Advances in Industrial Control

Qingsong Xu
Kok Kiong Tan

Advanced Control
of Piezoelectric
Micro-/Nano-
Positioning
Systems
Advances in Industrial Control

Series editors
Michael J. Grimble, Glasgow, UK
Michael A. Johnson, Kidlington, UK
More information about this series at http://www.springer.com/series/1412
Qingsong Xu Kok Kiong Tan
•

Advanced Control
of Piezoelectric
Micro-/Nano-Positioning
Systems

123
Qingsong Xu Kok Kiong Tan
Department of Electromechanical Department of Electrical and Computer
Engineering Engineering
University of Macau National University of Singapore
Macau Singapore
China Singapore

ISSN 1430-9491 ISSN 2193-1577 (electronic)

Advances in Industrial Control
ISBN 978-3-319-21622-5 ISBN 978-3-319-21623-2 (eBook)
DOI 10.1007/978-3-319-21623-2

Library of Congress Control Number: 2015944753

Mathematics Subject Classification: 93, 93B12, 93B30, 93B52, 93C10, 93C40, 93C55, 93C57, 93C62,
93C73, 93C83, 93C85, 93C99, 93D05, 93D21, 68T40, 70Q05

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2016
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.

MATLAB® and Simulink® are registered trademarks of The MathWorks, Inc., 3 Apple Hill Drive,
Natick, MA 01760-2098, USA, http://www.mathworks.com.
LabVIEW™ is a trademark of National Instruments Corporation, 11500 N Mopac Expwy, Austin, TX
78759-3504, USA, http://www.ni.com/.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
To Prof. Wei Zhao and my family for the
encouragement all the time

Qingsong Xu

To my family

Kok Kiong Tan

Series Editors’ Foreword

The series Advances in Industrial Control aims to report and encourage technology
transfer in control engineering. The rapid development of control technology has an
impact on all areas of the control discipline. New theory, new controllers, actuators,
sensors, new industrial processes, computer methods, new applications, new phi-
losophies and, new challenges. Much of this development work resides in industrial
reports, feasibility study papers, and the reports of advanced collaborative projects.
The series offers an opportunity for researchers to present an extended exposition of
such new work in all aspects of industrial control for wider and rapid dissemination.
This monograph reports the recent work of Qingsong Xu and Kok Kiong Tan in
the field of modelling, design, and control of piezoelectric systems. Qingsong Xu is
an Associate Professor in the Department of Electromechanical Engineering at the
University of Macau, China. His background is in mechatronics and electrome-
chanical systems. He has been well supported by his university to develop consid-
erable expertise in this field and since 2005 has published regularly in peer-refereed
journals on his electromechanical systems work. Author Kok Kiong Tan is a
well-established contributor to the Advances in Industrial Control monograph series.
He is an Associate Professor in the Department of Electrical and Computer
Engineering at the National University of Singapore and it is possible to appreciate
his extensive expertise and research activities through his monograph contributions
to the series:
• Advances in PID Control, Kok Kiong Tan, Qing-Guo Wang and Chang Chieh
Hang with Tore J. Hägglund (ISBN 978-1-85233-138-2, 1999);
• Precision Motion Control (2nd Edition), Kok Kiong Tan, Tong Heng Lee and
Sunan Huang (ISBN 978-1-84800-020-9, 2008);
• Applied Predictive Control, Sunan Huang, Kok Kiong Tan and Tong Heng Lee
(ISBN 978-1-85233-338-6, 2002); and
• Drives and Control for Industrial Automation, Kok Kiong Tan and Andi
Sudjana Putra (ISBN 978-1-84882-424-9, 2011).

vii
viii Series Editors’ Foreword

There are clear technical links between these monographs and the work reported
in Advanced Control of Piezoelectric Micro-/Nano-positioning Systems. This is a
work of 11 chapters and after an introductory chapter is partitioned into four parts:
• Part I—Chapters 2 and 3: these two chapters cover hysteresis modelling and
look at feedforward controller design, with and without recourse to a hysteresis
model;
• Part II—Chapters 4 and 5: the control approaches followed by these two
chapters use a fusion of model predictive and discrete sliding-mode control;
• Part III—Chapters 6–9: each chapter in this part investigates variants of
sliding-mode control and concludes with a model-reference adaptive control
approach; and
• Part IV—Chapters 10 and 11: the last part of the monograph is devoted to
micromanipulation and mini-gripper applications.
A full and useful description of the contents of individual chapters can be found
in the Preface. There, the reader will find there are some special features of the
monograph that make it a distinctive contribution to the piezoelectric control lit-
erature. Apart from the wide range of control approaches developed, all the chapters
after the introductory chapter contain experimental verification studies and results
sections. The range of control approaches, the inclusion of experimental results, and
the applications studied makes the monograph an important addition to the litera-
ture and to the Advances in Industrial Control series.
Recently, piezoelectric control studies have received some exposure in the
Advances in Industrial Control series and readers will find complementary material
to the work of Qingsong Xu and Kok Kiong Tan in the series monograph Design,
Modelling and Control of Nanopositioning Systems by Andrew J. Fleming and
Kam K. Leang (ISBN 978-3-319-06616-5, 2014).
Related studies of the modelling and control applications of “smart materials”
can be found in a very novel application described in the forthcoming monograph
by R. Ganguli, D. Thakkar, and S.R. Viswamurthy entitled Smart Helicopter
Rotors: Optimization and Piezoelectric Vibration Control that will appear in the
series in the near future.

Industrial Control Centre, Glasgow, Scotland, UK M.J. Grimble

M.A. Johnson
Preface

Micro-/nano-positioning systems are positioning devices that are capable of

producing displacement down to micro-/nano-meter resolution and accuracy. As
typical ultrahigh-precision positioning devices, piezoelectric micro-/nano-
positioning systems are driven by piezoelectric actuators (PEAs). They have
found extensive applications in fields involving scanning probe microscopy,
lithography manufacturing, and wafer alignment. In addition, these devices have
also been applied in precision micro-/nano-manipulation and assembly domains.
Piezoelectric micro-/nano-positioning systems usually exhibit low damping,
which causes the structural vibration problem. In addition, PEA introduces non-
linearity into the system in terms of piezoelectric hysteresis and drift effects. The
hysteresis is a nonlinear relationship between the applied voltage and output dis-
placement of the PEA and induces a severe open-loop positioning error as high as
10–15 % of the travel range. Thus, the hysteresis has to be suppressed in high
precision applications. Moreover, the nonlinearity poses a great challenge in
combined position and force control of piezoelectric micro-/nano-positioining
devices dedicated to micromanipulation tasks.
This book is focused on enabling technologies in the control implementation of
piezoelectric micro-/nano-positioning systems. The book provides a collation of the
state of the art of emerging techniques to precision motion control of micro-/
nano-positioning systems actuated by PEAs. It covers both feedforward and feed-
back control strategies for positioning and tracking control of piezoelectric actua-
tion micro-/nano-positioning systems. This book also demonstrates the joint
position and force control of piezoelectric micro-/nano-positioning systems in
micromanipulation applications. A comprehensive treatment of the subject matter is
provided in a manner amenable to readers ranging from researchers to engineers, by
providing detailed experimental verifications of the developed approaches.
The book begins with an introduction to piezoelectric micro-/nano-positioning
system and provides a brief survey of its development and applications. According
to different realizations of the control strategies, the remaining ten chapters are
divided into four parts.

ix
x Preface

Part I consists of Chaps. 2 and 3, which addresses the hysteresis mode-based

feedforward control technology. Chapter 2 extends least squares support vector
machines (LSSVM) to the domain of rate-dependent piezoelectric hysteresis
modeling and compensation. A LSSVM-based rate-dependent hysteresis model is
proposed for a PEA by introducing the current input value and input variation rate
as one data set to construct a one-to-one mapping. The adoption of the input
variation rate allows the capture of the rate dependency of the hysteresis. For
comparative studies, the widely used Bouc–Wen and modified Prandtl–Ishlinskii
(MPI) hysteresis models are implemented. The hysteresis nonlinearity is suppressed
by a hybrid control which employs an LSSVM inverse model-based feedforward
controller combined with a proportional-integral-derivative (PID) feedback
controller.
Chapter 3 addresses the identification and compensation of the rate-dependent
piezoelectric hysteresis using an intelligent hysteresis model, while without mod-
eling the hysteresis inverse. Generally, both a hysteresis model and an inverse
hysteresis model are required for hysteresis identification and compensation pur-
poses. In this chapter, an LSSVM-based hysteresis model is established and a
feedforward compensator is developed based on a single model, which provides a
computationally efficient way in hysteresis compensation.
Chapters 4 and 5 construct Part II, which presents the hysteresis mode-free, state
observer-based control schemes. The merit of such schemes lies in that no hysteresis
model is required. The unmodeled hysteresis is considered as an uncertainty or a
disturbance to the nominal system, which is tolerated by an advanced robust or
adaptive controller. Chapter 4 reports on a model predictive discrete-time
sliding-mode control (MPDTSMC) to achieve the advantages of both model pre-
dictive control (MPC) and discrete-time sliding-mode control (DTSMC). It is shown
that the proposed MPDTSMC with proportional-integral (PI) action drives the
system state to slide in a vicinity of the sliding surface with a boundary layer of
thickness O(T3) (T is the sampling time), which is much lower than a commonly
designed DTSMC with O(T) boundary layer in the sliding mode. In addition, the
state tracking error of the order O(T2) is achieved with the presented control scheme.
The theoretical analysis and effectiveness of the PI action in the control scheme are
verified by experimental studies carried out on a nanopositioning platform.
Chapter 5 presents the design and implementation of a more desirable
output-based DTSMC strategy, called model predictive output integral
discrete-time sliding-mode control (MPOIDSMC), for micro-/nano-positioning
applications. The presented strategy is capable of improving the system perfor-
mance by integrating MPC, output integral discrete-time sliding-mode control
(OIDSMC), and state observer techniques together. A sliding-mode observer is
adopted to estimate the system state and to suppress the spillover effects. It is
demonstrated that the integrated scheme eliminates the chattering effect by forcing
the system trajectory to a vicinity of the sliding surface with a boundary layer of
thickness O(T3) in an optimal manner, and achieves an output tracking error of the
order O(T2). The theoretical analysis and effectiveness of the proposed strategy are
validated by experimental studies.
Preface xi

Part III includes Chaps. 6–9, which addresses the strategy of hysteresis
model-free, state observer-free digital feedback control. Generally, the DTSMC can
be categorized into state-based and output-based methods. The implementation of
either method usually requires the state feedback of the system. A state observer is
indispensable for the practical realization of DTSMC, which complicates the
control design procedure. In addition, an improperly designed state observer may
cause instability of the system. Hence, it is desirable to eliminate the use of state
observers. Chapter 6 proposes the design of a digital sliding-mode control (DSMC)
for a piezoelectric micro-/nano-positioning system with a simple second-order plant
model. The local stability of the closed-loop system is proved theoretically and the
effectiveness of the proposed scheme is validated through experimental
investigations.
Chapter 7 reports an input–output-based DSMC (IODSMC) algorithm for pre-
cision motion tracking of a class of piezoelectric micro-/nano-positioning systems,
which can be described by a high-order linear model preceded by disturbances. Its
implementation requires an input–output data-based model only, whereas neither a
hysteresis model nor a state observer is needed.
Chapter 8 presents a digital sliding-mode prediction control (DSMPC) scheme
for precision motion tracking of a class of piezoelectric micro-/nano-positioning
systems, which can be represented by a high-order linear model preceded by dis-
turbances. Its implementation does not require the knowledge of system states. The
overall control action of the proposed DSMPC scheme involves equivalent control,
switching control, and predictive control.
Chapter 9 devises a model-reference adaptive control (MRAC) scheme to
compensate for the unmodeled hysteresis effect of a class of PEA-actuated systems
which possess a second-order nominal model. By treating the uncertainties as a
lumped perturbation to the nominal system, a scheme of MRAC with perturbation
estimation (MRACPE) is developed and validated on a micropositioning system.
Compared with the existing work, the presented scheme allows the predesign of the
maximum tracking error. It is capable of estimating the unmodeled perturbation
of the system.
Part IV is composed of Chaps. 10 and 11, which deal with the technique of
position and force joint control for micromanipulation applications. As a typical
micro-/nano-positioning device, microgripper is a crucial tool to realize the grasp–
hold–release operation in micromanipulation tasks. To avoid damaging the grasped
fragile microobjects and the microgripper itself, an interaction control is critical to
regulate the desired position and contact force simultaneously. Chapter 10 presents
an adaptive discrete-time sliding-mode generalized impedance control (ADSMGIC)
to realize an interaction control of a piezoelectric microgripper dedicated to
micromanipulation and microassembly. The control scheme regulates a desired
dynamics relation between the position and contact force. The chattering phe-
nomenon is suppressed by employing an adaptive law for the switching gain. The
stability of the closed-loop system is proved theoretically and the effectiveness
of the interaction control scheme is validated by conducting experimental studies on
a piezo-bimorph microgripper system.
xii Preface

Chapter 11 reports on a scheme of position/force switching control to adjust the

gripper tip position and gripping force in an alternate manner. An incremental
control framework is developed to achieve a smooth transition between the position
and force controls for the gripper system. Specifically, in the closing and opening
phases, an incremental digital sliding-mode control is devised to cater for the
position control. During the contact phase, an incremental-type digital PID force
control is adopted. The proposed control algorithm is deployed to a
field-programmable gate array (FPGA)-based digital control platform. The feasi-
bility of the scheme is examined by a typical grasp–hold–release operation of a
micro copper wire through experimental studies. Its superiority over the conven-
tional approach is demonstrated through experimental comparisons.
This book provides the state-of-the-art coverage of the methodology and algo-
rithms of precision motion control of piezoelectric micro-/nano-positioning systems
in the context of control engineering and soft computing. The control approaches
involve feedforward control based on intelligent hysteresis model, discrete-time
sliding-mode control, model predictive control, and model-reference adaptive
control. The book also touches on the combined position and force control in
micromanipulation, which is a typical application of micro-/nano-positioning sys-
tems. Detailed examples of their implementations are provided. Readers can expect
to learn how to design and apply new control approaches to precision motion
control of piezoelectric micro-/nano-positioning systems.
The authors would like to acknowledge the University of Macau (under Grants
SRG006-FST11-XQS, MYRG083(Y1-L2)-FST12-XQS, and MYRG078(Y1-L2)-
FST13-XQS) and the Science and Technology Development Fund (FDCT) of
Macao (under Grants 024/2011/A, 070/2012/A3, and 052/2014/A1) for co-funding
the projects. The authors are also grateful for the help provided by Editor Oliver
Jackson and Editor Karin de Bie from Springer.

Macau, China Qingsong Xu

Singapore Kok Kiong Tan
June 2015
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Micro-/Nano-positioning Technique. . . . . . . . . . . . . . . . . . . . 1
1.2 Actuators and Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Piezoelectric Nonlinearity. . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Feedforward Control Based on Hysteresis Models. . . . . . . . . . 5
1.4.1 Conventional Hysteresis Model. . . . . . . . . . . . . . . . . 6
1.4.2 Intelligent Hysteresis Model . . . . . . . . . . . . . . . . . . . 8
1.4.3 Feedforward Plus Feedback Control . . . . . . . . . . . . . 8
1.5 Robust Feedback Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5.1 Sliding-Mode Control . . . . . . . . . . . . . . . . . . . . . . . 10
1.5.2 Model Predictive Control . . . . . . . . . . . . . . . . . . . . . 11
1.5.3 Model-Reference Adaptive Control . . . . . . . . . . . . . . 12
1.5.4 Other Control Strategies . . . . . . . . . . . . . . . . . . . . . 13
1.6 Position/Force Control in Micromanipulation . . . . . . . . . . . . . 13
1.6.1 Hybrid Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.6.2 Impedance Control . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.6.3 Switching Control . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.7 Book Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

Part I Hysteresis-Model-Based Feedforward Control

2 Feedforward Control Based on Inverse Hysteresis Models . . . . . . 23

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 System Description and Hysteresis Characterization . . . . . . . . 24
2.2.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.2 Hysteresis Characterization. . . . . . . . . . . . . . . . . . . . 25
2.3 Hysteresis Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.1 Hysteresis Modeling with the Bouc–Wen Model. . . . . 28

xiii
xiv Contents

2.3.2 Hysteresis Modeling with the MPI Model . . . . . . . . . 29

2.3.3 Hysteresis Modeling with the LSSVM. . . . . . . . . . . . 32
2.4 Experimental Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4.1 Bouc–Wen Model Results . . . . . . . . . . . . . . . . . . . . 36
2.4.2 MPI Model Results . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4.3 LSSVM Model Results . . . . . . . . . . . . . . . . . . . . . . 40
2.4.4 Model Capability Comparison . . . . . . . . . . . . . . . . . 42
2.4.5 Generalization Study . . . . . . . . . . . . . . . . . . . . . . . . 44
2.5 Controller Design and Verification . . . . . . . . . . . . . . . . . . . . 47
2.5.1 Feedforward Controller Design . . . . . . . . . . . . . . . . . 47
2.5.2 Feedforward Plus Feedback Controller Design . . . . . . 49
2.5.3 Controller Verification. . . . . . . . . . . . . . . . . . . . . . . 49
2.6 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3 Feedforward Control Without Modeling Inverse Hysteresis . . . . . 57

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.2 Dynamics Modeling of Hysteretic System . . . . . . . . . . . . . . . 58
3.2.1 Dynamics Modeling with Bouc–Wen Hysteresis . . . . . 58
3.2.2 Dynamics Modeling with Intelligent
Hysteresis Model . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3 Hysteresis Modeling Using LSSVM . . . . . . . . . . . . . . . . . . . 60
3.3.1 Regression Model Establishment. . . . . . . . . . . . . . . . 60
3.3.2 LSSVM Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.4 Experimental Studies on Hysteresis Identification . . . . . . . . . . 62
3.4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.4.2 Dynamics Model Identification . . . . . . . . . . . . . . . . . 63
3.4.3 Bouc–Wen Model Results . . . . . . . . . . . . . . . . . . . . 64
3.4.4 LSSVM Model Results . . . . . . . . . . . . . . . . . . . . . . 67
3.5 Experimental Studies on Hysteresis Compensation . . . . . . . . . 71
3.5.1 Feedforward Compensation . . . . . . . . . . . . . . . . . . . 71
3.5.2 Feedforward Plus Feedback Control . . . . . . . . . . . . . 73
3.6 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

Part II Hysteresis-Model-Free, State-Observer-Based

Feedback Control

4 Model Predictive Discrete-Time Sliding-Mode Control ...... . . . 79

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . ...... . . . 79
4.2 Problem Formulation. . . . . . . . . . . . . . . . . . . . . ...... . . . 81
4.2.1 Dynamics Modeling of a Nanopositioning System . . . 81
Contents xv

4.2.2 Sliding-Mode Controller Design . . . . . . . . . . . . . . . . 83

4.2.3 Control Gain Design . . . . . . . . . . . . . . . . . . . . . . . . 85
4.3 DTSMC Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.3.1 Controller Design and Analysis . . . . . . . . . . . . . . . . 86
4.3.2 Tracking Error Bound Analysis . . . . . . . . . . . . . . . . 88
4.4 MPDTSMC Design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.4.1 MPDTSMC Controller Design . . . . . . . . . . . . . . . . . 89
4.4.2 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.4.3 State Observer Design . . . . . . . . . . . . . . . . . . . . . . . 92
4.4.4 Tracking and Estimation Error Bound Analysis . . . . . 93
4.5 Experimental Investigation . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.5.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.5.2 Hysteresis Characterization. . . . . . . . . . . . . . . . . . . . 94
4.5.3 Plant Model Identification . . . . . . . . . . . . . . . . . . . . 95
4.5.4 Controller Parameter Design. . . . . . . . . . . . . . . . . . . 96
4.5.5 Simulation Studies . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.5.6 Experimental Testing Results . . . . . . . . . . . . . . . . . . 99
4.5.7 Discussion on System Performance . . . . . . . . . . . . . . 102
4.6 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

5 Model Predictive Output Integral Discrete-Time

Sliding-Mode Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2 Problem Formulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.3 MPOIDSMC Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.3.1 OIDSMC Controller Design . . . . . . . . . . . . . . . . . . . 107
5.3.2 MPOIDSMC Controller Design . . . . . . . . . . . . . . . . 109
5.4 Experimental Investigations . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.4.2 Plant Model Identification . . . . . . . . . . . . . . . . . . . . 115
5.4.3 Controller Parameter Design. . . . . . . . . . . . . . . . . . . 116
5.4.4 Experimental Studies. . . . . . . . . . . . . . . . . . . . . . . . 118
5.4.5 Discussion on Controller Performance . . . . . . . . . . . . 121
5.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

Part III Hysteresis-Model-Free, State-Observer-Free

Feedback Control

6 Digital Sliding-Mode Control of Second-Order Systems . . . . . . . . 127

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.2 Dynamics Model and Problem Formulation . . . . . . . . . . . . . . 128
xvi Contents

6.3 DSMC Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

6.4 Experimental Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.4.2 Plant Model Identification . . . . . . . . . . . . . . . . . . . . 134
6.4.3 Experimental Results. . . . . . . . . . . . . . . . . . . . . . . . 135
6.4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

7 Digital Sliding-Mode Control of High-Order Systems . . . . . . . . . . 147

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
7.2 Problem Formulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.2.1 System Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.2.2 Disturbance Estimation . . . . . . . . . . . . . . . . . . . . . . 150
7.3 IODSMC Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.3.1 Sliding Function Definition . . . . . . . . . . . . . . . . . . . 151
7.3.2 Design of IODSMC . . . . . . . . . . . . . . . . . . . . . . . . 152
7.3.3 Robust IODSMC Design . . . . . . . . . . . . . . . . . . . . . 154
7.4 Experimental Setup and Controller Setup . . . . . . . . . . . . . . . . 156
7.4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 156
7.4.2 Plant Model Identification . . . . . . . . . . . . . . . . . . . . 157
7.4.3 Controller Parameter Design. . . . . . . . . . . . . . . . . . . 158
7.5 Experimental Results and Discussion. . . . . . . . . . . . . . . . . . . 159
7.5.1 Set-Point Positioning Results . . . . . . . . . . . . . . . . . . 159
7.5.2 Sinusoidal Positioning Results . . . . . . . . . . . . . . . . . 159
7.5.3 Bandwidth Testing Results. . . . . . . . . . . . . . . . . . . . 161
7.5.4 Robustness Testing Results . . . . . . . . . . . . . . . . . . . 162
7.5.5 Further Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 164
7.6 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

8 Digital Sliding-Mode Prediction Control . . . . . . . . . . . . . . . . . . . 167

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
8.2 Problem Formulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
8.2.1 System Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 168
8.2.2 Disturbance Estimation . . . . . . . . . . . . . . . . . . . . . . 169
8.3 DSMC Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
8.3.1 Sliding Function Definition . . . . . . . . . . . . . . . . . . . 170
8.3.2 Design of DSMC . . . . . . . . . . . . . . . . . . . . . . . . . . 170
8.3.3 Error Bound Analysis . . . . . . . . . . . . . . . . . . . . . . . 172
8.4 DSMPC Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
8.4.1 DSMPC Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
8.4.2 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 175
8.4.3 Error Bound Analysis . . . . . . . . . . . . . . . . . . . . . . . 177
Contents xvii

8.5 Experimental Studies and Discussion. . . . . . . . . . . . . . . . . . . 177

8.5.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 177
8.5.2 Plant Model Identification . . . . . . . . . . . . . . . . . . . . 178
8.5.3 Controller Parameter Design. . . . . . . . . . . . . . . . . . . 179
8.5.4 Experimental Studies. . . . . . . . . . . . . . . . . . . . . . . . 179
8.5.5 Further Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 183
8.6 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

