. Both the simulation and experiment results demonstrate that the mixed H2 /H∞ control design gives a signiﬁcant performance improvement of TMR over the conventional method of PID control combined with notch ﬁlters. zi (k) ∈ 115 . z1 (k) = Cz1 x(k) + Dz11 w(k) + Dz12 u(k).1 Introduction The mixed H2 /H∞ control problem is concerned with the design of a controller that minimizes the H2 norm of a certain closedloop transfer function while satisfying a H∞ norm constraint on the same or another closedloop transfer function. we employ two methods: the one described in Chapter 5 and an improved method. The two methods are applied to the control design in disk drives.6 Mixed H2/H∞ Control Design for Vibration Rejection 6. One of the important applications of this problem is to address the optimal nominal performance subject to a robust stability constraint. We consider the statespace representation for linear timeinvariant systems: x(k + 1) = Ax(k) + B1 w(k) + B2 u(k).2 Mixed H2 /H∞ control problem This section aims to derive an improved design method for the mixed H2 /H∞ control which will give rise to an equal or better performance than the method in Chapter 5. a mixed H2 /H∞ control can be formulated for disk drive control. In this chapter. In order to meet the robustness requirement against unmodeled high frequency dynamics of the VCM actuator in disk drives.2) (6. y(k) ∈ Rny is the measurement output.3) (6. y(k) = C2 x(k) + D21 w(k). Hence.4) where x(k) ∈ Rnx is the state.
T race(Π) < µ.116 Modeling and Control of Vibration in Mechanical Systems Rnz (i = 1. First. Cz2 . Given scalars µ > 0 and γ > 0. 6. B1 . and Tz2 w the transfer function matrix from w to z2 . Π Cz1 X + Dz12 E ∗ X + X T − P2 ∗ ∗ (6. B= Bc C2 Ac ¯ C1 = [Cz1 + Dz12 Dc C2 Dz12 Cc ] . C2 . design a controller of the form (6. denote ξ = [xT xT ]T . Cc . It follows from (6. (6. B2 .4). Let a controller of the same dimension as that of the system (6. B2 Dc D21 + B1 .5)−(6.5)−(6.6) where the matrices (Ac . Dz11 . and A. (6. 2) are the controlled outputs.1)−(6.9) Denote Tz1 w the transfer function matrix from w to z1 .7) (6. u(k) = Ccxc (k) + Dc y(k). Bc D21 ¯ D1 = Dz12 Dc D21 + Dz11 . Dz12 . a controller of the form (6.13) Cz1 + Dz12 Dc C2 > 0.5)−(6. the following solution to the mixed H2 /H∞ control follows directly. THEOREM 6. Tz1 w 2 is minimized subject to the constraint Tz2 w ∞ < γ.14) .4) and (6. (6. D21 . Dc ) are to be determined. Cz1 . ¯ D2 = Dz22 Dc D21 + Dz21 .6) that c ξ(k + 1) = z1 (k) = z2 (k) = where A + B2 Dc C2 B2 Cc ¯ ¯ A= .10) (6.1)−(6.6) such that the H2 norm.1)−(6.3 Method 1: slack variable approach Given the solutions of H2 control and H∞ control in Chapter 5.8) (6.11) (6. w(k) ∈ Rnw is the disturbance input. ¯ ¯ C1 ξ(k) + D1 w(k). The mixed H2 /H∞ control problem is then stated as: Given a positive scalar γ. I + ZT − J T Y +Y −H (6. ¯ ¯ C2 ξ(k) + D2 w(k). Bc .4) be of the form: xc (k + 1) = Ac xc(k) + Bc y(k). ¯ C2 = [Cz2 + Dz22 Dc C2 H22Cc ] .6) that solves the mixed H2 /H∞ control problem exists if the following conditions are satisﬁed.5) (6. Dz21 .1 Consider the system (6. Dz22 are of appropriate dimensions.12) ¯ ¯ Aξ(k) + Bw(k). u(k) ∈ Rnu is the control input.
6. a single Lyapunov matrix is adopted for both the H2 and H∞ performances. A signiﬁcant improvement was made in [90] where a slack variable technique is introduced which separates the Lyapunov matrices from the controller parameters and hence allows them to be different for the H2 and H∞ performances.4 Method 2: an improved slack variable approach Recall from Chapter 5 that for a given controller that stabilizes the system (6. J. −1 Bc = N1 (W − Y B2 Dc ). the H2 norm square. H.15) 0 T ∗ ∗ ∗ Y +Y −H 0 ∗ ∗ ∗ ∗ I Dz11 + Dz12 Dc D21 = 0. J and H are used for H∞ performance. W .19) −1 −1 Ac = N1 (U − Y (A + B2 Dc C2 )X − N1 Bc C2 X − Y B2 CcM1 ]M1 . Z. which is very conservative in general. Y . (6. E. U .18) (6. P∞.17) ˜ where the matrices X. A feasible mixed H2 /H∞ controller is obtained by choosing N1 and M1 nonsingular such that N1 M1 = Z − Y X and calculating −1 Cc = (E − Dc C2 X)M1 . (6. can be computed by the following minimization [95]: 2 (Q=QT . (6. This is observed in the above theorem where the Lyapunov matrices P2 .16) and P∞ ∗ ∗ ∗ ∗ ∗ ˜ J AX + B2 E A + B2 Dc C2 B1 + B2 Dc D21 0 ˜ H U Y A + W C2 Y B1 + W D21 0 T T T T T T ˜ ∗ X + X − P∞ I + Z − J 0 X Cz2 + E Dz22 T T T T ˜ ∗ ∗ Y +YT −H 0 Cz2 + C2 Dc Dz22 T T T T ∗ ∗ ∗ I Dz21 + D21 Dc Dz22 ∗ ∗ ∗ ∗ γI > 0. (6. Tz1 w 2 . to solve the problem in terms of LMIs. (6. H and Π are the variables. J and H are used for H2 performance and different matrices ˜ ˜ P∞ . Dc .Mixed H2 /H∞ Control Design for Vibration Rejection 117 P2 J AX + B2 E A + B2 Dc C2 B1 + B2 Dc D21 ∗ H U Y A + W C2 Y B1 + W D21 ∗ ∗ X + X T − P2 I + Z T − J > 0. Dc = Dc .21) .9). J and the symmetric matrices ˜ P2 .7)−(6.20) In the early development of the mixed H2 /H∞ control. (6.Π=ΠT ) min T race(Π).
For example.1 The characterization of the H2 norm in the above lemma has the advantage that it provides a uniﬁed treatment of H2 and H∞ designs via an LMI approach. if there exist Q and Π satisfying (6. we shall assume nw = nz . Π.25). the same Q and Π also satisfy (6. (6. D x z w z z w ˆ Dz21 = [Dz21 0nz ×(nz −nw ) ]. some simple modiﬁcation can render the requirement satisﬁed.1 The minimization of the H2 norm square in (6. Without loss of generality. then (6. ˆ if nw < nz . (6. . D (6.8).23). Dz11 and Dz21 can be augmented as B1 = ˆ 21 = [D21 0n ×(n −n ) ]. LEMMA 6. On the other hand.25) can be rewritten as ¯ ¯ ¯T ¯ ¯ ¯ ¯T ¯ AT QA − Q + C1 C1 AT QB + C1 D1 − Σ ¯ ¯ ¯T ¯ ¯ T QA + D1 C1 − ΣT B T QB + D1 D1 − Π < 0. the additional parameter Σ in (6. Σ) to (6. as seen later.25) Proof First of all. respectively.24) (Q=QT . Furthermore.7)−(6. Dz11 = [Dz11 0n ×(n −n ) ] and ˆ [B1 0nx ×(nz −nw ) ]. the disturbance input and the signal to be controlled have the same dimension.118 subject to Modeling and Control of Vibration in Mechanical Systems ¯ ¯ ¯T ¯ AT QA − Q + C1 C1 < 0.21) is equivalent to the following minimization: min T race(Π).26) It is then clear that if there exists a solution (Q.22) and (6. the matrices B1 .25) is also met by the same Q and Π and ¯ ¯ ¯T ¯ Σ = AT QB + C1 D1 . i. D21 .22) and (6.23). REMARK 6. T T ¯ ¯ ¯ ¯ B QB + D1 D1 < Π.Π=ΠT . Note that.25) will oﬀer an additional freedom in optimization of performance when a mixed H2 and H∞ control design is concerned.23) The following lemma leads to an alternative approach for computing the H2 norm of the system (6. (6.Σ) subject to ˜ ˜ AT diag{Q.22) (6. if this is not the case.e. ¯ ¯T ¯ B (6. I}A − where ˜ A= ¯ A ¯1 C Q Σ < 0. ΣT Π ¯ B ¯1 .
F. 3 as compared to the result of Theorem 6.1 which help reduce the design conservatism in Theorem 6. it is easy to know that (6. However.Mixed H2 /H∞ Control Design for Vibration Rejection 119 LEMMA 6.23) and Lemma 6. 2.2 There exists a solution (Σ. by applying the Schur complement.28) into (6.21)−(6. Substituting (6. G). the latter will result in a less or equally conservative design due to the additional variables F and G. we specialize the matrices F and G as follows.7)−(6. Π) with Q = QT to (6. respectively. as will be seen later. it contains three additional variables λi . G= . F = λ1 Φ 0 Φ 0 .2 give equivalent computations of the H2 norm of the system. where Γ= (6.8) when a controller (6.1. REMARK 6.25) if and only if there exist matrices (Σ.2. there is no advantage of using Lemma 6.2 can be applied to compute the H2 norm of the system (6.27) is satisﬁed with F = 0 and GT = G = diag{Q. 2.1. T ˜ T −F + G A diag{Q. Π. F. I}.27) from the left and from the right by ΓT and Γ. Q.25) holds for some Q > 0.28) I ˜ . if (6. While this specialization of F and G generally introduces some conservatism.27) leads to ¯ ¯ ¯T ¯ −Q + λ1 AT Φ + λ1 ΦT A λ2 C1 + λ1 ΦT B − Σ λ C + λ B T Φ − ΣT ¯1 ¯ ¯ T + D1 ¯ − Π + λ2 D1 1 2 T ¯ T ¯ −λ1 Φ + Φ A Φ B ¯ ¯ λ3 C1 −λ2 I + λ3 D1 . Q. i = 1. it is well known that (6. On the other hand. Q.6) is given. 0 λ2 I 0 λ3 I (6. A where Φ ∈ R2n×2n and λi . 3 are real scaling parameters. A. For systems without uncertainty.21)−(6. I} − (G + G ) (6. Lemma 6. Hence. G) with Q = QT and Π = ΠT such that Q Σ T ˜T F + F T A ˜ ˜T G −F + A − ΣT Π + A < 0. While Lemma 6. multiplying (6. We observe that when F = 0 and Σ = 0.27) holds for some (Σ.25) follows.23) can be applied to derive the optimal H2 controller [95].2 It should be pointed out that (6. it may not be directly applicable to the H2 ˜ control design problem due to the presence of the products of F with A and G with ˜ To overcome this difﬁculty.2 reduces to the result in Theorem 6. if (6.5)−(6.27) Proof First. when additional performances such as the H∞ performance are to be met in addition to the H2 performance. i = 1.
25) with Π = γ 2 I and Σ = 0. Y. Q12 . E. there exist matrices P∞ = P∞ > 0 and Φ such that ¯T ¯ ¯ ¯ −P∞ + ε1 (AT Φ + ΦT A) ε2 C2 + ε1 ΦT B 2 T ¯2 + ε1 B T Φ ¯ ¯ 2 + D2 ¯ ε2 C −γ I + ε2 D T ¯ T ¯ Φ B −ε1 Φ + Φ A ¯ ¯ ε3 C2 −ε2 I + ε3 D2 ¯ ¯T −ε1 ΦT + AT Φ ε3 C2 ¯ ¯T BT Φ −ε2 I + ε3 D2 < 0. setting F = 0) and ε3 = 1 and by some rowcolumn exchanges.29) T Q − (Φ + Φ ) 0 0 (1 − 2λ3 )I (6.30) Observe that (6. γ 2 I} < 0. where ˜ A= ¯ ¯ A B ¯2 D2 . 2. it can be shown that the system (6. Thus.7) and (6. λi . following a similar procedure for deriving (6. Q11 = QT . 3 are additional variables which can be exploited to alleviate the conservatism in the mixed H2 /H∞ design of Theorem 6. i = 1.7) and (6. I}A − diag{P∞ . C ¯ Modeling and Control of Vibration in Mechanical Systems ¯ ¯T −λ1 ΦT + AT Φ λ3 C1 ¯ ¯T BT Φ − λ2 I + λ3 D1 < 0. i = 1. (6. Σ1 . U. it is stable with its H∞ norm less than γ if and only if there exists a matrix P = P T > 0 such that ˜ ˜ AT diag{P∞. Z. 11 22 ¯ Σ2 . 3) = . 3. i = 1. Indeed.31) REMARK 6.120 A similar characterization for the H∞ performance can be derived.30) turns out to be a special case of (6.1. W. Denote ¯ ¯ ¯ ¯ ¯ ¯ Ω (Π. 2. Q22 = QT .29). T P − (Φ + Φ ) 0 0 (1 − 2ε3 )I (6. Dc. recall that when the system (6.9) has an H∞ performance γ if and only if for some real scalars T εi . X. 2.e.3 When setting ε1 = ε2 = 0 (i. Therefore. εi .9) is known. the above inequality reduces to ¯ ¯ P ∞ − Φ − ΦT ΦT A ΦT B 0 ¯ ¯T AT Φ −P∞ 0 C2 2 ¯T Φ ¯2 < 0 B 0 −γ I DT ¯ ¯ 0 C2 D2 −I which is the result in Chapter 5.
5)−(6. 3)Σ1=Σ2 =0 < 0. Σ2 . 2. X. P12. Q11. Π. Dc. U.6) that solves the mixed H2 /H∞ control problem exists if for some λi . 3. U. ¯ ¯ ¯ ¯ ¯ Ω (γ 2 I. P22. i = 1.33) (6. λi . Σ1 . i = 1. P22. Y. 3) < 0. Σ1 . Σ2 . εi . ∗ ∗ ¯ Q22 − (X + X T ) ∗ 0 (1 − 2λ3 )I ∗ ∗ ∗ ¯ Q11 − (Y + Y T ) ¯ QT − (I + Z T ) 12 0 (6.Mixed H2 /H∞ Control Design for Vibration Rejection T ¯ −Q11 + λ1 (AY + Y T AT + E T B2 + B2 E) T T ¯ T + λ1 (AT + C2 DcT B2 ) + λ1 U T −Q12 T T T λ2 (Cz1 Y + Dz21 E) + λ1 (B1 + D21 Dc T B2 ) − ΣT ¯1 T −λ1 Y + AY + B2 E −λ1 I + U λ3 (Cz1 Y + Dz21 E) 121 ∗ T ¯ −Q22 + λ1 (AT X + X T A) + λ1 (C2 W T + W C2 ) T ¯2 λ2 (Cz1 + Dz21 DcC2 ) + λ1 B T X + λ1 D21 W T − ΣT −λ1 Z + A + B2 Dc C2 −λ1 X + X T A + W C2 λ3 (Cz1 + Dz21 Dc C2 ) ∗ ∗ T T T −Π + λ2 (D21 Dc T Dz21 + Dz21 Dc D21 + Dz11 + Dz11 ) B1 + B2 Dc D21 X T B1 + W D21 −λ2 I + λ3 (Dz11 + Dz21 Dc D21 ) ∗ ∗ ∗ ∗ ∗ ∗ . ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ there exists a solution (P11 . P11 . Σ1 . X.34) . i = 1. εi . W. 2. A controller of the form (6. Z. Dc.1)−(6. Z.2 Consider the system (6. E. W . P12 . Q22 . ¯ = [Qlj ] = QT > 0 and P = [Plj ] = P T > 0 to the following ¯ ¯ ¯ ¯ ¯ Ψ. 2. E. Q11 . X. Y. W. Q22. Q12 .32) We have the following solution to the mixed H2 /H∞ control. Dc ) with Q optimization: min T race(Π) subject to ¯ ¯ ¯ ¯ ¯ Ω (Π. E. THEOREM 6. ¯ ¯ (6. Y. Z.4). Q12. Σ2 . U.
it follows from (6.2 reduces to Theorem 6. Thus.35) (6.5)−(6.4 It is worth noting that when setting λ1 = λ2 = ε1 = ¯ ¯ ε2 = 0. Hence. Multiplying the above from the left by [Y −T −I] and from the right by [Y −T −I]T .29) by J1 and J1 respectively and ¯ = [Qlj ] = J T QJ and [ΣT ΣT ] = ΣT J. T −1 Cc = (E − Dc C2 Y )N1 . U = X T (A + B2 Dc C2 )Y + T T X T B2 Cc N1 + M1 Bc C2 Y + M1 Ac N1 . ¯ ¯ ¯ applying (6. Φ−1 = . Certainly. Hence. I.10). denote E = Dc C2 Y + Cc N1 . T where M1 .31). it is easy to see that Y +YT I +Z I + ZT X + XT ¯ > Q > 0.34). λ3 = ε3 = 1 and Σ1 = Σ2 = 0. (6. M1 U N1 E and J= Further.40) (6. Dc = Dc . N1 0 (6.6) can be obtained as in (6. Theorem 6. J.33) where Q 2 1 By applying a similar procedure to (6.38)−(6.38) (6. we obtain (6.41) T Multiplying from the left and the right of (6. T Y In .36) (6. T Y T X + N1 M1 . there exist invertible matrices M1 T and N1 such that Z − Y T X = N1 M1 . (6.41) that the controller parameters of (6.39) (6. a less or equally conservative mixed H2 /H∞ design can be achieved. the . by exploring the freedoms oﬀered by these parameters.29) that Φ is invertible since Φ + ΦT > Q > 0. Bc = −T M1 (W − X B2 Dc ).35)−(6. J1 = diag{J.33) and (6. a mixed H2 /H∞ controller is given by −T Ac = M1 (U − X T (A + B2 Dc C2 )Y − X T B2 Cc N1 T −1 −M1 Bc C2 Y )N1 .1. W = Z= T X B2 Dc + M1 Bc . N1 satisfy N1 M1 = Z − Y T X.37). we obtain that (X − Y −T Z) + (X − Y −T Z)T > 0.122 Modeling and Control of Vibration in Mechanical Systems In this situation. I}. If there exists a solution for the LMIs (6.34) can be obtained. It is then clear that Z − Y T X is invertible. observe from (6.37) Proof First. REMARK 6. Denote X M Y H Φ= .
FIGURE 6. and N are the disturbance and noise models.5 Observe that for given λi . Q22. U. Σ1 . REMARK 6. P22. Q12 . i = 1. Z. εi . Π. . εi .1 Application in servo loop design for hard disk drives Problem formulation A blockdiagram representation of a typical HDD servo loop is shown in Figure 6. Σ2 . 3. Dc). we demonstrate that Method 2 can give a much better performance than Method 1 with Theorem 6.34) ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ are linear in (P11. P12 . (6.33) and (6.1 Mixed H2 /H∞ control scheme for HDD servo loop with disturbance models. n denotes the PES demodulation and measurement noise. v represents all torque disturbances. D2 . and e is the measured position error or measurement output y in (6. 2. and w1 . X. The problem of searching for the optimal scaling parameters λi . E. i = 1. 2.1 with disturbances injected. respectively.5. 3 in general may be numerically costly although the function fminsearch in MATLAB Optimization Toolbox may be applied.5 6. Q11.2).1 together with an iterative reﬁnement in the latter design. D1 .1 is derived by a convex optimization and can be further reﬁned by an iterative procedure. 6. z1 is the true position error. d represents disturbances that are due to nonrepeatable disk and suspension/slider motions. In the disk drive application to be presented later. w2 and w3 are white noises of zero mean and unit variance. P (z) and C(z) represent transfer functions of the plant and controller.Mixed H2 /H∞ Control Design for Vibration Rejection 123 controller from Theorem 6. and hence can be solved by employing the LMI Tool [69]. W. Y.
the corresponding zdomain models of the VCM and the bounding function.172 × 1012 s2 + 1.3.45) where q is the number of true PES samples. i. To achieve a high capacity disk drive.. s2 + 3.032 × 109 s3 + 1. and P (jω) is the frequency response of the model in (6. will not cause any major problem. As mentioned. . however.42).2.3. which is given in terms of the standard deviation of the true PES.3 shows the multiplicative uncertainty of the VCM actuator deﬁned by Nmea Pi (jω) − P (jω) . the total amount of random ﬂuctuation about the desired track location. P (z) and Wu (z) can be obtained. the frequency responses of the actual VCM is obtained and is shown in Figure 6.44) From Figure 6.e. It is clear that their difference is more signiﬁcant for the frequency range of over 4 kHz.42)−(2. An approximate bounding function Wu (s).82 × 1017 s + 3. one way in servo control is to minimize TMR. In fact.. We observe that the actual uncertainties at some frequencies below 5 kHz exceed the bounding function. i. D2 (s) and N (s) are given by (2. 1 q −1 q 3σz1 = 3 z1 (i)2 . (6.016 × 104 s + 1.421 × 109 Figure 6. it is clear that the uncertainty at frequency over 5 kHz is the major concern.e.e. By discretization using the zeroorder hold. which. the smooth line in Figure 6. To capture the unmodeled dynamics in high frequencies. one of the most important performance measures for HDDs is the track misregistration or TMR.587 × 1015 s + 7. The disturbance and noise models D1 (s).42) s5 + 2. D2 (z) and N (z) are their discretetime forms.44).345 × 1018 where Nmea is the number of measurements and Pi (jω) is the actual frequency response of the plant in the ith measurement.2 shows the comparison between the frequency responses of the actual data and those of P (s). and D1 (z). i=1 (6. dozens of frequency response measurements are carried out and Figure 6.906 × 1013 s2 +8. A 5th order model is used to approximate the actual frequency responses of the VCM actuator and is given by P (s) = 5. is obtained as Wu (s) = 3s2 + 2.124 Modeling and Control of Vibration in Mechanical Systems Through experiments.117 × 104 s4 + 1.. i. (6. TMR is used to judge the required accuracy of positioning and thus to scale the disk capacity.433 × 108 .267 × 1021 .903 × 104 s + 1.43) ∆(ω) = max i=1 P (jω) (6. we verify that the robust stability of our designed system is guaranteed even with the worst case of uncertainty.