9 Model-Reference Adaptive Control with Perturbation

Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
9.2 Dynamics Modeling and Perturbation Estimation . . . . . . . . . . 188
9.2.1 Dynamics Modeling . . . . . . . . . . . . . . . . . . . . . . . . 188
9.2.2 Perturbation Estimation . . . . . . . . . . . . . . . . . . . . . . 189
9.3 MRACPE Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . 190
9.3.1 MRACPE Controller Design . . . . . . . . . . . . . . . . . . 190
9.3.2 Dead-Zone Modification of Adaptive Laws . . . . . . . . 193
9.3.3 Overview of Control Scheme . . . . . . . . . . . . . . . . . . 193
9.4 Experimental Setup and Controller Setup . . . . . . . . . . . . . . . . 194
9.4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 194
9.4.2 Statics Testing and Dynamics Model Identification . . . 196
9.4.3 Controller Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
9.5 Experimental Results and Discussion. . . . . . . . . . . . . . . . . . . 198
9.5.1 Set-Point Positioning Testing . . . . . . . . . . . . . . . . . . 198
9.5.2 Sinusoidal Positioning Testing . . . . . . . . . . . . . . . . . 200
9.5.3 Control Bandwidth Testing . . . . . . . . . . . . . . . . . . . 200
9.5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
9.6 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

Part IV Applications to Micromanipulation

10 Adaptive Impedance Control of Piezoelectric Microgripper. . . . . . 209

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
10.2 Problem Formulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
10.2.1 Dynamics Modeling and Perturbation Estimation . . . . 212
10.2.2 Impedance Control Problem . . . . . . . . . . . . . . . . . . . 213
10.3 ADSMGIC Scheme Design . . . . . . . . . . . . . . . . . . . . . . . . . 214
10.3.1 Sliding Function Definition . . . . . . . . . . . . . . . . . . . 214
10.3.2 ADSMGIC Design . . . . . . . . . . . . . . . . . . . . . . . . . 215
10.3.3 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 216
10.3.4 Evaluation of Steady-State Errors . . . . . . . . . . . . . . . 218
xviii Contents

10.4 Experimental Setup and Controller Setup . . . . . . . . . . . . . . . . 219

10.4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 219
10.4.2 Force Observer Design . . . . . . . . . . . . . . . . . . . . . . 220
10.4.3 Controller Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
10.5 Experimental Results and Discussion. . . . . . . . . . . . . . . . . . . 224
10.5.1 Interaction Control Results . . . . . . . . . . . . . . . . . . . . 224
10.5.2 Discussion on Control Performance. . . . . . . . . . . . . . 229
10.6 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230

11 Position/Force Switching Control of a Miniature Gripper . . . . . . . 233

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
11.2 Experimental Setup and Calibration. . . . . . . . . . . . . . . . . . . . 234
11.2.1 Working Principle of the Gripper . . . . . . . . . . . . . . . 234
11.2.2 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 236
11.2.3 Calibration of Position and Force Sensors . . . . . . . . . 237
11.2.4 Gripping Range Testing. . . . . . . . . . . . . . . . . . . . . . 239
11.2.5 Frequency Response Testing . . . . . . . . . . . . . . . . . . 240
11.3 Position/Force Switching Control Scheme Design . . . . . . . . . . 241
11.3.1 Event-Based Switching Control Framework . . . . . . . . 241
11.3.2 Incremental DSMC Position Controller . . . . . . . . . . . 242
11.3.3 Incremental PID Force Controller . . . . . . . . . . . . . . . 245
11.3.4 Switching Criterion . . . . . . . . . . . . . . . . . . . . . . . . . 245
11.4 Experimental Investigations and Discussion . . . . . . . . . . . . . . 246
11.4.1 Controller Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
11.4.2 Position/Force Switching Control Results. . . . . . . . . . 247
11.4.3 Further Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 249
11.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Abbreviations

ADSMGIC Adaptive digital sliding-mode generalized impedance control

ANN Artificial neural networks
DSMC Digital sliding-mode control
DSMPC Digital sliding-mode prediction control
DSP Digital signal processor
DTSMC Discrete-time sliding-mode control
EDM Electrical discharge machining
FEA Finite-element analysis
FPGA Field-programmable gate array
GA Genetic algorithm
GIC Generalized impedance control
GMVC Generalized minimum variance control
IODSMC Input–output-based digital sliding-mode control
LSSVM Least squares support vector machines
LTI Linear time-invariant
LVDT Linear variable differential transformer
MAE Mean absolute error
MAXE Maximum error
MPC Model predictive control
MPDTSMC Model predictive discrete-time sliding model control
MPI Modified Prandtl–Ishlinskii
MPOIDSMC Model predictive output integral discrete-time sliding-mode
control
MRAC Model-reference adaptive control
MRACPE Model-reference adaptive control with perturbation estimation
OIDSMC Output integral discrete-time sliding-mode control
PC Personal computer
PEA Piezoelectric actuator
PI Proportional-integral
PID Proportional-integral-derivative
PRB Pseudo-rigid-body

xix
xx Abbreviations

PSA Piezoelectric stack actuator

PSO Particle swarm optimization
PZT Pb(Zr, Ti)O3 crystal
QP Quadratic programming
QSD Quasi-sliding domain
RMS Root-mean-square
RMSE Root-mean-square error
SISO Single-input-and-single-output
SMC Sliding-mode control
SNR Signal-to-noise ratio
SVM Support vector machines
Z–N Ziegler–Nichols
ZOH Zero-order-hold
Chapter 1
Introduction

This chapter provides an introduction to the piezoelectric micro-/nano-positioning

system and the concerned control problems. The hysteresis compensation is solved
using both feedforward and feedback control approaches to achieve a precision
motion control. As a typical application of micro-/nano-positioning device, the
micromanipulation application is also introduced, and the concerned position and
force control problem is addressed to facilitate a delicate automated manipulation
operation.

1.1 Micro-/Nano-positioning Technique

Micro-/nano-positioning concerns precision motion control with the resolution down

to micro-/nanometer scale. Micro-/nano-positioning system refers to a position-
ing device which is capable of producing displacement down to micro-/nanometer
resolution. Such positioning devices have been widely applied in fields including
scanning probe microscopy, lithography manufacturing, and wafer alignment. In
addition, these devices have also found extensive applications in precision micro-
/nano-manipulation and micro-/nano-assembly domains. For instance, in biological
cell manipulation or micro-component assembly, micro-/nano-positioning systems
are employed to realize a precise positioning of the end-effector, such as micro pipette
or gripper, for successful execution of the manipulation or assembly task.
Conventionally, the lead-screw mechanisms are usually employed to achieve a
precision positioning motion. However, these traditional mechanisms suffer from
the adverse effects of backlash, clearance, friction, and wear, which deteriorate the
positioning accuracy of the system. Alternatively, flexure hinge-based-compliant
mechanisms can be used to overcome these shortcomings. Different from traditional
sliding and rolling mechanisms, compliant mechanisms deliver motions by mak-
ing use of the elastic deformation of the material. As a consequence, they allow
the generation of several attractive merits in terms of no backlash and clearance,
© Springer International Publishing Switzerland 2016 1
Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2_1
2 1 Introduction

Fig. 1.1 Examples of micro-/nano-positioning devices. a A compliant micromanipulator. b A

compliant microgripper

no friction and wear, no need for lubrication, smooth and continuous motion, inher-
ently infinite resolution, vacuum and cryogenic compatibility, and monolithic manu-
facturing. Hence, various modern micro-/nano-positioning stages have been devised
based on compliant mechanisms. For example, a flexure hinge-based compliant
micromanipulator and microgripper are shown in Fig. 1.1a and b, respectively.

1.2 Actuators and Sensors

Different kinds of actuators can be employed to drive the positioning stages, such
as piezoelectric actuators, magnetic levitation motors, electromagnetic actuators,
and voice coil motors. In particular, smart actuators based on smart materials (e.g.,
piezoelectric materials, shape memory alloys, magnetostrictive materials, etc.) are
popularly employed for the actuation in various precision systems dedicated to
micro-/nano-positioning. As a typical smart actuator, piezoelectric actuator (PEA)
prevails in many industrial applications thanks to its merits of sub-nanometer posi-
tioning resolution, rapid response speed, large blocking force, and compact size.
As a typical flexure-based ultrahigh-precision positioning device, piezoelec-
tric micro-/nano-positioning system is driven by PEAs. Popular PEAs involve the
forms of piezo stacks, piezo benders, and piezo diaphragms as shown in Fig. 1.2.
Owing to the aforementioned advantages, PEAs are popularly applied in various
1.2 Actuators and Sensors 3

Fig. 1.2 Examples of popular piezoelectric actuators. a Piezo stacks. b Piezo bender. c Piezo
diaphragms

micro-/nano-positioning systems such as scanning probe microscopes, precision

alignment devices, e-beam lithography systems, and so on.
To measure the output displacement of micro-/nano-positioning stages, sensors
based on different principles have been exploited. The popular types of displacement
sensors include capacitive sensor, optical sensor, strain sensor, and inductance sensor.
As reviewed below, these sensors own different properties and can be employed in
different scenarios in accordance with specific requirements.
Capacitive sensors measure the displacement by detecting the capacitance change
between two metal plates. Generally, they can provide the sensing resolution down
to sub-nanometer level in a small measuring range (typically, tens of micrometers).
Optical sensors are capable of high resolution measurement in a large range. Popular
optical sensors involve laser interferometers and laser displacement sensors, where
the latter are based on the triangulation measurement method. Although these sensors
can provide nanometer-level resolution, they are bulky and not suitable for compact
system application. In contrast, strain sensors possess a small size and they are
most suitable for embedded design of micro-/nano-positioning system. Strain sensors
involve metal strain gauges and piezoresistive strain gauges. While the former is
capable of a larger strain deformation, the latter is able to deliver a higher sensitivity.
Usually, the strain sensors exhibit considerable noises, which require extra signal
processing so as to achieve a high sensing resolution. Commonly used inductance
sensors involve linear variable differential transformers (LVDT) and eddy-current
sensors. These sensors are sensitive to the motion in the axes orthogonal to the
measurement axis. Hence, these inductance sensors are more suitable for uniaxial
motion measurement to achieve nanometer-level resolution.
In addition, different microscopes in terms of optical microscopes and scanning
electron microscopes, etc., have also been adopted as position sensors to detect the
output displacement of micro-/nano-positioning devices. Yet, the disadvantage lies
in the relatively low sampling frequency of universal image processing unit.
4 1 Introduction

(a) (b) (c)

Unpolarized Polarized After polarization
+
− − −

displacement
+ +

displacement
+

Remanent
Actuator

Fig. 1.3 Polarization process for a piezoelectric actuator

1.3 Piezoelectric Nonlinearity

The piezoelectric effect is a fundamental process involving electro-mechanical inter-

actions and represents the conversion of energy. It relates the electric field to the
mechanical compression/elongation in a piezoelectric material. This fundamental
property of piezoelectricity has therefore led to the utilization of such materials in
the fabrication of various piezoelectric transducers including actuators and sensors.
Generally, PEAs are fabricated by piezoelectric ceramic materials such as lead
zirconate-titanate Pb(Zr, Ti)O3 crystal, commonly called PZT. Figure 1.3 illustrates
the polarization process for a piezoelectric actuator. In Fig. 1.3a, the electric dipoles
of ceramic materials are oriented randomly. Without an applied electric field, it does
not exhibit the function of piezoelectricity. Once an electric field is applied, the
electric dipoles align themselves in a direction close to the applied electric field,
and the crystals expand in a direction close to the applied electric field, as shown
in Fig. 1.3b. This process is called polarization. However, when the electric field is
removed, the electric dipoles will not return to the original position exactly, which
results in remanent polarization, as shown in Fig. 1.3c.
PEAs produce mechanical movement by making use of the inverse piezoelectric
effect of piezoelectric materials, which causes the shape change under the input of
voltage or current. The movement with the resolution of micro-/nanometer level is
desirable. At the same time, PEA introduces nonlinearity into the system due to
the piezoelectric hysteresis and creep effects. The creep effect causes the drifts of
PEA output when the PEA is driven by low-frequency input. Hence, the creep is
a low-frequency behavior, which can be modeled, for instance, by a linear transfer
function.
The hysteresis nonlinearity has some special features. First, it exhibits multi-
valued mapping, as shown in Fig. 1.4. That is, given the same input signal, the outputs
are different if the history output states are not identical. In other words, it has the
memory of history state. In order to determine the output of the system, both the input
1.3 Piezoelectric Nonlinearity 5

20 20

Displacement (µm)
15 15

10 10

5 5

0 0
0 10 20 0 0.5 1
0 Time (s)
Time (s)

0.5

1
0 10 20
Voltage (V)

Fig. 1.4 Illustration of the piezoelectric hysteresis. Given the same input voltage, the output dis-
placements are different due to the hysteresis effect

signal and the history state of system should be considered. In addition, when the
direction of the input change rate changes, the output signal shows non-differentiable
characteristics. The hysteresis nonlinearity induces a severe open-loop positioning
error as high as 10–15 % of the stage travel range. Moreover, the output signal is also
dependent on the frequency of the input signal. When the rate of the input signal is
increased, the error increases accordingly.
Therefore, the hysteresis has to be suppressed to cater for the accuracy requirement
of high precision applications. In the past two decades, extensive works have been
carried out for the compensation of the hysteretic behaviors. Although the hysteresis
can be greatly alleviated if a PEA is driven by a charge source [13], it is at the cost
of a reduced stroke. Thus, voltage actuation has been widely adopted in practice.

1.4 Feedforward Control Based on Hysteresis Models

By taking the driving voltage and the produced position as the input and output,
respectively, a piezoelectric micro-/nano-positioning system can be described by
a block diagram as shown in Fig. 1.5a. As a nonlinear system, the piezoelectric
micro-/nano-positioning system can be represented by a Hammerstein cascade model
[32, 33], as shown in Fig. 1.5b. The nonlinear hysteresis model H (.) is followed by a
linear dynamics model G(s). The dynamics model involves the vibration dynamics
and creep effects. The hysteresis nonlinearity can be compensated for by differ-
ent ways as discussed in the subsequent sections. Generally, the existing hysteresis
6 1 Introduction

(a)
U Piezoelectric micro/nano- X
positioning stage

(b)

U W X
H( ) G(s)

Fig. 1.5 a Piezo-actuated stage with voltage input U and position output X . b Hammerstein cascade
model of a nonlinear hysteresis model H (.) and linear dynamics model G(s). The intermediate
variable W is usually unmeasurable

Xd X
H −1 ( ) H( ) G ( s)

Fig. 1.6 Block diagram of feedforward control to compensate the nonlinear hysteresis effect H (.)
using an inverse hysteresis model H −1 (.)

compensation schemes fall into two categories in terms of hysteresis model-based

feedforward control and hysteresis model-free feedback control.
To compensate for the hysteresis nonlinearity, an inverse hysteresis model H −1 (.)
can be developed and used as a feedforward control, as shown in Fig. 1.6. By this
approach, the hysteresis is cancelled and a linear dynamics system is obtained approx-
imately. In order to construct the feedforward controller, a hysteresis model is usually
identified and then inverted to generate an inverse hysteresis model. Hence, the estab-
lishment of a suitable hysteresis model is required.

1.4.1 Conventional Hysteresis Model

Traditionally, the modeling of hysteresis nonlinearity is a time-consuming problem,

and there is no universal model. In the literature, a number of models have been
developed for describing the hysteresis nonlinearity, which can be classified into two
main types: physics-based models and phenomenological models. A classification
diagram is depicted in Fig. 1.7.
Based on the fundamental physical principle of piezoelectric material,
physics-based models try to establish the hysteresis model by making use of the
relationships of energy, displacement and so on. Generally, it is difficult to build a
physics-based model because the physical feature of the hysteretic system is usu-
ally very complicated. In addition, the physics-based model of one hysteretic system
1.4 Feedforward Control Based on Hysteresis Models 7

Jiles-Atherton model
Physics-based model
etc.

Duhem model
Hysteresis Differential equation Backlash-like model
model model
Bouc–Wen model

Preisach model
Phenomenological Operator-based
model model Prandtl–Ishlinskii model
Krasnosel'skii-Pokrovskii model

Polynomial model
Other model
etc.

Fig. 1.7 Classification diagram of some popular hysteresis models

usually cannot be applied to another one. Hence, the physics-based model is lack of
generalization.
In contrast, starting from the characteristic of hysteresis curve, phenomenologi-
cal models try to describe the hysteresis effect using effective mathematical model
directly [9, 21, 26, 30, 35, 38, 71]. Based on different mathematical interpretations,
phenomenological models have been established without considering the physical
meanings. These models include differential equation-based and operator-based hys-
teresis models.
Normally, a hysteresis model is identified using the excitation input and output
data of the system. Afterwards, an inverse hysteresis model is constructed and utilized
as a feedforward controller to cancel the hysteresis effect. With a perfect cancellation
of the hysteresis effect, the resulted linear dynamics system is easy to control.
A majority of the aforementioned models deal with rate-independent hysteresis.
Actually, the hysteresis effect is dependent not only on the amplitude but also on the
frequency of input signals. Fortunately, the rate-dependent behavior can be described
by the cascade of a rate-independent hysteresis followed by a linear dynamics model,
i.e., a Hammerstein model as described in Fig. 1.5b. Hence, a two-step process is
needed to accurately identify the hysteresis model and dynamics model, respectively.
Alternatively, the entire system can be modeled together using a rate-dependent
hysteresis model. However, it is difficult to capture the complicated rate-dependent
hysteretic behavior precisely. In addition, a majority of the existing models employ
a large number of parameters to describe the rate-dependent hysteresis [4, 78]. As
an adverse effect, it may block their applications in high-speed real-time control.
8 1 Introduction

1.4.2 Intelligent Hysteresis Model

Recently, it has been shown that artificial intelligence algorithms pave a promis-
ing way to model the complicated hysteresis nonlinearity. In particular, the artificial
neural networks (ANN) provide an efficient approach to model the nonlinear hys-
teresis [15, 76]. Nevertheless, there is no universal method to determine an optimal
ANN structure in terms of the number of hidden layers and the number of neurons
in each layer. Moreover, ANN exhibit the problems of overfitting and sinking into
local optima, which are their major drawbacks in practical implementation.
Alternatively, support vector machines (SVM) are a potential way to estimate non-
linear system models accurately. Based on statistical learning theory and structural
risk minimization principle, the SVM approach is capable of modeling nonlinear sys-
tems by transforming the regression problem into a convex quadratic programming
(QP) problem and then solving it with a QP solver. Compared with conventional
ANN, SVM have the major advantages of global optimization and higher general-
ization capability.
As a variant of SVM, least squares support vector machines (LSSVM) have been
introduced in 1999 [60]. LSSVM utilizes equality constraints instead of inequality
constraints as in the ordinary SVM. Hence, it simplifies the regression to a problem
that can be easily solved from a set of linear equations. As a result, the LSSVM has a
relatively low complexity and it is computationally more efficient than the standard
SVM [60]. Another advantage of LSSVM is that it has fewer parameters to be tuned.
More detailed information about LSSVM can be found in the book [59].
Although SVM techniques have been widely applied to solve classification and
regression problems, their applications in the treatment of hysteresis are still limited.
To date, the employment of SVM or LSSVM for the modeling and compensation of
hysteresis effects has been conducted by only a few works [74]. The major challenge
of developing an intelligent model lies in how to establish an effective one-to-one
mapping of the hysteresis nonlinearity.
In addition, a review of model-based hysteresis compensation approaches reveals
that a majority of the existing works compensate for the hysteresis effect by employ-
ing an inverse hysteresis model. Thus, both a hysteresis model and an inverse hystere-
sis model are required for the hysteresis identification and compensation purposes.
In this sense, to reduce the workload of precision control, it is desirable to develop a
sole hysteresis model for purposes of both hysteresis description and compensation.

1.4.3 Feedforward Plus Feedback Control

In practice, uncertainties always exist in some degree in the identified hystere-

sis model and dynamics model. To handle these uncertainties, feedback control is
required to make the system robust against the disturbances. Two popular schemes
of feedback combined with feedforward control are shown in Fig. 1.8. Figure 1.8a
1.4 Feedforward Control Based on Hysteresis Models 9

(a)
Xd + X
H −1 ( ) H( ) G(s)
+
+
C
−

(b)
Xd X
C H −1 ( ) H( ) G(s)
+
−

Fig. 1.8 Block diagram of feedforward plus feedback control scheme for precision motion control

illustrates a control scheme where the feedback control C is connected in parallel to

the feedforward control term H −1 (.) [21, 58], while Fig. 1.8b shows a strategy where
the feedback control C is applied once the nonlinear system is linearized by the feed-
forward control H −1 (.) [16]. It is notable that the purpose of the feedforward control
in both schemes is to compensate for the hysteresis effect. An alternative scheme is
presented in the book [17] (Fig. 9.6, page 259), where an offline feedforward control
is employed to compensate for the vibrational dynamics instead.
Targeting at specific applications, different feedback controllers can be designed
to realize such a precision motion control system. In the literature, classical feed-
back controls have been employed, such as proportional-integral-derivative (PID)
control, high-gain feedback control, state feedback control, etc., owing to their easy
implementation.

1.5 Robust Feedback Control

Considering that hysteresis modeling is usually time-consuming, it is attractive to

suppress the hysteresis nonlinearity by implementing a stand-alone feedback con-
trol scheme. Hence, the main property of hysteresis model-free control is that no
hysteresis model is required. Alternatively, the hysteresis effect is considered as an
uncertainty or a disturbance [75] to the nominal dynamics system, which is toler-
ated by an advanced robust or adaptive controller. For instance, the applications of
sliding-mode control [11, 68, 72, 77], H∞ robust control [63, 65], adaptive con-
trol [36], fuzzy logic control [24, 25, 62], and neural network control [49] have
been successfully reported in the literature. Three typical nonlinear control meth-
ods, namely sliding mode control, model predictive control, and model-reference
adaptive control, are discussed as follows.
10 1 Introduction

1.5.1 Sliding-Mode Control

Sliding-mode control (SMC) technique provides an effective and simple way to deal
with the model imperfection and uncertainties for nonlinear systems. As a variable
structure control, SMC is a nonlinear control approach which drives the nonlinear
system’s state trajectory onto a specified sliding surface (s = 0) and maintains the
trajectory on this surface for the subsequent time, as illustrated in Fig. 1.9. Because
the control can be as simple as switching between two states, it is not sensitive to
parameter variations and uncertainties once entering into the control loop. Thus, the
main attractive property of SMC lies in its robustness in the presence of modeling
imprecision and uncertainty.
In practice, the control scheme of a micro-/nano-positioning system is usually
realized using a sampled-data system, such as a computer or a digital signal proces-
sor (DSP). Deploying a continuous-time controller directly to a digital system may
cause instability of the control system. In addition, concerning SMC, the digital
implementation of a continuous-time control deteriorates the invariance property
exhibited by continuous-time SMC. Hence, to implement a reliable control scheme
on a sampled-data system, the discrete-time sliding-mode control (DTSMC) can be
adopted [7, 19].
Generally, the DTSMC can be categorized into state-based and output-based meth-
ods. The former is developed based on system state or state error [1, 66, 73], while
the latter is realized on the basis of system output or output error [68, 72]. Usually, the
implementation of either method requires the state feedback of the system. However,
in the majority of practical situations, only the position information of a piezoelec-
tric micro-/nano-positioning system is provided by the displacement sensor. Hence,
a state observer is indispensable for the practical realization of DTSMC, which

Fig. 1.9 Phase plane of the

sliding-mode control (SMC)
x
Reaching phase

Sliding phase
x

Sliding surface
( s = 0)
1.5 Robust Feedback Control 11

complicates the control design procedure. Furthermore, an improperly designed state

observer may cause instability of the system. In this sense, it is desirable to eliminate
the use of a state observer.
In previous works, an input–output based adaptive DTSMC has been developed in
[55], which is based solely on input and output data. However, the controller is dedi-
cated to a first-order model with long dead-time. It is not suitable for a piezoelectric
actuation system which typically features a plant model of higher order preceded by
complicated nonlinearity. To remove the restriction of the first-order plant model as
required in [55], a new input–output based DTSMC control has been figured out for
a velocity tracking system [56]. Adaptive coefficients of the switching hyperplane
are employed to tolerate the modeling error, and its effectiveness is verified by sim-
ulation studies. Nonetheless, due to a relatively long parameter adaptation process,
the adaptive control normally leads to a slow transient response before attaining the
steady state [11, 70]. On the other hand, the generalized minimum variance control
(GMVC) scheme can be employed to develop a DTSMC control without using a
state observer. To improve the tracking accuracy, a DTSMC scheme is presented
in [45], which is developed based on input/output measurements and the GMVC
technique. Nevertheless, the output tracking accuracy is unknown and filtering is
required to suppress the induced chattering phenomenon. To eliminate the chattering
completely, a fuzzy sliding mode control is demanded [44]. In addition, the multirate
output feedback technique has been developed to relax the need of the entire state
vector for the realization of DTSMC control scheme [5].
To this end, targeting at easy realization, it is desirable to develop a simple DTSMC
scheme without using the complicated hysteresis model and state observer for precise
motion control of a piezoelectric micro-/nano-positioning system.

1.5.2 Model Predictive Control

For a discrete-time system, the strategy of model predictive control (MPC) can be
employed to predict the system performance in a specified time in the future and to
produce an optimal control action with respect to a predefined cost criterion at each
time step [64], as illustrated in Fig. 1.10.
By combining the SMC and MPC techniques, the model predictive sliding-mode
control has been recently proposed to achieve the advantages of both methods
[20, 67]. Although the combined control has been adopted in a few previous works
[29, 47] for the control of piezoelectric actuator by suppressing the chattering phe-
nomenon, it remains unclear why the MPC methodology can eliminate the chattering
phenomenon in the sliding mode. Hence, more theoretical analysis is required to facil-
itate better understanding of the combination of SMC and MPC techniques, which
is addressed in Chaps. 4, 5, and 8 in this book.
12 1 Introduction

Past Future
Actual output

Predicted output
uk +1
uk Predicted input

k k+2 k + N −1
k +1 k +3

Prediction horizon

Fig. 1.10 Schematic diagram of the model predictive control (MPC)

1.5.3 Model-Reference Adaptive Control

Compared with the general robust control approach, the adaptive control does not
require prior information about the bounds on uncertain or time-varying items. Hence,
adaptive control paves a more straightforward way to the precision control of micro-
/nano-positioning systems.
In the literature, only limited work has been dedicated to the extension of adap-
tive controllers to micro-/nano-positioning system control. In addition, uncertainty
bounds are usually required to realize the control system. In previous work [34],
a model-reference adaptive control (MRAC) strategy is reported to compensate for
the hysteresis effect of a micropositioning stage. Although the adaptive controller is
realized without modeling the hysteresis effect nor acquiring the uncertainty bounds,
a Prandtl–Ishlinskii hysteresis model is required to convert the desired motion tra-
jectory into a voltage input. More recently, a MRAC scheme based on hyperstability
theory is presented for a piezoactuated system [39]. Nonetheless, a Bouc–Wen hys-
teresis model is still employed to identify the dynamics equation of the system.
From a practical point of view, it is preferable to develop a MRAC scheme without
modeling the complicated nonlinear effects. By considering the nonlinearity as a per-
turbation to the system, several perturbation estimation methods have been reported
to be integrated with MRAC schemes. To name a few, a MRAC with disturbance
rejection strategy is presented for the systems, which can be represented by parabolic
or hyperbolic partial differential equations along with known disturbance model or
constant disturbance [31]. However, these scenarios are different from the situation in
a PEA-driven system where the unmodeled disturbance involving hysteresis nonlin-
earity is not constant. In addition, an adaptive perturbation approximator is outlined
1.5 Robust Feedback Control 13

to estimate the structured perturbations of second-order systems online [14], and a

perturbation estimation mechanism is proposed to design a MRAC controller for
a class of multiple-input and multiple-output (MIMO) dynamic systems [12]. The
aforementioned two methods are implemented on the basis of the assumption that the
plant states are all measurable. Nevertheless, this assumption does not always hold
for micro-/nano-positioning systems which typically only offer displacement feed-
back. Moreover, the reference signals used in the preceding literature are all given
in terms of voltage. For practical applications, a desired motion trajectory instead of
voltage signal is preferable. These issues are overcome in Chap. 9 in the book.

1.5.4 Other Control Strategies

In addition to the aforementioned control schemes, H∞ robust control, repetitive

control, iterative learning control, disturbance observer, neural network control,
fuzzy control, etc., have also been extensively explored for piezo-driven micro-/nano-
positioning control.
Different control schemes own their specific properties and are suitable for differ-
ent application scenarios. For example, H∞ controller is usually of high order, which
requires excessive computational burden. Repetitive control and iterative learning
control are commonly applicable to the tracking of periodical motion trajectory.
Fuzzy control is heavily dependent on the finely tuned rule tables.

1.6 Position/Force Control in Micromanipulation

Micromanipulation and microassembly are important approaches to producing com-

plex micro/nano-systems which are composed of microparts fabricated from different
materials [6, 40, 51]. In view of the limitations of human operators in terms of accu-
racy and efficiency for executing such tasks, automated handling of the objects in
micro- and nanometer scales necessitates the development of robotic microassembly
systems. As a crucial micro-/nano-positioning device to realize a grasp operation
in the aforementioned application, microgripper has gained intensive attention in
the literature. Typically, the gripper is employed to execute an automated grasp-
hold-release operation for such applications as material characterization, biological
sample pick-and-place, and micro-object assembly.
To avoid the damage of the fragile microobjects and microgripper arms which
contact each other during microassembly process, the interaction control is critical
to regulate the desired position and contact force simultaneously. One of the major
challenges stems from the combination method of position and force control. Gen-
erally, there are two basic approaches to interaction control, i.e., hybrid control and
impedance control, which are presented below.
14 1 Introduction

1.6.1 Hybrid Control

The hybrid position/force control method decomposes the task space of the end-
effector into position subspace and force subspace. It allows the generation of com-
pliant motion by controlling the end-effector position in position subspace and the
interaction force in force subspace, respectively. Generally, a position control law
along the force constrained directions and a force control law along the position
constrained directions are designed through the use of selection matrices. However,
a switching of the control law is required to implement the position and force con-
trol alternately. In addition, this approach usually suffers from robustness problems
during the transition between unconstrained and constrained motions [54].

1.6.2 Impedance Control

The impedance control enables compliant motion of the end-effector in a unified

framework for both constrained and unconstrained directions [27]. The goal of
impedance control is to regulate a desired dynamics relation between the position and
contact force. That is, both position and force are indirectly controlled by regulating
the predefined dynamics response, i.e., the desired impedance.
Impedance control emerges as a promising approach to interaction control in
micro-/nanoscales [23, 42]. For example, a position-based impedance control of a
piezoelectric bimorph microgripper is reported in [53], where a state-feedback posi-
tion controller and a low-pass filtered force feedback controller are implemented.
However, the impedance control usually requires an exact dynamics model of the sys-
tem. Due to the uncertainties in dynamics model, it is practically difficult to achieve
the desired impedance. To remedy this shortcoming, adaptive [41] and robust [43]
impedance control algorithms have been developed to tolerate the dynamic parameter
uncertainties. Nevertheless, the conventional impedance control framework exhibits
a drawback of inability of tracking the reference force input. Several schemes have
been exploited to overcome this shortcoming. Typically, a generalized impedance
control (GIC) scheme has been established based on a generalized desired impedance
[10]. It integrates both position and force errors, and enables the force tracking ability.
Recently, intensive efforts have been made toward force-tracking impedance con-
trol, which is robust to parameter uncertainties in both dynamics and environment
models [28]. Although this robustness property appears attractive, it is at the expense
of implementation complication. Considering that the impedance control possesses
some inherent robustness to environment model errors [3, 52], a GIC with robustness
to dynamics model uncertainties is desirable for precise interaction control.
On the one hand, the control scheme of a practical microassembly system is usu-
ally realized on a digital system, e.g., a computer or DSP apparatus. As it is known,
applying a continuous-time controller directly to a digital system can induce instabil-
ity due to the computational delays [8]. This is true for a microassembly system, e.g.,
1.6 Position/Force Control in Micromanipulation 15

a visual servoing system [50], which typically has a low sampling rate, typically, of
the order of 10 Hz. Hence, to deploy a reliable interaction control to a sampled-data
system, a discrete-time impedance controller is more preferable. However, limited
work has been conducted on discrete-time sliding-mode impedance control scheme
in the literature. It is necessary to devise new discrete-time sliding-mode impedance
control to realize an interaction control of a piezoelectric microgripper dedicated to
micromanipulation and microassembly tasks.

1.6.3 Switching Control

To use the impedance control framework [37, 69], care must be taken for the posi-
tion and force trajectory planning to ensure a desired performance. Intuitively, a
position/force switching control can be adopted to adjust the gripper tip position and
gripping force in an alternate manner [57]. The concerned dominant issue is how to
realize a smooth transition when different control variables are switched [48].
In order to achieve a smooth transition between the position and force control,
an intermediate brake mode can be added between the two control modes [22, 61].
Furthermore, to accomplish a rapid intermediate control and to reduce the force
overshoot in the initial contact, an optimal velocity profile can be used [2]. However,
an extra control mode is required, whose execution increases the overall operation
time of the gripper. More recently, fuzzy control has been proposed to smooth the
transition [18, 46]. However, the membership functions have to be readjusted to adapt
to the gripping of different objects, which complicates the practical implementation
process. Hence, a simple and efficient approach is desirable to realize a smooth
transition between the position and force control.
Chapter 11 reports on a scheme of incremental control to implement a smooth
transition of the position/force switching control for a microgripping system. Specif-
ically, in the closing and opening phases, an incremental digital sliding-mode control
(DSMC) is devised to cater for the position control. During the contact phase, an
incremental-type digital PID force control is adopted. The incremental controller
only solves the change value of the control action at each time step. At the moment
of switching between two consecutive control phases, the controller adopts the con-
trol action in the previous phase as the base value for the next one. In this way, a
smooth transition between two adjacent phases is easily achieved. The feasibility of
the scheme is examined by a typical grasp-hold-release operation of a micro copper
wire through experimental studies. Moreover, its superiority over the conventional
approach is revealed by experimental demonstrations.
16 1 Introduction

1.7 Book Summary

This book mainly concerns hysteresis modeling and compensation of piezoelec-

tric micro-/nano-positioning systems. Both feedforward and feedback controls have
been employed in piezoelectric micro-/nano-positioning control. The approaches of
piezoelectric system control with and without using hysteresis model are two different
ways. By comparing the two approaches, it is observed that the feedforward con-
trol is capable of compensating the majority of the piezoelectric nonlinearity effect.
Through the combination of a simple feedback control with the feedforward control,
better tracking accuracy can be achieved. However, hysteresis or inverse hysteresis
modeling is a time-consuming work, which requires extensive experimental testing.
In addition, the couplings among hysteresis, creep, and vibration dynamics call for
a careful treatment. Such kind of computational burden can be avoided by develop-
ing a robust control which considers the unmodeled nonlinearity as a disturbance to
the nominal plant model. In this way, the time-consuming procedure of hysteresis
modeling is eliminated. At the same time, extensive knowledge of control engineer-
ing is needed to design such kind of control strategies. Therefore, these different
approaches possess specific advantages and disadvantages. It is suggested to select
a suitable control approach based on the control objective and on the background of
the reader.
As a typical application of the piezoelectric micro-/nano-positioning devices,
micromanipulation tasks have been demonstrated using piezoelectric grippers. When
targeting such an application, the implementation of joint position and force control is
very important. In this book, the robust impedance control and position/force switch-
ing control are carried out to show the effectiveness of the micro-/nano-positioning
systems dedicated to micromanipulation and microassembly scenarios.

References

1. Abidi, K., Xu, J.X., Yu, X.: On the discrete-time integral sliding mode control. IEEE Trans.
Automat. Control 52(4), 709–715 (2007)
2. Alkkiomaki, O., Kyrki, V., Kalviainen, H., Liu, Y., Handroos, H.: Smooth transition from
motion to force control in robotic manipulation using vision. In: Proceeding of the 9th Inter-
national Conference on Control, Automation, Robotics and Vision, pp. 1–6 (2006)
3. Almeida, F., Lopes, A., Abreu, P.: Force-impedance control: A new control strategy of robotic
manipulators. In: Kaynak, O., Tosunoglu, S., Ang, M. (eds.) Recent Advances in Mechatronics,
pp. 126–137 (1999)
4. Ang, W.T., Khosla, P.K., Riviere, C.N.: Feedforward controller with inverse rate-dependent
model for piezoelectric actuators in trajectory-tracking applications. IEEE/ASME Trans.
Mechatron. 12(2), 134–142 (2007)
5. Bandyopadhyay, B., Fulwani, D.: High-performance tracking controller for discrete plant using
nonlinear sliding surface. IEEE Trans. Ind. Electron. 56(9), 3628–3637 (2009)
6. Bargiel, S., Rabenorosoa, K., Clévy, C., Gorecki, C., Lutz, P.: Towards micro-assembly of
hybrid moems components on a reconfigurable silicon free-space micro-optical bench. J.
Micromech. Microeng. 20(4), 045012 (2010)
References 17

7. Bartoszewicz, A.: Discrete-time quasi-sliding-mode control strategies. IEEE Trans. Ind. Elec-
tron. 45(5), 633–637 (1998)
8. Bonitz, R.G., Hsia, T.C.: Internal force-based impedance control for cooperating manipulators.
IEEE Trans. Robot. Automat. 12(1), 78–89 (1996)
9. Boukari, A.F., Carmona, J.C., Moraru, G., Malburet, F., Chaaba, A., Douimi, M.: Piezo-
actuators modeling for smart applications. Mechatronics 21(1), 339–349 (2011)
10. Chen, S.P., Liaw, H.C.: Generalized impedance control of robot for assembly tasks requiring
compliant manipulaiton. IEEE Trans. Ind. Electron. 43(4), 453–461 (1996)
11. Chen, X., Hisayama, T.: Adaptive sliding-mode position control for piezo-actuated stage. IEEE
Trans. Ind. Electron. 55(11), 3927–3934 (2008)
12. Cheng, C.C., Chang, C.C., Su, T.M.: Design of model reference adaptive tracking controllers
for mismatch perturbed nonlinear systems with input nonlinearity. In: Proceeding of 17th IFAC
World Congress, pp. 5974–5979. Seoul, Korea (2008)
13. Clayton, G.M., Tien, S., Fleming, A.J., Moheimani, S.O.R., Devasia, S.: Inverse-feedforward
of charge-controlled piezopositioners. Mechatronics 18(5–6), 273–281 (2008)
14. Demetriou, M.A., Fahroo, F.: Model reference adaptive control of structurally perturbed
second-order distributed parameter systems. Int. J. Robust Nonlinear Control 16(16), 773–
799 (2006)
15. Dong, R., Tan, Y., Chen, H., Xie, Y.: A neural networks based model for rate-dependent
hysteresis for piezoelectric actuators. Sens. Actuator A-Phys. 143(2), 370–376 (2008)
16. Esbrook, A., Tan, X., Khalil, H.K.: Control of systems with hysteresis via servocompensation
and its application to nanopositioning. IEEE Trans. Control Syst. Technol. 21(3), 725–738
(2013)
17. Fleming, A.J., Leang, K.K.: Design. Springer, Modeling and Control of Nanopositioning Sys-
tems (2014)
18. Gao, B., Shao, J., Han, G., Sun, G., Yang, X., Wu, D.: Using fuzzy switching to achieve the
smooth switching of force and position. Appl. Mech. Mater. 274, 638–641 (2013)
19. Gao, W., Wang, Y., Homaifa, A.: Discrete-time variable structure control systems. IEEE Trans.
Ind. Electron. 42(2), 117–122 (1995)
20. Garcia-Gabin, W., Zambrano, D., Camacho, E.F.: Sliding mode predictive control of a solar
air conditioning plant. Control Eng. Pract. 17(6), 652–663 (2009)
21. Ge, P., Jouaneh, M.: Tracking control of a piezoceramic actuator. IEEE Trans. Control Syst.
Technol. 4(3), 209–216 (1996)
22. Hara, S., Yamada, Y.: A control method switching from servo automatic transfer to force
sensorless impedance control manual positioning. In: Proceeding IECON 2007—33rd Annual
Conference of the IEEE Industrial Electronics Society, pp. 292–298 (2007)
23. Huang, H.B., Sun, D., Mills, J.K., Cheng, S.H.: Robotic cell injection system with position and
force control: Toward automatic batch biomanipulation. IEEE Trans. Robot. 25(3), 727–737
(2009)
24. Huang, H.P., Yan, J.L., Cheng, T.H.: Development and fuzzy control of a pipe inspection robot.
IEEE Trans. Ind. Electron. 57(3), 1088–1095 (2010)
25. Hwang, C.L.: Microprocessor-based fuzzy decentralized control of 2-D piezo-driven systems.
IEEE Trans. Ind. Electron. 55(3), 1411–1420 (2008)
26. Janaideh, M.A., Rakheja, S., Su, C.Y.: Experimental characterization and modeling of rate-
dependent hysteresis of a piezoceramic actuator. Mechatronics 19(5), 656–670 (2009)
27. Jung, S., Hsia, T.C., Bonitz, R.G.: Force tracking impedance control for robot manipulators
with an unknown environment: Theory, simulation, and experiment. Int. J. Robot. Res. 20(9),
765–774 (2001)
28. Jung, S., Hsia, T.C., Bonitz, R.G.: Force tracking impedance control of robot manipulators
under unknown environment. IEEE Trans. Control Syst. Technol. 12(3), 474–483 (2004)
29. Kim, B., Washington, G.N.: Nonlinear position control of smart actuators using model predic-
tive sliding mode control. In: Proceeding of ASME Conference on Smart Materials, Adaptive
Structures and Intelligent Systems, pp. 511–522 (2008)
18 1 Introduction

30. Kim, J., Kang, B.: Micro-macro linear piezoelectric motor based on self-moving cell. Mecha-
tronics 19(7), 1134–1142 (2009)
31. Kim, J.Y., Bentsman, J.: Disturbance rejection in robust model reference adaptive control of
parabolic and hyperbolic systems. In: Proceeding of 45th IEEE Conference on Decision and
Control, pp. 3083–3088. San Diego, CA, USA (2006)
32. Leang, K.K., Zou, Q., Devasia, S.: Feedforward control of piezoactuators in atomic force
microscope systems: Inversion-based compensation for dynamics and hysteresis. IEEE Control
Syst. Mag. 19(1), 70–82 (2009)
33. Li, G., Wen, C., Zheng, W.X., Chen, Y.: Identification of a class of nonlinear autoregressive
models with exogenous inputs based on kernel machines. IEEE Trans. Signal Process. 59(5),
2146–2159 (2012)
34. Li, Y., Xu, Q.: Hysteresis modeling and compensation for an XY micropositioning stage with
model reference adaptive control. In: Proceeding of 48th IEEE Conference on Decision and
Control, pp. 5580–5585. Shanghai, China (2009)
35. Li, Y., Xu, Q.: Adaptive sliding mode control with perturbation estimation and PID sliding
surface for motion tracking of a piezo-driven micromanipulator. IEEE Trans. Control Syst.
Technol. 18(4), 798–810 (2010)
36. Liaw, H.C., Shirinzadeh, B.: Enhanced adaptive motion tracking control of piezo-actuated
flexure-based four-bar mechanisms for micro/nano manipulation. Sens. Actuator A-Phys. 147,
254–262 (2008)
37. Liaw, H.C., Shirinzadeh, B.: Robust adaptive constrained motion tracking control of piezo-
actuated flexure-based mechanisms for micro/nano manipulation. IEEE Trans. Ind. Electron.
58(4), 1406–1415 (2011)
38. Lin, C.J., Chen, S.Y.: Evolutionary algorithm based feedforward control for contouring of a
biaxial piezo-actuated stage. Mechatronics 19(6), 829–839 (2009)
39. Liu, Y.T., Chang, K.M., Li, W.Z.: Model reference adaptive control for a piezo-positioning
system. Precis. Eng. 34(1), 62–69 (2010)
40. Lopez-Walle, B., Gauthier, M., Chaillet, N.: Principle of a submerged freeze gripper for
microassembly. IEEE Trans. Robot. 24(4), 897–902 (2008)
41. Lu, W.S., Meng, Q.H.: Impedance control with adaptation for robotic manipulations. IEEE
Trans. Robot. Automat. 7(3), 408–415 (1991)
42. Lu, Z., Chen, P.C.Y., Lin, W.: Force sensing and control in micromanipulation. IEEE Trans.
Syst. Man Cybern. Part C, Appl. Rev. 36(6), 713–724 (2006)
43. Lu, Z., Kawamura, S., Goldenberg, A.A.: An approach to sliding-mode based control. IEEE
Trans. Robot. Automat. 11(5), 754–759 (1995)
44. Mitic, D., Milojkovic, M., Antic, D.: Tracking system design based on digital minimum variance
control with fuzzy sliding mode. In: Proceeding of 8th International Conference on Telecommu-
nications in Modern Satellite, Cable and Broadcasting Services (TELSIKS 2007), pp. 494–497.
Nis, Serbia (2007)
45. Mitic, D., Milosavljevic, C.: Sliding mode-based minimum variance and generalized minimum
variance controls with O(T 2 ) and O(T 3 ) accuracy. Electr. Eng. 86(4), 229–237 (2004)
46. Navarro, C., Treesatayapun, C., Baltazar, A.: Determination of the instantaneous initial contact
point on a parallel gripper using a multi input fuzzy rules emulated network controller with
feedback from ultrasonic and force sensors. In: Batyrshin, I., Gonzalez Mendoza, M. (eds.)
Advances in Artificial Intelligence, Lecture Notes in Computer Science, vol. 7629, pp. 261–
272. Springer, Berlin (2013)
47. Neelakantan, V.A., Washington, G.N., Bucknor, N.K.: Model predictive control of a two stage
actuation system using piezoelectric actuators for controllable industrial and automotive brakes
and clutches. J. Intell. Mater. Syst. Struct. 19(7), 845–857 (2008)
48. Niksefat, N., Wu, Q., Sepehri, N.: Stable control of an electro-hydraulic actuator during contact
tasks: Theory and experiments. In: Proceeding of the American Control Conference, pp. 4114–
4118 (2000)
49. Orlowska-Kowalska, T., Dybkowski, M., Szabat, K.: Adaptive sliding-mode neuro-fuzzy con-
trol of the two-mass induction motor drive without mechanical sensors. IEEE Trans. Ind.
Electron. 57(2), 553–564 (2010)
References 19

50. Reddy, A.N., Maheshwari, N., Sahu, D.K., Ananthasuresh, G.K.: Miniature compliant grippers
with vision-based force sensing. IEEE Trans. Robot. 26(5), 867–877 (2010)
51. Sariola, V., Jaaskelainen, M., Zhou, Q.: Hybrid microassembly combining robotics and water
droplet self-alignment. IEEE Trans. Robot. 26(6), 965–977 (2010)
52. Schutter, J.D., Bruyninckx, H., Zhu, W.H., Spong, M.W.: Force control: A bird’s eye view. In:
Siciliano, B. (ed.) Control Problems in Robotics and Automation: Future Directions, pp. 1–17
(1998)
53. Seki, H.: Modeling and impedance control of a piezoelectric bimorph microgripper. In:
Proceeding of IEEE/RSJ International Conference on Intelligent Robots and Systems,
vol. 2, pp. 958–965. Raleigh, USA (1992)
54. Seraji, H., Colbaugh, R.: Force tracking in impedance control. Int. J. Robot. Res. 16(1), 97–117
(1997)
55. Sha, D., Bajic, V.B.: Robust discrete adaptive input-output-based sliding mode controller. Int.
J. Syst. Sci. 31(12), 1601–1614 (2000)
56. Sha, D., Bajic, V.B., Yang, H.: New model and sliding mode control of hydraulic elevator
velocity tracking system. Simul. Pract. Theory 9(6), 365–385 (2002)
57. Shimada, N., Yoshioka, T., Ohishi, K., Miyazaki, T.: Novel force-sensor-less contact motion
control for quick and smooth industrial robot motion. In: Proceeding IECON 2011—37th
Annual Conference on IEEE Industrial Electronics Society, pp. 4238–4243 (2011)
58. Song, G., Zhao, J., Zhou, X., De Abreu-Garcia, J.: Tracking control of a piezoceramic actuator
with hysteresis compensation using inverse preisach model. IEEE/ASME Trans. Mechatron.
10(2), 198–209 (2005)
59. Suykens, J.A.K., Gestel, T.V., Brabanter, J.D., Moor, B.D., Vandewalle, J.: Least Squares
Support Vector Machines. World Scientific Publishing Co., Singapore (2002)
60. Suykens, J.A.K., Vandewalle, J.: Least squares support vector machine classifiers. Neural
Process. Lett. 9(3), 293–300 (1999)
61. Takahashi, T., Tsuboi, T., Kishida, T., Kawanami, Y., Shimizu, S., Iribe, M., Fukushima, T.,
Fujita, M.: Adaptive grasping by multi fingered hand with tactile sensor based on robust force
and position control. In: Proceeding of the International Conference on Robotics and Automa-
tion, pp. 264–271 (2008)
62. Wang, C.H., Huang, D.Y.: A new intelligent fuzzy controller for nonlinear hysteretic electronic
throttle in modern intelligent automobiles. IEEE Trans. Ind. Electron. 60(6), 2332–2345 (2013)
63. Wang, R., Liu, G.P., Wang, W., Rees, D., Zhao, Y.B.: H∞ control for networked predictive
control systems based on the switched Lyapunov function method. IEEE Trans. Ind. Electron.
57(10), 3565–3571 (2010)
64. Wills, A.G., Bates, D., Fleming, A.J., Ninness, B., Moheimani, S.O.R.: Model predictive control
applied to constraint handling in active noise and vibration control. IEEE Trans. Control Syst.
Technol. 16(1), 3–12 (2008)
65. Wu, Y., Zou, Q.: Robust-inversion-based 2-DOF control design for output tracking: Piezoelec-
tric actuator example. IEEE Trans. Control Syst. Technol. 17(5), 1069–1082 (2009)
66. Xi, Z., Hesketh, T.: Discrete time integral sliding mode control for overhead crane with uncer-
tainties. IET Control Theory Appl. 4(10), 2071–2081 (2010)
67. Xiao, L., Su, H., Chu, J.: Sliding mode prediction tracking control design for uncertain systems.
Asian J. Control 9(3), 317–325 (2007)
68. Xu, J.X., Abidi, K.: Discrete-time output integral sliding-mode control for a piezomotor-driven
linear motion stage. IEEE Trans. Ind. Electron. 55(11), 3917–3926 (2008)
69. Xu, Q.: Adaptive discrete-time sliding mode impedance control of a piezoelectric microgripper.
IEEE Trans. Robot. 29(3), 663–673 (2013)
70. Xu, Q., Jia, M.: Model reference adaptive control with perturbation estimation for a micropo-
sitioning system. IEEE Trans. Control Syst. Technol. 22(1), 352–359 (2014)
71. Xu, Q., Li, Y.: Dahl model-based hysteresis compensation and precise positioning control of an
XY parallel micromanipulator with piezoelectric actuation. J. Dyn. Syst. Meas. Control-Trans.
ASME 132(4), 041011 (2010)
20 1 Introduction