2 Frequency responses of the VCM actuator. .Mixed H2 /H∞ Control Design for Vibration Rejection 125 60 40 Magnitude(dB) 20 0 −20 −40 10 2 measured modeled 10 3 10 4 0 −200 Phase(deg) −400 −600 −800 −1000 −1200 2 10 3 4 10 Frequency(Hz) 10 FIGURE 6.
3 Multiplicative uncertainty of the VCM actuator. .126 Modeling and Control of Vibration in Mechanical Systems 20 VCM uncertainty bounding curve 10 0 −10 Magnitude(dB) −20 −30 −40 −50 −60 10 2 10 3 10 4 Frequency(Hz) FIGURE 6.
the constraint T Wu ∞ < 1 is to be met. Dz11 = 0 −Dd2 0 0 . 0 Au (6. and xp. Dp ).50) (6. (Ad2 . Bd1 . Bd2 .49) Bp 0 B2 = 0 . . (6. we have the mixed H2 /H∞ control scheme as shown in Figure 6. 0 0 Bn 0 0 Bu Dd2 0 0 C2 = −Cp 0 −Cd2 Cn 0 .1. xd2 . (Ad1 . and xu are respectively the state variables of the VCM actuator P (z).53) (6.52) (6. xn .1. a disturbance input or a reference.1)−(6. the input disturbance model D1 (z). i. ∞). Dd2 ). Bn . We now derive the statespace representation (6. ˜ ˜ ˜ When q is large enough. nz = ny = 1. Note that for the servo control shown in Figure 6..51) (6. the control design problem to minimize TMR can be treated as an H2 optimal control problem. Dz12 = 0. Cz2 = Du Cp 0 Du Cd2 0 Cu . where T is the closedloop transfer function and Wu is the bounding function of the unmodeled dynamics which was derived earlier. the output disturbance model D2 (z). Clearly.54) D21 = 0 −Dd2 Dn 1 . 0 0 (6. Dn ) and (Au . Ap Bp Cd1 0 0 0 Ad1 0 0 0 Ad2 0 A= 0 0 0 0 An Bu Cp 0 Bu Cd2 0 Bp Dd1 0 0 0 Bd1 0 0 0 0 B1 = Bd2 0 0 . On the other hand.e. where w4 ∈ ℓ2 [0.46) Thus. xd1 . D2 (z). and the uncertainty Wu (z). we need to ensure the system stability against the unmodeled high frequency dynamics of the VCM actuator. Dd1 ). N (z). (Ap . the measurement noise model N (z).Mixed H2 /H∞ Control Design for Vibration Rejection 127 Let w = [w1 w2 w3 ]T and Tz1 w denote the transfer function matrix from w to z1 . i=1 (6. the H2 norm of Tz1 w is given by [88] ˜ 1 q−1 q Tz1 w ˜ 2 ≈ z1 (i)2 .48) w = w1 w2 w3 w4 . Dz21 = 0 Du Dd2 0 0 . Cu .1 with x = xT xT1 xT2 xT xT p n u d d T T . Bu . Cz1 = −Cp 0 −Cd2 0 0 .4) for the system in Figure 6. Therefore. Cn . and Wu (z). 0 0 0 . Cd2 .47) (6. (An . the transfer function from w4 to z2 is T Wu . Bp . Du ) are respectively the statespace models of P (z). D1 (z). Cd1 . Cp. Dz22 = 0.
08nm with an unchanged H∞ norm of 0. except that its hump is slightly higher. λ2 = 0. Although the H∞ norm of T Wu of the improved design is slightly higher than that of the design using Method 1.1.2 Design results In this section.1 and the minimum H2 norm. the σ value of the true PES z1 . i.5 shows the testing result of sensitivity functions. and (5.79. the H2 norm is reduced to 10.128 Modeling and Control of Vibration in Mechanical Systems 6.9.1. the slight higher peak does not degrade the disturbance rejection performance much.3% more improvement on TMR than the design method in Theorem 6. i.78. Step 3: With the obtained P and X. stop. 0.. The comparison of control performances obtained by the improved method and that of Theorem 6.1 is negligible. .28. By applying Theorem 6.3.14)−(5. a slightly reduced H2 norm of 10. we will apply the mixed H2 /H∞ control to hard disk drive servo formulated previously. C = C. The sampling frequency being used is 20 kHz. no obvious improvement on the H2 performance has been observed. The iterative procedure gives a controller that produces a slight improvement of the H2 norm. B. hence.1 together with an iterative reﬁnement. B = B. with γ = 1. This means that with a larger γ.1 may be obtained by a γ value of greater than one as the actual H∞ norm is 0.4 shows the comparison of sensitivity functions. Starting from this controller. Step 2: solve LMIs (5. (5. For the purpose of comparison. Figure 6. Step 4: go to Step 1. ε3 = 1. Further. Bc .e.11 nm is obtained whereas the actual H∞ norm is 0. the disturbance mainly concentrates on frequencies below 1 kHz. ε2 = 0. solve LMIs (5. and minimize T race(Π). Bc . one may argue that a reduced H2 norm for the method of Theorem 6. The minimum H2 norm of 0. which is consistent with the simulation results in Figure 6.15). where the bandwidth from the improved method is much higher.78. we obtain λ1 = 0. Figure 6.01002 µm. Dc ). we carried out a further iterative procedure between controller variables and Lyapunov parameters: ¯ ¯ ¯ ¯ Step 1: obtain the closedloop system (A. lower than that of the improved method.4.5. it is below one as required.14)−(5.1.23. and D = D. C.00748 µm. the improved approach represents about 25.01013 µm is obtained. λ3 = 0. go to step 3.6 From Table 6. based on our simulations. from Figure 2.. ε1 = 0. we also design a mixed H2 /H∞ controller for the disk drive using the approach in Theorem 6.3. where we can see that the sensitivity function designed based on the improved design method is better than that from Theorem 6. Otherwise.31.1.20) for (Ac .e.2 and searching for the optimal scaling parameters. Dc ).1 is given as in Table 6. and let A = A. Cc .15). REMARK 6. With γ = 5. For example. D) with the controller parameters ¯ ¯ ¯ ¯ (Ac .5.20) for P and X. implying the designed controller makes the closedloop system robustly stable in the presence of uncertainty bounded by Wu . Cc . Hence. the improvement on the H2 norm for the method of Theorem 6. However. of 0. If T race(Π) does not differ from the previous value.
we shall calculate the H2 norm using the measured plant frequency response as in Figure 6.4 Frequency response of sensitivity functions. the sensitivity function by PID control has a lower bandwidth and almost the same peak value. This demonstrates that the parameter Σ in Lemma 6.2) blocks in (6.1) and (3.1 is useful in achieving a better performance. leading to a worse TMR.5.32) are set to zeros. ¯ ¯ The previous simulation is carried out with Σ1 and Σ2 as solutions to the LMI (6.32).3 and sensitivity function as in Figure 6.4 and Figure 6. which gives a higher TMR as listed in Table 6. As shown in Figure 6.1. it is found that the designed controller gives a 9% lower bandwidth for the closedloop system. Next. 10 Method 1 Method 2 PID 5 0 −5 Magnitude(dB) −10 −15 −20 −25 −30 10 3 10 4 Frequency(Hz) FIGURE 6.Mixed H2 /H∞ Control Design for Vibration Rejection 129 The improved method is also compared with conventional PID design. The spectrum of z1 is given by z1 (fk )2 = P (fk )S(fk )2 D1 (fk )2 + S(fk )2 D2 (fk )2 +1 − S(fk )2 N (fk )2 (6. When (3.55) .5.
130 Modeling and Control of Vibration in Mechanical Systems 10 5 0 −5 −10 Magnitude(dB) −15 −20 −25 −30 −35 −40 −45 Method 1 Method 2 PID 10 3 10 4 Frequency(Hz) FIGURE 6.5 Frequency response of sensitivity functions. .
The experimental result validates the advantage of the improved design. P (fk ) represents the measured frequency response of the plant. 2.. . K) are frequency points.78 0. and another one is a less or equally conservative design in terms of LMIs that contain more free variables than the conventional approaches.89 0.02714 with the controller designed by the method in Theorem 6.01155. Those variables offer additional freedoms in optimization.48 10.Mixed H2 /H∞ Control Design for Vibration Rejection TABLE 6. resulting in a less or equally conservative control design. The improved H2 /H∞ control design has been applied to hard disk drives to minimize the track misregistration while guaranteeing the system robustness in the presence of actuator uncertainties. Compared with the slack variable method. 3.3% in the H2 norm or the TMR performance while guaranteeing the system robustness by satisfying the required H∞ constraint. and S(fk ) is the frequency response of the sensitivity function.13 12.27 where fk (k = 1.4k 969 976 7. The resultant σ value of z1 is 0..1 131 Control performance comparison Method Openloop crossover frequency (Hz) H2 norm (σ (nm)) T Wu ∞ Method 2 Method 1 PID 1..11 0. the improved design result for hard disk drives has indicated a marked improvement of 25.1. 6. better than 0. One is a slack variable approach.6 Conclusion This chapter has presented two design methods for the mixed H2 and H∞ control.
.
7
LowHump Sensitivity Control Design for Hard Disk Drive Systems
7.1
Introduction
In feedback control systems, sensitivity functions are critical to the determination of their ability in disturbance and noise rejections. However, Bode has shown the limitation of using a feedback structure in terms of an integral constraint on the sensitivity function, as discussed in Chapter 5. Brieﬂy speaking, the Bode integral theorem implies that we cannot have a sensitivity function less than unity at all frequencies using output feedback with a ﬁnitebandwidth controller. Such a sensitivity function must amplify the disturbances existing in frequencies higher than the system bandwidth. In view of this, we shall employ the special structure of a secondary actuator system and design appropriate controllers for primary and secondary actuators such that the hump of the sensitivity function comes as low as possible without the cost of lowfrequency performance. This optimized sensitivity function is expected to minimize the ampliﬁcation of highfrequency disturbances while attenuating lowfrequency and midfrequency disturbances. With this lowhump sensitivity, the dualstage control system is able to reduce the contribution from all existing disturbances to the error. Two types of microactuator models are considered in this chapter: a MEMS actuator [102] and a PZT actuator [103]. The purpose is to design controls for the primary and the secondary actuators such that a low hump of the sensitivity function can be achieved with the help of secondary actuators. A comparison will be made to evaluate the effectiveness of the proposed method for the two microactuators. Besides simulations, an implementation with a PZT microactuator veriﬁes that the HDD servo loop design method leads to a lowhump sensitivity function.
7.2
Problem statement
Figure 7.1 shows a dualstage actuation system with one primary actuator Pv (s) and one secondary actuator Pm (s), and two parallel controllers Cv (s) and Cm (s). With disturbances and noise injected, the error is contributed by the disturbances and noise
133
134 in terms of
Modeling and Control of Vibration in Mechanical Systems
Se = Pv (fk )S(fk )2 d1 (fk )2 + S(fk )2 d2 (fk )2 + S(fk )2 n(fk )2 , (7.1) where Se is the power spectrum of the error e, S is the sensitivity function, and fk (k = 1, 2, · · · , N ) are frequency points. Figure 7.2 shows an example of main disturbances existing in disk drive systems, where the low frequency components are represented by d1 , the higher frequency portions such as disk vibration and windage are lumped as d2 , and the base line of the spectrum stands for the noise n.
FIGURE 7.1 Parallel structure of a dualstage actuation system with disturbances and noise injected.
Equation (7.1) implies that the sensitivity function S is important in determining the disturbance rejection of the dualstage closed loop control system. In a conventional singlestage system using one primary actuator it is difﬁcult to have a lowhump sensitivity function. Thus the dualstage structure is applied to increase the system bandwidth and lower the sensitivity function peak such that sup S(jω) ≤ 1 + τ, ω = 0, · · · , ∞ (7.2)
where 0 < τ << 1 is a sufﬁciently small tolerance. The sensitivity function satisfying (7.2) is called a nearly “ﬂat” sensitivity function. A dualstage actuation system uses a microactuator to increase the system bandwidth. Among many control schemes for the dualstage control loop, Figure 7.1 is one of the most popular ones [106]. The overall sensitivity function S(s) is S(s) = 1 . 1 + Pv (s)Cv (s) + Pm (s)Cm (s) (7.3)
Another popular control scheme is the decoupled structure as shown in Figure 7.3 [101], where the overall sensitivity function of the closedloop system is given by
LowHump Sensitivity Control Design for Hard Disk Drive Systems cascading each sensitivity function, i.e., S(s) = Sm (s)Sv (s), 1 Sv (s) = , 1 + Pv (s)Cv (s) 1 . Sm (s) = 1 + Pm (s)Cm (s)
135
(7.4)
In this chapter we are mainly concerned with the parallel structure as in Figure 7.1 since it is a basic structure for dualstage control systems and can be converted to the decoupled masterslave structure.
FIGURE 7.2 Power spectrum of PES nonrepeatable runout in open loop.
136
Modeling and Control of Vibration in Mechanical Systems
FIGURE 7.3 Decoupled structure of dualstage actuation systems.
FIGURE 7.4 Structure of H∞ loop shaping.
LowHump Sensitivity Control Design for Hard Disk Drive Systems
137
7.3
7.3.1
Design in continuoustime domain
H∞ loop shaping for lowhump sensitivity functions
The H∞ loop shaping method is used to design controllers for the primary actuator and the microactuator to achieve a lowhump sensitivity function for the dualstage actuator system. The structure of the H∞ loop shaping method is depicted in Figure 7.4, where W (s) is a weighting function of the desired sensitivity function. For a plant model P (s), a controller C(s) is to be designed such that the closedloop system is stable and Tzw
∞
<1
(7.5)
is satisﬁed, where Tzw is the transfer function from w to z, i.e., S(s)W (s). Clearly, (7.5) means that the sensitivity function S(s) can be shaped similarly to the inverse of the chosen weighing function W (s). A simple form of W (s) is W (s) =
1 2 s + 2ωζ √1 s + ω2 M √ M , s2 + 2ω εs + ω2 ε
(7.6)
where ω is valued by the desired bandwidth, ε is to determine the lowfrequency level of the desired sensitivity function, and ζ is the damping ratio. Associated with the weighting function, Figure 7.4 can be formulated as follows. x(t) = Ax(t) + B1 w(t) + B2 u(t), ˙ z(t) = C1 x(t) + D11 w(t) + D12 u(t), y(t) = C2 x(t) + D21 w(t) + D22 u(t), where A= Ap 0 0 Bp , B1 = , B2 = , Bw Bw Dp Bw Cp Aw (7.7) (7.8) (7.9)
C1 = Dw Cp Cw , D11 = Dw , D12 = Dw Dp , C2 = Cp 0 , D21 = 1, D22 = Dp , (Ap , Bp , Cp , Dp ) and (Aw , Bw , Cw , Dw ) are respectively the statespace realizations of plant P (s) and weighting function W (s). Let (Ac , Bc , Cc, Dc ) be the state space description of C(s). Then (Ac , Bc , Cc , Dc ) is to be designed such that (7.5) is satisﬁed. An LMI approach stated in Theorem 5.4 is used to design the controller. It is known that MATLAB functions, say “hinfsyn.m,” are available to design the controller. However, numerical errors will occur due to the large gain of VCM actuator and will be the hindrance for running the function. Thus, we would rather use the LMI approach in our application to the VCM actuator. There always exists a minimum level γ that makes the LMI (5.99) solvable, which gives a sensitivity
138
Modeling and Control of Vibration in Mechanical Systems
function closer to the inverse of its weighting function than that given by a larger γ. Certainly, the solvability is also related to the chosen weighting W (s), which must be realistic due to the Bode limitation. Bode’s integral theorem allows the possibility of a “ﬂat” sensitivity function up to a frequency of our concern in the continuous time domain, since the integral constraint in (5.136) is deﬁned on the frequency range from 0 to inﬁnity. Equation (7.2) can be achieved by choosing appropriate weighting functions for Sv and Sm separately. One way is to put the peak of Sv within the reduction band of Sm , as in Figure 7.5, and lower the highfrequency hump of Sm . This would need to decrease the bandwidth of the primary actuator loop and increase the bandwidth of the microactuator loop, which can be realized by adjusting the weighting functions Wv (s) for the primary actuator and Wm (s) for the microactuator. Here, the primary actuator loop and the microactuator loop are designed separately for the dualstage parallel structure. An additional reason for this is to meet the goals of ensuring the stability of the separate primary and microactuator loops and as in Figure 7.6, letting the primary actuator open loop have a higher gain at low frequencies and the microactuator open loop have a higher gain at high frequencies. The dualstage parallel control scheme could be formulated as an MIMO problem, however, to satisfy these speciﬁc requirements, the controller design would be more complicated.
.LowHump Sensitivity Control Design for Hard Disk Drive Systems 139 FIGURE 7.5 Frequency responses of Sv (s) and Sm (s).
6 Frequency responses of Pv (s)Cv (s) (solid line) and Pm (s)Cm (s)(dotted line).140 Modeling and Control of Vibration in Mechanical Systems FIGURE 7. .
5560].3.9570j.0421j. FIGURE 7.7862j. zeros and gain of its model are listed in (7.10) . ε = 10−3.1150 ± 38947.7 shows the frequency response of the VCM actuator and the poles. Figure 7.2745j.2122 ± 57759. With ω = 2π350. The dualstage HDD uses a microactuator as a ﬁne positioner to increase the positioning accuracy.2 Application examples This section will apply the proposed control design method to a dualstage actuation system in hard disk drives with the structure as in Figure 7.6). ζ = 0. The microactuator piggyback on a VCM actuator is driven jointly with the VCM actuator through suspension. poles = [−2312. slider or head [101].11).10)−(7.LowHump Sensitivity Control Design for Hard Disk Drive Systems 141 7. two different microactuator models will be studied and they are coupled separately with one VCM actuator. Due to various microactuator designs.8 using the H∞ loop shaping method described in the previous section. The sensitivity function Sv (s) is shown in Figure 7. −219.3538 ± 53340. −18849.4778 ± 31401.5.7 Frequency response of VCM actuator P (s). −2670. −942. −779.2 .9115 ± 588. VCM controller Cv is designed as in Figure 7.1 such that the designed controllers can produce a lowhump sensitivity function.5772j.4 in (7. (7.
Gain = 2. With ε = 10−0.883 × 107 (7. which are designed by different mechanism and possess different dynamics.9481j. ζ = 1 and ω = 2π2391 in (7.12) According to their physical behavior and the locations of the poles.142 Modeling and Control of Vibration in Mechanical Systems zeros = −2356. including the MEMS actuator and PZT actuator.9. the designed controller Cm (s) is shown in Figure 7. (7.11) FIGURE 7.1945 ± 47064.6).12) without delay.13) is a special case of (7.8 Frequency response of VCM controller Cv (s). .13) Equation (7.3. They can generally be characterized by a pade delay and a 2pole rolloff model: P adeDelay × O(s2 ) (7. The lowhump design method in Section 7. Various microactuators have been proposed for HDD dualstage servo systems. microactuators are categorized as two cases.7s + 8.6 .1 will be applied to the two cases to investigate their effect on the ultimate dualstage sensitivity function. such as a MEMS actuator [101] Pm (s) = 2 × 106 . s2 + 282.127723028788067 × 1040 . Case 1: Poles are located at low frequency.