72. Xu, Q., Li, Y.: Micro-/nanopositioning using model predictive output integral discrete sliding
mode control. IEEE Trans. Ind. Electron. 59(2), 1161–1170 (2012)
73. Xu, Q., Li, Y.: Model predictive discrete-time sliding mode control of a nanopositioning
piezostage without modeling hysteresis. IEEE Trans. Control Syst. Technol. 20(4), 983–994
(2012)
74. Xu, Q., Wong, P.K.: Hysteresis modeling and compensation of a piezostage using least squares
support vector machines. Mechatronics 21(7), 1239–1251 (2011)
75. Yi, J., Chang, S., Shen, Y.: Disturbance observer-based hysteresis compensation for piezoelec-
tric actuators. IEEE/ASME Trans. Mechatron. 14(4), 456–464 (2009)
76. Yu, S., Alici, G., Shirinzadeh, B., Smith, J.: Sliding mode control of a piezoelectric actuator with
neural network compensating rate-dependent hysteresis. In: Proceeding of IEEE International
Conference on Robotics and Automation, pp. 3641–3645 (2005)
77. Yu, X., Kaynak, O.: Sliding mode control with soft computing: A survey. IEEE Trans. Ind.
Electron. 56(9), 3275–3285 (2009)
78. Yu, Y., Xiao, Z., Naganathan, N.G., Dukkipati, R.V.: Dynamic Preisach modelling of hysteresis
for the piezoceramic actuator system. Mech. Mach. Theory 37(1), 75–89 (2002)
 Part I
Hysteresis-Model-Based Feedforward
Control
Chapter 2
Feedforward Control Based on Inverse
Hysteresis Models

This chapter presents the rate-dependent hysteresis compensation of a piezoelectric

nanopositioning stage using the feedforward control based on an inverse hystere-
sis model. Three different controllers are realized and compared, which employ
Bouc–Wen model, modified Prandtl–Ishlinskii (MPI) model, and least squares sup-
port vector machines (LSSVM)-based intelligent model, respectively. Experimental
studies demonstrate the superiority of LSSVM model in hysteresis modeling and
compensation tasks.

2.1 Introduction

In order to compensate for the hysteresis effect, the hysteresis behavior is usually
modeled by Preisach model [3, 6], Prandtl–Ishlinskii model [7], Bouc–Wen model
[11, 12], Maxwell-based model [8], Dahl model [18], polynomial model [9], etc.
Then, an inverse hysteresis model is constructed and utilized as a feedforward con-
troller to cancel the hysteresis effect. However, the hysteresis effect is dependent not
only on the amplitude but also on the frequency of input signals. It is very difficult
to capture the complicated rate-dependent hysteretic behavior precisely. In addition,
majority of the existing models employ a large number of parameters to describe the
rate-dependent hysteresis [1, 21], which may block their applications in high-speed
real-time control as an adverse effect.
Recently, it has been shown that ANN provides an efficient way to model the
nonlinear hysteresis [5, 20]. Nevertheless, there is no universal method to deter-
mine an optimal ANN structure in terms of the number of hidden layers and number
of neurons in each layer. Moreover, ANN exhibits the problems of overfitting and
sinking into local optima, which are their major drawbacks in practical implementa-
tion. Alternatively, SVM gives a promising way to estimate nonlinear system models
accurately. Based on the statistical learning theory and structural risk minimization
principle, the SVM approach is capable of modeling nonlinear systems by trans-
© Springer International Publishing Switzerland 2016 23
Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2_2
24 2 Feedforward Control Based on Inverse Hysteresis Models

forming the regression problem into a convex quadratic programming (QP) problem
and then solving it with a QP solver. As compared with the conventional ANN,
SVM exhibits the major advantages of global optimization and higher generaliza-
tion capability. As a variant of SVM, LSSVM is introduced in 1999 [13]. LSSVM
utilizes equality constraints instead of inequality constraints as in the ordinary SVM.
Hence, it simplifies the regression to a problem that can be easily solved from a set
of linear equations [17]. As a result, the LSSVM has a relatively low complexity and
are more computationally efficient than the standard SVM [13]. Another advantage
of LSSVM lies in that it has fewer parameters to tune [19]. For more details about
LSSVM algorithm, the reader may refer to the book [14].
In this chapter, the LSSVM is introduced to the domain of hysteresis modeling
and compensation for a nanopositioning stage, which is driven by piezoelectric stack
actuators. A LSSVM model is proposed and trained by introducing the current input
value and input variation rate as the input data set to formulate a one-to-one map-
ping. By adopting the radial basis function (RBF) as kernel function, the LSSVM
model only has two free hyperparameters, which are optimally tuned by resorting to
Bayesian inference framework. The effectiveness of the presented model is verified
by carrying out experimental studies. For a comparison study, two state-of-the-art
approaches, namely, Bouc–Wen model and MPI model, are employed. Both Bouc–
Wen and MPI models are identified by resorting to a global search approach of
particle swarm optimization (PSO).
In addition, a LSSVM inverse model-based feedforward control in combination
with a proportional-integral-derivative (PID) feedback control is designed to com-
pensate for the hysteresis nonlinearity. The feasibility of the presented modeling and
control approaches are validated by a series of simulation and experimental studies.

2.2 System Description and Hysteresis Characterization

The test bed as employed in this work and its hysteresis properties are described in
this section.

2.2.1 Experimental Setup

The experimental setup is shown in Fig. 2.1. A commercial two-axis nanoposition-

ing stage (model: P-734.2CL, from Physik Instrumente Co., Ltd.) is adopted. It is
driven by two piezoelectric stack actuators (PSAs) through a high-voltage amplifier
(model: E-503.00, from Physik Instrumente Co., Ltd.). The integrated capacitive
sensors are used for displacement measurement. The displacement sensors provide
the accuracy and repeatability in nanometer and sub-nanometer levels, respectively.
The output signals of the two capacitive sensors are processed by a signal condi-
tioner module (model: E-509.C2A, from Physik Instrumente Co., Ltd.). In addition,
2.2 System Description and Hysteresis Characterization 25

Fig. 2.1 Experimental setup of a piezoelectric nanopositioning system

the control hardware consists of a digital signal processor (DSP)-based real-time

controller (model: DS1103, from dSPACE GmbH) with equipped 16-bit A/D and
D/A converters. Control algorithms are developed with MATLAB software and
then downloaded to DS1103 board via the ControlDesk interface to implement the
real-time control.

2.2.2 Hysteresis Characterization

By applying a triangular-wave voltage with the constant feed-rate and variable ampli-
tudes to PSA #1, the open-loop position responses of the XY nanopositioning stage
in the two working axes are shown in Fig. 2.2. It is seen that the cross-talk between
the two axes is about 1.17 %, which indicates that the output motions in the two axes
are approximately decoupled. Owing to the decoupling property of the two work-
ing axes, the motion control in one axis (x-axis) of the system is carried out in the
following discussion.
The x-axis hysteresis curves in Fig. 2.2c reveal that the shape of the hysteresis
loop is dependent on the amplitude of the input. The larger the amplitude of the input
signal, the wider the output hysteresis loop. In addition, by applying sine waves with
the same amplitude and varying frequencies (1–10 Hz) as shown in Fig. 2.3a, the
x-axis displacement output is depicted in Fig. 2.3b. It can be observed from Fig. 2.3c
that the hysteresis shape also relies on the input rate. The higher the frequency of the
input signal, the larger the width of the output hysteresis.
The foregoing open-loop testing results indicate that the complicated hysteresis
effects are dependent not only on the amplitude but also on the frequency of the input
signal. The modeling and compensation of the hysteretic nonlinearity are carried out
in the following sections.
26 2 Feedforward Control Based on Inverse Hysteresis Models

(a) 8

Input voltage (V) 6

0
0 2 4 6 8 10 12
Time (s)

(b) 100
Displacement (µm)

80 x
60

40

20 y

0 2 4 6 8 10 12
Time (s)

(c) 100
x
80
Displacement (µm)

60

40

20 y

0 1 2 3 4 5 6 7 8
Input voltage (V)

Fig. 2.2 Hysteresis curves of the XY nanopositioning stage with triangular-wave inputs. a Input
voltage to the voltage amplifier. b Time history of the displacements x and y. c Output displacements
x and y versus input voltage
2.2 System Description and Hysteresis Characterization 27

(a) 8

Input voltage (V) 6

0
0 2 4 6 8 10
Time (s)

(b) 100
Displacement (µm)

80

60

40

20

0
0 2 4 6 8 10
Time (s)

(c) 100
Displacement (µm)

80 1−10 Hz

60

40

20

0
0 1 2 3 4 5 6 7 8
Input voltage (V)

Fig. 2.3 Hysteresis curves of the XY nanopositioning stage with sinusoidal-wave inputs. a Input
voltage to the voltage amplifier. b Time history of the displacement x. c Output displacement x
versus input voltage
28 2 Feedforward Control Based on Inverse Hysteresis Models

2.3 Hysteresis Modeling

The purpose of hysteresis modeling is to capture the hysteresis behavior of the system.
First, the hysteresis is modeled using the popular Bouc–Wen, MPI, and LSSVM
models. In addition, the corresponding inverse hysteresis models are also derived for
the purpose of hysteresis compensation.

2.3.1 Hysteresis Modeling with the Bouc–Wen Model

Thanks to a fewer number of parameters, Bouc–Wen model has been widely

employed in piezoelectric hysteresis modeling. In particular, the entire dynamics
model of the piezoelectrically driven system with Bouc–Wen type of hysteresis can
be established as follows [11, 12]:

m ÿ(t) + b ẏ(t) + ky(t) = k[du(t) − h(t)] (2.1)

ḣ(t) = αd u̇(t) − β|u̇(t)|h(t)|h(t)|n−1 − γ u̇(t)|h(t)|n (2.2)

where t is the time variable; parameters m, b, k, and y represent the mass, damping
coefficient, stiffness, and displacement response of the piezostage, respectively; d
is the piezoelectric coefficient; u denotes the input voltage; and h indicates the hys-
teretic loop in terms of displacement whose magnitude and shape are determined by
parameters α, β, γ , and the order n, with n governing the smoothness of the transi-
tion from elastic to plastic response. For the elastic structure and material, n = 1 is
assigned in Eq. (2.2).
To determine the model parameters, the seven parameters (m, b, k, d, α, β, and γ )
of the Bouc–Wen model are identified by minimizing the following fitness function
[11]:

1  2
N
F(m, b, k, d, α, β, γ ) = xi − xiBW (2.3)
N
i=1

where N denotes the total number of samples, and xi − xiBW represents the error
of the i-th sample which is calculated as the deviation of Bouc–Wen model output
(xiBW ) from experimental result (xi ).
The hysteresis model describes the relationship between the input voltage and
output displacement of the system. The dynamics model with Bouc–Wen hysteresis
model can be realized using MATLAB /Simulink blocks as shown in Fig. 2.4.
On the other hand, the input voltage used to produce a desired displacement value is
solved by the inverse hysteresis model [11]. Specifically, the inverse dynamic model
can be derived from Eqs. (2.1) and (2.2). It is realized using MATLAB/Simulink
blocks as shown in Fig. 2.5, where the “Bouc–Wen model” subsystem is depicted in
Fig. 2.4b.
2.3 Hysteresis Modeling 29

Fig. 2.4 Dynamic model implemented with MATLAB/Simulink blocks. a Dynamic model of the
entire system. b Bouc–Wen hysteresis model

Fig. 2.5 The inverse dynamic model implemented with MATLAB/Simulink blocks

2.3.2 Hysteresis Modeling with the MPI Model

The hysteresis effect is described using MPI model, which has more parameters than
the Bouc–Wen model. Then, an inverse hysteresis model is derived to compensate
for the hysteresis effect.
30 2 Feedforward Control Based on Inverse Hysteresis Models

(a) (b)

Fig. 2.6 Illustrations of elementary PI operators. a Backlash operator Hr . b Dead-zone operator Sd

2.3.2.1 PI Model

First, the Prandtl–Ishlinskii (PI) model is described briefly. As a subclass of the

Preisach hysteresis model, the PI model is a superposition of elementary back-
lash or play operators. The backlash operator Hr is illustrated in Fig. 2.6a, which
is defined by

y(t) = Hr [u, x0 ](t)

= max{u(t) − r, min{u(t) + r, y(t − T )}} (2.4)

where u is the voltage control input, y is the displacement response of the stage, r is
the control input threshold value or magnitude of the backlash, and T is the sampling
time interval.
The initial condition of Eq. (2.4) is given by

y(0) = max{u(0) − r, min{u(0) + r, y0 }} (2.5)

where y0 is usually but not necessarily initialized to zero.

A generalized backlash operator can be obtained by multiplying Hr with a weight
value wh , i.e.,

y(t) = wh Hr [u, y0 ](t) (2.6)

where the weight wh describes the gain of the backlash operator.

Using a linearly weighted superposition of multiple backlash operators with dif-
ferent thresholds and weights, the complex hysteresis can be modeled by

y(t) = whT Hr [u, y0 ](t) (2.7)

where whT = [wh0 , wh1 , . . . , whn ] is the weight vector, Hr [u, y0 ](t) = [Hr [u, y00 ]
(t), Hr [u, y01 ](t), . . . , Hr [u, y0n ](t)]T with the threshold vector r = [r0 , r1 , . . . ,
2.3 Hysteresis Modeling 31

rn ]T and the initial state vector x0 = [x00 , x01 , . . . , x0n ]T , for 0 = r0 < r1 <
· · · < rn < +∞. The control input thresholds ri can be chosen as equal intervals
between the minimum and maximum voltage control input values of the piezoelectric
actuation stage.

2.3.2.2 MPI Model

The PI operator has the same symmetric property as the backlash operator with
respect to the center point of the loop formed by the operator. Hence, the model
accuracy of PI operator will be reduced for the situations where the hysteresis loops
are not symmetric. In order to overcome this issue, a saturation operator is adopted
to connect in serial with the hysteresis operator.
Specifically, the saturation operator is a weighted linear superposition of linear-
stop or one-sided dead-zone operators given as follows:

max{y(t) − d, 0}, d > 0
Sd [y](t) = (2.8)
y(t), d=0

where the dead-zone operator Sd is depicted in Fig. 2.6b.

Based on the aforementioned dead-zone operator, the saturation operator is
expressed by

z(t) = wsT Sd [y](t) (2.9)

where y is the output of the hysteresis operator, z is the stage response, wsT =
[ws0 , ws1 , . . . , wsm ] is the weight vector, Sd [y](t) = [Sd0 [y](t), Sd1 [y](t), . . . , Sdm
[y](t)]T with the threshold vector d = [d0 , d1 , . . . , dm ]T , for 0 = d0 < d1 < · · · <
dm < +∞.
Therefore, the modified PI operator is derived as follows:
 
z(t) = Γ [u](t) = wsT Sd whT Hr [u, y0 ] (t). (2.10)

A realization of the MPI model using MATLAB/Simulink blocks is shown in

Fig. 2.7.

Fig. 2.7 MPI model implemented with MATLAB/Simulink blocks

32 2 Feedforward Control Based on Inverse Hysteresis Models

2.3.2.3 Inverse MPI Model

In order to use the MPI model for the feedforward hysteresis compensation, an inverse
model is required to express the voltage as a function of the position.
It has been shown that the inverse MPI model is also PI-type, which can be
calculated by

u(t) = Γ −1 [z](t) = w∗h T Hr∗ w∗s T S∗d [z], y∗0 (t) (2.11)

where w∗h T = [wh0 ∗ , w∗ , . . . , w∗ ] and w∗ T = [w∗ , w∗ , . . . , w∗ ] are the weight

h1 hn s s0 s1 sm
vectors, y∗0 = [y00 ∗ , y ∗ , . . . , y ∗ ]T is the initial state vector, and the threshold vectors
01 0n
are r∗ = [r0∗ , r1∗ , . . . , rn∗ ]T and d∗ = [d0∗ , d1∗ , . . . , dm∗ ]T , respectively.
The inverse MPI model parameters are given as follows [1, 10]:

∗ 1
wh0 = (2.12)
wh0
∗ −whi
whi = 
 
 , i = 1, . . . , n (2.13)
i i−1
j=0 wh j j=0 wh j


i
ri∗ = wh j (ri − r j ), i = 0, 1, . . . , n (2.14)
j=0


i 
n
∗
y0i = wh j y0i + wh j y0 j , i = 0, 1, . . . , n (2.15)
j=0 j=i+1

∗ 1
ws0 = (2.16)
ws0
−ws j
ws∗j = 
 
, j = 1, . . . , m (2.17)
j j−1
k=0 wsk k=0 wsk


j
d ∗j = wsk (d j − dk ), j = 0, 1, . . . , m. (2.18)
k=0

2.3.3 Hysteresis Modeling with the LSSVM

In this section, the hysteresis modeling is treated as a nonlinear regression problem,

and the LSSVM is employed to model the piezoelectric hysteresis for the piezo-
electric nanopositioning system. Due to the hysteresis effects, one input voltage can
produce multiple values of the output displacements. Thus, one challenge lies in
how to convert the one-to-many mapping into a one-to-one mapping. Here, both the
current input and input variation rates are introduced to form the input data set, which
2.3 Hysteresis Modeling 33

determines a unique output value. By making use of the input variation rate, the rate
dependency of the hysteretic behavior can be captured.
Specifically, the hysteresis model is identified using the input voltage (u) and
voltage variation rate (u̇) as the inputs and the displacement (y) as the output to train
the LSSVM.

2.3.3.1 LSSVM Modeling

LSSVM maps the input data into a high-dimensional feature space and constructs a
linear regression function therein. The unknown hysteresis function is approximated
by the equation

y(x) = wT ϕ(x) + b (2.19)

with the given training data set {xi , yi }i=1

N , where N represents the number of training

data set, xi = {u i , u̇ i } ∈ R are the input data, and yi ∈ R are the output data.
2

Additionally, the weight vector w ∈ Rn h , the nonlinear mapping ϕ(·) : R2 → Rn h

denotes a map from the input space to a feature space, and b is the bias.
The LSSVM approach formulates the regression as an optimization problem in
the primal weight space:

1  2
N
1 T
min J (w, e) = w w+ γ ei (2.20)
w,b,e 2 2
i=1

subject to the equality constraints

yi = wT ϕ(xi ) + b + ei , i = 1, 2, . . . , N (2.21)

where ei is the residual error.

The weight vector w may be in high dimension, and it cannot be solved directly. In
order to solve the optimization problem, a Lagrangian function is defined as follows:


N
L(w, b, e; α) = J (w, e) − αi [wT ϕ(xi ) + b + ei − yi ] (2.22)
i=1

where αi are the Lagrange multipliers which can be either positive or negative val-
ues. The conditions for optimality can be obtained by solving the following partial
derivatives:

∂L  N
=0→w= αi ϕ(xi ) (2.23)
∂w
i=1
34 2 Feedforward Control Based on Inverse Hysteresis Models

∂L  N
=0→ αi = 0 (2.24)
∂b
i=1
∂L
= 0 → αi = γ ei , i = 1, 2, . . . , N (2.25)
∂ei
∂L
= 0 → wT ϕ(xi ) + b + ei − yi = 0, i = 1, 2, . . . , N (2.26)
∂αi

which can be used to construct the dual formulation by eliminating w and ei , i.e.,
    
0 1TN b 0
= (2.27)
1 N Ω + γ −1 I N α y
 
Φ

where α = [α1 , α2 , . . . , α N ]T is called the support vector, whose support values

αi are proportional to the residual errors ei by the regularization factor γ ∈ R. In
addition, 1 N = [1, 1, . . . , 1]T , y = [y1 , y2 , . . . , y N ]T , and I N is an identity matrix.
Besides, the kernel trick is employed to derive that:

Ωi j = ϕ(xi )T ϕ(x j ) = K (xi , x j ), i, j = 1, 2, . . . , N (2.28)

where K is a predefined kernel function. The purpose of introducing the kernel

function is to avoid the explicit computation of the map ϕ(·) in dealing with the
high-dimensional feature space.
It is observed that the LSSVM approach utilizes the equality constraints instead of
the inequality constraints as in the ordinary SVM. Thus, it simplifies the regression
to a problem that can be easily solved from a set of linear equations. Assume that Φ
is invertible, then b and α can be calculated from Eq. (2.27):
   
b 0
= Φ −1 . (2.29)
α y

Then, in view of Eq. (2.23), the solution for the regression problem can be derived:


N
y(x) = αi K (x, xi ) + b (2.30)
i=1

where K is the kernel function satisfying Mercer’s condition, xi is the training data,
and x denotes the new input data.
By adopting the RBF kernel function
 
x − xi 2
K (x, xi ) = exp − (2.31)
σ2
2.3 Hysteresis Modeling 35

with σ > 0 denoting the width parameter (which specifies the kernel sample variance
σ 2 ) and · representing the Euclidean distance, the LSSVM model for the hysteresis
model estimation becomes


N  
x − xi 2
y(x) = αi exp − + b. (2.32)
σ2
i=1

Once the regularization parameter γ and kernel parameter σ are assigned, a train-
ing process is needed to determine the support values αi and the bias b.

2.3.3.2 Hyperparameters Tuning with Bayesian Inference

The above modeling procedure shows that there are two hyperparameters (γ and σ )
to be adjusted. The high generalization ability of the LSSVM model relies on the
appropriate tuning of the two parameters. To select the best hyperparameters, the
approach of tenfold cross validation is usually adopted. However, it is a very time-
consuming method. Alternatively, it has been shown that the Bayesian framework
with three levels of inference is an effective way to infer the optimal hyperparameters
of the LSSVM regressors [15].
Given a set of training data, Bayesian inference is a robust framework to determine
the distribution of the estimated model parameters based on the data sets. Afterward,
the optimal model parameters can be predicted. The basic idea of the hyperparameters
inference procedure using Bayesian framework arises from a modified version of the
LSSVM regression problem as shown below:

min J (w, e) = μE W + ζ E D (2.33)

w,b,e

subject to:

ei = yi − [wT ϕ(xi ) + b], i = 1, 2, . . . , N (2.34)

with
1 T
EW = w w (2.35)
2
1 2 1  2
N N
ED = ei = yi − [wT ϕ(xi ) + b] (2.36)
2 2
i=1 i=1

where μ is the new regularization factor and ζ denotes the variance of the noise for
the residual error ei .
36 2 Feedforward Control Based on Inverse Hysteresis Models

The dual program of the above optimization problem is the same as Eq. (2.27).
The hyperparameter γ is related to μ and ζ by γ = ζ /μ. It is noticeable that, by
substituting Eq. (2.35) and the above relationship into Eq. (2.33), the same problem as
described by Eq. (2.20) is generated. Given a training data set, the Bayesian inference
algorithm for the calculation of the two optimal hyperparameters is presented in [16].
Here, the procedure is implemented by resorting to a MATLAB toolbox [4].