FIGURE 7.12 even though the Bode integral limitation (5. and its open loop frequency response is shown in Figure 7. . A “ﬂat” sensitivity function is achieved as shown in Figure 7.11.9 Frequency response of microactuator controller Cm (s). The dualstage servo system has a gain margin of 17 dB and a phase margin of 92◦ .136) has to be fulﬁlled in this case.10 implies that the resultant sensitivity function Sm (s) follows its weighting function closely.LowHump Sensitivity Control Design for Hard Disk Drive Systems 143 Figure 7.
. FIGURE 7.144 Modeling and Control of Vibration in Mechanical Systems FIGURE 7.11 Open loop frequency response of the dualstage system.10 Sensitivity function Sm (s) (solid) and its weighting function inverse (dashed).
14). Note that a nonminimum phase zero is included in the microactuator model and a nearly “ﬂat” sensitivity function is still achieved.13.14.934 × 109 ) (s + 6.195 × 109 ) (7. The designed Sv (s) and Sm (s) are matched as in Figure 7.16. not hamper the practical usability of the resulting controllers due to the large gain and phase margins. We can observe that a “ﬂat” sensitivity function has been achieved.438572836 × 109 (s − 6.7s + 1. Case 2: Poles are located around 10 kHz.1 The control design is based on the relevant sensitivity function only and does not consider robustness to plant pole/zero variations. The frequency response of the open loop dualstage system is shown in Figure 7. REMARK 7. ε is kept unchanged to have the same lowfrequency level of the sensitivity function as in Case 1. which subsequently leads to the overall sensitivity as in Figure 7.7s + 1.157 × 105 )(s2 + 923. For the control design of (7. and ω is adjusted to be 2π2107 to have a similar openloop bandwidth. . where the gain margin is 31 dB and the phase margin is 88◦ . such as the active piezoelectric suspension modeled by Pm (s) = −1.12 Sensitivity and complementary sensitivity functions. This will. however.LowHump Sensitivity Control Design for Hard Disk Drive Systems 145 FIGURE 7.157 × 105 )(s2 + 791.517 × 109 )(s2 + 5089s + 7.14) with the frequency response in Figure 7. The resultant Cm (s) is shown in Figure 7.5.15.
1 Control performance comparison.13 Microactuator frequency response. from a manufacturing point of view. looks better than Case 1. the MEMS actuator. .1 for the two dualstage systems consisting of different microactuators. Case 2.146 Modeling and Control of Vibration in Mechanical Systems TABLE 7. the PZT actuated suspension. the MEMS actuator beneﬁts from its mass production and low cost. whereas its behavior is to be improved to possess a characteristic as simple as a padedelay which relies on its manufacturing process. Also. in the sense that its gain margin is much higher and phase margin is slightly lower. It can be seen that they both achieve a “ﬂat” sensitivity function with bandwidth of 2 kHz. Microactuator model type Case 1 Gain Margin (dB) 17 Phase Margin (deg) 90 Sensitivity function Flat Case 2 31 88 Flat Control performances are summarized in Table 7. FIGURE 7.
15 Open loop frequency response of the dualstage system. FIGURE 7.LowHump Sensitivity Control Design for Hard Disk Drive Systems 147 FIGURE 7. .14 Frequency response of microactuator controller Cm (s).
5326 0.9z 2 − 1.7 and the PZT microactuator in Figure 7. The controllers are implemented with dSpace 1103 on TMS320C240 DSP board.14 and sampling rate of 40 kHz. where the rough line is the tested result and the smooth line is the simulation result.8 and 7. The VCM and PZT controllers are obtained by discretizing the designed controllers with frequency responses shown in Figure 7. The displacement range used is 2 µm/V.16) The tested and simulation results of the dualstage open loop are shown in Figure .4081z + 0. i. The resultant sensitivity functions are shown in Figure 7.148 Modeling and Control of Vibration in Mechanical Systems FIGURE 7.2627 Cm (z) = 2 .e. An LDV is used to measure the position of the dualstage actuator as the feedback signal.13. We can observe that the hump of the sensitivity function is lower than 3 dB.209z + 0.532z + 0. The frequency response of points B over A is the sensitivity function. the dualstage system consisting of the VCM in Figure 7.18.3. z 2 − 1.8788 .16 Sensitivity and complementary sensitivity functions. A DSA is used to measure the frequency response of the sensitivity function by injecting a swept sine signal at point A in Figure 7.778z + 0. 7. z − 1.3 Implementation on a hard disk drive The implementation is carried out for Case 2 in the previous section.2631 Cv (z) = (7.15) (7.17. which is better than that with the following PIDlike controllers designed by the leadlag method: 0..
which is consistent with Figure 7.21 shows the 3σ value of position error versus frequencies.19.5 kHz. The Parseval’s formula stated in Chapter 1 addresses how to calculate σ value in the frequency domain. Moreover. Note that the hump of the sensitivity function S(z) is tested to be 3 dB.LowHump Sensitivity Control Design for Hard Disk Drive Systems 149 7.1). According to (7. Channel 2 is the control signal of VCM and Channel 3 is the control signal of PZT.20. instead of 0 dB as in Figure 7. We know that a pole at 20 kHz may not affect 2 kHz bandwidth that much and thus it is not needed to include a pole far beyond 20 kHz in the actuator modeling for 2 kHz servo bandwidth. The lowhump design outperforms the PID design for disturbance rejection with almost the same performance at frequencies lower than 1.1) with the sensitivity function in Figure 7.16. FIGURE 7. Figure 7. an actuator with highfrequency poles is helpful to the lowhump sensitivity function design.18. It is expected that the microactuator could be modeled as a padedelay within the frequency range of our interest.17 Experimental structure. the sensitivity function as in Figure 7.18 must amplify the corresponding highfrequency disturbances due to the hump above 0 dB. The position error is evaluated according to (7. The step response of the dualstage system is shown in Figure 7. This is mainly due to the locations of the microactuator poles which are around 10 kHz and not sufﬁciently far away from the bandwidth of 2 kHz. where Channel 1 is the output corresponding to the reference Channel 4. As such. In the next section we discuss the possibility of achieving lowhump sensitivity functions in the discrete time domain.2. and certainly the one with the lower hump will result in less ampliﬁcation.18 associated with the PES spectrum in Figure 7. .
18 Sensitivity function of the dualstage system (smooth line: simulation result. rough line: testing result. FIGURE 7. rough line: testing result.19 Open loop frequency response of the dualstage system (smooth line: simulation result. dotted line: PID design).150 Modeling and Control of Vibration in Mechanical Systems FIGURE 7.) .
.20 Step response of the dualstage system.21 3σ of PES NRRO versus frequencies.LowHump Sensitivity Control Design for Hard Disk Drive Systems 151 FIGURE 7. FIGURE 7.
the sensitivity function is bounded by 0 dB. −2 ≤ kg < 0 and kg = −1. . b).1 Synthesis method for lowhump sensitivity function We recall a simpliﬁed Bode’s theorem which is helpful to analyze the sensitivity function of the dualstage servo system.152 Modeling and Control of Vibration in Mechanical Systems 7. ω ∈ [0. c). When the orders of the denominator and numerator of G(z) are the same and kg < −2 or kg > 0. a). (7. then π −π ln S(ejω )dω ≥ 0. Thus. π]. Before the control design.4 Design in discretetime domain This part is concerned with servo loop control design in the discretetime domain that aims to achieve a lowhump sensitivity function for dualstage actuator systems. When G(z) is strictly proper.e. we have the following statements. i. THEOREM 7.4. then π −π ln S(ejω )dω = 2π(−lnkg + 1). then π −π ln S(ejω )dω < 0. 7. When the orders of the denominator and numerator of G(z) are the same. sensitivity functions for a discretetime dualstage system will be analyzed on the basis of the Bode integral theorem.17) Based on the discrete Bode’s theorem. The sensitivity function of the system is given by S(z) = 1/(1 + G(z)). If the closedloop system is stable and kg = limz→∞ G(z). we can conclude that only if G(z) is nonstrictly proper and kg < −2 or kg > 0. then S(ejω ) ≤ 1.. then limz→∞ G(z) = 0 and π −π ln S(ejω )dω = 0.1 Discrete Bode’s Theorem [104] Consider a discretetime SISO LTI system whose openloop transfer function G(z) does not have unstable poles.
2 An application example The VCM actuator and the microactuator are the same as those in Case 2 in Section 7. Bc .19) (7. C2 = Cp 0 .23) C1 = Dw Cp Cw .4. (7. In the discrete time case. y(k) = C2 x(k) + D21 w(k) + D22 u(k). Let (Ac . the discrete H∞ loop shaping method will be applied to the control designs of the VCM loop and the microactuator loop such that the sensitivity functions of the two loops are coupled and achieve a lowhump overall sensitivity. Dc ) be a state space description of C(z). 7. Cc . Notice that the PZT microactuator is represented using a Padedelay with two 2nd order resonance terms. G(z) = Gv (z) + Gm (z). Bp . The form (7.18) (7.LowHump Sensitivity Control Design for Hard Disk Drive Systems 153 The openloop transfer function of the dualstage parallel system is the sum of each path. Figure 7. Dc ). (Ap .21) (7.. When Gm (z) is strictly proper.2. z(k) = C1 x(k) + D11 w(k) + D12 u(k). k/s2 ) combined with some resonance modes and is often described by a strictly proper model.20) Note that a VCM actuator can be approximately represented by a double integrator (i. x(k + 1) = Ax(k) + B1 w(k) + B2 u(k). Gm (z) = Pm (z)Cm (z). D21 = 1.e. It is strictly proper and thus according to the analysis in . In what follows. These are necessary conditions and could be used as a criterion to examine the closedloop design.14) can be regarded as a general model of PZT actuated suspensions.3. B1 = . Cc .22) (7. i. An LMI approach stated in Chapter 5 will be used to design the controller C(z) : (Ac . Dw ) are respectively the statespace realizations of plant P (z) and weighting function W (z).. where A= Ap 0 0 Bp . Dp ) and (Aw . D22 = Dp . B2 = Bw Cp Aw Bw Bw Dp (7. Cp .4. Only a nonstrictly proper model of microactuator could possibly produce a “ﬂat” sensitivity function for dualstage servo systems. D11 = Dw . D12 = Dw Dp . Gv (z) = Pv (z)Cv (z). It can be concluded that the Bode’s integral of the overall dualstage loop is determined by the microactuator loop.4 is formulated as follows. Bw . G(z) is also strictly proper and when Gm (z) is nonstrictly proper. G(z) is nonstrictly proper. Bc . Cw . The structure of the H∞ loop shaping method is the same as in Figure 7. An additional condition for the “ﬂat” sensitivity is kg < −2 or kg > 0.e.
REMARK 7.2.2 If manufacturing processes could produce an ideal actuator that can be modeled by a Padedelay only. the controller Cm (z) for the microactuator is designed as in Figure 7.49 kHz.22 using the H∞ loop shaping method in the previous section.24.38.14).3. s + 1. can be obtained with the same weighting function as that for (7. as shown as in Figure 7.1.071/2 in (7. a VCM controller Cv (z) is designed as in Figure 7.4. ζ = 0. the phase margin of 64◦ and the bandwidth of 2. The sensitivity function with the hump lower than 2 dB.257 × 105 (7. The used sampling rate for the controller design is 40 kHz. it is possible to have a “ﬂat” sensitivity like that in Figure 7. The openloop system has the gain margin of 8. The sensitivity function of the dualstage system is shown in Figure 7. It can be described by a 1pole rolloff model Pm (s) = 1257 .25 with kg > 0.4 in (7. a “ﬂat” sensitivity function is impossible. ζ = 1.24) which is strictly proper and does not satisfy the necessary conditions in Section 7.6). which is nonstrictly proper.26.6).1. which resembles the ideal case. is the PZT microactuator in [107]. . With ω = 2π350.154 Modeling and Control of Vibration in Mechanical Systems Section 7. Also applying the LMI approach.23 with ω = 2π3700. An actual microactuator.4 dB. and M = 0. The discretetime actuator models are obtained using “zeroorderhold” method to ensure that the designed controllers are implementable. ε = 10−1. where we can observe that the hump is below 3 dB. ε = 10−3.
FIGURE 7. .23 Microactuator controller Cm (z).LowHump Sensitivity Control Design for Hard Disk Drive Systems 155 FIGURE 7.22 VCM controller Cv (z).
FIGURE 7.25 Sensitivity function.24 Sensitivity function of the dualstage system.156 Modeling and Control of Vibration in Mechanical Systems FIGURE 7. .
FIGURE 7. .27 Frequency responses of Pv (z)Cv (z) (solid curve) and Pm (z)Cm (z) (dashed curve).26 Sensitivity function of the dualstage system.LowHump Sensitivity Control Design for Hard Disk Drive Systems 157 FIGURE 7.
1) with the designed sensitivity functions. and less vibration ampliﬁcation in a high frequency range. The resultant sensitivity function is shown in Figure 7. The position error is evaluated from (7.28. The system is veriﬁed to maintain stability in spite of the variation of resonance frequency.158 Modeling and Control of Vibration in Mechanical Systems 7. where we can see that the performance varies slightly. a swept sine signal is injected at point A.30 shows that the system is stabilized and working in real time. however. which is better than that by a PID design as shown by the dotted line in Figure 7. The sensitivity function as in Figure 7. and thus is effective in achieving higher positioning accuracy. . e. The proposed control design is based on the sensitivity weighting function only and does not consider robustness to plant parameter variations. which is consistent with Figure 7.30 is the control signal of the VCM actuator and Channel 3 is the control signal of the PZT microactuator.31.4.17.3 Implementation on a hard disk drive The experimetal setup is the same as in Figure 7. The 3σ value of the position error versus frequencies is shown in Figure 7. Channel 2 in Figure 7. The testing and simulation results of the dualstage open loop system are shown in Figure 7. The performance change with the presence of the system parameter uncertainty is illustrated in Figure 7. When the dualstage loop is closed and stabilized with the designed controllers.29.28. 7. This.g. Controllers are implemented with dSpace 1103 on a TMS320C240 DSP board.. will not hamper the practical application of the resulting controllers due to the large gain and phase margins.2 due to the hump above 0 dB. The step response in Figure 7.28 will amplify the corresponding highfrequency disturbances shown in Figure 7.5 Conclusion An H∞ method has been proposed in both continuous and discrete time domains to achieve a lowhump sensitivity function for dualstage HDD systems using an LMI approach. We can observe that the hump of the sensitivity function is lower than 3 dB. An LDV with a range of 2 µm/V is used to measure the position of the dualstage actuator. Such a design process can generate a robust servo controller with high disturbance rejection in a low frequency range. With the proposed selection of sensitivity weighting functions. where we can see that the lowhump design outperforms the PID design for disturbance rejection after 2.4 kHz.5 kHz. the sensitivity function with a hump below 3 dB has been achieved in both simulations and experiments.28. ±5% shift of PZT resonance frequency around 13.31. A DSA is then used to measure the frequency response of points B over A and obtain the sensitivity function. Two different microactuator models have been studied. which are represented by a MEMS actuator and a PZT actuated suspension. where the rough line is the testing result and the smooth line is the simulation result.
29 Open loop frequency responses of the dualstage system.28 Sensitivity function of the dualstage system. FIGURE 7.LowHump Sensitivity Control Design for Hard Disk Drive Systems 159 FIGURE 7. .
160 Modeling and Control of Vibration in Mechanical Systems FIGURE 7.30 Step response of the dualstage system. .31 3σ value of PES NRRO versus frequency. FIGURE 7.
increase the system and controller complexity since the weighting functions usually have to be of high orders in order to capture the desired speciﬁcations accurately. H∞ optimization together with frequency weighting is a commonly used method. it is very suitable for analysis and synthesis problems in practical applications where different speciﬁcations over different frequency ranges are usually required.8 Generalized KYP LemmaBased Loop Shaping Control Design 8. the Youla parametrization approach is used to parameterize the closedloop transfer function. In this chapter. This is especially so when a controller is to be designed. The KYP Lemma [61]. Hence. the process of choosing appropriate weights is tedious and timeconsuming. are satisﬁed. Unlike the standard KYP Lemma. aiming at achieving a wider bandwidth while simultaneously suppressing disturbances of particular frequencies within or beyond the servo bandwidth. however. To overcome this difﬁculty. It allows us to characterize various properties of dynamic systems in the frequency domain in terms of linear matrix inequalities. being one of the most fundamental results in systems theory and control. the matrix inequality in the generalized KYP Lemma involves a matrix variable which is not necessarily positive deﬁnite and thus the Schur complement cannot be applied to convexify the controller design. The search for the coefﬁcients of the parameter Q(z) is then converted to a linear matrix inequality problem within the generalized KYP Lemma framework. The standard KYP Lemma is only applicable for the inﬁnite frequency range. Further. The additional weight functions. the generalized KYP Lemma is applied to design a feedback control such that the speciﬁcations of the sensitivity function. while the generalized KYP Lemma [62] establishes the equivalence between a frequency domain property and a linear matrix inequality over a ﬁnite frequency range. establishes the equivalence between a frequency domain inequality (FDI) for a transfer function and a linear matrix inequality associated with its state space realization. allowing designers to impose performance requirements over chosen ﬁnite or inﬁnite frequency ranges. required to suppress some speciﬁc frequency disturbances. An application of the method in the rejection of narrowband highfrequency and midfrequency distur 161 .1 Introduction To shape frequency responses of closedloop transfer functions such as sensitivity/complementary sensitivity functions.
it is impossible to achieve disturbance rejection at all frequencies higher than the bandwidth of the control loop for actuators or microactuators used in mechanical motion systems such as hard disk drives. Smaller ri means the less contribution of the disturbance in frequency range (fi1 . which is however possible to be achieved through shaping the sensitivity function. In order to convexify the matrix inequality. In view of the constraint stated in the Bode integral theorem. . it may be applied to address the above design problem. The above design problem may be approached by selecting a proper frequency weighting function and carrying out an H∞ optimization. However. the Youla parametrization approach as shown in Figure 8. N . The speciﬁcation (8. In what follows. fi2 ). 8.2) is considered for a speciﬁc frequency range. the problem of how to select a proper frequency weighting function that can give an accurate shaping of the sensitivity function is generally difﬁcult and timeconsuming.. S(f) < ri . the resultant controller order will depend on the order of the weighting function and the plant. Thus.162 Modeling and Control of Vibration in Mechanical Systems bances is presented to demonstrate the simplicity of the design and the improvement of the positioning accuracy by the resultant controller. fi2 ) to the error.2 Problem description It is known that the power spectrum of the error e in Figure 2. we employ the generalized KYP Lemma to design a feedback control such that the sensitivity function satisﬁes the required speciﬁcations so as to reject disturbances at speciﬁc frequencies.1) which implies that the sensitivity function S(z) is important in determining the disturbance rejection of a control loop.. The purpose here is then stated as: to design a dynamic feedback controller C(z) for plant P (z) such that the closedloop system is stable and for some prescribed positive scalars ri and frequency ranges (fi1 . The generalized KYP Lemma [62] gives a necessary and sufﬁcient condition for a given transfer function to satisfy a required frequency domain property over a ﬁnite frequency range in terms of a matrix inequality condition. fi1 ≤ f ≤ fi2 (8.. Thus. Note that a more accurate frequency shaping usually requires a higher order weighting function.1 is used to design a controller with the generalized KYP Lemma.20 is given by Se = P (z)S(z)2 d1 2 + S(z)2 d22 + S(z)2 n2 (8. .2) where S(f) = 1/(1 + C(f)P (f)). the design of a controller that gives rise to a sensitivity function which can reject speciﬁc disturbances with known frequencies becomes rather signiﬁcant. Further. i = 1.
B. where ˜ A= ˜ B= Ap − Bp D−1 Dc Cp Bp D−1 Cc . D).1 Q parameterization for control design. Dc ). 8. (8. Bc . Cp .3 Generalized KYP lemmabased control design method The previous analysis indicates that the sensitivity function plays a key role in disturbance rejection. (8. Cc . a proper shaping of the sensitivity function can be carried out. D = 1 − Dp D−1 Dc . Then. First. Let a statespace representation of the controller C(z) be given by (Ac .1. z(k) = −Cp x(k) + w(k) − Dp u(k). we shall present a frequency shaping method based on the generalized KYP Lemma. C.4) where x ∈ Rnx is the state.6) A special case of the generalized KYP Lemma that relates the bounded realness of the sensitivity function over ﬁnite frequency ranges to its state space representation is given below. it is easy to see that the sensitivity function S(z) is equal to the transfer function from w to z in Figure 8.3) (8.5) ˜ C = −Cp + Dp D−1 Dc Cp ˜ − Dp D−1 Cc . −1 −Bc Cp + Bc Dp D Dc Dp Ac − Bc Dp D−1 Cc Bp D−1 Dc . In this section. Bp .1 is given by x(k + 1) = Ap x(k) + Bp u(k). . Dp ).Generalized KYP LemmaBased Loop Shaping Control Design 163 FIGURE 8. a statespace representation of the system in Figure 8. The statespace model of the plant under consideration is denoted as (Ap . then a statespace representation ˜ ˜ ˜ ˜ of the sensitivity function can be given by (A. To reject disturbance of frequency within a certain frequency range. Bc − Bc Dp D−1 Dc (8. Assuming that D = 1 + Dc Dp is invertible.
θd = (θ2 − θ1 )/2. (8. Further. Σ= −U V V U − (2cosθl )V .8) (ii) for middle frequency range θ1 ≤ θ ≤ θ2 . ˆ ˆ ˆ u(k) = −M x(k).12) Then.1. the set of sensitivity functions can be parameterized as S(z) = T11 (z) + T12 (z)Q(z)T21 (z) (8. Cc and Dc .164 Modeling and Control of Vibration in Mechanical Systems LEMMA 8. Bc . Cc. (iii) for high frequency range θ ≥ θh . Dc ) ˜ ˜ is to be designed since A and B involve the unknown parameters Ac . Bc .11) (8. (8. (8. the above gives a necessary and sufﬁcient condition for evaluating if S(z) ≤ r over some given frequency range in terms of LMI. Σ= −U e−jθc V ejθc V U − (2cosθd )V .9) θc = (θ1 + θ2 )/2. (8. for a given scalar r > 0. it is not possible to be converted to an LMI by Schur complement since the matrix Σ is not deﬁnite. Let K(z) be a given observer based controller that can be designed using approaches such as the LQG control: x(k + 1) = Ap x(k) + Bp u(k) + L(z(k) + Cpx(k)). S(e ) ≤ r over a ﬁnite frequency range if and only if there exist Hermitian matrices U and V ≥ 0 such that ˜ ˜ AB I 0 ∗ Σ ˜ ˜ 0 0 AB + 0 −r 2 I 0 ˜ ˜ CD ˜ ˜ CD −I ∗ ≤ 0. Σ= −U −V −V U + (2cosθh )V .10) For a given controller C(z). To overcome this difﬁculty. ˆ (8. (8.7) where (i) for low frequency range θ ≤ θl . as stated in Chapter 5. in the following we apply a Youla parametrization approach where the controller C(z) is now of the structure as shown in Figure 8. Then. However.7) is no longer an LMI when the controller C(z) = (Ac .1 ˜ ˜ ˜ ˜ ˜ [62] Consider the sensitivity function S(z) = C(zI − A)−1 B + D with A being jθ stable.13) .
Bt . V . 0 0 1 Cq = [qτ qτ−1 · · · q1 ] . where Iτ−1 is the identity matrix of dimension (τ − 1) × (τ − 1) and 0(τ−1)×1 is the zero matrix of dimension (τ − 1) × 1.. Hence. Therefore. Dt11) and (At . Denote the state space representations of T11 (z) and T12 (z)T21 (z) by (At11 . Ct11 . Ct . Cc = −M + Dq Cp Cq . Dt ).. q = [q0 q1 q2 . Bq . It is known that a state space realization for Q(z) can be given by Aq = 0 Iτ−1 0 . and the design parameter q can be computed via a convex optimization. from (8. Dq ). B = Bt . we know that q exists ˜ ˜ in C and D only. . then the designed controller C(z) is given by Ac = Bc = Ap − Bp M + LCp + Bp Dq Cp Bq Cp Bp Cq . Bq (8.7) deﬁnes an LMI in terms of the variables U .16) Dc = Dq .20) which is to be designed so that the required bounded realness of the sensitivity function is satisﬁed.14) If Q(z) has the statespace realization (Aq . + qτ z −τ . U . and q.19) (8. Bq = (τ−1)×1 . T21 (z) 0 Ap −Bp M 0 Bp −LCp Ap − Bp M + LCp L Bp AT BT .18). (8. In this case. qτ ] (8.17) 0 Bq Ct Aq Bq Dt ˜ ˜ C = Ct11 Dq Ct Cq . D = Dt11 + Dq Dt . Note that the ﬁlter parameter q to be designed only appears in Cq and Dq . . Dq = q0 ..13) a state space model of S(z) can be written as At11 0 0 Bt11 ˜ ˜ At 0 . respectively. Aq (8. Cq .17)−(8. . then from (8.15) L + Bp Dq . = −Cp CT DT Dp M 1 −Dp −Cp Cp 1 0 165 (8.. (8.18) Let Q(z) be an FIR ﬁlter: Q(z) = q0 + q1 z −1 + q2 z −2 + .Generalized KYP LemmaBased Loop Shaping Control Design where Q(z) is a stable transfer function to be designed and T11 (z) T12 (z) = CT (zI − AT )−1 BT + DT . Bt11 . A= 0 (8. V .
Based on disturbance spectrum. where R is the weighting for the control input in the cost function J= (ˆT Cp Cp x + uT Ru) x T ˆ for linear quadratic regulator design. Here in the KYP Lemmabased control design. U and V by solving these LMIs using MATLAB LMI toolbox.16). N ) for the sensitivity function S(fi ) < ri .. 8. Obtain q. Compute T11 (z). B..22) . Obtain the controller parameters from (8.2 with its discretized form F (z). − Cp . Obtain the state space model (A..21) where fi1 and fi2 deﬁne the frequency range. Wd . R.166 Modeling and Control of Vibration in Mechanical Systems The KYP Lemmabased control design can be carried out following the steps: Procedure 8. the speciﬁcations given in Step 4 are to be adjusted. Such a compensator has high gains at speciﬁc frequency. Bp . fi1 ≤ fi ≤ fi2 (8. It is named as peak ﬁlter according to its shape of frequency response. For each speciﬁcation on the resultant sensitivity function in the frequency range. and q with r = ri .15)−(8.18). Wd . R) and L = Ap · dlqe(Ap .17)−(8. Step 2. The peak ﬁlter F (s) works in the control loop as in Figure 8. 8. C. construct the LMI (8.4. ˜ ˜ ˜ ˜ Step 3. specify the positive scalars ri and frequency points fi (i = 1.4 Peak ﬁlter Many other control methods are available to reject narrowband disturbance. . Compute M and L using MATLAB commands M = dlqr(Ap . Step 4. D) in (8. V . Step 5. Step 6.14).1 Conventional peak ﬁlter 2 s2 + 2ξ1 ωp s + ωp . Cp Cp . and Wv can be chosen as identity matrices. and Wd and Wv are the variance matrices of process noise and measurement noise for the Kalman estimator design. It has been shown that most of these methods would result in a compensator with lightly damped complex poles at the center frequency of the disturbance. and adaptive feedforward methods using higher harmonic control and LMS algorithm [57]. T12 (z) and T21 (z) from (8.7) in terms of the variables U . Wv ). 2 s2 + 2ξ2 ωp s + ωp Conventionally. such as linear time invariant feedback methods using classical design and modern frequency shaping and ﬁlter shaping. the peak ﬁlter is described as the following model F (s) = (8.1 T Step 1. Bp . If the LMIs are not solvable.
3 Peak ﬁlter in the frequency domain. FIGURE 8.Generalized KYP LemmaBased Loop Shaping Control Design 167 FIGURE 8.2 Peak ﬁlter F in the nominal feedback loop. .
m ξ1 = (8. 1) is the damping ratio. ∆ stands for the peak width corresponding to N .25) (8.4. 2 m (8. Let T0 (s) = P (s)C(s) .2 Phase lead peak ﬁlter The phase ﬁlter is improved by adding a differentiator to provide additional phase lead such that the phase margin is preserved and the sensitivity function curve is smoothly shaped. Let m = 10M/20 . When M ≫ N .3. The peak ﬁlter F (s) can be designed according to its Bode plot. particularly for the disturbance near the 0 dB crossover frequency. and 0 < K < γ is the positive ﬁlter gain. ξ ∈ (0.23) The maximum phase loss θ caused by the pair of the lightly damped complex poles is given by m−1 θ = arctan √ . as illustrated in Figure 8. 2 s2 + 2ξω0 s + ω0 (8. 8. 2∆ + 2 ξ1 ξ2 = .30) (8. φ is the phase angle determined by φ = arg[T0 (jω0 )] ∈ [−π. a phaselead peak ﬁlter [118] was proposed.27) The peak ﬁlter adopts the following form F (s) = K −sin(φ)s2 + ω0 cos(φ)s . ξ1 and ξ2 are determined approximately by ∆2 + 2∆ √ n − 1.26) It is noticed that the conventional version of the peak ﬁlter induces additional phase loss. 1 + P (s)C(s) (8. K) = 1 + T0 (jω)F (jω). To overcome the drawback.168 Modeling and Control of Vibration in Mechanical Systems where ωp = 2πfp is the center frequency in radian/sec and ξ1 and ξ2 are the damping ratios with ξ1 > ξ2 . The phase loss negatively impacts the phase margin and distorts the gain of the sensitivity function around the disturbance frequency.29) .28) where ω0 is the disturbance frequency at which high disturbance rejection is required. n = 10N/20 . (8.24) (8. M and N denote the peak height. π]. Let G(ω.
An illustration example is shown in Figure 8.31) (8. 8. The ﬁlters associated with solid and dashed curves have different parameters. γ)] = 0.28) is given by ∆d (ω0 + 0.28) with the corresponding ﬁlter parameters (ωi . γ)] = 0. φi . K = (10M/20 − 1) T0 (jω0 ) 169 (8.5 Application in high frequency vibration rejection In this section.33) (8. Im[G(ω. The previously introduced peak ﬁltering method ﬁnds it difﬁcult to deal with disturbances in high frequency range.28) is a general highgain controller to reject narrowband disturbances in a speciﬁc frequency range because the ﬁlter zero location can be automatically shifted according to the disturbance frequency associated with the baseline servo system with C(s) and P (s).3 Group peak ﬁlter The group ﬁltering scheme is used to compensate multiple frequency narrowband disturbance. An estimate of ξ and K in the ﬁlter (8. 8.5∆d) . (8. we shall provide an example to demonstrate the design and effectiveness of the proposed KYP based control to reject narrowband disturbances at high frequencies. especially near the actuator resonance modes.32) ξ= (8.4.4 with the sensitivity functions plotted before and after activating the group ﬁlter with two subﬁlters at ω1 = 2π700 and ω2 = 2π2000. The disturbance ﬁlter in (8. Ki ). 2 4ω0 2ξ . ξi .35) where the subﬁlter F i (s) is given by (8.Generalized KYP LemmaBased Loop Shaping Control Design and γ is the minimal positive real solution of the following two equations: Re[G(ω. .34) where the disturbance bandwidth ∆d is deﬁned as the frequency difference between √ the two points whose magnitudes are 1/ 2 times of the peak value. and M in unit dB is the desired reduction ratio of the narrowband disturbance. A parallel structure of the group peak ﬁlter is described as n F (s) = i=1 F i (s).
.4 Sensitivity functions before and after group peak ﬁltering activated.5 PZT microactuator attached to VCM actuator arm. FIGURE 8.170 Modeling and Control of Vibration in Mechanical Systems 10 5 0 −5 Magnitude (dB) −10 −15 −20 −25 −30 −35 with filter with filter without filter 10 2 10 Frequency (Hz) 3 10 4 FIGURE 8.
Spec. The desired speciﬁcations for the sensitivity function S(z) are set as Spec. (a) means to guarantee a 2 kHz bandwidth at least.41 dB. the system suffers from the high frequency disturbances around 8 kHz and 10 kHz.6 shows an example of the main disturbances that exist in the hard disk drive servo system.7s + 1. 9950 Hz ≤ f ≤ 10050 Hz.Generalized KYP LemmaBased Loop Shaping Control Design 3 x 10 −3 171 2.257 × 105 )(s2 + 791.35 dB.5 10 kHz 0 0 2000 4000 6000 Frequency(Hz) 8000 10000 12000 FIGURE 8. Spec. these disturbances appear more prominent and create a signiﬁcant impact on the positioning accuracy of the R/W head.(b) S(f ) < −0.36) . Here we employ the generalized KYP Lemma to design a feedback control C(z) for the microactuator such that the sensitivity function S(z) satisﬁes the required speciﬁcations so as to reject disturbances around 8 kHz and 10 kHz. (s + 1. 7950 Hz ≤ f ≤ 8050 Hz.567 × 109 )(s2 + 5089s + 7.(c) S(f ) < −1. The transfer function of the PZT microactuator is given by P zt(s) = 180828605599509(s2 + 3079s + 1. A. f ≤ 2 kHz.195 × 109 ) (8. For higher rotational speed HDDs.5 shows one kind of microactuators for HDDs.934 × 109 ) . Besides the disturbances of low frequencies. Spec.5 1 8 kHz 0.5 PES NRRO power sepctrum(µm) 2 1.(a) S(f ) < 0 dB.6 Power spectrum of the position error before servo control. which is a PZT actuated suspension attached to a VCM actuator arm. System models Figure 8. Figure 8. which are induced by air turbulence to suspensions or sliders in HDDs.
(b) and (c) can be satisﬁed.172 Modeling and Control of Vibration in Mechanical Systems which is obtained by curveﬁtting to its frequency response measured via DSA. 20 10 Magnitude(dB) 0 −10 −20 −30 −40 3 10 4 measured modeled 10 200 100 Phase(deg) 0 −100 −200 −300 3 10 Frequency(Hz) 10 4 FIGURE 8. ξ1 < ξ2 .3 kHz and 13. and ω2 and ω1 should be chosen to be close to each other so that the resultant notch ﬁlter will . The notch ﬁlter is designed as the following form: N otch = 2 2 ω2 s2 + 2ξ1 ω1 s + ω1 · 2 2 2.3 kHz and 13. The sampling frequency being used is 40 kHz. ω1 s + 2ξ2 ω2 s + ω2 (8. z−1 (8.38) where ω1 is the frequency of the resonance to be suppressed . Controller design and LDV based experiment results The plant P zt(s) is precompensated by the integrator: Int(z) = 6.3z . where it is noticed that there are two dominant resonance modes at 6.37) and two notch ﬁlters for suppressing the two main resonances at 6. B.5 kHz. P zt(s) is discretized using the zeroorderhold method to obtain P zt(z).7.7 PZT micro actuator frequency response. A comparison between the measured and the modeled frequency responses is shown in Figure 8.5 kHz before the KYP Lemma approach is applied to design a controller so that the speciﬁcations (a).
9. From (8.4011z 2 + 0. Wd .3)−(8. Observed from Figures 8. As mentioned previously in Step 1 of the control design procedure.19).7) with Σ in (8.1391z + 0. Channel 2 over Channel 1 of DSA is the measured frequency response of the sensitivity function with a swept sine signal as the reference. z 2 − 0.6371.8.9259 0. The resulting controller for the PZT microactuator is C(z) · Int(z) · N otch1(z) · N otch2(z). M and L are T obtained using MATLAB commands. C1 .8. the sensitivity function is less than −4 dB and −2 dB at 8 kHz and 10 kHz respectively. which means that speciﬁcations (a)−(c) have been satisﬁed by the searched Q(z). Next we use the KYP Lemma to search for the coefﬁcients of Q(z). Bp . Three LMIs of the form (8. 1. The obtained q0 and q1 are q0 = −0. . C1 C1 .4117. (b) and (c). C(z) can be obtained via (8. as seen from Figure 8. (a). respectively. It can be observed in Figure 8.8) and (8.16).8. the σ values of the position errors before and after the KYP Lemmabased control versus frequencies are obtained and shown in Figure 8. and slightly worse before 5 kHz. q1 = 0.8751z + 0. the gain margin and the phase margin are 12 dB and 67 deg. higher than 7. which results in the reduction of the sensitivity function from 8 to 10 kHz. and an LDV was used to measure the actuator displacement.36) the notch ﬁlters after discretization are as follows. Moreover.9.8 that with Q(z). and q0 and q1 are to be solved via the KYP Lemma.8 and 8.12.15)−(8. As such.1). Cp . In the experiment. z 2 + 0. M = dlqr(A. and the sensitivity function is shown in Figure 8.3532 N otch2(z) = .10.8 where it can be seen that the required speciﬁcations are not met. The price for this compensation is a bit lower loop gain at lower frequencies.177z 2 − 1.007922 N otch1(z) = (Ap . i.11 shows the experimental sensitivity functions.4) is a state space description of the combined system P (z) = Int(z) · N otch1(z) · N otch2(z) · P zt(z). R) and L = A · dlqe(A.279z + 1. B2 .e. Figure 8. Dp ) in (8. B2 .9) need to be solved in order to achieve Spec. Here for the model (8. which agree with the simulation results in Figure 8. which is consistent with the sensitivity functions in Figure 8. which is consistent with the Bode Integral constraint for sensitivity function.3972z + 0. Wv ). as shown in Figure 8. the KYP design increases the loop gain around 9 kHz. It can be seen that the performance with the KYP Lemmabased control is much better from 5 kHz onwards.Generalized KYP LemmaBased Loop Shaping Control Design 173 not inﬂuence the system stability.7 dB and 63 deg before the KYP design. Q(z) is chosen as the 1st order FIR ﬁlter (8. the dSpace 1103 on TMS320C240 DSP board was used to implement the controller. The sensitivity function before the KYP Lemma based design is also shown in Figure 8.154 .
8 Sensitivity functions before and after the KYP lemmabased design: simulation result.174 Modeling and Control of Vibration in Mechanical Systems 10 Before KYP design After KYP design 0 −10 Magnitude(dB) −20 −30 −40 −50 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 8. .
.10 Structure of experimental setup. FIGURE 8.Generalized KYP LemmaBased Loop Shaping Control Design 175 50 40 Magnitude(dB) 30 20 10 0 −10 −20 1 10 After KYP optimization Before KYP optimization 10 2 10 3 10 4 0 −200 Phase(deg) −400 −600 −800 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 8.9 Openloop Bode plot before and after the KYP lemmabased design.
004 0.01 0.002 0 0 2000 4000 6000 Frequency(Hz) 8000 10000 12000 FIGURE 8.11 Sensitivity functions before and after the KYP lemmabased design: experimental results.012 Before KYP design After KYP design 0.12 σ value of PES versus frequency. . 0.006 0.176 10 Before KYP design After KYP design 5 Modeling and Control of Vibration in Mechanical Systems 0 Magnitude(dB) −5 < 0dB −10 −15 −20 −25 2 10 10 Frequency(Hz) 3 10 4 FIGURE 8.008 PES σ(µm) 0.
7 and 15 kHz are included in the model. After that. The resonance mode at 12.13.54 dB. f ≤ 500 Hz. 4.4. (a). the phaselead peak ﬁlter (PLPF) of the form in (8. 610 Hz ≤ f ≤ 670 Hz. with this sampling rate the platform can support up to 10th order controller. and ξ = 0. is also applied to .13.6 Application in midfrequency vibration rejection The frequency responses of the microactuator are shown in Figure 8.9. 45 kHz sampling rate is used in the servo control design. The 4. The objective here is to use the above KYP method to design a linear dynamic output feedback controller C(z) for the microactuator in Figure 8. have relatively small magnitudes and can be ignored as long as they are not excited in the control loop.39) which is precompensated by the proportionalintegral (PI) controller: Int(z) = 0.7. (b) means to attenuate the disturbances centering at 650 Hz by 10 dB at least.9 and 6.7. z − 0.7) corresponding to Spec.28) with values K = 0.Generalized KYP LemmaBased Loop Shaping Control Design 177 8. the computation of controller can be very efﬁcient. Spec. The parameters of Q(z) in (8.7 kHz is not considered in the control design as it is not excited easily and does not affect the whole loop stability when the 15 kHz mode is compensated. which will not change a lot the obtained performance of the low frequency part.(a) S(f ) < 0 dB.13 such that its closedloop system is stable and the disturbance centering at 650 Hz is suppressed sufﬁciently. The pade delay model is given by Ppade−delay = −5. Spec.584. The resultant C(z) is a 10th order controller.9. s + 2 · π · 17000 (8. Six resonance modes at 3. The desired speciﬁcations for the sensitivity function S(z) are set as: Spec.5).0632. 12. notch ﬁlters for the resonance modes at 3. due to the limitation by the DSP speed. we just involve the static part of the microactuator represented by a pade delay in the control design with the KYP Lemma. ω0 = 2π650. Because 650 Hz is at a relative low frequency range.19) with τ = 1 are attained by solving three LMIs of the form (8. 9 and 15 kHz will be used to compensate the dynamic part.(c) S(f ) < 9. Currently.40) Due to the ﬁrst order padedelay model used in the computation of LMIs. 6.14. It is noticed that there is a vibration mode at 650 Hz due to disk vibration. φ = −0. For the sake of comparison. The control algorithm is implemented with the digital position error signal generated from DSP TMS320C6711. Spec.(b) S(f ) < −10 dB.999 (8. The disturbance distribution is reﬂected in the nonrepeatable runout power spectrum of the measured PES in Figure 8. (b) and (c). seen in Figure 8.027(− z + 0.9 kHz resonance modes. 9. f ≥ 19 kHz.6234 s − 2 · π · 17000 .