2.4 Experimental Studies

In this section, the hysteresis modeling for a piezo-driven nanopositioning stage using
the Bouc–Wen, MPI, and LSSVM models is carried out by experimental studies.
Without loss of generality, the input voltage signal is chosen below in order to
identify the hysteresis model:

u(t) = 5e−0.11t [sin(6π te−0.345t − 1.55) + 1.0] (2.37)

which is depicted in Fig. 2.8a. In addition, the experimental output data are obtained
as depicted in Fig. 2.8b.

2.4.1 Bouc–Wen Model Results

By setting a time interval of 0.02 s, 500 training data sets are obtained as shown in
Fig. 2.9a. Then, the Bouc–Wen model is identified off-line by optimizing the seven
parameters to minimize the fitness function in Eq. (2.3).
In view of the superior performance of PSO over alternative methods such as
the direct search approach and genetic algorithm (GA), the PSO is adopted for the
parameter optimization. The optimization is carried out with a PSO toolbox [2]
running in MATLAB environment. The identified model parameters are shown in
Table 2.1.
The experimental result and the simulated Bouc–Wen model output are compared
in Fig. 2.9a. The plots indicate that the Bouc–Wen model cannot exactly represent
the complicated hysteresis of the system. A relative large error exists between the
identified model output and the experimental result as shown in Fig. 2.9b. Specifically,
the maximum model error is 4.70 µm, which accounts for 5.3 % of the concerned
travel range of the nanopositioning stage. It is observed that a smaller model error
is obtained when the input has lower magnitude and frequency (2–4 s). Hence, the
model error varies greatly at different amplitudes and frequencies of the input signal,
which indicates that the Bouc–Wen model cannot capture the rate dependency of the
hysteresis precisely.
2.4 Experimental Studies 37

(a) 8

6
Input voltage (V)
5

0
0 2 4 6 8 10
Time (s)

(b) 90

80

70
Displacement (µm)

60

50

40

30

20

10

0
0 2 4 6 8 10
Time (s)

Fig. 2.8 Time history of the training data sets. a Input voltage to the high-voltage amplifier.
b Output displacement of the nanopositioning stage

2.4.2 MPI Model Results

In order to identify the MPI model, the key step lies in the weight parameter deter-
mination for matching the model output to the experimental hysteresis data. Based
on the input and output data sets, the thresholds ri and di∗ are assigned as follows:
38 2 Feedforward Control Based on Inverse Hysteresis Models

(a) 90
Experiment
80 Bouc−Wen model

70

Displacement (µm) 60

50

40

30

20

10

0
0 2 4 6 8 10
Time (s)

(b) 5

4
Displacement error (µm)

−1

−2

−3
0 2 4 6 8 10
Time (s)

Fig. 2.9 Results of the identified Bouc–Wen model. a Experimental result and Bouc–Wen model
output. b Bouc–Wen model output errors

i
ri = max |u(t)|, for i = 0, 1, . . . , n (2.38)
n+1
j
d ∗j = max |y(t)|, for j = 0, 1, . . . , m (2.39)
m+1

where u and y denote the voltage input and displacement output data of the piezostage
system, respectively. In addition, the initial states are assigned as y0i = 0, i = 0, 1,
. . ., n, for simplicity.
2.4 Experimental Studies 39

Table 2.1 Parameters of the identified Bouc–Wen hysteresis model

Parameter Search space Identified value
m [0, 1] 0.2789 kg
b [102 , 106 ] 8.9055 × 105 N s/m
k [103 , 109 ] 3.5244 × 108 N/m
d [0, 0.1] 1.1351 × 10−5 m/V
α [0, 1] 0.4754
β [0, 1] 0.8718
γ [0, 1] 0.9925

Then, the weights wh and w∗s are identified by solving an optimization problem
of minimizing the model error:

E[u, y](wh , w∗s , t) = whT Hr [u, y0 ](t) − w∗s T S∗d [y](t). (2.40)

Here, instead of using L 22 -norm optimization [10], the weight parameters are
identified by resorting to the PSO approach. Additionally, n = 12 and m = 6 are
assigned, and the thresholds ri and d ∗j are allocated according to Eqs. (2.38) and
(2.39). Using the same training data sets as shown in Fig. 2.8a, b, the MPI model is
identified off-line by optimizing a total of n + m + 2 = 20 weight parameters (wh
and w∗s ) within the search space of [0, 1] using the PSO algorithm [2].
The identified MPI model and inverse MPI model parameters are described in
Tables 2.2 and 2.3, respectively. The presence of zero weight values of whi and ws∗j
indicates that the selected n and m are large enough for the modeling process.
The comparisons of the experimental output and simulated MPI model output
are illustrated in Fig. 2.10. It is observed from Fig. 2.10a that the MPI model cannot

Table 2.2 Parameters of the identified MPI hysteresis model

i ri whi j dj ws j
0 0 0.0677 0 0 99.9999
1 0.6036 0.0228 1 0.1276 −1.0010 × 10−5
2 1.2072 6.5979 × 10−6 2 0.2552 −3.3798 × 10−5
3 1.8108 0.0196 3 0.3827 −1.5611 × 10−5
4 2.4145 0.0037 4 0.5103 0
5 3.0181 4.0252 × 10−6 5 0.6379 −2.6054 × 10−4
6 3.6217 1.0681 × 10−6 6 0.7655 −6.8291 × 10−5
7 4.2253 6.7711 × 10−8
8 4.8289 6.1372 × 10−8
9 5.4325 3.8439 × 10−8
10 6.0362 2.9690 × 10−8
11 6.6398 8.4374 × 10−9
12 7.2434 0
40 2 Feedforward Control Based on Inverse Hysteresis Models

Table 2.3 Parameters of the inverse MPI hysteresis model

i ri∗ ∗
whi j d ∗j ws∗j
0 0 14.7797 0 0 0.0100
1 0.0408 −3.7195 1 12.7581 1.0010 × 10−9
2 0.0954 −8.0705 × 10−4 2 25.5163 1.3798 × 10−9
3 0.1500 −1.9672 3 38.2744 1.5611 × 10−9
4 0.2164 −0.2980 4 51.0326 0
5 0.2850 −3.1129 × 10−4 5 63.7907 2.6054 × 10−8
6 0.3537 −8.2597 × 10−5 6 76.5488 6.8292 × 10−9
7 0.4223 −5.2362 × 10−6
8 0.4909 −4.7460 × 10−6
9 0.5596 −2.9725 × 10−6
10 0.6282 −2.2959 × 10−6
11 0.6969 −6.5247 × 10−7
12 0.7655 0

exactly represent the complicated hysteresis of the nanopositioning system. A relative

large error exists between the identified model output and the experimental result as
shown in Fig. 2.10b. Specifically, the maximum model error is 6.35 µm which occurs
along with the input component of the highest frequency (0–0.38 s). It accounts
for 7.1 % of the concerned travel range of the nanopositioning stage. On the other
hand, the model error is reduced at lower frequency with smaller amplitude of the
input signal, that is, the model errors vary significantly at different frequencies and
amplitudes of the input signal, which means that the MPI model cannot capture the
rate dependency of the hysteresis either. Due to the existence of zero weights in
wh and w∗s , the increases of m and n will not cause significant improvement on the
model accuracy. Hence, the capability of the MPI model is limited in modeling the
rate-dependent hysteresis.

2.4.3 LSSVM Model Results

To identify the LSSVM model, the 500 sets of input and output data sets as shown in
Fig. 2.8 are also adopted to train the LSSVM. To capture the rate dependency behav-
ior, the input rate is numerically calculated by the backward difference equation:

u(kT ) − u(kT − T )
u̇(kT ) = (2.41)
T
where T is the sampling time and k denotes the index of the time series. The two
input variables are shown in Fig. 2.11.
2.4 Experimental Studies 41

(a) 90
Experiment
80 MPI model

70

Displacement (µm) 60

50

40

30

20

10

0
0 2 4 6 8 10
Time (s)

(b) 8

6
Displacement error (µm)

−2

−4

−6
0 2 4 6 8 10
Time (s)

Fig. 2.10 Results of the identified MPI model. a Experimental result and MPI model output.
b MPI model output errors

To facilitate the LSSVM modeling, the data sets are normalized before the training
process. Specifically, both the input and output data are normalized within the range
of [0, 1] by the following transformation [17]:

v − vmin
N (v) = v∗ = (2.42)
vmax − vmin

where N represents the normalization operation, and vmin and vmax denote the
minimum and maximum values of the input or output data v, respectively. It is
notable that, after the training process, the LSSVM output y ∗ is re-transformed into
the actual value y using the inverse normalization operation N −1 .
42 2 Feedforward Control Based on Inverse Hysteresis Models

Input voltage (V)

6

0
0 2 4 6 8 10
Time (s)

100
Voltage rate (V/s)

50

−50

−100
0 2 4 6 8 10
Time (s)

Fig. 2.11 Training data sets of the two input variables for the LSSVM model

Given arbitrary initial values (γ0 = 100 and σ0 = 1), the optimized LSSVM
model hyperparameters are obtained by the Bayesian framework automatically. The
optimal hyperparameters are γ = 741.57 and σ = 7.39. They are used to train the
LSSVM to obtain the values of αi and b, which minimizes the objection function
(2.22).
Once the training process is completed, the LSSVM model produces the outputs
as illustrated in Fig. 2.12. It is seen that the maximum model error is only 1.27 µm,
i.e., 1.4 % of the motion range. Thus, with the same number of training data sets, the
LSSVM model error has been reduced by 79 and 80 % as compared with the Bouc–
Wen and MPI model errors, respectively. In addition, Fig. 2.12b exhibits that the
LSSVM model errors are more uniformly distributed in comparison with Bouc–Wen
and MPI model results as shown in Figs. 2.9b and 2.10b, respectively. Relatively, the
LSSVM model errors are not dependent on neither the amplitude nor the frequency
of the input signals. Therefore, the trained LSSVM model captures the amplitude-
and rate-dependent hysteresis accurately.

2.4.4 Model Capability Comparison

Based on the displacement error e = yd − y, where yd and y represent the desired

and actual displacements, respectively, the mean absolute error (MAE) and root-
mean-square error (RMSE) are defined as follows:
2.4 Experimental Studies 43

(a) 90
Experiment
80 LSSVM model

70

Displacement (µm) 60

50

40

30

20

10

0
0 2 4 6 8 10
Time (s)

(b) 2

1.5

1
Model error (µm)

0.5

−0.5

−1

−1.5

−2
0 2 4 6 8 10
Time (s)

Fig. 2.12 Results of the trained LSSVM model. a Experimental result and LSSVM model output.
b LSSVM model output errors

1 
Nt
MAE = |ei | (2.43)
Nt
i=1


 1  Nt
RMSE =  ei2 (2.44)
Nt
i=1

where Nt = 500 is the number of test data sets in the experiment.

The modeling performances of the three models are tabulated in Table 2.4. Due to
the fewer number of parameters, the Bouc–Wen model takes 19.1 % shorter time than
44 2 Feedforward Control Based on Inverse Hysteresis Models

Table 2.4 Modeling performances of Bouc–Wen, MPI, and LSSVM models

Specification Bouc–Wen MPI LSSVM
No. of model parameters 7 20 2
Model identification time (s) 614.89 759.91 2.56
Modeling error
MAE (%) 1.276 1.358 0.364
RMSE (%) 1.560 1.692 0.470
Generalization error
MAE (%) 1.253 1.318 0.621
RMSE (%) 1.530 1.621 0.858

the MPI model to identify all the parameters by resorting to the same optimization
approach (PSO) in 1500 epoches. On the other hand, the LSSVM model needs
a significantly shorter time than both Bouc–Wen and MPI models to identify the
model parameters.
Concerning the modeling error, it is observed that the Bouc–Wen model produces
6.0 % lower MAE and 7.8 % lower RMSE than the MPI model. In contrast, the
LSSVM model achieves the best results among the three methods. Regarding MAE,
the LSSVM has improved the accuracy by 71.5 and 73.2 % in comparison with Bouc–
Wen and MPI models, respectively. As far as RMSE is concerned, the LSSVM has
enhanced the accuracy by 69.9 and 72.2 % as compared with Bouc–Wen and MPI
models, respectively.

2.4.5 Generalization Study

To test the generalization of the obtained models, a new input signal is chosen as
follows:

u(t) = 5e−0.14t [cos(3π te−0.092t − 3.15) + 1.0] (2.45)

which is depicted in Fig. 2.13a. In addition, the displacement output of the nanopo-
sitioning stage is shown in Fig. 2.13b.
The Bouc–Wen, MPI, and LSSVM model outputs are depicted in Fig. 2.14. The
model output errors with respect to the actual output obtained by experiments are
illustrated in Fig. 2.15. Table 2.4 describes the generalization testing results of the
three types of models.
It is found that the Bouc–Wen model produces a MAE of 1.086 µm and a RMSE
of 1.326 µm, which accounts for 1.25 and 1.53 % of the motion range, respectively.
Using the MPI model, the MAE and RMSE are 1.143 and 1.405 µm, which are
equivalent to 1.32 and 1.62 % of the overall motion range, respectively. As compared
2.4 Experimental Studies 45

(a) 8

Input voltage (V) 6

0
0 2 4 6 8 10
Time (s)

(b) 90
80

70
Displacement (µm)

60

50

40

30

20

10

0
0 2 4 6 8 10
Time (s)

Fig. 2.13 Time history of the testing data sets. a Input voltage to the high-voltage amplifier.
b Output displacement of the nanopositioning stage

with MPI model, the Bouc–Wen model reduces the MAE and RMSE by 4.9 and 5.6 %,
respectively. Thus, the generalization ability of the Bouc–Wen model is slightly better
than that of the MPI model.
In contrast, the LSSVM model produces a MAE of 0.538 µm and a RMSE of
0.744 µm, i.e., 0.62 and 0.86 % of the overall motion range, which have been reduced
by 50.5 and 43.9 % in comparison with the Bouc–Wen model results and by 52.9
46 2 Feedforward Control Based on Inverse Hysteresis Models

Fig. 2.14 Model (a) 90

generalization testing results.
80
a Boun-Wen model output.
b MPI model output. 70
c LSSVM model output

Displacement (µm)
60

50

40

30

20

10 Experiment
Bouc–Wen model
0
0 1 2 3 4 5 6 7 8
Input voltage (V)

(b) 90

80

70
Displacement (µm)

60

50

40

30

20

10 Experiment
MPI model
0
0 1 2 3 4 5 6 7 8
Input voltage (V)

(c) 90

80

70
Displacement (µm)

60

50

40

30

20

10 Experiment
LSSVM model
0
0 1 2 3 4 5 6 7 8
Input voltage (V)
2.4 Experimental Studies 47

6
Bouc–Wen model
5 MPI model
LSSVM model
4
Displacement error (µm) 3

−1

−2

−3

−4
0 2 4 6 8 10
Time (s)

Fig. 2.15 Displacement errors of the Bouc–Wen model, MPI model, and LSSVM model

and 47.0 % in comparison with the MPI model outputs, respectively. Therefore,
the experimental results demonstrate a good generalization capacity of the LSSVM
hysteresis model.

2.5 Controller Design and Verification

Based on the established hysteresis models, a feedforward plus feedback control

scheme is implemented and verified for the hysteresis compensation in this section.
Both the inverse Bouc–Wen model and inverse MPI model are adopted for compar-
isons with the LSSVM inverse model-based control scheme.

2.5.1 Feedforward Controller Design

To compensate for the hysteresis nonlinearity, a feedforward (FF) control based on

the inverse hysteresis model can be constructed. The FF control effort u FBW
F is created
by the inverse Bouc–Wen model. Similarly, the inverse MPI model is used to produce
the FF control signal u M PI
F F . The overall control scheme is shown in Fig. 2.16a, where
the feedback (FB) control u F B is adopted to suppress the residual control error due
to the model errors of the identified Bouc–Wen and MPI models.
48 2 Feedforward Control Based on Inverse Hysteresis Models

(a)
BW ( MPI )
yd (kT ) Inv. BW or Inv. MPI uFF (kT )
model
+ u (kT ) Piezostage y (kT )
+ e(kT ) plant
PID +
controller u FB (kT )
−

(b) u ( kT ) Piezostage y (kT )

plant
+ Zero-phase
filter
−
u SVM (kT ) LSSVM +
inv. model 1/T −1
− z
Off-line training
yd (kT ) Online control
SVM
u ( kT )
+ LSSVM FF
z −1 − 1/T inv. model +
u (kT ) Piezostage y ( kT )
plant
+ e(kT ) +
PID
controller u FB (kT )
−

Fig. 2.16 Block diagram of the feedforward (FF) plus feedback (FB) hybrid controller for a
piezostage system. a Bouc–Wen model or MPI model-based control. b LSSVM model-based control

To establish an inverse LSSVM hysteresis model, the displacement (y) and veloc-
ity ( ẏ) are selected as the inputs and the corresponding voltage (u) is assigned as the
output for the LSSVM training. Preliminary investigations indicate that the noisy sen-
sor readings produce slightly oscillating voltage input by the inverse LSSVM model.
Thus, to alleviate the noises in the displacement sensor readings, the displacement
data y are filtered by adopting a zero-phase digital filter with 50-Hz cutoff frequency.
The filter is implemented with the command “filtfilt” in MATLAB. Afterward, the
input data sets are formed by the filtered displacement data and the calculated veloc-
ity. Once trained off-line using the scheme as shown in Fig. 2.16b, the LSSVM inverse
model provides the online FF control signal u SV M
FF .
2.5 Controller Design and Verification 49

2.5.2 Feedforward Plus Feedback Controller Design

Due to the presence of modeling error, the hysteresis cannot be completely eliminated
by the stand-alone inverse model-based FF compensator. Therefore, a FB control u F B
is employed to create a hybrid control as described in Fig. 2.16.
Here, the PID control algorithm is implemented as a FB control due to its robust-
ness and popularity. By adopting an incremental PID algorithm, the overall control
input can be derived in the discretized form:

u(kT ) = u F F (kT ) + u F B (kT )

= u F F (kT ) + u F B (kT − T )
+K p [e(kT ) − e(kT − T )] + K i e(kT )
+K d [e(kT ) − 2e(kT − T ) + e(kT − 2T )] (2.46)

where e represents the displacement error, u F B (kT − T ) is the FB control command

in the previous time step, and the FF term u F F (kT ) is given by the inverse MPI
model or LSSVM inverse hysteresis model. Additionally, K p , K i , and K d are the
proportional, integral, and derivative gains, respectively.

2.5.3 Controller Verification

In the control experiments, the sampling time interval is assigned as T = 0.001 s.

Actually, a PI control is employed. The control gains are tuned as K p = 0.0864,
K i = 31.1040, and K d = 0 using the Ziegler–Nichols (Z–N) method.
For a reference displacement as shown in Fig. 2.17a, the Bouc–Wen (BW) and
MPI model-based FF control results are both shown in Fig. 2.17. The two FF plus
PID FB control results are depicted in Fig. 2.18. In addition, the LSSVM inverse
hysteresis model is trained with the hyperparameters γ = 741.82 and σ = 7.69,
which are optimized by the Bayesian inference framework. Using the same reference
input as shown in Fig. 2.17a, the control results and tracking errors of the LSSVM
model-based FF and FF+FB control are also shown in Figs. 2.17 and 2.18.
Figure 2.17b reveals that the inverse Bouc–Wen model-based FF gives a MAE of
1.199 % and a RMSE of 1.660 % with respect to the motion range. On the other hand,
the FF control using the inverse MPI model produces a MAE of 1.080 % and a RMSE
of 1.813 %. It is seen that the Bouc–Wen model and MPI model produce similar
results. While the former achieves a slightly better RMSE than the latter, the latter
produces a better MAE than the former. By contrast, the LSSVM inverse model-based
FF approach further reduces the MAE and RMSE to 0.416 and 0.537 %, respectively,
which have been significantly improved by 65.3 and 67.7 % in comparison with the
inverse Bouc–Wen model results and by 61.5 and 70.4 % in comparison with the
inverse MPI model results, respectively.
50 2 Feedforward Control Based on Inverse Hysteresis Models

(a) 90 Reference
BW FF
80 MPI FF
LSSVM FF
70
Displacement (µm)
60

50

40

30

20

10

0
0 2 4 6 8 10
Time (s)

(b) 6

4
Displacement error (µm)

−2

−4

BW FF
−6
MPI FF
LSSVM FF
−8
0 2 4 6 8 10
Time (s)

Fig. 2.17 Bouc–Wen, MPI, and LSSVM-based FF control results. a Feedforward (FF) control
results. b Control errors of the three FF methods

Moreover, the hybrid control results as shown in Fig. 2.18b indicate that the inverse
Bouc–Wen model-based FF+FB control scheme leads to a MAE of 0.024 % and a
RMSE of 0.048 %. In addition, the inverse MPI model-based FF+FB control scheme
results in a MAE of 0.023 % and a RMSE of 0.052 %, while the LSSVM-based
hybrid control substantially suppresses the MAE to 0.015 % and RMSE to 0.038 %.
As compared to the inverse Bouc–Wen hysteresis model, the LSSVM inverse model-
based FF+FB control reduces the tracking errors in terms of MAE and RMSE by 37.5
and 20.8 %, respectively. Moreover, the LSSVM inverse model-based FF+FB control
2.5 Controller Design and Verification 51

(a)
90 Reference
BW FF+FB
80 MPI FF+FB
LSSVM FF+FB
Displacement (µm) 70

60

50

40

30

20

10

0
0 2 4 6 8 10
Time (s)

(b) 0.4

0.3
Displacement error (µm)

0.2

0.1

−0.1

−0.2
BW FF+FB
−0.3 MPI FF+FB
LSSVM FF+FB
−0.4
0 2 4 6 8 10
Time (s)

Fig. 2.18 Bouc–Wen, MPI, and LSSVM-based FF plus FB control results. a FF plus feedback
(FB) control results. b Control errors of the three hybrid approaches

suppresses the MAE and RMSE by 34.8 and 26.9 %, respectively, as compared with
the inverse MPI model-based control. This is caused by the fact that the model
accuracy of LSSVM is superior to that of both Bouc–Wen and MPI models. The
effectiveness of the LSSVM-based control is evident from the control results.
In addition, to reveal the efficiency of the hybrid control, the control results of
the stand-alone FF and FB and the combined FF+FB control based on the LSSVM
model are illustrated in Fig. 2.19. It can be observed from Fig. 2.19b that the FF+FB
produces the MAE and RMSE of 0.015 and 0.038 %, respectively, which have been
52 2 Feedforward Control Based on Inverse Hysteresis Models

(a)
90 90 Reference
85 FF
80 80 FB
75 FF+FB
70 0.15 0.2 0.25
Displacement (µm)
60

50 20

40
15
2.5 3 3.5
30

20

10

0
0 2 4 6 8 10
Time (s)

(b) 2.5

1.5
Displacement error (µm)

0.5

−0.5

−1

−1.5
FF
−2 FB
FF+FB
−2.5
0 2 4 6 8 10
Time (s)

Fig. 2.19 LSSVM-based stand-alone and hybrid control results. a FF, FB, and FF+FB results.
b Tracking errors of the three methods

significantly improved by 96.5 and 92.9 % in comparison with the FF control results,
and substantially enhanced by 93.7 and 91.7 % as compared with the stand-alone FB
outputs, respectively. As a result, the hysteresis effects have been suppressed by the
FF+FB approach to a negligible level as indicated in Fig. 2.20a. Although the PID
FB control results can be improved by tuning the control gains larger, it is at the risk
of oscillations in the hybrid control results. With the hybrid control, the individual
control input components are illustrated in Fig. 2.20b. It is found that the overall
control effort is dominated by the FF control, whereas the FB control effort is minor.
2.5 Controller Design and Verification 53

(a) 100
30
90

80 25

Actual displacement (µm) 70 20

60 15
15 20 25 30
50

40

30

20
FF
10 FB
FF+FB
0
0 20 40 60 80 100
Reference displacement (µm)

(b) 8
FF
7 FB
FF+FB
6

5
Control effort (V)

−1
0 2 4 6 8 10
Time (s)

Fig. 2.20 LSSVM-based control results. a Actual versus reference displacements obtained by the
three approaches. b Control signal components of the FF+FB hybrid control

Even so, the FB control is capable of suppressing the residual tracking errors and
leading to a better result than the stand-alone FF method, which is verified by the
tracking errors as compared in Fig. 2.19b.
For the purpose of comparison, the control results obtained by the Bouc–Wen,
MPI, and LSSVM inverse models are tabulated in Table 2.5, where the MAE and
RMSE are calculated as percentage values with respect to the motion range. It is
obvious that the FF control accuracy of the LSSVM model has been enhanced by over
65 and 60 % as compared with those of the Bouc–Wen and MPI models, respectively.
54 2 Feedforward Control Based on Inverse Hysteresis Models

Table 2.5 Control performances using Bouc–Wen, MPI, and LSSVM models
Controller Performance (%) Bouc–Wen MPI LSSVM
FF MAE 1.199 1.080 0.416
RMSE 1.660 1.813 0.537
FF+FB MAE 0.024 0.023 0.015
RMSE 0.048 0.052 0.038

With the FF+FB hybrid control, the LSSVM approach has suppressed the control
errors by more than 20 and 25 % in comparison with the Bouc–Wen and MPI models,
respectively. Therefore, the superiority of the LSSVM model over the popular Bouc–
Wen and MPI models is verified for the hysteresis compensation task.