. it is known that the spectrum of the true PES y is given by Sy = P (z)S(z)2 × d1 2 + S(z)2 d2 2 + T (z)2 × n2 = Se − S(z)2 × n2 + T (z)2 × n2 . In the open loop comparison in Figure 8. the phase margin (PM) with the PLPF method is much higher. Thus the 3σ value of the true PES can be assessed from the power spectrum Se in Figure 8.178 Modeling and Control of Vibration in Mechanical Systems suppress the low frequency disturbances around 650 Hz.1). As a result it is improved from 6.7 Conclusion This chapter has applied the generalized KYP Lemma in the microactuator closedloop design to suppress the narrow band disturbances. and the sensitivity function comparison is shown in Figure 8. the resultant sensitivity function has to comply with the Bode integral theorem.42) where Se is in (8. Consistent with the sensitivity functions in Figure 8. It can be seen that the KYP method achieves better disturbance rejection from 60 Hz to 1 kHz. The resultant controller veriﬁes that the desired speciﬁcations to reject the disturbances have been satisﬁed via the search for the coefﬁcients of Q(z) in the Youla parametrization approach.15. meaning that it is not possible to achieve disturbance rejection across the entire frequency range. In the above application. we shall incorporate a nonlinear compensation in Chapter 13. However. The Youla parametrization approach has been used in the feedback controller design.17 which clearly shows that the KYP based design gives a better disturbance rejection around 650 Hz than the PLPF although at 650 Hz they offer a similar performance. A higher order Q(z) offers more design freedom and has the potential of achieving better results.16. only the ﬁrst order Q(z) is used. and T (z) = 1 − S(z) is the closed loop transfer function. while the bandwidth is lower and the gain margin (GM) is comparable with the KYP Lemma method.15.20. The system design problem with multiple speciﬁcations on the gain properties of the sensitivity function over several frequency ranges has been solved by the LMI optimization based on the KYP Lemma. To further improve the disturbance rejection at low frequency for the KYP Lemmabased design. From Figure 2. (8. the KYP method gives a poorer disturbance rejection performance for frequency below 60 Hz than the PLPF method. However. the PES NRRO power spectrum comparison is shown in Figure 8. 8.4 nm with the PLPF method to 6 nm with the KYP Lemma method. Practical applications have been demonstrated for narrowband high frequency and midfrequency disturbance rejection.17 with the known level of noise n.41) (8. whatever Q(z) is used. although they have almost the same rejection capability in the very narrow band around 650 Hz.
Generalized KYP LemmaBased Loop Shaping Control Design 179 40 30 Magnitude(dB) 20 10 0 2 3 4 Measured Modeled 10 10 10 100 0 Phase(deg) −100 −200 −300 −400 −500 −600 2 10 10 3 10 4 Frequency(Hz) FIGURE 8. .13 Frequency response of the PZT microactuator.
180 Modeling and Control of Vibration in Mechanical Systems x 10 −3 4 3.5 0 0 1000 2000 3000 4000 Frequency(Hz) 5000 6000 FIGURE 8.2 nm).5 2 1.14 PES NRRO power spectrum calculated from measured PES signal without servo control.5 NRRO magnitude(µm) 3 2. reﬂecting the vibration distribution of the system (3σ = 21 nm including the noise 3σ = 15.5 1 0. .
.15 Comparison of sensitivity functions. KYP(GM: 6 dB.4kHz)). .. PM: 34 deg.7 kHz))). Bandwidth 1.Generalized KYP LemmaBased Loop Shaping Control Design 10 KYP PLPF 0 181 −10 Magnitude(dB) −20 −30 −40 −50 −60 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 8.16 Open loop frequency responses (PLPF (GM: 6 dB. PM: 50 deg. FIGURE 8. Bandwidth: 1.
.6 650Hz 0.17 NRRO power spectrum with PLPF and KYP (50% reduction before 1 kHz).8 NRRO magnitude(µm) 0.182 Modeling and Control of Vibration in Mechanical Systems x 10 −3 KYP PLPF 1 0.2 0 0 1000 2000 3000 Frequency(Hz) 4000 5000 6000 FIGURE 8.4 0.
the KYP Lemmabased loop shaping method does not count for overall positioning error minimization which can be translated into the H2 optimal control problem by taking into consideration the disturbance and noise models. With the selected speciﬁc disturbances handled by the KYP Lemmabased design. Thus it usually does not have the ﬂexibility to speciﬁcally reject disturbances at certain frequency ranges. The design procedure will be illustrated and the resultant controller will be veriﬁed via an experiment. excluding the disturbances covered in the above design. Subsequently. 183 . This motivates us to incorporate the KYP Lemmabased method with the H2 control method in this chapter. In this chapter.1 Introduction As a closedloop shaping method. Therefore there is a need to suppress disturbances of speciﬁc frequencies when minimizing the positioning error. However. On the other hand. A series of simulation and experimental results will show the effectiveness of the control design method in terms of enhancing positioning accuracy. the H2 control design which incorporates all disturbance and noise models can result in an average performance across the entire frequency range and a high order controller. the positioning accuracy can be improved to some extent. which however may be dominant factors that inﬂuence the overall performance. the KYP Lemmabased approach allows designers to impose performance requirements over selected ﬁnite frequency ranges so as to have the desired sensitivity function that is able to reject the disturbances in these speciﬁc frequency ranges. Then one more disturbance at 2 kHz near the servo bandwidth 1kHz is also considered as a speciﬁc disturbance to be rejected via the the KYP Lemmabased design.9 Combined H2 and KYP LemmaBased Control Design 9. we will apply the combined control design method to a PZT microactuator such that a disturbance at 650 Hz is rejected with the KYP Lemmabased design and at the same time overall positioning error is minimized via the H2 control design. the H2 control is formulated with a lower order disturbance model. This will not only release the computation burden in the H2 control design but also result in a lower order controller.
m (9.2 Problem formulation FIGURE 9. The problem of the speciﬁc disturbance rejection can be solved by imposing such performance speciﬁcations in (9. In the previous chapter..2) is required. Such an upperbound speciﬁcation as in (9. This speciﬁcation helps to ﬁx the bandwidth or 0dB crossover frequency of S(z). i = 1. S(fi ) ≥ 1.1 H2 control scheme with Q parametrization for controller design.14.1) and then using the KYP Lemmabased control design method in Chapter 8. The 0dB crossover frequency of S(z) will be pushed away towards a higher frequency. In view of this.1) will lead to a problem when the frequency fi is larger than and especially near the desired bandwidth or 0dB crossover frequency of S(z). and fi1 and fi2 deﬁne the frequency range. a lowerbound speciﬁcation. as seen in Figure 9.e.184 Modeling and Control of Vibration in Mechanical Systems 9. we also . fi1 ≤ fi ≤ fi2 (9. the servo mechanical system suffers from various kinds of disturbances and sensing noise.9. fi1 < fi < fi2 . However. which tends to damage the system stability and deteriorate the system highfrequency performance. The KYP Lemmabased control design cannot include all disturbances and noises which contribute to the position error. 2. which will be seen later in the application results. i.1 associated with Figure 8.1) where ri < 1 is a positive scalar. · · · . In view of this. speciﬁcations on sensitivity function S(z) are described as S(fi ) < ri . as shown in Figure 9.
which is repreT sented as the socalled track misregistration (TMR) induced by w = w1 w2 w3 passing through D1 (s).Combined H2 and KYP LemmaBased Control Design 185 take into account the overall performance of the servo control system. Dc ) respectively be the statespace model of plant P (z) and controller C(z).14) that a set of sensitivity functions S(z): (A. K(z) is an observer based controller that can be designed using the LQG method as in (8. Π∗ Π22 12 (9. Let (Ap . the speciﬁcation S(z) ≤ r is written as σ(S.2).3.1) and (9.6) . − r2 (9. D) can be Qparameterized. Cc. I stands for an identity matrix and Π a Hermitian matrix of the form Π= Π11 Π12 . B. Speciﬁcation (9.4) where S(z) = S(ejθ ). 9. D2 (s).1 Q parametrization to meet speciﬁc speciﬁcations A.5). Dp ) and (Ac . we proceed to the controller design to achieve the speciﬁcations in (9. It is expressed by the standard deviation σz of z.4)−(9.1) ˜ ˜ ˜ ˜ Recall from (8. For the presentation of the KYP Lemma.5) which speciﬁes the frequency domain property to be investigated.3 Controller design for speciﬁc disturbance rejection and overall error minimization The generalized KYP Lemmabased design method in Chapter 8 is used to design a controller for speciﬁc disturbance attenuation. Bc .3). Π) := S I ∗ Π S I (9. C.12). Bp .11)−(8. Cp . we denote σ(S. According to the denotation (9.13)−(8.3) when w is a white noise with zero mean and identity covariance matrix. where Tzw is the transfer function from w to z. (9. and N (s). and Tzw 2 = σz . Π) ≤ 0 with Π= Π11 Π12 1 = Π∗ Π22 0 12 0 . and meanwhile to optimize (9. 9.1. In the next section. In order to convexify matrix inequalities. the Youla parametrization approach with the Q(z) in a FIR ﬁlter form is applied and the controller structure is shown in Figure 9.
Π= (9.1.7) is not possibly convexiﬁed according to the method in Section 9. b = −1. Hence we resort to the following speciﬁcation 0 a + jb a − jb 2c σ (S.8) e −jθc V θc = (θ1 + θ2 )/2. Speciﬁcation (9. (9. In this situation. b and c are properly selected. Denote the statespace representation of T11 (z) and T12 (z)T21 (z) by (At11 . Ct11 .2) Again. achieving S ejθ ≤ r for the frequency range θ1 ≤ θ ≤ θ2 can be obtained by solving the following matrix inequality ˜ ˜ AB I 0 ∗ ˜ ˜ ˜ ˜ AB CD Σ + I 0 0 I ∗ Π ˜ ˜ CD ≤ 0. Π) = aR (S) + bI (S) + c. Bt . ejθc V U − (2 cos θd ) V (9.14) . Dt ). 0 I (9. we shall give a state space realization of S(z) = T11 (z) + T12 (z) Q (z) T21 (z).3. equivalent to ˜ ˜ ∗ ˜ ˜ 0 0 AB AB I 0 Σ I 0 + 0 −r 2 ˜ ˜ Π11 C D where Σ= −U ˜ ˜ CD ∗ −Π11 Π11 .18). c = r.8).9) (9. Ct . θd = (θ2 − θ1 )/2. When a. Dt11 ) and (At . Π) = −I (S) + r. Π∗ Π22 0 r2 12 (9.11) However. Π) ≤ 0 means I (S) ≥ r. Π∗ Π22 j 2r 12 (9. Π := (9. A simple selection is a = 0.17)−(8.7) which is. Bt11 . Π) ≤ 0 with Π= Π11 Π12 −1 0 = . and σ (S. (9. since Π11 > 0.7) can not be converted equivalently to (9. S(z) ≥ r can be achieved. and subsequently S(z) ≥ r. To convexify the matrix inequality (9. according to the denotation (9.10) U and V are Hermitian matrices and V ≥ 0.8).186 Modeling and Control of Vibration in Mechanical Systems Thus based on the KYP Lemma in Chapter 8. Then a statespace model of S(z) can be written as (8. B. which means (9. because Π11 < 0. Π11 Π12 0 −j = .5) the speciﬁcation S(z) ≥ r is equivalent to σ(S.12) where R and I denote the real and the imaginary parts of S(ejθ ). ≤ 0.4)−(9. respectively.13) Thus σ(S.
It should be mentioned that R (S) ≥ r can also be used to achieve S(z) ≥ r.15) ˜ ˜ which is a linear matrix inequality with unknown variables in C and D only.7) is equivalent to ˜ ˜ AB I 0 ∗ 187 Σ ˜ ˜ ˜ 0 C ∗Π12 AB + ∗ ˜ ˜ ∗ Π12 + Π∗ D + Π22 ˜ I 0 Π12 C D 12 ≤ 0. and can be solved using the same method as in Section 9. (A2 . (9. B1 . D1 ). respectively. Π∗ Π22 −1 2r 12 (9.15) remains applicable. Ω=ΩT >0) min T race (Ω) (9.Combined H2 and KYP LemmaBased Control Design where Π11 = 0 and (9. and (A3 .20) 0 B3 ˜ ˜ BD2 −BD3 ˜ ˜ ˜ DD1 DD2 −DD3 + D3 . B2 . C3 .3. D3 ).18) that ¯ ¯ x (k + 1) = Ax (k) + Bw(k).23) .19) A1 0 ¯ A= 0 ˜ BC1 ¯ ˜ C = DC1 B1 0 0 0 A2 0 0 ¯ 0 . 9. C1 . From Figure 9.17) Denote a statespace realization of P (z)D1 (z).22) (9.1A. D = 2 0 0 B2 0 . It follows from (8.21) subject to ¯ ¯ ¯ ¯ AT ΞA − Ξ + C T C < 0 T ¯ ¯ ¯ ¯ B ΞB + D T D < Ω (9. (9.3. B = 0 0 A3 0 ˜ ˜ 2 −BC3 A ˜ ˜ BD1 BC ¯ ˜ ˜ ˜ DC2 −DC3 + C3 C . if it is suitable for a speciﬁc application. D2 (z) and N (z) by (A1 . B3 .1 we have −z = N (z) w3 + S (z) [P (z) D1 (z) w1 + D2 (z) w2 − N (z) w3 ] . D2 ). C2 . (9.13) and (8. The H2 norm Tzw can be minimized as follows: (Ξ=ΞT >0.2 Q parametrization to minimize H2 performance Next we focus on the design of Q(z) to minimize the H2 norm Tzw 2 . (9.18) (9. where.17)−(8. Π= Π11 Π12 0 −1 = . In this case. ¯ ¯ −z (k) = Cx (k) + Dw(k).16) and the linear matrix inequality (9.
25) deﬁne LMIs in terms of the and D. from (8. (9.188 or equivalently.18) and (9. D. . the controller C(z) is then given by Ap − Bp M + LCp + Bp Dq Cp Bp Cq .24)−(9. t11 t11 t11 Bt D1 Bt D2 −Bt D3 Bq Dt D1 Bq Dt D2 −Bq Dt D3 ¯ = (Dt11 + Dq Dt ) C1 (Dt11 + Dq Dt ) C2 C (Dt11 + Dq Dt ) C3 Ct11 Dq Ct Cq .17)−(8. Modeling and Control of Vibration in Mechanical Systems ¯ ¯ ¯ AT ΞA − Ξ C T ¯ C −I ¯ ¯ ¯ −Ω + B T ΞB DT ¯ D −I where <0 (9. V .8). ˜ ˜ ¯ Therefore. Bq . .26) Note that the Q(z) coefﬁcients qi (i = 0. Bq Cc = −M + Dq Cp Cq .24) <0 (9. ¯ = (Dt11 + Dq Dt ) D1 (Dt11 + Dq Dt ) D2 D − (Dt11 + Dq Dt ) D3 + D3 ] .26). Bq Cp Aq L + Bp Dq Bc = . τ ) only appear in Cq and Dq . variables U . Dq ). 0 0 0 .25) A1 0 0 0 0 0 A2 0 0 0 0 0 A3 0 0 ¯ A= Bt11 C1 Bt11 C2 −Bt11 C3 At11 0 Bt C1 Bt C2 −Bt C3 0 At Bq Dt C1 Bq Dt C −Bq Dt C3 0 Bq Ct B1 0 0 0 B2 0 0 0 B3 ¯ B= B D1 B D2 −B D3 .27) 9. Dc = Dq . With the solved Q(z): (Aq . Ξ. Hence. (9. . Ω and qi . . Ac = (9. . a design procedure for controller C(z) is given as follows.15) and (9. we know that qi exists only in C. C ¯ In this case. the Q(z) coefﬁcients qi can be computed via a convex optimization.3. 1.3 Design steps To summarize. 0 0 Aq (9. Cq .
− 0. Obtain the controller C(z) from (9. D) in (8.29) (9.8168].6275j. . T12 (z) and T21 (z) from (8. − 0. 9. Design K(z) from (8.11)−(8.12).7116 ± 0. Step 5. S (fi ) < ri . B. .4. n) for the sensitivity function S(z). fj1 ≤ fj ≤ fj2 .2 for comparison and it is subsequently discretized in MATLAB using the “zoh” method with a sampling rate of 40 kHz. 11.0093 ± 0.4927j. Based on disturbance spectrum and bandwidth requirement.0255 ± 1. − 0. Bq .0110 ± 1. 0. shown in Figure 9. Obtain Q(z) : (Aq . .18). − 0. Cq . . . Gain = −0. − 0. Cq and Dq . . Step 3.24)−(9. 9.. C.4071j.3 kHz.3 for a PZT microactuator to separately reject one or two speciﬁc disturbances and meanwhile minimize the H2 norm of the PES.1 System models The frequency response of the PZT microactuator. fi1 ≤ fi ≤ fi2 . Cq and Dq . Step 2. Ω. − 0.4 Simulation and implementation results This section will apply the control design method in Section 9.0964j. i. specify the positive scalars ri and rj . Step 6.Combined H2 and KYP LemmaBased Control Design 189 Step 1. 9. was obtained using a LDV and a DSA. P oles = 105 × [−0. Construct the LMIs (9.15) in terms of the variables U .5981j. .5 kHz.17)−(8.5 kHz. Step 4.4819. Compute T11 (z).27).0296 ± 0. The frequency response of the plant model is plotted against the measured data in Figure 9.6198j. The main resonance modes of the plant are at frequencies 6.0048 ± 0. . . Dq ) by solving the above LMIs using the MATLAB LMI toolbox.0245 ± 0. and S (fj ) > rj .2733j. V .25) in terms of the variables Ξ. and the frequency points fi (i = 1.0050 ± 0. The identiﬁed plant model of the microactuator P (s) has the following parameters: Zeros = 105 × [−0.14).28) For each speciﬁcation.7100j. − 0.8) and (9. and 20 kHz. (9. m) and fj (j = 1. and obtain the state space ˜ ˜ ˜ ˜ model (A.2.e. construct the LMIs (9.8168].
A suitable feedback controller.24) and (9. It is seen that the hump of S(z) is reduced to about 3 dB with the controller designed by the combined method.e. C(z) has to be designed for the system so that the overall system is stable and the disturbance around 650 Hz is suppressed sufﬁciently. On the other hand. The controllers are implemented using dSPACE 1103 on a TMS320C240 DSP board and the structure of the experimental setup is the same as in Figure 8. With each of the two designed controllers.. Hence.2 Rejection of speciﬁc disturbance and H2 performance minimization Consider the disturbances in Figure 8. Experiments are carried out for the KYP Lemmabased controller and the KYP+H2 controller to verify the simulation results. The sensitivity function obtained from the .14.190 40 Modeling and Control of Vibration in Mechanical Systems Measured Modeled Magnitude (dB) 20 0 −20 −40 2 10 10 Frequency (Hz) 3 10 4 50 0 Phase (deg) −50 −100 −150 −200 −250 2 10 3 4 10 Frequency (Hz) 10 FIGURE 9.4. the desired speciﬁcation of the sensitivity function S(z) is S (f) < −10 dB for 610 Hz ≤ f ≤ 670 Hz and at the same time the position error signal is to be minimized. The parameters of a ﬁrstorder FIR Q(z) are obtained by solving the three LMIs (9. the disturbance around 650 Hz is due to the disk vibration.2 Frequency response of a PZT microactuator. (9.8).10. where a swept sine signal is the reference input. the performance speciﬁcations listed in Table 9.3 and Figure 9.25). i. The sensitivity function of the system is obtained by the ratio of the measurements in Channel 2 and Channel 1 of the DSA. 9. while ensuring that the H2 norm of the position error signal is minimized.4. the combined H2 optimization and speciﬁc disturbance rejection method. the frequency response of the openloop C(z) × P (z) and the sensitivity function S(z) are depicted in Figure 9.1 show that the proposed method offers better stability margins although the openloop crossover frequency is a bit lower. For comparison. another controller is designed without H2 minimization and just to suppress the vibration around 650 Hz.
as shown in Figure 9. 50 KYP+H2 KYP Magnitude(dB) 0 −50 1 10 10 2 10 Frequency(Hz) 3 10 4 −50 −100 Phase(deg) −150 −200 −250 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 9.6.5.9 12. the position error is reduced and its σ value has a 4% improvement. As a result.9 experiment. demonstrates the effectiveness of the proposed method.1 191 Comparison of performance speciﬁcations Method KYP KYP+H2 Crossover frequency (kHz) 1.78 Gain margin (dB) 7.4 45.Combined H2 and KYP LemmaBased Control Design TABLE 9. . as seen in Figure 9.3 Openloop frequency responses.85 1.6 Phase margin (deg) 35.
10 KYP+H KYP 5 2 0 −5 Magnitude (dB) −10 −15 −20 −25 −30 −35 2 10 10 Frequency (Hz) 3 10 4 FIGURE 9.4 Designed sensitivity functions.5 Comparison of sensitivity functions obtained from experiment. .192 Modeling and Control of Vibration in Mechanical Systems 10 KYP+H2 KYP 0 −10 Magnitude(dB) −20 −30 −40 −50 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 9.
20 0. By solving the LMIs (9.35 KYP+H KYP 0. Ξ.9. V . the disturbance around 2 kHz is also to be suppressed. a controller C(z) which leads to the S(z) satisfying the speciﬁcations can be obtained with a fourthorder FIR Q(z).8 x 10 2 4 FIGURE 9.2 0.3 Rejection of two disturbances with H2 performance minimization In what follows.10 0. The experimental results of the openloop frequency response and sensitivity function are . Note that the third speciﬁcation aims to ﬁx the bandwidth.2 1.30 2 193 0. where in addition to the rejection of disturbance at 650 Hz. we include one more speciﬁc disturbance rejection at 2 kHz (see Figure 9.8 1 Frequency (Hz) 1. the speciﬁcations on S(z) are set as: i. (9.6 0.15 0. Ω.24)−(9.6 1. the openloop frequency response and the sensitivity function are shown in Figures 9.5 degrees.8 and 9. (ii) and (iii) have been achieved and the closedloop system remains stable. S (f) > 0 dB for 990 Hz ≤ f ≤ 1010 Hz. ii. S (f) < −10 dB for 610 Hz ≤ f ≤ 670 Hz.8). Cq and Dq .15) and (9. S (f) < −5 dB for 1950 Hz ≤ f ≤ 2050 Hz. respectively.05 0 0 0. 9. With the resultant controller. where the speciﬁcations (i).4 1. The openloop gain and phase margins are 14.6 NRRO power spectrum with KYP Lemmabased controller with and without H2 minimization.4.6 dB and 49.4 0. iii.Combined H2 and KYP LemmaBased Control Design 0. In this case.25) in terms of the variables P . which is caused by air ﬂow and is near the desired servo bandwidth 1 kHz.25 NRRO magnitude (nm) 0.7). We have also carried out experiments to verify the designed controller.
Based on Figure 9. 9. respectively.7 PES NRRO spectrum without servo control. the attenuation of disturbance centering at 2 kHz has improved by 35%. The NRRO spectrum of the system with the designed controller is shown in Figure 9.8 and Figure 9.11.5 Conclusion This chapter has addressed a combined control design method that incorporates the generalized KYP Lemmabased design and the H2 optimization.194 3 x 10 −3 Modeling and Control of Vibration in Mechanical Systems 2.5 1 0. The method has been applied to design a controller for a PZT microactuator to attenuate disturbances at 650 Hz and 2 kHz where disk vibrations are dominant and minimize the 3σ value of the overall PES. Simulation and experimental results have demonstrated the effectiveness of the proposed method and veriﬁed that the positioning accuracy has been improved.11.10. With the incorporated control design method.5 2 NRRO magnitude (µm) 1. speciﬁc narrowband disturbances have been attenuated and simultaneously the positioning error of the control system has been minimized. .5 0 0 1000 2000 3000 4000 5000 6000 Frequency (Hz) 7000 8000 9000 10000 FIGURE 9. plotted in Figure 9.
(ii) and (iii).9 Resultant sensitivity function (Solid line: with Spec. Dashed line: with Spec.Combined H2 and KYP LemmaBased Control Design 40 20 Magnitude (dB) 0 −20 −40 −60 195 10 2 10 3 10 4 200 0 Phase (deg) −200 −400 −600 −800 Simulation Experiment 10 2 10 3 10 4 Frequency (Hz) FIGURE 9. . (i). 10 0 Magnitude (dB) −10 Bandwidth is increased −20 −30 −40 −50 1 10 10 2 10 Frequency (Hz) 3 10 4 FIGURE 9. (i) and (ii)).8 Openloop frequency response.
.2 1 0.6 1.6 0.11 NRRO power spectrum with rejection of two speciﬁc disturbances at 0.2 0 0 1000 2000 3000 4000 5000 6000 Frequency (Hz) 7000 8000 9000 10000 FIGURE 9.8 1.196 Modeling and Control of Vibration in Mechanical Systems 10 5 0 −5 Magnitude (dB) −10 −15 −20 −25 Simulation Experiment −30 10 2 10 Frequency (Hz) 3 10 4 FIGURE 9. 2 x 10 −3 Without control Closed loop 1.4 NRRO magnitude (µm) 1.65 and 2 kHz.4 0.10 Resultant sensitivity function with all the three requirements fulﬁlled.8 0.
In this chapter. Each individual controller can be a static or a dynamic feedback controller derived by means of standard optimal and robust control methods such as the LQG/LTR control. H2 or H∞ control. (2) reject three disturbances with frequencies higher than bandwidth.10 Blending Control for MultiFrequency Disturbance Rejection 10.8inch HDD VCM actuator in three cases.2 Control blending Control blending accounts for the problem of simultaneous performance achievements. 10. If several targeted transfer functions are achieved respectively by their designed controllers. In this chapter.1 Introduction The blending control technique [120] aims to design a controller capable of simultaneously coping with different optimality criteria deﬁned for various inputoutput channels. Control design for each disturbance rejection is carried out by using the H2 optimal control method. it is shown that under some mild assumptions there exists a unique controller capable of replicating these transfer functions and thus simultaneously achieving the performance attained by each individual controller. which leads to the achievement of a satisfactory rejection of multifrequency narrow band disturbances. Note that this chapter presents a systematic approach using disturbance models only. The ultimate controller is obtained by blending all these H2 controllers and is expected to be able to reject all these disturbances simultaneously. the control blending is respectively applied in the control design for a 1. The blended controller can be easily computed by a procedure based on simple linear algebra. (1) reject two disturbances of frequencies higher than bandwidth. we shall apply the blending control idea to deal with the problem of rejecting several narrowband disturbances. (3) reject one disturbance with frequency lower than bandwidth and another one higher than bandwidth. and. The multifrequency disturbance rejection is formulated as a control blending problem. It involves designing an individual controller for each performance speciﬁca 197 .
has full column rank.1: (A. u(t) ∈ Rnu . ··· yr (t) = Cr x(t) + Dr u(t). where x(t) ∈ Rnx .1) (10. Ci .198 Modeling and Control of Vibration in Mechanical Systems tion and blending all the controllers to obtain an ultimate controller. (10. (10. Assumption 10. . and A. and De are matrices of appropriate dimensions.2: The nx × s matrix. Ce . yi (t) ∈ Rpi . e(t) ∈ R.1 Blending control scheme.5) y1 (t) = C1 x(t) + D1 u(t). B) is stabilizable. where s = r i=1 qi . B. Consider the system in state space: r x(t) = Ax(t) + Bu(t) + ˙ i=1 Ei wi (t). Di . e(k) = Ce x(k) + De u(k) + v(k).6) FIGURE 10. wi (t) ∈ Rqi .3) (10. Ei . Assumption 10. y2 (t) = C2 x(t) + D2 u(t).2) (10.4) (10. The following two assumptions are necessary in the control blending scheme. E = E1 E2 · · · Er .
. (10.11) (10.e. . the closedloop system is stable with spectrum ¯ σ(Acl ) = ∪ σ(Ai ). (10. r. · · · .1. . we consider the static state feedback control u(t) = Ki x(t). 2. Ci = Ci + Di Ki .7) (10. i = 1. then there exists a single dynamic compensator which simultaneously achieves the same performances. i = 1.1)−(10.2. . ··· .8) can be found following the procedure below.5). i (10. 2. . . the transfer function Ti (s) from the ith input to the ith output coincides with the one given by the ith gain Ki . (10. . r is then described by A + BDs BCs E1 E2 · · · Er Bs As 0 0 · · · 0 C1 + D1 Ds D1 Cs 0 0 · · · 0 Acl Ecl (10.14) The signiﬁcance of the result lies in the fact that if Ki .13) and 2). . i = 1. 2. .8) The closedloop system from wi to yi .7)−(10. . The individual closedloop transfer function is given by ¯ ¯ ¯ Ti (s) = Ci (sI − Ai )−1 Ei . ¯ Ti (s) = Ti (s). The single dynamic compensator of the form in (10. i. · · ·. . for each i. Cr + Dr Ds Dr Cs 0 0 · · · 0 For the system (10. with ¯ ¯ Ai = A + BKi .. r. . .1 State feedback control blending The feedback control scheme of the above system is shown in Figure 10. Procedure 10.9) = C2 + D2 Ds D2 Cs 0 0 · · · 0 . where the controller is expressed as xs (t) = As xs (t) + Bs x(t) ˙ u(t) = Cs xs (t) + Ds x(t).8) of order µ = nx(r − 1).10) The result in [120] reads that there exists a single dynamic compensator of the form (10. Ccl 0 .12) (10.7)−(10.1 . · · · . such that 1).Blending Control for MultiFrequency Disturbance Rejection 199 10. are the static compensators each of which optimizes some performance criteria.
10. Ai = A + BKi is asymptotically stable.20) .. Cs . i. • Deﬁne ˜ ˜ Zi = [0 Zi ][Ei Ei ]−1 . This is possible under Assumption 10. 2. In practical application.15) • Compute the controller matrices As . where yp ∈ Rp . Theoretically to match r different transfer functions Ti .16) (10.22) (10. Ds as Ds Cs Bs As = K1 K2 · · · Kr Ξ−1 . Bs . • Form the square matrix Ξ as Ξ= I I ··· I .17) (10.e. choose a matrix Ei of dimension nx × (nx − qi ) such that [Ei Ei ] is square and invertible. Z1 Z2 · · · Zr (10. ˜ • Choose matrices Zi of dimension µ × (nx − qi ) (µ = nx (r − 1)) such that the matrix ˜ ˜ ˜ E1 E2 · · · Er E1 E2 · · · Er ˜ ˜ ˜ 0 0 · · · 0 Z1 Z2 · · · Zr is invertible. and ¯ Vi = Zi Ai .2. The output feedback controller takes the form xc (t) = Ac xc(t) + Bc y(t) ˙ u(t) = Ccxc (t) + Dc y(t). however. i = 1. (10. V1 V2 · · · Vr (10.2. ˜ ˜ • For each i.18) (10. · · · . nx (r − 1) + nx = nx r different poles have to be allocated and a full order controller is required. r be the static state feedback controllers each of which ¯ asymptotically stabilizes the plant.19) The order of the blended controller is nx (r − 1) which is very high. the feedback control has to rely on some measurements given by yp (t) = Cy x(t).21) (10.2 Output feedback control blending When the full state variables of the plant are not available. reducing the order is needed as long as it is possible.200 Modeling and Control of Vibration in Mechanical Systems • Let Ki .
(10. such that 1). .2 • Let Ki . has full column rank. . .21)−(10.22) of order µ = nx r − p. ˜ ˜ • For each i. choose a matrix Ei of dimension n × (n − qi ) such that [Ei Ei ] is square and invertible. .22) can be found following the procedure below.e. . . · · · . the spectrum of the closedloop system satisﬁes ¯ ∪ σ(Ai ) ⊆ σ(Acl ). ··· ..28) and 2). This is possible under Assumption 10. Cr + Dr Dc Cy Dr Cc 0 0 · · · 0 E = Cy E1 Cy E2 · · · Cy Er . i.e.29) The single dynamic output controller of the form in (10. the matrix Ai = A + BKi is asymptotically stable. with ¯ ¯ Ai = A + BKi . i = 1.21)−(10.23) = C2 + D2 Dc Cy D2 Cc 0 0 · · · 0 . the transfer function Ti (s) from the ith input wi to the ith output yi coincides with the one given by the ith gain Ki . for each i. · · · . . Ci = Ci + Di Ki . . r. 2. i.26) (10. . r The closedloop system from wi to yi . . r is then described by A + BDc Cy BCc E1 E2 · · · Er Bc Cy Ac 0 0 · · · 0 C1 + D1 Dc Cy D1 Cc 0 0 · · · 0 Acl Bcl (10. r be the static state feedback controllers each of which ¯ asymptotically stabilizes the plant.. Ccl 0 . Consider the static state feedback control as stabilizing compensators u(t) = Ki x(t).3.Blending Control for MultiFrequency Disturbance Rejection 201 Assumption 10. . (10.27) (10. .24) where s = i=1 qi . ¯ Ti (s) = Ti (s). 2. The corresponding closedloop transfer function from wi to yi is given by ¯ ¯ ¯ Ti (s) = Ci (sI − Ai )−1 Ei . i = 1. . (10. · · · . . 2. i (10. i = 1.3: The composed p × s matrix. Procedure 10.25) The theorem on output blending control in [120] says that there exists a single dynamic compensator of the form (10.
and ¯ Vi = Zi Ai . If E is square then the controller matrix is stable if and only if system (10. ··· . and it is proved in [120] that the matrix Ξ is indeed an invertible matrix. . = . ¯ N 0 . . Bc . . Dc as Dc Cc Bc Ac = K1 K2 · · · Kr Ξ−1 . . . • Form the square matrix Ξ as Ξ= Cy Cy · · · Cy . . iii). Proposition 10.34) The procedure is quite simple to implement.30) • Compute the controller matrices Ac .35) is of minimum phase. ··· . Cc . V1 V2 · · · Vr (10. 0 Cy · · · 0 E1 ··· 0 0 . . ii). ¯ · · · Ar 0 · · · Cy 0 ··· 0 E2 · · · 0 .35) is of minimum phase. The controller matrix Ac is not stable if system (10.32) (10. . The controller matrix can be generically chosen as a stability matrix if system (10. Z1 Z2 · · · Zr (10. .1 [120] i). .33) (10. 0 Cy 0 ¯ A2 . • Deﬁne ˜ ˜ Zi = [0 Zi ][Ei Ei ]−1 . .31) (10.35) has unstable invariant zeros.35) The following Proposition advises on the stability of the controller matrix Ac . . Consider the system ¯ A1 0 ¯ ¯ AE . 0 · · · Er 0 0 ··· 0 (10.202 Modeling and Control of Vibration in Mechanical Systems ˜ • Choose matrices Zi of dimension µ × (nx − qi ) (µ = nxr − p) such that the matrix ˜ ˜ ˜ Cy E1 Cy E2 · · · Cy Er Cy E1 Cy E2 · · · Cy Er ˜ ˜ ˜ 0 0 ··· 0 Z1 Z2 · · · Zr is invertible.
(10. · · · . + 2ξi ωi s + ωi (10. and (Adi . · · · .36) The multifrequency disturbance rejection problem is stated as follows: For narrowband disturbances di . r. C(z) is the feedback controller. fiH ] can be suppressed simultaneously. r. we ﬁrst formulate the problem of multifrequency disturbance rejection into the above control blending framework. 10. ζi used to adjust the damping of Di . r x(k + 1) = Ax(k) + Bu(k) + i=1 Ei wi (k). where P (s) is the actuator model. Bp . and gain kdi . Ddi ) as the statespace realizations of Di (z) with state vector xdi ∈Rndi . Ddi = 0.1 Problem formulation The closed control loop of a mechanical actuation system is shown in Figure 10. 2. v is the measurement noise with σv as its standard deviation.2.39) (10.38) Denote the output due to di by yi . we focus on the rejection of output disturbance Di (s). i = 1. To ensure necessary rejection of lowfrequency disturbances. Dp ) as the statespace realizations of P (z) with state vector xp ∈Rnp . a full state observer is required. i = 1. Then. ˜ ˜ (10. Cdi . considering that an input disturbance can be converted to an output disturbance. with Ref = 0 we have y = S(z)(D1 (z)w1 + D2 (z)w2 + · · · + Dr (z)w r ) + (1 − S(z))v. Bdi . To facilitate the single dynamic state feedback controller. (10. 2.37) with frequency ωi . damping ratio ξi . r are disturbance models and wi are white noises with unity variances. Cp . denote (Ap . 2.Blending Control for MultiFrequency Disturbance Rejection 203 10. e is the measured error signal. Clearly. In Figure 10. Di (s) is proposed to have the form: Di (s) = s2 kdi (s + 2ζi ωi ) 2 .2. Then the combined system is given as follows.2. and the disturbances di in different frequency ranges [fiL . a precompensation integrator Int(z) is introduced which is shown in Figure 10. design a feedback controller C(z) such that the closedloop system is stable. Here.40) y1 (k) = C1 x(k) + D1 u(k). . i = 1. · · · .1 is used to design a blended controller associated with the static state feedback controller designed via the H2 optimal control method. Di (s).3 Control blending application in multifrequency disturbance rejection In this section.3.3 and given by ˜ u(k + 1) = ke(k) + u(k). From Figure 10. Procedure 10.
FIGURE 10.3 Control structure. .204 Modeling and Control of Vibration in Mechanical Systems FIGURE 10.2 Control loop with injected disturbances at different frequencies.
. .47) (10. .3.50) . where x = xT uT xT1 xT2 · · · xTr p ˜ d d d r i=1 T 205 (10. 0 0 0 0 · · · Adr 0 0 0 0 −kDd ˜ −kDd ˜ −kDd ˜ 1 2 r Bd 0 0 1 (10.43) includes all disturbance outputs yi . . r. the SISO (single input and single output) output feedback controller Cb (z) will be given based on the dynamic state feedback controller and the state observer. . . ··· yr (k) = Cr x(k) + Dr u(k). wi (k) ∈ ndi . . u(k) ∈ Rnu .43) Rqi . · · · . Before the blending. 0 0 Bdr C2 = −Cp −Dp 0 −Cd2 · · · 0 . C1 = −Cp −Dp −Cd1 0 · · · 0 . D2 = −Dp . (10. . (10. De = −Dp . . . . e(k) = Ce x(k) + De u(k) + v(k).49) Ce = −Cp −Dp −Cd1 −Cd2 · · · −Cdr .2 Controller design via the control blending technique In this section. qi = 1. . . Next a state observer will be designed to facilitate the state feedback controller with the measurement signal. D1 = −Dp . (10. nu = 1. the blending technique will be utilized to yield one single dynamic state feedback controller that is able to reject all disturbances. Er = 0 . Denote the dynamic state feedback controller by Σ(z) : xs (k + 1) = As xs (k) + Bs x(k). .45) E1 = 0 .41) (10. nx = np + 1 + ∈ Rnx .46) (10. yi (k) ∈ Rpi . 2 . 10. . .48) (10. Ap Bp 0 0 ··· 0 Bp ˜ ˜ ˜ ˜ ˜ ˜ −kCp 1 − kDp kCd1 kCd2 · · · kCdr −kDp 0 0 0 Ad1 0 · · · 0 A= 0 0 0 Ad2 · · · 0 . Cr = −Cp −Dp 0 0 · · · −Cdr . the H2 optimal control method will be used to design a static state feedback controller to minimize the error caused by each disturbance. 2. . .42) (10. we present how to design the controller C(z) by using the control blending technique. . Note that the measurement error signal e in (10. i = 1. . . E2 = Bd . . Dr = −Dp . Subsequently. .Blending Control for MultiFrequency Disturbance Rejection y2 (k) = C2 x(k) + D2 u(k). . . B = 0 . Finally. .44) . . . pi = 1. · · · .