2.6 Chapter Summary

The results presented in this chapter show that the rate-dependent hysteresis of a
piezoelectric nanopositioning stage can be accurately modeled and effectively sup-
pressed by the LSSVM regression model and LSSVM inverse model, respectively.
By selecting the input variation rate as an auxiliary input variable, the multi-valued
mapping due to the hysteresis nonlinearity is converted into a one-to-one mapping,
and the LSSVM is trained to capture the rate-dependent hysteretic behavior. Its supe-
rior modeling performance over the popular Bouc–Wen and MPI hysteresis models
has been revealed by experimental studies. Results also demonstrate that the hybrid
control using the LSSVM inverse model-based feedforward control combined with
a PID control is capable of compensating the hysteresis nonlinearity effectively. Due
to a simple structure of the presented modeling and control framework, it can be eas-
ily extended to the hysteretic systems driven by shape-memory alloy or other types
of smart actuators as well. As a future work, an incremental LSSVM model can be
established to automatically update the model parameters online once new training
data occur.

References

1. Ang, W.T., Khosla, P.K., Riviere, C.N.: Feedforward controller with inverse rate-dependent
model for piezoelectric actuators in trajectory-tracking applications. IEEE/ASME Trans.
Mechatron. 12(2), 134–142 (2007)
2. Birge, B.: PSOt—a particle swarm optimization toolbox for use with Matlab. In: Proceedings
of IEEE Swarm Intelligence Symposium, pp. 182–186. Indianapolis, Indiana, USA (2003)
3. Boukari, A.F., Carmona, J.C., Moraru, G., Malburet, F., Chaaba, A., Douimi, M.: Piezo-
actuators modeling for smart applications. Mechatronics 21(1), 339–349 (2011)
References 55

4. De Brabanter, K., Karsmakers, P., Ojeda, F., Alzate, C., De Brabanter, J., Pelckmans, K.,
De Moor, B., Vandewalle, J., Suykens, J.A.K.: LS-SVMlab toolbox user’s guide version 1.7.
Internal Report 10-146, ESAT-SISTA, K.U.Leuven, Leuven, Belgium (2010)
5. Dong, R., Tan, Y., Chen, H., Xie, Y.: A neural networks based model for rate-dependent
hysteresis for piezoelectric actuators. Sens. Actuator A-Phys. 143(2), 370–376 (2008)
6. Ge, P., Jouaneh, M.: Tracking control of a piezoceramic actuator. IEEE Trans. Contr. Syst.
Technol. 4(3), 209–216 (1996)
7. Janaideh, M.A., Rakheja, S., Su, C.Y.: Experimental characterization and modeling of rate-
dependent hysteresis of a piezoceramic actuator. Mechatronics 19(5), 656–670 (2009)
8. Juhasz, L., Maas, J., Borovac, B.: Parameter identification and hysteresis compensation of
embedded piezoelectric stack actuators. Mechatronics 21(1), 329–338 (2011)
9. Kim, J., Kang, B.: Micro-macro linear piezoelectric motor based on self-moving cell. Mecha-
tronics 19(7), 1134–1142 (2009)
10. Kuhnen, K.: Modeling, identification and compensation of complex hysteretic nonlinearities:
a modified Prandtl-Ishlinskii approach. Eur. J. Control 9(4), 407–421 (2003)
11. Li, Y., Xu, Q.: Adaptive sliding mode control with perturbation estimation and PID sliding
surface for motion tracking of a piezo-driven micromanipulator. IEEE Trans. Contr. Syst.
Technol. 18(4), 798–810 (2010)
12. Lin, C.J., Chen, S.Y.: Evolutionary algorithm based feedforward control for contouring of a
biaxial piezo-actuated stage. Mechatronics 19(6), 829–839 (2009)
13. Suykens, J.A.K., Vandewalle, J.: Least squares support vector machine classifiers. Neural
Process. Lett. 9(3), 293–300 (1999)
14. Suykens, J.A.K., Gestel, T.V., Brabanter, J.D., Moor, B.D., Vandewalle, J.: Least Squares
Support Vector Machines. World Scientific Publishing Co., Singapore (2002)
15. Van Gestel, T., Suykens, J.A.K., Baestaens, D.E., Lambrechts, A., Lanckriet, G., Vandaele, B.,
De Moor, B., Vandewalle, J.: Financial time series prediction using least squares support vector
machines within the evidence framework. IEEE Trans. Neural Netw. 12(4), 809–821 (2001)
16. Vong, C.M., Wong, P.K., Li, Y.P.: Prediction of automotive engine power and torque using
least squares support vector machines and Bayesian inference. Eng. Appl. Artif. Intell. 19(3),
277–287 (2006)
17. Wong, P.K., Vong, C.M., Tam, L.M., Li, K.: Data preprocessing and modelling of electronically-
controlled automotive engine power performance using kernel principal components analysis
and least-square support vector machines. Int. J. Veh. Syst. Model. Test. 3(4), 312–330 (2008)
18. Xu, Q., Li, Y.: Dahl model-based hysteresis compensation and precise positioning control of an
XY parallel micromanipulator with piezoelectric actuation. J. Dyn. Syst. Meas. Control-Trans.
ASME 132(4), 041,011 (2010)
19. Yang, J., Bouzerdoum, A., Phung, S.L.: A training algorithm for sparse LS-SVM using com-
pressive sampling. In: Proceedings of IEEE International Conference on Acoustics Speech and
Signal Processing, pp. 2054–2057 (2010)
20. Yu, S., Alici, G., Shirinzadeh, B., Smith, J.: Sliding mode control of a piezoelectric actuator
with neural network compensating rate-dependent hysteresis. In: Proc. of IEEE Int. Conf. on
Robotics and Automation, pp. 3641–3645 (2005)
21. Yu, Y., Xiao, Z., Naganathan, N.G., Dukkipati, R.V.: Dynamic Preisach modelling of hysteresis
for the piezoceramic actuator system. Mech. Mach. Theory 37(1), 75–89 (2002)
Chapter 3
Feedforward Control Without Modeling
Inverse Hysteresis

This chapter presents an approach of hysteresis identification and compensation of

piezoelectric actuators by resorting to an intelligent hysteresis model. In particular, a
least squares support vector machine (LSSVM)-based hysteresis model is developed
and used for both purposes of hysteresis identification and hysteresis compensa-
tion. By this way, the inverse hysteresis is not needed in the feedforward hysteresis
compensator because the hysteresis model is directly used. The effectiveness of the
presented idea is validated by a series of experimental studies on a piezoactuated
micro-/nano-positioning system.

3.1 Introduction

Concerning the approach of model-based hysteresis compensation, the hysteresis

effect is generally compensated for by employing an inverse hysteresis model. It
follows that both a hysteresis model and an inverse hysteresis model are needed for
the purposes of hysteresis identification and compensation. Hence, it is desirable to
mitigate the hysteresis effect by adopting a hysteresis model directly without using
the inverse hysteresis [5].
In addition, it has been shown that the piezoelectric hysteresis effect is rate-
dependent [2]. The hysteresis behavior is dependent not only on the amplitude but
also on the frequency of input voltage signals. As compared with ANN and SVM,
LSSVM is superior in terms of modeling accuracy and computational efficiency.
Thus, LSSVM is selected for the hysteresis modeling process. Owing to the hysteresis
nonlinearity, one value of input voltage corresponds to multiple position outputs for a
piezoactuated device. Therefore, one challenge in identifying the hysteresis behavior
with LSSVM lies in how to transfer the multivalued mapping into a single-valued one.
In previous work, a one-to-one mapping was established by introducing the current
input and input variation rate as one data set [7, 8]. However, in case that the input
data are accompanied with noises, the variation rate is non-smooth, which causes
© Springer International Publishing Switzerland 2016 57
Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2_3
58 3 Feedforward Control Without Modeling Inverse Hysteresis

modeling error. Besides, it is unknown how many orders of the variation rates are
sufficient to construct the mapping. Additionally, a regression-type hysteresis model
was developed in [6] by employing the current and previous inputs and previous
outputs as exogenous inputs. However, this work considered the position and voltage
as output and input variables, respectively, and both direct and inverse hysteresis
models are required, which implies a time-consuming process.
This chapter presents the identification and compensation of rate-dependent piezo-
electric hysteresis using an intelligent hysteresis model without modeling the inverse
hysteresis. A LSSVM-based hysteresis model is established, and a feedforward com-
pensator is developed based on the single model, which provides a computationally
efficient approach in hysteresis compensation. To establish the LSSVM model, the
problem of how to select input variables to convert the multivalued mapping into a
single-valued one is addressed in detail. The effectiveness of the idea is verified by a
series of experimental studies. It will be shown that the established LSSVM model
is more effective than Bouc–Wen model in terms of hysteresis identification as well
as hysteresis compensation performances.

3.2 Dynamics Modeling of Hysteretic System

3.2.1 Dynamics Modeling with Bouc–Wen Hysteresis

Owing to a fewer number of parameters, the Bouc–Wen model has been exten-
sively applied in piezoelectric hysteresis modeling. The entire dynamics model of a
piezoactuated system can be established as follows [3, 4]:

M ÿ(t) + B ẏ(t) + K y(t) = K [Du(t) − H (t)] (3.1)

Ḣ (t) = α D u̇(t) − β|u̇(t)|H (t) − γ u̇(t)|H (t)| (3.2)

where t is the time variable; parameters M, B, K , and y represent the mass, damp-
ing coefficient, stiffness, and displacement response of the piezoactuated system,
respectively; D is the piezoelectric coefficient; u denotes the input voltage; and H
indicates the hysteretic loop in terms of displacement whose magnitude and shape
are determined by parameters α, β, and γ .
Once the dynamics parameters M, B, and K are determined, the four parameters
(D, α, β, and γ ) of the Bouc–Wen model can be identified by minimizing an objective
function:

1 
N
2
f (D, α, β, γ ) = yi − yBWi (3.3)
N
i=1
3.2 Dynamics Modeling of Hysteretic System 59

where N denotes the total number of samples, and yi − yBWi represents the residual
error of the ith sample which is the discrepancy between Bouc–Wen model output
(yBWi ) and experimental result (yi ).
Given the input voltage u and Bouc–Wen hysteresis model H , the output dis-
placement of the system is governed by the model (3.1). On the other hand, given a
desired displacement yd , the required input voltage can be determined via Eq. (3.1):

1
u FF (t) = [M ÿd (t) + B ẏd (t) + K yd (t) + K H (t)]. (3.4)
KD
It is found that the feedforward controller (3.4) is obtained using the hysteresis
term H without solving the inverse hysteresis model.

3.2.2 Dynamics Modeling with Intelligent Hysteresis Model

By extending the aforementioned idea of hysteresis compensation using Bouc–Wen

model where an inverse hysteresis model is not needed, an intelligent hysteresis
model is proposed as follows.
First, the dynamics model (3.1) of the system is rewritten into the following form:

ÿ(t) + 2ξ ωn ẏ(t) + ωn2 y(t) = du(t) + h(t) (3.5)

where ξ and ωn denote the damping ratio and natural frequency of the piezo-actated
system, respectively; d is a positive parameter; and h represents the hysteresis effect
in terms of acceleration.
Then, in view of Eq. (3.5), the hysteresis term can be generated below:

h(t) = ÿ(t) + 2ξ ωn ẏ(t) + ωn2 y(t) − du(t). (3.6)

In practice, the system parameters and output positions are not always available.
Under such scenarios, it is necessary to establish an intelligent model to identify and
compensate for the hysteresis term h.
Once an intelligent model of the hysteresis term ĥ is obtained, a feedforward
hysteresis compensator can be developed to create a desired output position yd :

1
u FF (t) = [ ÿd (t) + 2ξ ωn ẏd (t) + ωn2 yd (t) − ĥ(t)] (3.7)
d
which uses the hysteresis model directly without solving the inverse hysteresis model.
An intelligent hysteresis model is established based on LSSVM technique in the
subsequent section.
60 3 Feedforward Control Without Modeling Inverse Hysteresis

3.3 Hysteresis Modeling Using LSSVM

LSSVM maps the input data into a high-dimensional feature space and constructs
a linear regression function therein [7]. As is known, the hysteresis exhibits multi-
valued and non-smooth features [1]. Generally, the LSSVM is only able to estimate
high-dimensional smooth functions with a satisfactory accuracy. Therefore, only the
multivalued nonlinearity of hysteresis effect is handled here, that is, it is assumed
that the signal of input command is smooth. With such a hypothesis, a LSSVM-based
regression model is established to represent the piezoelectric hysteresis effect.

3.3.1 Regression Model Establishment

Equation (3.6) indicates that the output variable is the hysteresis term h, whereas
input variables include both input voltage u and output position y.
Using LSSVM, a nonlinear regression model is formulated to describe the hys-
teresis effect as follows:

ĥ k = f (xk ) (3.8)

where

xk = [u k , . . . , u k−m , yk , . . . , yk−n , h k−1 , . . . , h k−l ] (3.9)

with ĥ k denoting the hysteresis term predicted by LSSVM at the current time instant
k. u k−1 , yk−1 , and h k−1 are the input voltage, output position, and hysteresis term at
the previous time instant k − 1, respectively. In addition, the order of the model is
defined by m (m ≥ 0), n (n ≥ 0), and l (l ≥ 1). The problem of how to select these
three orders is discussed later.

3.3.2 LSSVM Modeling

Assume that the training data set is given by {xk , h k }k=1

N , where N represents the

number of training data set, xk ∈ Rm+n+l+2 is an input vector as shown in Eq. (3.9),
and h k ∈ R are the output data. The unknown hysteresis function can be approxi-
mated by the following equation:

ĥ(x) = wT ϕ(x) + b (3.10)

where w is a weight vector, ϕ(·) denotes a nonlinear mapping from the input space
to a higher dimensional feature space, and b is the bias.
3.3 Hysteresis Modeling Using LSSVM 61

The LSSVM approach formulates the regression as an optimization problem in

the primal weight space. Then, the conditions for optimality are obtained by solving
a series of partial derivatives, which are used to construct the dual formulation as
follows:
    
0 1TN b 0
= (3.11)
1 N Ω + Γ −1 I N α h

where α = [α1 , α2 , . . . , α N ]T is called the support vector, and αk = Γ ek are

support values with Γ ∈ R denoting the regularization factor. In addition, 1 N =
[1, 1, . . . , 1]T , h = [h 1 , h 2 , . . . , h N ]T , and I N is an identity matrix.
Then, the kernel trick is employed to derive the relationship:

Ωk j = ϕ(xk )T ϕ(x j ) = K (xk , x j ), k, j = 1, 2, . . . , N (3.12)

where K (·) is a predefined kernel function. The role of the kernel function is to avoid
explicit computation of the map ϕ(·) in dealing with the high-dimensional feature
space.
Calculating b and α from Eq. (3.11), the solution to the regression problem is
obtained below:


N
ĥ(x) = αk K (x, xk ) + b (3.13)
k=1

where K (·) is the kernel function satisfying Mercer’s condition, xk is the training
data, and x denotes the new input data.
By employing the radial basis function (RBF) as kernel function,
 
x − xk 2
K (x, xk ) = exp − (3.14)
σ2

where σ > 0 denotes the width parameter (which specifies the kernel sample variance
σ 2 ) and  ·  represents the Euclidean distance, and the LSSVM model for the
hysteresis model estimation becomes


N  
x − xk 2
ĥ(x) = αk exp − + b. (3.15)
σ2
k=1

Assigning the regularization parameter Γ and kernel parameter σ , the objective

of training process is to determine the support values αk and the bias b. The high
generalization ability of the LSSVM model relies on appropriate tuning of these two
hyperparameters (Γ and σ ). Here, the leave-one-out cross-validation approach is
adopted to infer the values of the hyperparameters.
62 3 Feedforward Control Without Modeling Inverse Hysteresis

3.4 Experimental Studies on Hysteresis Identification

In this section, the hysteresis identification processes with Bouc–Wen model and
LSSVM model are carried out through experimental studies.

3.4.1 Experimental Setup

The experimental setup is depicted in Fig. 3.1. A four-layer piezoelectric bimorph

actuator (model: T434-A4-201, from Piezo Systems, Inc.) with the dimension of
28 × 5 × 0.86 mm3 is selected. The actuator is fabricated from the piezo-material of
Industry Type 5A (Navy Type II), and its key parameters are shown in Table 3.1. A
USB-6259 board (from National Instruments Corp.) equipped with 16-bit D/A and
A/D converters is adopted to produce an analog voltage, which is then amplified by
a high-voltage amplifier (model: EPA-104, from Piezo Systems, Inc.) to provide a
voltage of ±200 V for the drive of the piezoelectric actuator. The output displace-
ment at the end point of piezo-bimorph is measured by a laser displacement sensor
(model: LK-H055, from Keyence Corp.). The analog output voltage of the sensor

Fig. 3.1 Experimental setup of a piezoactuated system

Table 3.1 Main parameters Specification Parameter

of the piezoelectric bimorph
actuator Mass 1.8 × 10−3 kg
Stiffness 1.4 × 103 N/m
Capacitance 77 nF
Maximum voltage ±90 V
Resonant frequency 440 Hz
Free deflection ±262 µm
Blocked force ±0.36 N
3.4 Experimental Studies on Hysteresis Identification 63

signal conditioner is acquired by a PC through one A/D channel of the USB-6259

TM
board. LabVIEW software is adopted to implement a real-time control of the
piezoactuated system.

3.4.2 Dynamics Model Identification

Prior to the development of hysteresis models, a linear dynamics model of the system
plant is identified by the frequency response approach. Specifically, swept-sine waves
with an amplitude of 0.5 V and the frequency range of 1–1000 Hz are produced
to drive the piezo-bimorph. The position responses of the piezoelectric actuator
are recorded using a sampling rate of 2 kHz. The input–output data sets are used
to identify the plant transfer function by estimating the model from the frequency
response data. Figure 3.2 depicts the results of frequency responses obtained by
experiment and the identified model (3.16):

1.247 × 107
G(s) = . (3.16)
s 2 + 1.847s + 6.477 × 106

It is found that the first resonant mode occurs around 404 Hz. The identified
second-order model well captures the system dynamics in terms of magnitude and
phase at frequencies below 600 and 100 Hz, respectively, that is, the second-order
model cannot properly describe the phase behavior about 100 Hz. It implies that
a model of much higher order is required to capture the high-frequency dynamics

Fig. 3.2 System frequency 100

response obtained by
Magnitude (dB)

experimental study and the

50
identified second-order
model
0

−50 0 1 2 3
10 10 10 10

0
Phase (degrees)

−200

−400

−600 Experiment
Identified model
−800 0 1 2 3
10 10 10 10
Frequency (Hz)
64 3 Feedforward Control Without Modeling Inverse Hysteresis

more accurately. Here, a simple second-order model is employed to demonstrate the

effectiveness of the reported hysteresis identification scheme.

3.4.3 Bouc–Wen Model Results

Regarding the dynamics parameters, the mass is taken as the nominal value M =
1.8 × 10−3 kg as shown in Table 3.1. Additionally, by neglecting the nonlinear
term H in Eq. (3.1), taking the Laplace transform and comparing the parameters
with Eq. (3.16), the damping coefficient and stiffness values are calculated as B =
3.3 × 10−3 N · s/m and K = 1.1659 × 104 N/m, respectively. Other parameters are
identified below.

3.4.3.1 Bouc–Wen Model Identification

To identify the Bouc–Wen hysteresis model, various types of signals can be utilized.
Without loss of generality, an input voltage signal as shown in Fig. 3.3a is chosen
for the purpose of model identification:

u(t) = 50e−0.13t [cos(3π te−0.09t − 3.15) + 1.0]. (3.17)

By applying the signal to the piezoelectric bimorph actuator, the position output is
generated as depicted in Fig. 3.3b.
By selecting a time interval of 0.02 s, 500 training data sets are acquired as
shown in Fig. 3.3a, b. The Bouc–Wen model is then identified by optimizing the
four parameters so as to minimize the objective function (3.3). In particular, the
particle swarm optimization (PSO) is adopted for the function minimization, and
the optimum parameters are obtained as D = 2.4373 × 10−6 m/V, α = 0.1947, β =
3.3626, and γ = −2.8526.
It is notable that the optimized piezoelectric coefficient D = 2.4373 µm/V deviates
a little from the nominal value D = 2.91 µm/V calculated from Table 3.1. The
discrepancy can be explained by examining the experimental setup as shown in
Fig. 3.1. Owing to an offset of the laser measuring point away from the free end of
the bimorph, the actual length of the bimorph is less than the nominal value of 28
mm. Thus, the optimized D is slightly less than the nominal value.

3.4.3.2 Modeling Results

The results obtained by experiment and the Bouc–Wen model simulation are com-
pared in Fig. 3.3b–d. It is observed that the Bouc–Wen model cannot describe the
complicated hysteresis behavior of the piezoactuated system exactly. A relatively
large error exists between the model output and experimental result as shown in
3.4 Experimental Studies on Hysteresis Identification 65

(a) (b)
100 250
Experiment
Bouc−Wen model

Displacement (μm)
80 200
Input voltage (V)

60 150

40 100

20 50

0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)

(c) (d)
20 250
Displacement error (μm)

Displacement (μm) 200

10

150
0
100

−10
50
Experiment
Bouc−Wen model
−20 0
0 2 4 6 8 10 0 20 40 60 80 100
Time (s) Input voltage (V)

Fig. 3.3 Results of the identified Bouc–Wen model. a Input voltage. b Experimental result and
Bouc–Wen model output. c Bouc–Wen model output errors. d Displacement–voltage hysteresis
loops

Fig. 3.3c. The maximum model error is 15.56 µm, which accounts for 6.6 % of the
travel range of the piezoelectric actuator. Based on the displacement error e = yd − y,
the root-mean-square error (RMSE) is defined as follows:

1 Nt
RMSE =  ei2 (3.18)
Nt
i=1

where Nt = 500 is the number of test data sets.

The identified Bouc–Wen model produces a RMSE of 4.43 µm, which accounts
for 1.9 % of the travel range of the piezo-actuator. When the input has lower magni-
tude and frequency, a smaller model error is produced. Hence, the model error varies
greatly at different amplitudes and frequencies of the input signal, which indicates
that the Bouc–Wen model cannot capture the rate dependency of the hysteresis.
66 3 Feedforward Control Without Modeling Inverse Hysteresis

(a) (b)
100 250
Experiment
Bouc−Wen model
80 200

Displacement (μm)
Input voltage (V)

60 150

40 100

20 50

0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)

(c) (d)
20 250
Displacement error (μm)

Displacement (μm) 200

10

150
0
100

−10
50
Experiment
Bouc−Wen model
−20 0
0 2 4 6 8 10 0 20 40 60 80 100
Time (s) Input voltage (V)

Fig. 3.4 Testing result #1 of the Bouc–Wen model. a Input voltage. b Experimental result and
Bouc–Wen model output. c Bouc–Wen model output errors. d Displacement–voltage hysteresis
loops

3.4.3.3 Generalization Study

In order to test the generalization capability of the Bouc–Wen model, two new input
signals are selected as shown in Figs. 3.4a and 3.5a.
For the test signal #1, the model output is depicted in Fig. 3.4b–d. The discrep-
ancy between the model result and the obtained actual output (yd ) by experiment is
illustrated in Fig. 3.4c. It is observed that the Bouc–Wen model produces a RMSE
of 6.78 µm, which accounts for 2.8 % of the motion range. Hence, as compared with
the training error (1.9 %), the Bouc–Wen model gives a much worse testing result
even with a simple test signal.
With a more complex test signal #2, the model output is described in Fig. 3.5b–d.
It is found that the Bouc–Wen model creates a RMSE of 4.73 µm, i.e., 1.9 % of the
motion range. Hence, as compared with the training result, the Bouc–Wen model
obtains a slightly larger testing error for the test signal #2.
3.4 Experimental Studies on Hysteresis Identification 67

(a) (b)
100 250
Experiment
Bouc−Wen model

Displacement (μm)
80 200
Input voltage (V)

60 150

40 100

20 50

0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)

(c) (d)
20 250
Displacement error (μm)

Displacement (μm) 200

10

150
0
100

−10
50
Experiment
Bouc−Wen model
−20 0
0 2 4 6 8 10 0 20 40 60 80 100
Time (s) Input voltage (V)

Fig. 3.5 Testing result #2 of the Bouc–Wen model. a Input voltage. b Experimental result and
Bouc–Wen model output. c Bouc–Wen model output errors. d Displacement–voltage hysteresis
loops

3.4.4 LSSVM Model Results

Taking the Laplace transform of Eq. (3.5) and comparing its parameters with
Eq. (3.16), the following parameters are obtained: ω = 2.5450 × 103 rad/s,
ξ = 3.6287 × 10−4 , and d = 12.47 m/s2 − V.

3.4.4.1 LSSVM Modeling

For the training of LSSVM model, the same exciting voltage signal as shown in
Eq. (3.17) is employed, which is depicted in Fig. 3.3a. In order to identify the hys-
teresis behavior accurately based on LSSVM model, a suitable input vector (3.9)
requires to be determined.
By selecting three types of input variables with different orders defined by m, n,
and l, three LSSVM models are trained using the corresponding input and output
variables. The output variable is taken as the hysteresis term as shown in Fig. 3.6a,
which is obtained by resorting to Eq. (3.6).
68 3 Feedforward Control Without Modeling Inverse Hysteresis

(a) 8
x 10
(b) 8
4 x 10
4
Hysteresis term h (μm/s2)
Experiment

Hysteresis term h (μm/s2)

LSSVM model
3 3
Model error ×1000

2 2

1 1

0 0
Experiment
(b)
LSSVM model
−1 −1
0 2 4 6 8 10 0 20 40 60 80 100
Time (s) Input voltage (V)

(c) (d)
250 250
Experiment
200 LSSVM model Displacement (μm)
Displacement (μm)

200
Model error
150
150
100
100
50

0 50
Experiment
LSSVM model
−50 0
0 2 4 6 8 10 0 20 40 60 80 100
Time (s) Input voltage (V)

Fig. 3.6 Training result of LSSVM model III. a Experimental result and LSSVM model output
for hysteresis term h. b h versus input voltage. c Experimental result and LSSVM model result for
displacement y. d Displacement y versus input voltage

The LSSVM model I only takes the voltage and hysteresis terms as the input
variables with m = 2, n = 0, and l = 2. Similarly, LSSVM model II only considers
the position and hysteresis terms as the input variables with m = 0, n = 2, and
l = 2. In contrast, by choosing the voltage, position, and hysteresis terms as input
variables, LSSVM model III is constructed with m = 2, n = 2, and l = 2. Results
show that further increase of the orders does not lead to significant improvement on
the modeling accuracy.
For example, the training results of LSSVM model III are shown in Fig. 3.6. Using
the test signal #1 as shown in Fig. 3.4a and test signal #2 as shown in Fig. 3.5a, the
testing results are depicted in Figs. 3.7 and 3.8, respectively. In addition, the training
and testing results of Bouc–Wen model and the three LSSVM models are compared
in Table 3.2. Evidently, each LSSVM model achieves a better identification result
than the Bouc–Wen model.
By comparing the results of models I and II, it is found that model II produces
much better testing results than model I for both test signals. This indicates that
the position is better than voltage as input variables. Moreover, among the three
3.4 Experimental Studies on Hysteresis Identification 69

(a) 8
x 10
(b) 8
4 x 10
4
Hysteresis term h (μm/s )
Experiment

Hysteresis term h (μm/s2)

2

LSSVM model
3 3
Model error ×10

2 2

1 1

0 0
Experiment
LSSVM model
−1 −1
0 2 4 6 8 10 0 20 40 60 80 100
Time (s) Input voltage (V)

(c) (d)
250 250
Experiment
200 LSSVM model
Displacement (μm)
Displacement (μm)

200
Model error ×5
150
150
100
100
50

0 50
Experiment
LSSVM model
−50 0
0 2 4 6 8 10 0 20 40 60 80 100
Time (s) Input voltage (V)

Fig. 3.7 Testing result #1 of LSSVM model III. a Experimental result and LSSVM model output
for hysteresis term h. b h versus input voltage. c Experimental result and LSSVM model result for
displacement y. d Displacement y versus input voltage

LSSVM models, the LSSVM model III produces the best results in terms of lower
modeling errors for both hysteresis term h and output position y. Specifically, the
model III produces a negligible training RMSE for the hysteresis term h, which leads
to a percent RMSE of 1.05 % for the output position y. With the test signal #1, the
LSSVM model III gives a RMSE of 0.27 % for h, which results in a 0.35 % RMSE
for the output position y. Using the test signal #2, it produces 0.15 % and 0.50 %
RMSE for the hysteresis term h and output position y, respectively.
Owing to a better performance, the LSSVM model III is chosen for a further
comparison study with respect to the identified Bouc–Wen model. Results show
that the LSSVM model has reduced the testing errors of output position by 87.4 %
and 74.3 % in comparison with the Bouc–Wen model for the test signals #1 and #2,
respectively. Hence, the hysteresis identification results demonstrate the effectiveness
of LSSVM model.
70 3 Feedforward Control Without Modeling Inverse Hysteresis

(a) 8
(b) 8
x 10 x 10
4 4
Hysteresis term h (μm/s2)

Experiment

Hysteresis term h (μm/s2)

LSSVM model
3 3
Model error ×10

2 2

1 1

0 0
Experiment
LSSVM model
−1 −1
0 2 4 6 8 10 0 20 40 60 80 100
Time (s) Input voltage (V)
(c) (d)
250 250
Experiment
200 LSSVM model Displacement (μm) 200
Displacement (μm)

Model error ×5
150 150

100 100

50 50

0 0 Experiment
LSSVM model
−50 −50
0 2 4 6 8 10 0 20 40 60 80 100
Time (s) Input voltage (V)

Fig. 3.8 Testing result #2 of LSSVM model III. a Experimental result and LSSVM model output
for hysteresis term h. b h versus input voltage. c Experimental result and LSSVM model result for
displacement y. d Displacement y versus input voltage

Table 3.2 Training and testing results of Bouc–Wen model and LSSVM models
Model Input Training RMSE (%) Testing RMSE (%)
variable
Trajectory 1 Trajectory 2
h y h y h y
Bouc– – – 1.88 – 2.80 – 1.95
Wen
model
LSSVM m = 2, 0.0391 1.05 3.62 0.79 3.04 0.62
model I n = 0,
l=2
LSSVM m = 0, 0.0075 1.05 0.19 0.36 0.21 0.51
model II n = 2,
l=2
LSSVM m = 2, 0.0005 1.05 0.27 0.35 0.15 0.50
model III n = 2,
l=2
3.5 Experimental Studies on Hysteresis Compensation 71

3.5 Experimental Studies on Hysteresis Compensation

In this section, the control schemes based on the developed Bouc–Wen model and
LSSVM model are realized to compensate for the rate-dependent hysteresis effect.
The real-time control in the experimental studies is implemented with a sampling
time of 0.004 s.

3.5.1 Feedforward Compensation

To compensate for the hysteresis nonlinearity, a feedforward (FF) control (3.4) based
on the Bouc–Wen model is first implemented. The block diagram of LSSVM model-
based FF control is depicted in Fig. 3.9, where the control signal u FF is obtained
by Eq. (3.7). It is seen that only the hysteresis model is needed, whereas no inverse
hysteresis model is required to implement the FF compensation.
In order to demonstrate the performances of the Bouc–Wen model and LSSVM
model for the purpose of hysteresis compensation, a series of experimental studies are
performed. For instance, concerning a desired 2-Hz sinusoidal position trajectory as
shown in Fig. 3.10a, the FF control results of Bouc–Wen model and LSSVM model
are illustrated in Fig. 3.10a, c, and the tracking errors are compared in Fig. 3.10b.
The control actions are shown in Fig. 3.10d. The control results are summarized in
Table 3.3, where the hysteresis width is defined as the ratio between the maximum
discrepancy of the ascending and descending curves of hysteresis loop and the output
motion range.

− PID uFB
+ controller
+ 1 u
System y
2
yd plant
ωn2 + + 3
+ 1
d / dt + d uFF
−
2ξωn
Control (u ) switch:
d / dt 1. FB control
2. FF+FB control
3. FF control
LSSVM ĥ
hyst. model

Fig. 3.9 Block diagram of LSSVM hysteresis model-based control schemes

72 3 Feedforward Control Without Modeling Inverse Hysteresis

(a) (b)
150 15
B−W model FF

Displacement error (μm)

LSSVM FF
Displacement (μm)

10
100

5
Reference
50 B−W model FF
LSSVM FF 0

0 −5
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time (s) Time (s)

(c) (d)
150
Actual displacement (μm)

60
Control effort (V)
100
40
B−W model FF
LSSVM FF
20
50

B−W model FF 0
LSSVM FF
0
0 50 100 150 0 0.5 1 1.5 2
Reference displacement (μm) Time (s)

Fig. 3.10 Motion tracking results of the FF compensation. a Reference and experimental results
of Bouc–Wen and LSSVM models. b Tracking errors. c Actual–reference displacement hysteresis
loops. d Control actions

Table 3.3 Hysteresis compensation results of different control schemes

Controller RMSE Hysteresis width RMS control
action
(µm) (%) (%) (V)
PID FB 0.04 1.05 14.29 3.62
Bouc–Wen FF 6.76 4.98 10.20 3.47
LSSVM FF 2.55 1.88 4.29 3.72
Bouc–Wen 1.55 1.14 2.23 3.65
FF+FB
LSSVM FF+FB 0.62 0.46 0.45 3.64

Results show that the Bouc–Wen FF control produces a RMSE of 6.76 µm, i.e.,
5.0 % of motion range, and the LSSVM-based FF control produces a 2.55-µm RMSE,
i.e., 1.9 % of motion range. As compared with Bouc–Wen model, the LSSVM model
suppresses the tracking error furthermore by 62 %, which leads to a 58 % reduction
3.5 Experimental Studies on Hysteresis Compensation 73

of the hysteresis width. Hence, the superiority of LSSVM over Bouc–Wen model in
FF hysteresis compensation is validated by the experimental results. Moreover, the
LSSVM model requires 7.2 % higher RMS control action than Bouc–Wen model for
FF compensation.

3.5.2 Feedforward Plus Feedback Control

In order to further suppress the residual tracking errors of the feedforward compensa-
tion, a FF augmented with feedback (FB) control is realized. Owing to its popularity,
the digital PID algorithm is employed as the FB control:


k
u FBk = K p ek + K i e j + K d (ek − ek−1 ) (3.19)
j=0

where the displacement error ek = ydk − yk with ydk and yk representing the desired
and actual system outputs at the kth time step, respectively. In addition, K p , K i , and
K d denote the proportional, integral, and derivative gains, respectively. The control
gains are tuned by Ziegler–Nichols method through experiments.
Using the LSSVM hysteresis model, the three control schemes of FF, FB, and
FF+FB can be switched easily as shown in Fig. 3.9. For the same position reference
trajectory as shown in Fig. 3.10a, the tracking results of the Bouc–Wen model and
LSSVM model-based FF+FB control are depicted in Fig. 3.11a–c. In addition, the
stand-alone PID FB tracking error is also illustrated in Fig. 3.11b, and the control
actions of the three schemes are shown in Fig. 3.11d. For a clear comparison, the
control results are tabulated in Table 3.3.
Although the Bouc–Wen model-based FF+FB control reduces the tracking error
to 1.55 µm (i.e., 1.14 % of motion range) and hysteresis width to 2.23 %, the LSSVM
model-based one further suppresses the RMSE and hysteresis to negligible levels of
0.62 µm (i.e., 0.46 % of motion range) and 0.45 %, respectively. Thus, the LSSVM
model-based hybrid control has improved the tracking accuracy and hysteresis com-
pensation by 60 % and 80 %, respectively.
Comparing the control actions, the LSSVM model-based FF+FB control needs
0.17 % lower control voltage than the Bouc–Wen model-based one. The slight lower
control effort of the former is contributed by the employed PID FB control, which
produces a 2.9 % lower control action than the LSSVM model-based FF control
approach, as shown in Table 3.3.
The experimental results demonstrate the effectiveness of the LSSVM model over
Bouc–Wen model for the task of hysteresis compensation. The performance of the
LSSVM model can be improved to achieve a better generalization ability by selecting
more comprehensive input signal to excite the piezo-actuator for the model training.
Moreover, to further reduce the tracking error for periodic reference inputs, repetitive
control or iterative learning control algorithms can be employed.
74 3 Feedforward Control Without Modeling Inverse Hysteresis

(a) (b)
150 15
PID FB

Displacement error (μm)

10 B−W model FF + FB
Displacement (μm)

LSSVM FF + FB
100 5

Reference 0
B−W model FF + FB
50 LSSVM FF + FB −5

−10

0 −15
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time (s) Time (s)

(c) (d)
150
Actual displacement (μm)

60
Control effort (V)
100 40
PID FB
20 B−W model FF + FB
50 LSSVM FF + FB
PID FB
B−W model FF + FB 0
LSSVM FF + FB
0
0 50 100 150 0 0.5 1 1.5 2
Reference displacement (μm) Time (s)

Fig. 3.11 Motion tracking results of the FF plus FB compensation. a Reference trajectory.
b Tracking errors. c Actual–reference displacement hysteresis loops. d Control actions

3.6 Chapter Summary

This chapter is concentrated on hysteresis identification and compensation of a piezo-

electric actuator. The nonlinear hysteresis behavior has been well identified by resort-
ing to a developed LSSVM-based intelligent hysteresis model. It has been shown
that the position information is better than voltage information as input variables
in hysteresis model development. Moreover, the model can be directly employed to
compensate for the hysteresis effect without modeling the inverse hysteresis. This
indicates that the established single hysteresis model is used for both hysteresis
identification and compensation, which is more computationally efficient than the
existing approaches where both a hysteresis model and an inverse hysteresis model
are employed. Experimental results demonstrate that the LSSVM model is superior
to Bouc–Wen model in terms of both hysteresis identification accuracy and hysteresis
compensation effectiveness. The presented idea can also be extended to hysteresis
identification and compensation of other types of smart actuators.
References 75

References

1. Chen, X., Hisayama, T.: Adaptive sliding-mode position control for piezo-actuated stage. IEEE
Trans. Ind. Electron. 55(11), 3927–3934 (2008)
2. Janaideh, M.A., Rakheja, S., Su, C.Y.: Experimental characterization and modeling of rate-
dependent hysteresis of a piezoceramic actuator. Mechatronics 19(5), 656–670 (2009)
3. Li, Y., Xu, Q.: Adaptive sliding mode control with perturbation estimation and PID sliding
surface for motion tracking of a piezo-driven micromanipulator. IEEE Trans. Control Syst.
Technol. 18(4), 798–810 (2010)
4. Lin, C.J., Chen, S.Y.: Evolutionary algorithm based feedforward control for contouring of a
biaxial piezo-actuated stage. Mechatronics 19(6), 829–839 (2009)
5. Rakotondrabe, M.: Bouc-Wen modeling and inverse multiplicative structure to compensate hys-
teresis nonlinearity in piezoelectric actuators. IEEE Trans. Autom. Sci. Eng. 8(2), 428–431
(2011)
6. Wong, P.K., Xu, Q., Vong, C.M., Wong, H.C.: Rate-dependent hysteresis modeling and control
of a piezostage using online support vector machine and relevance vector machine. IEEE Trans.
Ind. Electron. 59(4), 988–2001 (2012)
7. Xu, Q., Wong, P.K.: Hysteresis modeling and compensation of a piezostage using least squares
support vector machines. Mechatronics 21(7), 1239–1251 (2011)
8. Yu, S., Alici, G., Shirinzadeh, B., Smith, J.: Sliding mode control of a piezoelectric actuator with
neural network compensating rate-dependent hysteresis. In: Proceedings of the IEEE Interna-
tional Conference on Robotics and Automation, pp. 3641–3645. Barcelona, Spain (2005)
 Part II
Hysteresis-Model-Free,
State-Observer-Based Feedback Control
Chapter 4
Model Predictive Discrete-Time
Sliding-Mode Control

This chapter presents a scheme of model predictive discrete-time sliding mode control
(MPDTSMC) with proportional-integral (PI) sliding function and state observer for
the motion tracking control of a nanopositioning system driven by piezoelectric
actuators. One distinct advantage of the controller is that its implementation only
requires a simple second-order model of the system, whereas it does not need to
know the hysteresis model and the bounds on system uncertainties. The reasons why
the model predictive control methodology and PI action can eliminate the chattering
effects and produce a low level of tracking error are addressed in the state-space
framework. Experimental investigations are carried out to verify the feasibility of
the control scheme.

4.1 Introduction

Due to the low damping of the piezoelectric micro-/nano-positioning system, a num-

ber of vibration modes exist and a high-order model is commonly identified, which
results in a high-order controller as a consequence. From the practical point of view,
a linear model of low order, e.g., second order, is very desirable for easy implemen-
tation of control strategies. However, the adoption of a low-order model means that
the residual modes will not be considered in the controller and observer design. The
neglect of residual modes may cause both control spillover and observation spillover
[17]. Spillover is undesirable because it may induce system instability and perfor-
mance degradation [20]. In view of these issues, it is a major challenge in piezostate
control by employing a low-order model without modeling the complicated hystere-
sis effects.
In the literature, some pioneer works have been conducted to deal with the piezo-
electric micro-/nano-positioning control problem by employing a low-order linear
model. For example, by identifying a second-order model of a piezostage, a precise

© Springer International Publishing Switzerland 2016 79

Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2_4
80 4 Model Predictive Discrete-Time Sliding-Mode Control

control is implemented in [26] based on the integral resonant control method, in

which the charge actuation is employed to reduce piezoelectric hysteresis. Although
the hysteresis becomes almost negligible if a piezoelectric actuator is driven by a
charge source, the stroke will be reduced as the cost. Hence, voltage actuation is
still widely adopted in practice. An enhanced adaptive sliding-mode motion track-
ing control is reported in [15] for a piezoelectrically actuated system, which only
uses a second-order dynamics model. However, the sliding-mode control (SMC) is
designed based on the continuous-time dynamics model, and the uncertainty bounds
are required to implement the controller. Based on the second-order dynamics model
of a piezoelectric actuator (PEA), a model predictive sliding-mode control is pre-
sented in [9], where a nonlinear, energy-based model is employed to construct an
inverse compensator of the hysteresis. The same model predictive-based control is
also adopted in [16] for the force control of a piezoelectric actuator. In the pre-
vious works, a SMC controller is designed based on a proportional-type sliding
function, whereas no stability analysis of the control system is given. Besides, the
spillover effects are not treated explicitly. Based on a second-order mass–spring–
damper model of a nanopositioning system, this work presents a MPDTSMC with
PI type of sliding function, and provide stability analysis of the closed-loop system in
detail. The observation spillover is suppressed by designing a proper state observer.
It is known that SMC is a popular nonlinear control approach to deal with distur-
bance. The discrete-time sliding-mode control (DTSMC) is presented for the imple-
mentation on sampled-data systems [3, 7]. For a discrete-time system, the strategy
of model predictive control (MPC) can be employed to predict the system perfor-
mance in a specified time in the future and to produce an optimal control action
with respect to a predefined cost criterion at each time step [21]. By combining the
SMC and MPC techniques, model predictive sliding-mode control has been recently
proposed to achieve the advantages of both methods [8, 23]. Although the combined
control has been adopted in a few of previous works [9, 16] for the control of PEA
by suppressing the chattering phenomenon, it still remains unclear why the MPC
methodology can eliminate the chattering phenomenon in the sliding mode.
This chapter shows that the MPDTSMC with the PI action drives the system
state to slide in a vicinity of the sliding surface with a boundary layer of thickness
O(T 3 ) [1], which is much lower than a commonly designed DTSMC with O(T )
boundary layer in the sliding mode. In addition, the state tracking error of the order
O(T 2 ) is achieved with the presented control scheme. The theoretical analysis and
the effectiveness of the PI action in the control scheme are verified by conducting
experimental studies on a nanopositioning platform.
4.2 Problem Formulation 81

4.2 Problem Formulation

4.2.1 Dynamics Modeling of a Nanopositioning System

The entire dynamics model of a piezoelectric nanopositioning system with nonlinear

hysteresis can be established as follows:

m ẍ(t) + b ẋ(t) + kx(t) = du(t) − f h (t) (4.1)

where t is the time variable, parameters m, b, k, and x represent the mass, damping
coefficient, stiffness, and output displacement of the nanopositioning stage, respec-
tively; d is the piezoelectric coefficient, u denotes the input voltage, and f h describes
the hysteretic effect of the system in terms of force which is not modeled in this
research.
The dynamics model (4.1) can be expressed in the state-space form:

Ẋ(t) = Ac X(t) + Bc u(t) + Dc f h (t) (4.2)

Y(t) = C X(t) (4.3)

where the state vector X(t) = [x(t), ẋ(t)]T , the unmodeled hysteresis effect f h is
considered as a disturbance to the system, and the four matrices are
     
0 1 0 0
Ac = , Bc = , Dc = , C = [1 0] . (4.4)
− mk − mb d
m − m1

It is observed that the disturbance f h satisfies the matching condition, i.e., Dc =

δd Bc where δd = − d1 is a scalar.
Using a sampling time T , the continuous-time system model given by Eqs. (4.2)
and (4.3) can be discretized as:

Xk+1 = AXk + Bu k + fk (4.5)

Yk = CXk (4.6)

where the notation Xk = X(kT ) and the corresponding matrices are calculated by:

A = e Ac T (4.7)
 T
B= eAc τ dτ Bc (4.8)
0
 T
fk = eAc τ δd Bc f h ((k + 1)T − τ )dτ (4.9)
0

which indicate that both B and fk are O(T ), i.e., the order of T .
82 4 Model Predictive Discrete-Time Sliding-Mode Control

In addition, two useful properties are employed to facilitate the control design
[24, 25].
Property 4.1 dk = O(T ), dk − dk−1 = O(T 2 ), and dk − 2dk−1 + dk−2 = O(T 3 ).
Property 4.2 Assume E k+1 = Ξ E k + δk where matrix Ξ is asymptotically stable
(Ξ  < 1), and the magnitude of δk is of the order O(T r ) where r is an integer.
Then the ultimate bound of E k is of the order O(T r −1 ) when k → ∞.
The proof of Property 4.2 is given as follows.
Proof Assume that m eigenvalues of Ξ are deadbeat, i.e., Ξ has m poles locating at
the origin. Then, Ξ can be written as [1]

Ξ = PJP−1 (4.10)

where P is a transformation matrix and J is the Jordan matrix of the eigenvalues

of Ξ .
The matrix J can be expressed as
 
J1 0
J= (4.11)
0 J2

where J1 ∈ R m×m and J2 ∈ R (n−m)×(n−m) taking on the forms:

 
0 Im−1
J1 = (4.12)
0 0
⎡ ⎤
λm+1 0
⎢ .. ⎥
J2 = ⎣ . ⎦ (4.13)
0 λn

where λ j denote the poles of Ξ .