55) (10. Ki is obtained as Ki = Wi Pi−1 . E. is needed to facilitate the controller Σ(z). as in Figure 10. i = 1. and R = σv . Dc = 0. Since only the measurement is available for feedback. R).53) < 0. E is as in (10. e(k) u(k) . e(k) (10. . Ce . · · · . x(k + 1) = Ao x(k) + [Bo1 Bo2 ] ˆ ˆ yo (k) = Co x(k) + Do ˆ where Ao = A − ALCe . 2.52) (10. Q. The overall controller C(z) is then given by u(k) . Ki can therefore be obtained by the following optimization: min T race(Si ) subject to the LMIs −Pi ∗ ∗ Pi AT + WiT B T −Pi ∗ < 0. Q = I.1. (10.206 Modeling and Control of Vibration in Mechanical Systems u(k) = Cs xs (k) + Ds x(k).6).57) (10.51) is designed based on Procedure 10. 2 L = dlqe(A. Bo1 Cs Ao + Bo1 Ds Co Bo2 (10. (10.3.59) (10. a dynamic state feedback controller of the form (10. Co = I.56) (10. Bc = . where Ki in step 1 is designed based on the H2 optimal control method. Bo1 = B. r be the static state feedback controllers each of which asymptotically stabilizes the plant and minimizes the H2 norm of the transfer function from wi to yi . Let Ki .50)−(10. With the solved Wi and Pi .58) Ac = As Bs Co 0 . which is described as follows. Bo2 = AL. a state observer. 0 Ci P + Di Wi −1 −Si ∗ Ei −Pi (10.51) This dynamic state feedback controller is designed following Procedure 10.1. Do = 0. The state observer O(z) is designed as follows.54) Next.60) Cc = Cs Ds Co .
964 × 108 ) 1 .61) × 2 (s + 2815s + 2.4 shows the frequency response of the VCM actuator. controller C(z) will be designed to reject a few disturbances with frequencies higher than bandwidth.326628 × 1017 (s2 + 1081s + 7. 10.527 × 109 )(s2 + 1. . First.131 × 104 s + 3. Figure 10.4 Simulation and experimental results The disk drive under consideration is a 1. The VCM actuator model P (s) is described by 8.3 × 108 )(s2 + 6635s + 2. (10.948 × 109 ) P (s) = 60 40 Gain [dB] 20 0 −20 −40 −60 1 10 10 2 Measured Modeled 10 3 10 4 0 −100 Phase [degree] −200 −300 −400 −500 −600 1 10 10 2 10 Frequency [Hz] 3 10 4 FIGURE 10.4 Frequency response of the VCM actuator. and next.852 × 109 ) (s2 + 552.Blending Control for MultiFrequency Disturbance Rejection 207 10.777 × 105 )(s2 + 1056s + 6. two disturbances will be rejected.1 Rejecting highfrequency disturbances In this section. three disturbances are to be attenuated simultaneously.4.9s + 4.8inch small hard disk drive with a spindle motor rotational speed of 4200 RPM.
085. More simulations show that the change of S(z) for D1 at 4 kHz leads to slight variation of S(z) for D2 at 8 kHz. With Σ(z). By ﬁtting the model in (10.55)−(10. ζ2 = 0. C(z) can be reduced to 14th order. and ω2 = 2π8000. The openloop bandwidth is 1 kHz. Moreover. Take ki = 100 · Ts . After simple order reduction by canceling close zeros and poles. changing kd2 will affect rejection at 4 kHz more than changing damping ξ2 . 80 Magnitude(dB) 60 40 20 0 −20 1 10 200 Phase(deg) 100 0 −100 −200 1 10 10 2 10 3 10 4 10 2 10 Frequency(Hz) 3 10 4 FIGURE 10. S(z) < 1) frequency ranges around 4 kHz and 8 kHz.6..7.e.5. Rejection capability in terms of S(z) for Di (i = 1.5 Openloop frequency response with disturbance rejection at 4 and 8 kHz. while the change of S(z) for D2 at 8 kHz affects more on S(z) around 4 kHz. the observer O(z) (10. ξ2 = 0. The simulated sensitivity function S(z) = 1/(1 + P (z)C(z)) is shown in Figure 10. kd2 = 43 for D2 (s).37).208 Modeling and Control of Vibration in Mechanical Systems We consider two disturbances at frequencies 4 kHz and 8 kHz which are the resonance frequencies of the plant as shown in Figure 10. kd1 = 20 for D1 (s). which agrees well with the simulated one (the dashed curve). ξ1 = 0.19) of Σ(z) are obtained by the design procedure outlined in the last section.4.6.01.6. where there are two obvious rejection (i. we have ω1 = 2π4000. The sensitivity function S(z) of the closed control loop is measured via DSA and plotted in Figure 10. The controller is implemented via dSpace 1103 with the utilization of LDV to measure the displacement of the actuator. 2) can be changed by adjusting kdi and ξi . Incorporated with an integrator. a 27th order C(z) is obtained. we have controller Cb (z) by (10.59)−(10.56). The parameters in (10. ζ1 = 0.60). . The ﬁnal controller C(z) leads to the frequency response of the openloop P (z)C(z) as shown in Figure 10.
ζ3 and kd3 of D3 for 10 kHz disturbance will affect the rejection for the other two disturbances more than itself.37) with ω1 = 2π3000. The rejection capabilities for these three disturbances are further examined via more simulations. ζ2 = 0. Next. kd3 = 52 for D3 (s).5 kHz. With the lower order C(z). ξ3 = 0. assume that disturbances at three frequencies are to be rejected simultaneously. 6. the compensated openloop frequency response is shown in Figure 10. and ω3 = 2π10000. Changing parameters ξ3 .6. After order reduction. and 10 kHz are clearly seen.5 kHz disturbance by changing kd2 .8. The order of the designed C(z) is 46.6 Simulated sensitivity function with disturbance rejection at 4 and 8 kHz.61. ξ2 or ζ2 will not lead to much change to other two disturbance rejections. The rejection change for 6.Blending Control for MultiFrequency Disturbance Rejection 10 209 0 −10 4 kHz 8 kHz Magnitude(dB) −20 −30 −40 −50 −60 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 10.3. It is shown that changing rejection capability for 3 kHz disturbance by adjusting kd1 and/or ξ1 will change that for 6.085. while will not affect much on that for 10 kHz. The simulated and the measured sensitivity functions are shown in Figure 10. ξ2 = 0. kd2 = 50 for D2 (s). ζ1 = 0.9 where three rejection frequency bands around 3 kHz.02.5 kHz. ζ3 = 0. a 31st order C(z) is obtained. The disturbances are modeled as (10. ω2 = 2π6500. . and the frequencies are not the plant resonance frequencies. kd1 = 40 for D1 (s).7. ξ1 = 0.
80 Magnitude(dB) 60 40 20 0 −20 1 10 200 Phase(deg) 100 0 −100 −200 1 10 10 2 10 3 10 4 10 2 10 Frequency(Hz) 3 10 4 FIGURE 10. .8 Openloop with disturbance rejection at 3.7 Measured (solid curve) sensitivity function with disturbance rejection at 4 and 8 kHz. and 10 kHz.210 20 Modeling and Control of Vibration in Mechanical Systems 10 0 −10 Magnitude(dB) −20 −30 −40 −50 −60 −70 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 10. 6.5.
In this section. ξ2 = 0.11. Changing ζ2 will easily cause phase loss at 2 kHz. the phases are both lifted. As observed in Figure 10. increasing kd2 in D2 (s) for 2 kHz disturbance will not affect much on S(z) at 650 Hz.9 Sensitivity function with disturbance rejection at 3.4. The disturbance models are as in (10. kd2 = 0.Blending Control for MultiFrequency Disturbance Rejection 10 211 0 −10 Magnitude(dB) −20 −30 −40 −50 −60 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 10. ζ2 = 0. By close zeropole cancelations. 6.10. kd1 = 77 for D1 (s) at 650 Hz. It is thus suggested to change kd2 and ξ2 to have a satisfactory sensitivity function S(z) at 2 kHz without causing phase loss. and is able to accomplish the task of rejecting D1 and D2 . S(z) < 0 at 2 kHz. a 21st order controller can be obtained. the frequencies of the disturbances that need to be rejected are higher than the bandwidth. In Figure 10. but also lower it at 2 kHz. On the other hand. and ω2 = 2π2000. higher than the bandwidth. 10. ζ1 = 0. Increasing kd1 will lower S(z) at 650 Hz. lower than the bandwidth.1 for D2 (s) at 2 kHz. see the rough curve in Figure 10.10 and 10.85.8. we consider two disturbances near the bandwidth: one at 2 kHz.37) where ω1 = 2π650. and the other at 650 Hz. The experimentally measured S(z) is plotted.05.0046.5. Additional simulations show that decreasing damping ξ1 will not only lower S(z) at 650 Hz. Using the method of phase lead peak ﬁlter in Chapter 8. ξ1 = 0.11. but will increase the hump of S(z) and cause slight increase of S(z) at 2 kHz. which is desired in order not to lose the phase margin. and 10 kHz.2 Rejecting a combined mid and high frequency disturbance Previously. the rejection capability .11. which is illustrated in Figures 10. The resultant controller C(z) is of 27th order. a signiﬁcant point is that corresponding to the peaks at 650 Hz and 2 kHz.
that however have been treated in this chapter.1 From the above application results and analysis. a baseline controller needs to be designed ﬁrst. the phase lead peak ﬁlter method is difﬁcult to deal with multifrequency vibrations in higher frequency range and at frequencies near plant resonance modes. Moreover. which are individually designed to deal with the two vibrations.65 and 2 kHz. i. . it is noted that to deal with lower (say ≤ 2 kHz) frequency disturbances. then the two peak ﬁlters are obtained based on the predesigned baseline controller. 60 Magnitude(dB) 40 20 0 −20 2 10 200 Phase(deg) 100 0 −100 −200 2 10 10 3 10 4 10 Frequency(Hz) 3 10 4 FIGURE 10. This feature is beneﬁcial to the stability of the closedloop system. For higher (say > 2 kHz) frequency disturbances.. 1 + pf1 + pf2 . say pf1 and pf2 . In this design method.212 Modeling and Control of Vibration in Mechanical Systems for vibrations at 650 Hz and 2 kHz can be achieved with two peak ﬁlters. it works like phasestabilized control [86]. REMARK 10. and then the two ﬁlters are connected in parallel. Thus the rejection performance of the system for vibration in other frequency ranges depends on the predesigned baseline controller.10 Openloop with disturbance rejections at 0. the proposed control design method based on the blending technique works like the phase lead peak ﬁltering method.e.
10. The control blending method has been respectively applied to design a controller for a 1. In addition. (1) rejecting two disturbances of frequencies higher than bandwidth and close to actuator resonance frequencies. A controller for each disturbance rejection has been designed individually by using the H2 optimal control method. . (3) rejecting one disturbance with frequency lower than bandwidth and another higher than bandwidth.65 and 2 kHz.Blending Control for MultiFrequency Disturbance Rejection 10 213 0 −10 Magnitude(dB) −20 −30 −40 −50 −60 −70 10 1 10 2 10 Frequency(Hz) 3 10 4 FIGURE 10.11 Sensitivity function with disturbance rejections at 0. it is worth noting that the method is able to prevent phase loss when it is used to deal with disturbances near the bandwidth. Simulation and experimental results have shown that the control blending technique results in a simultaneous attenuation for these disturbances.5 Conclusion The rejection problem of several disturbances around different frequencies has been formulated as a control blending problem.8inch HDD VCM actuator in three cases. The ultimate controller has been obtained by blending all these H2 controllers so that all these disturbances can be rejected simultaneously. and. (2) rejecting three disturbances with frequencies higher than bandwidth and different from resonance frequencies.
.
it requires the model of the controlled plant to be accurately known and invertible. 215 . at least within the bandwidth of interest [129]. which does not need to solve the plant model inverse. they could be estimated from the system output measurement by an asymptotic estimator (Luenberger observer) and the effect of the disturbances could be neutralized by feeding the disturbance estimates back into the system [124]. will be introduced. it has been proven that under certain assumptions imposed on the plant and disturbance models. The designed disturbance observer is applied to an HDD servo system and its effectiveness in disturbance attenuation is demonstrated by simulations and experiments. Subsequently. If majority of the disturbances are of relatively lower frequency. it is not always easy to identify the disturbance model. However. Over the years. It was suggested that if the disturbances were supposedly generated by a linear dynamic system and the model of the system was known. we introduce the conventional disturbance observer ﬁrst. In this chapter. This new method does not require control designers to have the full information of the disturbance model and does not need the assumption that the disturbance model is linear timeinvariant.1 Introduction The idea of observing disturbance to improve the performance of a servomechanism was ﬁrst introduced in [123]. the two different methods are equivalent in that the original disturbance observer introduced in [123] is actually a generalization of the latter method [126]. In this chapter. the method has been modiﬁed and applied. a new type of disturbance observer (DOB) has been introduced [128]. and then present a general form of disturbance observer. Further. when a DOB is added on to attenuate the effect of disturbances. An H∞ control based method is applied to the design of a Qﬁlter.11 H∞ Based Design for Disturbance Observer 11. the standard and conventional way of designing a Qﬁlter is to design it to be a lowpass ﬁlter with unity DC gain [128]. it is not always true that the disturbance model is linear timeinvariant. Recently. we study the latter method where a general form of disturbance observers. However.
d1 is the input disturbance.216 Modeling and Control of Vibration in Mechanical Systems 11. FIGURE 11. P (z) = z Pn (z). τ represents a delay in the plant P . when Q = 1. However.2) that an ideal solution of Q is zero. but this will mean Tyd1 ≈ P and hence the disturbance will be ampliﬁed.2) To reject the disturbance d1 . Thus. it is known from (11. and Pn is the −τ inverse of Pn .2 Conventional disturbance observer Figure 11. Tyn ≈ −1 which means the measurement noise n is not attenuated. Ignoring the nominal feedback loop with C. Theoretically. the transfer functions from the disturbance d1 and noise n to the output y are given by Tyd1 = Tyn P Pn (1 − Qz −τ ) . Pn is the nominal model of the plant −1 ˆ P . n is the noise. where C is the feedback controller.1) (11. d1 is an estimate of d1 . Pn + Q(P − Pn z −τ ) −P Q = .1 Block diagram of the control loop with a conventional disturbance observer. Q can be set as unity because Tyd1 ≈ 0 when the delay is negligible. Pn + Q(P − Pn z −τ ) (11.1 shows the block diagram of the conventional disturbance observer (DOB) structure. . to eliminate the noise effect.
1. CM . it is needed to solve the inverse Pn of the nominal plant model. B(z) = bm z m + bm−1 z m−1 + · · · + b0 . DM ). 11.4) which is stable as long as Q is stable since the loop is stable before the disturbance observer is added on. yq (k) = Cy x(k) + Dyw w(k) + Dyu uq (k). 1 + PC (11. the disturbance can be rejected because S(z) → 0.2. x(k + 1) = Ax(k) + B1 w(k) + B2 uq (k). Bc . A(z) = z n + an−1 z n−1 + · · · + a0 (11. the plant model is P = Pn z −τ .6) and d2 is the output disturbance. (11.2 displays a general form of disturbance observer. With the same notations as in Figure 11. N (z) = Pn (z). y(k) = Cz x(k). Moreover.1. In what follows. −1 M (z) = z −τ .3) becomes S(z) = 1 − Qz −τ . Dc ). where M (z) = A(z) B(z) .H∞ Based Design for Disturbance Observer 217 Considering the overall system in Figure 11. it is deduced from (11. (11. N (z) = d .3) Theoretically. a general form of the disturbance observer is proposed and the plant model inverse is not needed. In the conventional disturbance observer in Figure 11.3) that when Qz −τ = 1 with zero phase around the disturbance frequency.10) . z dm z n (11. From Figure 11. thus (11. Dp ). + P C + P QPn (11. the sensitivity function is given by S(z) = 1− Qz −τ 1 − Qz −τ −1 . Bp . CN .1. Hence Q is designed such that the phase of Qz −τ is almost zero degree and the magnitude is close to one in the frequency range where the disturbance d1 −1 dominates. BN .8) (11. Figure 11. Note that in Figure 11. BM . M (z) : (AM . C(z) : (Ac . Cc .1.9) (11. DN ).7) Denote the statespace descriptions P (z) : (Ap . N (z) : (AN .3 Let A general form of disturbance observer Pn (z) = B(z)/A(z). Cp .5) be the nominal model of the plant P .
Here the models for disturbances d1 . and N (z). and xT (k) = [xT (k) xT (k) xT (k) xT (k)] is the augmented state of P (z). ∞ (11. where Ap − Bp Dc Cp Bp Cc 0 0 −Bc Cp Ac 0 0 A= −BM Dc Cp BM Cc AM 0 .12) (11.2 Block diagram of the control loop with a general disturbance observer. . The H∞ optimization method will be applied to design Q(z) to minimize the H∞ norm Wd1 Tyd1 Wd2 Tyd2 Wn Tyn .14) Wd1 . Dyu = −DM . However. Denote the transfer function from w to y as Tyw = [Tyd1 Tyd2 Tyn ].11) Dyw = 0 −(DN + Dm Dc ) DN + DM Dc Cy = −DN Cp − DM Dc Cp DM Cc CM CN . B2 = −BM . C(z). Dzu = 0. −BN Cp 0 0 AN −Bp Bp −Bp Dc Bp Dc 0 0 −Bc Bc B1 = 0 −BM Dc BM Dc . Dzw = [0 − 1 0]. p c M N M (z). 0 0 −BN BN (11. and w T (k) = [d1 (k) d2 (k) n(k)].13) Cz = −Cp 0 0 0 . d2 and noise n are not needed. in order to have a desired suppression of the . (11.218 Modeling and Control of Vibration in Mechanical Systems FIGURE 11. Wd2 and Wn are weightings.
P = Pn . Different from the conventional disturbance observer. The objective of the general disturbance observer design is then stated as: Given a positive scalar γ and appropriate weightings Wd1 . design a stable Q(z) : (AQ . CQ . meaning that the conventional disturbance observer is a special case of the general disturbance observer.16) With the conventional disturbance observer in (11. (11. Wd2 and Wn are required which relies on our knowledge of the disturbances. BQ .20) .19) z dn .7). DQ ) such that [Wd1 Tyd1 Wd2 Tyd2 Wn Tyn ]T ∞ < γ.16) Q(z) = Qn (z) × then S(z) = 1 + Qn z dn /z dm . (11.4. and derived as Pequ = P y = . Wd2 and Wn . REMARK 11. the general disturbance observer does not need the inverse of the nominal plant model. if the loop is cut at u. appropriately chosen weightings Wd1 .H∞ Based Design for Disturbance Observer 219 disturbances d1 . u 1 − (M − P N )Q (11. 1 + P C + Qz −d − P QPn (11.17) which recovers the form in (11.18) 1 + Qz −d −1 . d2 and the noise n. As shown later.16) is equal to S(z) = Assume that in (11.2 In Figure 11. B(z) (11.1 The sensitivity function of the control system in Figure 11. the transfer function from u to e or ˜ y is considered as a new plant.15) The H∞ control design problem can be solved via Theorem 5.17) when dm ≥ dn .2 with the general disturbance observer is given by S(z) = 1 + QM . 1 + P C + Qn z dn /z dm − Qn (11. it is able to suppress the disturbances in a low frequency range without much performance degradation to higher frequency disturbances and noise. REMARK 11.2. 1 + P C + QM − P QN (11.
(11.02025. and the compensation of VCM pivot friction [130].517 × 10−5 z 4 − 6. due to its cost effectiveness and easy “addon” implementation with minimal change required to the existing feedback controller. 1 − (M − P N )Q (11.474 × 10−5 z 6 − 4.7765.59 × 10−5 z 5 + 8.997z 2 − 2.081 × 10−5 z 4 − 4.979z 3 + 3.20) becomes Pequ = P . 11.621 × 10−5 z 3 +5.03483z 3 − 0.4 Application results It has been shown that a disturbance observer is capable of estimating disturbances and modeling error [128]. Consider the VCM plant with model P (s) given in Chapter 10. (11. N (z) = 6 .657z 4 − 4.5) with B(z) = 0. 3.1868z 4 − 0.234 × 10−6 z Let M (z) = A(z) B(z) . The discretized model with the sampling time Ts = 1/30000 sec is given by (11. 6 z z (11. such as the attenuation of disturbances [127].24) .545 × 10−5 z 2 − 5. disturbance observers can be used to increase the R/W headpositioning accuracy in hard disk drives (HDDs) by using its estimation results to cancel the effect of the disturbances and modeling error.21) Assume ideally that Pn = P and dm = dn .02165z 2 + 0. and TEQ−OL (z) = Pequ C = PC . the disturbance observer without using additional sensors is frequently used to enhance the tracking performance of a hard disk drive servo system.02399z 5 + 0.220 Modeling and Control of Vibration in Mechanical Systems An equivalent open loop is denoted by TEQ−OL (z).23) The controller C(z) is the combination of a PID controller and two notch ﬁlters.333 × 10−5 z 6 − 7.846 × 10−5 z 3 +1. which implies that the proposed general disturbance observer does not inﬂuence the characteristics of openloop P (z)C(z) greatly and thus the stability and performance achieved by the nominal control loop with controller C(z) are maintained.25) (11.872 × 10−5 z + 9. Further.053 × 10−6 C(z) = .416 × 10−5 z 5 + 6.015 × 10−5 z 2 − 2. Hence. Equation (11. This is similar to the conventional disturbance observer.1728z + 0.052z 5 + 4.22) A(z) = z 6 − 3.399z + 0. and given by 1.
626 × 106 ) . which agree with the simulation results in Figure 11.9654)(z 2 + 0. Wd2 and Wn are selected as 1 and 1. where D1 (s) = 0.14) and its frequency response is shown in Figure 11.4. and wi (i = 1. such as bandwidth.60) and (2. the attenuation to d2 and n will be degraded.0004(s2 − 83. from which it can be seen that the designed disturbance observer is able to suppress disturbance with frequency lower than 1 kHz without causing much degradation for rejection of higher frequency disturbance. we assume that the disturbances d1 and d2 and the noise n are generated by d1 = D1 (s)w 1 .9501)(z + 0. The servo performance. the designed disturbance observer will result in a desired reduction rate of the error. The general disturbance observer is more effective to compensate for the input disturbance d1 than d2 and n. which is similar to the one from the general disturbance observer.3.4. This beneﬁt is of great signiﬁcance.3. 50% reduction is achieved. the plant model inverse needs to be calculated. TEQ−OL is measured with the . A stable Q(z) is then obtained via the H∞ optimization in (11.61).1836z + 0. Wd2 = 1 and Wn = 1.H∞ Based Design for Disturbance Observer 221 Wd1 = 0. However.5. the conventional disturbance observer is designed for comparison. It follows the original d1 approximately. will change with different weightings Wd1 . To demonstrate the effectiveness of the disturbance estimation in the time domain.7. Thus by adjusting the weightings according to the weights of d1 . Note that the plant model inverse is not required in the general disturbance observer. With lower Wd2 and Wn and higher Wd1 .5 shows the resultant sensitivity function.011 × 106 ) D2 (s) and Nn (s) are in (2. although more suppression to d1 will be attained by using the disturbance observer.983z + 0.39s + 9.8. −1 Pn (z) 2 (11.5 is to keep the attenuation to d2 and n achieved by the nominal feedback controller C(z). the error signal is shown in Figure 11. d2 = D2 (s)w2 and n = Nn (s)w3 .7s + 3. d2 .116)(z 2 − 1. ˆ With the designed general disturbance observer.1259)(z + 0. = z(z + 0.22z + 0.9834)(z 2 − 1. 3) are independent white noises with unity variance.741 × 105 )(s2 + 1616s + 9.6.253z + 0.2494(z − 1. (s2 + 125. especially for nonminimum phase plant. Moreover. and noise n in the position error signal. The sensitivity function S(z) is plotted in Figure 11. Experiment has been done with a LDV and a dSpace 1103. Figure 11. To evaluate the effect of the disturbance observer on the stability and performance achieved by the nominal controller C(z). M (z) and N (z) are given by M (z) = z −1 . 2.8179) . the estimate d1 of d1 is shown in Figure 11. The measured sensitivity functions are shown in Figure 11.05s + 1.948 × 105 )(s2 + 10. As a result.9646) (11.26) N (z) = 5.27) A stable Q(z) for the conventional disturbance observer is designed with the H∞ control method. Wd2 and Wn .
As expected.10. especially for nonminimum phase plant.222 Modeling and Control of Vibration in Mechanical Systems general disturbance observer and the conventional disturbance observer. and shown in Figure 11. and will not sacriﬁce the stability and performance of the nominal feedback control loop. and thus its design is simpliﬁed and has great advantages over the conventional disturbance observer.9 and Figure 11. . performance measures such as gain margin and phase margin are not affected. 11.5 Conclusion A general form of disturbance observer has been presented and designed based on the H∞ control method to achieve desired disturbance/noise rejection. The simulation and implementation results show that the general disturbance observer designed using the method employed in this chapter is able to effectively improve the attenuation of disturbance in low frequency.3 Frequency response of the designed Q(z). The disturbance observer does not need to solve the plant model inverse. Bode Diagram 40 35 30 25 Magnitude (dB) Phase (deg) 20 15 10 5 0 −5 −10 315 270 225 180 135 90 45 10 1 10 2 10 Frequency (Hz) 3 10 4 FIGURE 11.
10 0 −10 Magnitude(dB) −20 −30 −40 Conventional DOB General DOB −50 1 10 10 2 10 Frequency(Hz) 3 10 4 FIGURE 11.H∞ Based Design for Disturbance Observer 10 223 0 −10 Magnitude(dB) −20 −30 −40 No DOB With DOB −50 10 2 10 Frequency(Hz) 3 10 4 FIGURE 11.5 The sensitivity function comparison with the general and the conventional disturbance observers. .4 The sensitivity functions without and with the general disturbance observer.
7 Error signal e.2 0.03 No DOB With DOB 0.6 Disturbance d1 .03 0 0. .25 0.02 −0.01 0 −0.05 0.224 Modeling and Control of Vibration in Mechanical Systems 8 Amplitude of disturbance d1 and its estimate (µm) 6 4 2 0 −2 −4 −6 −8 x 10 −3 d1 estimate of d1 0 0.1 0.15 Time(sec) 0. 0.05 0.3 FIGURE 11.15 Time(sec) 0.01 −0.3 FIGURE 11.25 0.2 0.02 Amplitude of error e (µm) 0.1 0.
H∞ Based Design for Disturbance Observer 225 10 0 −10 Magnitude(dB) −20 −30 −40 No DOB With general DOB −50 10 2 10 Frequency(Hz) 3 10 4 FIGURE 11. .9 Comparison of TEQ−OL about the general disturbance observer. 60 Magnitude(dB) 40 20 0 −20 −40 10 0 −100 Phase(deg) −200 −300 −400 −500 1 10 10 2 1 Nominal with the general DOB 10 2 10 3 10 4 10 Freuqency(Hz) 3 10 4 FIGURE 11.8 Measured sensitivity functions without and with the general disturbance observer.
.226 Modeling and Control of Vibration in Mechanical Systems 60 Magnitude(dB) 40 20 0 −20 −40 10 0 −100 Phase(deg) −200 −300 −400 −500 1 10 10 2 1 Nominal with the conventional DOB 10 2 10 3 10 4 10 Freuqency(Hz) 3 10 4 FIGURE 11.10 Comparison of TEQ−OL about conventional disturbance observer.
Recently. track shape errors may be ampliﬁed from one track to the next through the closedloop response when writing the propagation tracks. the H2 control problem has been studied for 2D systems and a sufﬁcient condition for the evaluation of 2D system H2 performance in terms of LMIs is derived [139]. Instead of the conventional feedforward control. Selfservo track writer (SSTW) for data storage devices is one of these systems [17]. The developed 2D H2 control design method is of great importance to those systems that have 2D behavior and can be modeled using 2D linear system models. each step in the process carries a “memory” of all preceding track shape errors. Thus selfservo writing systems must provide a means of accurately writing servopatterns while controlling the propagation of track shape errors. Then the error propagation containment problem of the SSTW process is formulated as a 2D stabilization problem. Furthermore. In this chapter we describe the SSTW process with a twodimensional (2D) model. This “memory” depends on the particular closedloop response of the servo loop. Using the condition. so that the new bursts are written at locations reﬂecting the errors present in the preceding track via the closedloop response of the servo loop. The target of preventing error propagation is to reject the track shape error due to track noncircularity recorded in propagated tracks so that the circular concentric tracks are achieved in every propagation trace. the 2D stabilizing control is able to prevent the error propagation. equivalently. Its objective is to minimize the error energy of the system when the system is subject to a unit impulse input or. due to vibrations and noise the servo controller causes the actuator to follow the resulting noncircular trajectory in the next burst writing step. Because of the interdependency of propagation tracks. a white noise input of unit variance. Because of this analytically and practically meaningful speciﬁcation. In the selfservo track writing. as well as in the present track. the 227 . a systematic method for the design of the H2 controller for 2D systems in terms of LMIs has been developed. error propagation containment is critically important. the H2 problem and solution has been well studied and applied for several decades. Therefore.1 Introduction The H2 optimal control for 1D systems is a classical problem in linear systems theory.12 TwoDimensional H2 Control for Error Minimization 12. Consequently.
Let x(i. A2 = A. 12. j + 1) + B22 u(i + 1. xv (i. j). j) ∈ Rl are. j) + D12 u(i. j + 1) + Bc2 e(i + 1. j) xh (i. j) = C2 xh (i. j) + D22 u(i. the disturbance input. j).9) (12. u(i. xv (i. 2.5) (12. u(i. j + 1) + B12 w(i + 1. 0 0 0 In2 (12. j). j) + D11 w(i.7) Introduce the following 2D output feedback controller C(z1 . j) = C1 x(i.2) (12. e(i. k = 1. j) + D22 u(i.3) where xh ∈ Rn1 .1) (12. xv (i. j) where A1 = Bk1 = In1 0 0 0 A. j). the horizontal state. j + 1) + A2 x(i + 1. j) = Ccxc (i. and the measurement of the plant. y(i. j) +B11 w(i. 0 0 0 In2 In1 0 0 0 Bk . j).8) xh (i. y(i. the vertical state. j + 1) = Ac1 xc(i. j) y(i. j) = the above system is equivalent to x(i + 1. j).10) . j) ∈ Rq . the controlled output. respectively. j) . j) + D12 u(i. (12. j) ∈ Rm . z2 ): xc(i + 1. j) We consider the following 2D system model [138]: (12. j) ∈ Rp and e(i. j) +Bc1 e(i. j). the control input. Bk2 = Bk . j) (12. w(i. j) + Dc e(i. j + 1) + Ac2 xc (i + 1. A 2D H2 controller is designed which is able to prevent the error propagation and minimize the TMR. xv ∈ Rn2 . j + 1) = A1 x(i. j) xh (i. j) + D11 w(i.2 2D stabilization control xh (i + 1.228 Modeling and Control of Vibration in Mechanical Systems TMR minimization problem of the SSTW process is formulated as a 2D H2 control problem. j) = C2 x(i. j) +B21 u(i. j + 1) x (i. j) + D21 w(i.6) (12. j) =A v + B1 w(i. j) = C1 e(i. j) + B2 u(i. xv (i.4) (12.
A1 R + B21 Z (12.13) A2 ˜ A1 ˜ A2 ˜ Ω Ω −ΩF where S I SA1 + V1 C ˜ ˜ ΩF = . (12. The H2 norm of Tyw is approximately given by Tyw = 1 1 LK−1 L.5)−(12. 2) and Z such that the following LMI holds: ˜ ˜ −ΩX 0 ΩT A1 ˜ ˜ ˜ 0 −(ΩF − ΩX ) ΩT < 0 (12. 2. Uk .K−1 2 y(i. (12. Vk (k = 1.1)−(12. Bck = Λ−1 (Vk − SB2k Dc ). Vk = SB2k Dc + ΛBck .3 2D H2 control Let Tyw : w → y denote the closedloop system subject to the white noise w. there exists a full order output feedback controller of the form in (12.7) if there exist matrices R > 0.7). A2 R + B22 Z U1 .17) .1 [136] Consider the 2D system (12.10) can be given by Ack = Λ Cc = (Z − Dc CR)Ξ−T .9)−(12.9)−(12.16) 12. the 2D stabilization problem is stated as follows: for the 2D system (12. ˜ ΩX > 0.14) In this situation.TwoDimensional H2 Control for Error Minimization 229 Associated with the 2D controller (12. j)2 i. design a dynamic output feedback controller of the form in (12.5)−(12.9)−(12. −1 T (12. k = 1. 2.3) or (12. the controller parameters of (12. S > 0. j) = 0.9)−(12. Deﬁne Z = Dc CR + Cc ΞT . ΩA1 = I R A1 + B21 Dc C ˜ ΩA2 = SA2 + V2 C A2 + B22 Dc C U2 .5)−(12. j=1 (12. Dc . S > 0 and Ξ and Λ are invertible matrices satisfying ΞΛT = I − RS. THEOREM 12.10) that asymptotically stabilizes the system (12.7) with w(i.11) Uk = S(Ak + B2k Dc C)R + SB2k Cc ΞT + ΛBck CR + ΛAck ΞT .10) such that the resulting closedloop SSTW servo system is asymptotically stable.10).12) where R > 0. k = 1.15) −T [Uk − S(Ak + B2k Dc C2 )R − SB2k Cc Ξ − ΛBck C2 R]Ξ . (12. Then.
V v } satisfy XY T + U V T = M . The control design problem to minimize the 2D H2 norm is stated as follows: ﬁnd a 2D output feedback controller of the form in (12.20) are satisﬁed. THEOREM 12. D11 + D12 Dc D21 0 0 I T race(S) < λ2 . j).23) Cc = (Λ − Dc C2 Y T )V −T . Bc = U −1 (Γ − XB2 Dc ).5)−(12.18) S ∗ ∗ ∗ XB1 + ΓD21 X + X T − H11 ∗ ∗ (12. xh ∈ Rnh . H12}. (12.24) where U = diag{U h . H22 = diag{H22 .22) . of appropriate dimensions. j) c = Ac c + Bc e(i. v h v h v H11 } > 0.1)−(12. (12.3) such that the closedloop system is stable and the error is minimized.19) T B1 + B2 Dc D21 I + M T − H12 Y + Y T − H22 ∗ > 0. H22} > 0. j) xh (i. j) c u(i.1)−(12. An LMI approach will be given as follows to design a 2D H2 controller for the 2D system (12.1)−(12. j + 1) xv (i. j). If the above stated conditions are satisﬁed.7) such that the closedloop system is stable and the H2 performance Tyw 2 is minimized. a feasible H2 controller is given by xh (i + 1.230 Modeling and Control of Vibration in Mechanical Systems where L and K are large enough.9)−(12.1 The solution of the 2D controller in the above theorem is in terms of LMIs which can be eﬃciently solved by convex optimization just xh (i. j) = Cc with Ac = U −1 [Θ − X(A + B2 Dc C2 )Y T − XB2 CcV T − U Bc C2 Y T ]V −T .10) for the 2D system (12. T A + B2 Dc C2 AY T + B2 Λ I + M T − H12 Y + Y T − H22 ∗ C1 + D12 Dc C2 C1 Y T + D12 Λ 0 0 I (12. U v } and V = diag{V h . REMARK 12. j) c (12. xv ∈ Rnv . Γ.21) v c c xc (i. M = diag{M h .3) or (12. (12. M v }. Θ. j) c + Dc e(i. Dc and blockdiagonal matrices X = h diag{X h .3) is solvable if there exist matrices S > 0.(12. xv (i. Λ. Y = diag{Y h .2 [139] The 2D H2 control problem for the plant (12. such that H11 ∗ ∗ ∗ ∗ T H12 H22 ∗ ∗ ∗ XA + ΓC2 Θ X + X T − H11 ∗ ∗ > 0. and H12 = diag{H12. Y v }. H11 = diag{H11. X v }.
it is proved that if there exist solutions X. H22 .18)(12. M .1 and is generally known to involve the following distinct steps [135]: writing some tracks or at least one track called seed tracks. In [139]. Bc can be carried out by solving (12.23) and (12.20). and H12 for the LMIs (12.TwoDimensional H2 Control for Error Minimization 231 like other 1D control problems involving LMIs. S > 0. Θ. H11.24). Dc . writing servopattern for . M − XY T is invertible. The process of selfservo writing is shown in Figure 12. Γ. nonsingular matrices U and V can be computed. the computation of the controller parameters Cc .1 SSTW process. From U V T = M − XY T . Λ. Then. Y . Ac . reader reads back the seed track and writer writes actual product servopattern for the next track based on the readback signal.4 SSTW process and modeling FIGURE 12. 12.
232 Modeling and Control of Vibration in Mechanical Systems the next track based on the readback signal from the previous written track till the whole process is completed.2 shows the SSTW servo loop with disturbances and noises. D1 (s). d2 and noise n. each step in the process carries a “memory” of all preceding track shape errors. as well as in the present step. The radial error in track N is inevitably compounded to the following tracks: tracks N + 1. y(k − K) represents the track proﬁle of the previous track. · · ·.. In selfservo track writing. track shape errors such as noncircularity are introduced by mechanical disturbances. Selfservo writing systems must provide a means of accurately writing servopatterns while controlling the propagation of track shape errors. Consequently. Figure 12. Similarly. This leads to error propagation in SSTW. and N (s) are respectively the models of disturbances d1 . Let Tp be the rotational period of the disk. During the process. SSTW generates radial information progressively to deploy servopattern.1 SSTW servo loop FIGURE 12.2 SSTW servo loop with disturbances and noise models.e. one revolution of y(k) becomes the reference of the next written track . spindle motor vibration and other factors when writing the propagation tracks. This “memory” depends on the particular closedloop response of the servo loop.4. so that the new bursts are written at locations reﬂecting the errors present in the preceding step via the closedloop response of the servo loop. D2 (s). P ES(k −K) represents the position error when writing the previous track. Then the sector number K = Tp /Ts . 12. i. and Ts be the sampling rate of the position error signal. The read head follows on the track y(k − K) which is the reference input for the SSTW servo system. The servo controller causes the actuator to follow the resulting noncircular trajectory in the next burst writing step. N + 2. y(k) is the position of the write head with respect to a perfectly circular track on the disk and P ES(k) is the position error signal.
these frequency components in disturbances or noise will be ampliﬁed during propagation. while others will compound to following tracks and decay gradually. (The time sequence and the σ value increase with the track number.04 0.02 PES and σ of PES 0 −0. The typical closedloop transfer function will amplify the error at the frequency where its magnitude is more than 0 dB. T (z) = 1 − S(z) is the closedloop transfer function.3 shows PES NRRO in the time domain and its σ value versus track number when a 1D feedback controller C(z) is used in the closed loop.06 Position error signal Sigma value 0. Hence the error propagation problem must be addressed in the servo control design for SSTW. Figure 12. With the same disturbance models D1 (s) and D2 (s) and noise model N (s) as in Chapter 11. 0.3 PES NRRO and its σ values versus track number during propagation.TwoDimensional H2 Control for Error Minimization or a few subsequent tracks due to the action of selfservo writing.2 Twodimensional model For the 2D modeling. the following notations are used: .04 −0. (12.25) where S(z) = 1/(1 + P (z)C(z)) is the sensitivity function.) 12. In other words. The error propagation mentioned previously is clearly observed.4. It follows from Figure 12.02 −0.2 that 233 y(k) = T (z)y(k − K) + P (z)S(z)d1 (k) + S(z)d2 (k) + T (z)n(k).06 0 1 2 3 4 5 Track number 6 7 8 9 10 FIGURE 12.
y(i − 1. (12. j · Ts : time with j = 0. j) + Bp (u(i. j) is the position of the write head at track i in the radial dimension and time j · Ts in the axial dimension and P ES(i. y(i. 1. Similarly. j) represents the position error of the (i − 1)th track. xv (i. we have xp (i. i = 0. j) =A v + B1 w(i. xd2 and xn the corresponding state vectors of P (z). Let xh (i + 1. j) = Cp xp (i. j). j).27) (12.30) (12.e. xh ∈ Rnh . 1. j + 1) x (i. j + 1) = Ap xp (i. P ES(i. j). j) becomes the reference of the next written track due to the action of selfservo track writing. j) the position error. Based on Figure 12.26) (12. e(i.29) it follows from (12. j) (i. xv ∈ Rnv . j)T ]. The read head follows the track y(i − 1. j + 1)T ]. j) + B2 u(i.4. j) + d1 (i. L: the ith track. j).234 Modeling and Control of Vibration in Mechanical Systems FIGURE 12. w(i. j) xh (i. j). j). j) + n(i. j)T = [w1 (i. As shown in Figure 12. j) + d1 (i. · · · . j)T w2 (i. 2. j) (12.28) Denote xp .28) that xh (i + 1. v x (i. j) + Dp (u(i. j) + D12 u(i. j) which is the reference input for the SSTW servo system. j) = y(i. · · ·. j) = P ES(i. i. j) y(i. xd1 . j)). j) + D11 w(i. D1 (z). P ES(i − 1. one revolution of y(i. xv (i. j + 1)T xd1 (i. y(i.e. j + 1)T xd2 (i. K − 1. j) = y(i − 1.. D2 (z) and N (z).26)−(12.4. j)T w3 (i. j) = C1 xh (i. j + 1)T xn (i.31) . y(j − K)) represents the track proﬁle of the (i − 1)th track. respectively. 2.. (12. j + 1)T = [xp (i. j)) + d2 (i.4 SSTW servo loop modeling in two dimensions. j) − y(i.
j).36) (12. The target of error propagation containment is thus to reject the writtenin error due to track noncircularity recorded in propagated tracks so that the circular concentric . j) + D22 u(i. it