For simplicity, it is assumed that the nonzero poles are designed to be distinct
and their continuous-time counterparts are of the order O(1). Then, the solution of
E k+1 = Ξ E k + δk can be derived as:
k−1

−1 −1
Ek = P J Pk
E0 + P i
J P δk−i−1 (4.14)
i=0

which can be rewritten as

⎛ ⎞ ⎛ ⎞
  Ji 0 
k−1 0 0 
k−1
E k = P J k P −1 E 0 + P ⎝ 1 P −1 δk−i−1 ⎠ + P ⎝ P −1 δk−i−1 ⎠ .
0 0 0 J2i
i=0 i=0
(4.15)
4.2 Problem Formulation 83

In view of J1i = 0 for i ≥ m, Eq. (4.15) can be expressed into the following form
for k ≥ m:
⎛ ⎞ ⎛ ⎞
m  i  0 0 
k−1
J1 0
E k = P J k P −1 E 0 + P ⎝ −1
P δk−i−1 ⎠ + P ⎝ −1
P δk−i−1 ⎠ .
0 0 0 J2i
i=0 i=0
(4.16)

Taking into account J1  = 1 and J2  = λmax = max{λm+1 , . . . , λn } with  · 

denoting  · 2 , Eq. (4.16) leads to

m  
  J1 0 i
E k  ≤ P   −1
 0 0  P  δk−i−1 
i=0
k−1  
  0 0 i
+   −1
 0 J2  P  δk−i−1  . (4.17)
i=0

Considering that the matrix Ξ is asymptotically stable, λmax < 1 holds, then


m ∞
 1
J1 i = m, J2 i = . (4.18)
1 − λmax
i=1 i=1

Using the Tustin approximation

2+ Tp
λmax =
2− Tp
1 1 2− Tp
⇒ = = = O(T −1 ) (4.19)
1 − λmax 1− 2+T p −2T p
2−T p

where p = O(1) is the corresponding pole in continuous-time.

Assuming that m = O(1) and considering P −1  = P−1 , the ultimate bound
of E k can be deduced from Eq. (4.17):

lim E k  ≤ O(1) · O(T r ) + O(T −1 ) · O(T r )

k→∞
≈ O(T r ) + O(T r −1 )
≈ O(T r −1 ). (4.20)

4.2.2 Sliding-Mode Controller Design

Most of the existing works construct a discrete-time sliding-mode controller using

a proportional sliding function defined based on the system state only [3, 7, 19],
84 4 Model Predictive Discrete-Time Sliding-Mode Control

which leads to the order reduction of the system. A few of previous studies employ
the integral term in the sliding surface [1, 24]. In this work, a PI-type sliding surface
is defined in order to speed up the system response.
By assigning a vector of state error

Ek = Xk − Xdk (4.21)

where Xdk denotes the desired system state, a PI-type sliding function is defined as
follows:

Sk = GEk + K I ξ k (4.22)

where

S = {Sk |Sk = 0, for k = 0, 1, . . . } (4.23)

represents the sliding surface or sliding hyperplane, and the integral error vector is
defined as follows:

ξ k = Ek + ξ k−1 . (4.24)

In addition, G and K I are constant proportional and integral gain vectors, which are
designed to assign appropriate eigenvalues for the system.
In view of Eqs. (4.21) and (4.5), one gets

Ek+1 = Xk+1 − Xdk+1

= AXk + Bu k + fk − Xdk+1
= AEk + Bu k + fk + AXdk − Xdk+1
= AEk + Bu k + dk (4.25)

where dk = fk + AXdk − Xdk+1 denotes the overall disturbance.

eq
Considering that the equivalent control u k is the solution to ΔS = Sk+1 − Sk = 0
[6], it can be deduced that

GEk+1 + K I ξ k+1 = Sk
⇒ GEk+1 + K I (Ek+1 + ξ k ) = Sk
⇒ (G + K I )Ek+1 + K I ξ k = Sk
⇒ (G + K I )(AEk + Bu k + dk ) + K I ξ k = Sk (4.26)

which allows the generation of an equivalent controller:

u k = −(Ga B)−1 (−Sk + Ga AEk + Ga dk + K I ξ k )

eq
(4.27)
4.2 Problem Formulation 85

where

Ga = G + K I (4.28)

and the disturbance dk is estimated by its one-step delayed value dk−1 .

Thus, Eq. (4.27) becomes

u k = −(Ga B)−1 (−Sk + Ga AEk + Ga dk−1 + K I ξ k )

eq
(4.29)

where

dk−1 = Ek − AEk−1 + Bu k−1 (4.30)

is derived by noting Eq. (4.25).

4.2.3 Control Gain Design

To implement the PI sliding function, the gain vectors G and K I should be designed
such that Ga B is invertible. Moreover, the problem of gain vector design is addressed
in the following discussion to guarantee the stability of the closed-loop system.
In the ideal sliding mode, Sk+1 = Sk = 0 holds. Then, Eq. (4.27) reduces to

u k = −(Ga B)−1 (Ga AEk + Ga dk + K I ξ k ).

eq
(4.31)

Substituting Eq. (4.31) into Eq. (4.25) leads to the closed-loop state error dynamics
equation:

Ek+1 = AEk + Bu k + dk
= AEk − B(Ga B)−1 (Ga AEk + Ga dk + K I ξ k ) + dk
= UEk − Vξ k + Wdk (4.32)

where
 
U = I − B(Ga B)−1 Ga A (4.33)
V = B(Ga B)−1 K I (4.34)
−1
W = I − B(Ga B) Ga (4.35)

with I denoting an identity matrix.

The eigenvalues of the matrices U, V, and W are the solutions of det(zI − U) = 0,
det(zI−V) = 0, and det(zI−W) = 0, respectively, which determine the closed-loop
86 4 Model Predictive Discrete-Time Sliding-Mode Control

dynamics behavior. To ensure the stability of the closed-loop system, the eigenvalues
of the three matrices should be located inside the unit circle in the complex z-plane.
For example, the eigenvalues of the matrix U can be calculated as the solutions
to the equation [4]:
    
 z − u 11 −u 12  ρ1
 ρ1 
 ρ1  = z z − u 11 − ρ u 12 = 0. (4.36)
ρ2 u 11 z − ρ2 u 12 2

It is found that there is one fixed pole (z 1 = 0) at the origin. The other pole lies on
the real axis (z 2 = u 11 − ρρ21 u 12 ), whose position can be adjusted by tuning the values
of Ga = [ρ1 ρ2 ] to make the pole locate inside the unit circle. In theory, a faster
response can be produced by placing the pole close to 0 and a better robustness can
be obtained by locating the pole near 1 for the controller. Similarly, the values of Ga
and K I should also be designed to locate the eigenvalues of the other two matrices
inside the unit circle.

4.3 DTSMC Design

The equivalent controller as designed in the previous section takes effect in the sliding
mode when the system state trajectory is kept on the sliding surface. However, if the
initial state of the system does not lie on the sliding surface S or external disturbances
occur during the sliding mode motion, the stand-alone equivalent control cannot drive
the state toward the sliding surface. Thus, a switching control is necessary. In the
following discussions, a DTSMC with switching control is designed.

4.3.1 Controller Design and Analysis

To devise a robust controller, a commonly used method is to augment the equivalent

control with a discontinuous switching control u sw sw
k . The role of u k is to force the
system state to reach the sliding mode. Then, the overall control action is derived as:
eq
u k = u k + u sw
k . (4.37)

Theorem 4.1 For the system (4.5) with the switching function (4.22), if the following
controller (4.38) is employed, then the discrete sliding mode will occur after a finite
number of steps

u k = −(Ga B)−1 [−Sk + Ga AEk + Ga dk−1 + K I ξ k + Ga Ks sgn(Sk )] (4.38)

where Ks is a positive switching gain vector and the last term in the right-hand side
represents the switching control action.
4.3 DTSMC Design 87

Proof In view of Eqs. (4.26) and (4.28), the following expression can be derived:

Sk+1 = Ga (AEk + Bu k + dk ) + K I ξ k . (4.39)

Substituting Eq. (4.38) into Eq. (4.39) yields

Sk+1 = Sk − Ga Ks sgn(Sk ) + Ga ε k (4.40)

where

εk = dk − dk−1 ≈ fk − fk−1
 T  (k+1)T −τ
= e Ac τ
δd Bc f˙h (σ )dσ dτ = O(T 2 ) (4.41)
0 kT −τ

is the estimation error vector of the disturbance, which has a magnitude of O(T 2 )
and is assumed to be bounded, i.e.,

|εk | = |dk − dk−1 | ≤ Δd . (4.42)

The switching control gain vector Ks is designed to meet the condition:

Ks > Δd + σ (4.43)

where σ is an arbitrary positive constant. Then, in the case of Sk ≥ 0, one has

Sk+1 = Sk − Ga Ks + Ga ε k
< Sk − Ga [Δd + σ − ε k ]
< Sk . (4.44)

On the other hand, if Sk < 0, it can be derived that

Sk+1 = Sk + Ga Ks + Ga ε k
> Sk + Ga [Δd + σ + ε k ]
> Sk . (4.45)

Considering Eqs. (4.44) and (4.45) together, yields

|Sk+1 | < |Sk | for k ≥ k0 (4.46)

which indicates that |Sk | decreases monotonously, and the discrete sliding mode is
reached after a finite number (k0 ) of steps.
It has been shown in [18] that the relationship (4.46) represents a sufficient con-
dition for the existence of discrete sliding mode.
88 4 Model Predictive Discrete-Time Sliding-Mode Control

Remark 4.1 Theorem 4.1 gives a sufficient condition for the existence of the discrete
sliding mode. Due to the discontinuity of the signum function sgn(s), chattering may
occur in the control input. To alleviate the chattering phenomenon, the boundary
layer technique is adopted by replacing the signum function in Eq. (4.38) with the
saturation function:

sgn(Sk ) if |Sk | > δ
sat(Sk ) = (4.47)
Sk /δ if |Sk | ≤ δ

where the positive constant δ denotes the boundary layer thickness, which ensures
that Sk is bounded by ±δ. In the selection of parameter δ, a tradeoff between the
chattering and tracking error should be taken.

It is observed from Eq. (4.43) that the selection of the control gain Ks is dependent
on the upper bound of the disturbance estimation error Δd . Alternatively, this term is
the upper bound of the disturbance in the classical SMC. Generally, the disturbance
estimation error is much lower than the disturbance itself. Hence, a much smaller
switching gain vector Ks can be designed in comparison with the classical SMC.
This is one of the advantages of the constructed DTSMC with one-step delayed
disturbance estimation over the conventional SMC.

4.3.2 Tracking Error Bound Analysis

Taking into account the relationship (4.40) and Property 4.2, it can be deduced that
Sk has a magnitude of the order O(T ). It means that the ultimate state will stay in
the neighborhood of S with a boundary layer of the thickness O(T ).
Considering Eq. (4.22), the error dynamics equation (4.32) can be further exp-
ressed below after a necessary calculation:

Ek+1 = PEk − QSk + Wdk (4.48)

where the three matrices are

P = A − B(Ga B)−1 [Ga A − G] (4.49)

−1
Q = B(Ga B) (4.50)
W = I − B(Ga B)−1 Ga . (4.51)

In view of Sk = O(T ) and dk = O(T ), it can be deduced that Ek = O(1), i.e.,

the ultimate state error is of the order O(1).
Due to the existence of unmodeled dynamics of the system including the actuator
nonlinearities, sensor noises, and high-frequency dynamics, the switching control
constructed above may still induce chattering phenomenon. Chattering is the oscil-
lations of the system states around the sliding surface with a certain amplitude and
4.3 DTSMC Design 89

frequency [11–14]. As a result, chattering leads to a poor system performance and

reduces usable life of the actuator. In the next section, the model predictive control
scheme is employed to eliminate the chattering effect.

4.4 MPDTSMC Design

A MPDTSMC scheme, which integraties MPC and DTSMC control strategies, is

constructed to handle the nonlinearities and disturbances existing in the system.
mp
Specifically, the MPC is used to generate an optimal control (u k ) to drive the system
eq
state to the sliding surface, and the equivalent control (u k ) is employed to maintain
the state trajectory on the sliding surface for the subsequent time. It is shown that
the system state is forced to reach the sliding mode smoothly in an optimal manner
without chattering effects.

4.4.1 MPDTSMC Controller Design

Substituting the overall control action

eq mp
uk = uk + uk (4.52)

into Eq. (4.39) and noting Eq. (4.29) gives

 eq mp 
Sk+1 = Ga AEk + B(u k + u k ) + dk + K I ξ k
mp
= Sk + Ga Bu k + Ga εk (4.53)

which describes the dynamics of the sliding mode and also represents a one-step
prediction of the sliding-mode dynamics.
Based on Eq. (4.53), an N -step prediction of the sliding-mode dynamics can be
obtained as follows:
mp mp mp
Sk+N = Sk + Ga B(u k + u k+1 + · · · + u k+N −1 ) +
(Ga ε k+N −1 + · · · + Ga ε k+1 + Ga ε k ) (4.54)

where N is called the prediction horizon.

The N prediction functions can be stacked together into the form:
mp
Sk = ΛSk + Φuk−1 + Γ ζ k−1 (4.55)
90 4 Model Predictive Discrete-Time Sliding-Mode Control

where the vectors for the sliding surface, future control, and disturbance estimation
errors are:
⎡ ⎤ ⎡ mp ⎤ ⎡ ⎤
Sk+1 uk εk
⎢ Sk+2 ⎥ ⎢ u mp ⎥ ⎢ εk+1 ⎥
⎢ ⎥ mp ⎢ k+1 ⎥ ⎢ ⎥
Sk = ⎢ . ⎥ , uk−1 = ⎢ . ⎥ , ζ k−1 = ⎢ .. ⎥. (4.56)
⎣ .. ⎦ ⎣ .. ⎦ ⎣ . ⎦
mp
Sk+N u k+N −1 ε k+N −1

Additionally, the matrices take on the following forms:

⎡ ⎤
I
⎢I⎥
⎢ ⎥
Λ = ⎢.⎥ (4.57)
⎣ .. ⎦
I
⎡ ⎤
Ga B 0 . . . 0
⎢ Ga B Ga B . . . 0 ⎥
⎢ ⎥
Φ=⎢ . .. . . .. ⎥ (4.58)
⎣ .. . . . ⎦
Ga B Ga B . . . Ga B
⎡ ⎤
Ga 0 . . . 0
⎢ Ga Ga . . . 0 ⎥
⎢ ⎥
Γ =⎢ . . . . ⎥. (4.59)
⎣ .. .. . . .. ⎦
Ga Ga . . . Ga

Using the above notations, a cost function for minimization can be expressed by:
 mp T mp
J = SkT Sk + λ uk−1 uk−1 (4.60)

where λ is the weighting parameter to limit the partial control effort. It is notable that
mp
uk−1 approaches to zero as Sk tends to zero. Thus, the cost function J will vanish at
the steady state.
Substituting Eq. (4.55) into the cost function (4.60) and applying the optimization
criterion by setting the condition:

∂J
mp = 0 (4.61)
∂uk−1

yields

uk−1 = −(Φ T Φ + λ I)−1 Φ T (ΛSk + Γ ζ k−1 )

mp
(4.62)
4.4 MPDTSMC Design 91

where the future disturbance values in ζ k−1 are unknown. Thus, they are estimated
as follows

ζ̂ k−1 = [εk−1
T
ε k−1
T
· · · ε k−1
T
]T . (4.63)

Considering that only the first value of the predicted control sequence is used, the
optimal control action is obtained as:

= −c(Φ T Φ + λ I)−1 Φ T (ΛSk + Γ ζ̂ k−1 )

mp
uk (4.64)

where the vector c = [1, 0, 0, · · · , 0].

Therefore, the total control action can be expressed as:
eq mp
uk = uk + uk
= −(Ga B)−1 [−Sk + Ga AEk + Ga dk−1 + K I ξ k ]
−c(Φ T Φ + λ I)−1 Φ T (ΛSk + Γ ζ̂ k−1 ). (4.65)

4.4.2 Stability Analysis

Theorem 4.2 For the system (4.5) with the switching function (4.22), if the distur-
bance change rate is limited, i.e., the inequality (4.42) is satisfied, then the controller
(4.65) leads to a stable closed-loop control system.

Proof Substituting Eq. (4.65) into Eq. (4.39) results in Eq. (4.53). Then, inserting
Eq. (4.62) into the stacked equation (4.55) yields

Sk = ΛSk − Φ(Φ T Φ + λ I)−1 Φ T (ΛSk + Γ ζ̂ k−1 ) + Γ ζ k−1 (4.66)

where λ is defined in the cost function (4.60) to restrict the control input. For sim-
plicity, it is assumed that there is no penalty for the control effort, i.e., λ = 0.
Then, Eq. (4.66) reduces to

Sk = ΛSk − (ΛSk + Γ ζ̂ k−1 ) + Γ ζ k−1

= Γ (ζ k−1 − ζ̂ k−1 ). (4.67)

Considering only the first value of the predicted sliding hyperplane vector in
Eq. (4.67) and taking into account Eqs. (4.56) and (4.63) yields

Sk+1 = Ga (ε k − ε k−1 )
= Ga (dk − 2dk−1 + dk−2 ) = O(T 3 ) (4.68)

which indicates that the ultimate magnitude of Sk+1 is of the order O(T 3 ).
92 4 Model Predictive Discrete-Time Sliding-Mode Control

Because the disturbance change rate εk is bounded as supposed in Eq. (4.42), it

is reasonable to deduce that the rate of εk is also limited:

|εk − ε k−1 | ≤ Δε . (4.69)

Thus, it can be generated that

|Sk+1 | ≤ |Ga Δε | = Δ S (4.70)

where 2Δ S is defined as the quasi-sliding mode band width.

According to Definition 1 and Definition 2 in [3], it can be concluded that the
system (4.5) satisfies the reaching condition of the quasi-sliding mode in the Δ S
vicinity of the sliding surface S in a finite number of steps. Therefore, the closed-
loop control system is stable.

Remark 4.2 The essence of the presented MPDTSMC is to drive the system state to a
quasi-sliding mode as shown in Eq. (4.70). One advantage of this quasi-sliding model
[3] lies in that it does not require the system state to cross the sliding hyperplane
S in each successive control step as in [7]. Therefore, the chattering phenomenon
is eliminated. As a result, it leads to a reduction of the control effort and improved
control performance [3].

Remark 4.3 It is observed from Eq. (4.68) that Sk+1 has an ultimate magnitude of the
order O(T 3 ). Therefore, the controller (4.65) drives the system state to slide in the
vicinity of S with a boundary layer of thickness O(T 3 ), whose magnitude is much
lower than the O(T ) thickness achieved by the DTSMC controller designed earlier.
Therefore, a much lower tracking error is expected for the proposed MPDTSMC
controller.

4.4.3 State Observer Design

The controller equation (4.65) reveals that its implementation needs the full state
(Xk ) feedback of the system. However, only partial states are available in practice.
Thus, a state observer is designed to estimate the full state by making use of the only
measurable position information of the nanopositioning system.
Although various methods are available [10], a Luenberger-like observer is
adopted here owing to its simplicity. The state observer takes on the form:

X̂k+1 = AX̂k + Buk + L(Yk − Ŷk ) + f̂k (4.71)

Ŷk = C X̂k (4.72)
4.4 MPDTSMC Design 93

where the notation X̂k represents the estimate of Xk , and the value of the disturbance
fk in Eq. (4.5) is obtained by its one-step delayed estimation:

f̂k = fk−1 = X̂k − AX̂k−1 − Buk−1 . (4.73)

Subtracting Eq. (4.71) from Eq. (4.5) allows the derivation of

Xk+1 = (A − LC)
 Xk + (fk − fk−1 ) (4.74)

where  Xk = Xk − X̂k denotes the estimation error of the state Xk . It is seen that
the state observer is independent of the controller. To guarantee the stability of the
observation unit, the observer gain L should be designed to make all the eigenvalues
of the matrix A − LC locate inside the unit circle.

Remark 4.4 The convergence speed of the observer relies on the location of the
poles of the matrix A − LC. The poles can be arbitrarily placed by the gain vector
L. Generally, small pole value will give rapid convergence, and the convergence
speed of the observer should be faster than the system response so that the observer
dynamics is insignificant as compared to the system dynamics. However, too small
pole produces instability of the system, which arises from the noise and spillover
effect of the neglected system dynamics. Hence, a trial-and-error approach is usually
required to place the poles appropriately.

4.4.4 Tracking and Estimation Error Bound Analysis

In order to evaluate the tracking error bound, noting that Sk = O(T 3 ) and dk =
O(T ), it is deduced from Eq. (4.48) that Ek = O(1), i.e., the ultimate state error is
of the order O(1).
In the case of m = n, Ga and B are nonsingular, i.e., the number of inputs is equal
to the number of system states, the matrix W as shown in Eq. (4.51) will vanish. In
such a situation, Eq. (4.48) reduces to

Ek+1 = PEk − QSk . (4.75)

Then, by noting Sk = O(T 3 ) as shown in Eq. (4.68) and Property 4.2, it can be derived
that Ek = O(T 2 ), which is much lower than the O(1) tracking error obtained with
the DTSMC designed earlier.
Furthermore, considering that fk − fk−1 is of the order O(T 2 ), it can be deduced
from Eq. (4.74) that the ultimate bound of Xk is O(T ). For the systems with relative
degrees greater than 1, it has been shown that the state estimation error will be reduced
to O(T 2 ) for the closed-loop system [24] .
94 4 Model Predictive Discrete-Time Sliding-Mode Control

4.5 Experimental Investigation

The designed controllers are validated by a series of simulations and experimental

studies in this section.

4.5.1 Experimental Setup

The experimental setup is depicted in Fig. 4.1. A flexure-based nanopositioning stage

is fabricated from a plate of Al-7075 alloy by the wire electrical discharge machin-
ing (EDM) process. Driven by a piezoelectric stack actuator (PSA) with the stroke
of 14.5 µm (model: TS18-H5-202, from Piezo Systems, Inc.), the stage delivers
an amplified translational motion through elastic deformations of the bridge-type
displacement amplifier. The output displacement is measured by a capacitive sen-
sor (model: D-510.050, from Physik Instrumente Co., Ltd.). A dSPACE DS1103
(from dSPACE GmbH) rapid prototyping system equipped with 16-bit A/D and
D/A converters are employed to implement the controller. The D/A channel pro-
duces a voltage control signal which is then amplified by a high-voltage amplifier
(model: EPA-104, from Piezo Systems, Inc.) to provide a voltage ranging between
−100 and 100 V for driving the PSA. In addition, the sensor output voltage signal
is passed through a signal conditioner (model: E-852, from Physik Instrumente Co.,
Ltd.) and then acquired by the A/D channel. Control algorithms are developed with
MATLAB /Simulink software and downloaded to the DS1103 board to realize
the real-time control.

4.5.2 Hysteresis Characterization

By applying a 0.5-Hz sinusoidal wave with the amplitude of 10 V to the PSA, the
output versus input curve is generated as shown in Fig. 4.2, which exhibits a nonlin-
ear relationship between the output displacement and input voltage. The maximum

Fig. 4.1 Experimental setup of a flexure-based nanopositioning stage driven by a piezoelectric

stack actuator (PSA)
4.5 Experimental Investigation 95

Fig. 4.2 The output versus

80
input hysteresis loops which h/H ×100% = 14.4%
are obtained by applying a 70
0.5-Hz sine wave input
60

Displacement (μm)

H = 78.4 μm
h = 11.3 μm
50

40

30

20
With proposed control
10
Without compensation
0
0 2 4 6 8 10
Input voltage (V)

output displacement is 78.4 µm, and the maximum deviation between the ascending
and descending curves is 11.3 µm. Hence, the output hysteresis accounts of 14.4 %
of the travel range, which reveals a big challenge to the controller design task.
In this work, the hysteresis is treated as a bounded disturbance to the nanoposi-
tioning system, which is considered as a second-order linear system as described in
Eq. (4.1).

4.5.3 Plant Model Identification

The linear plant model is identified by the frequency response approach through
experimental study. Specifically, swept sine waves with the amplitude of 0.1 V and
frequency range of 1–2000 Hz are produced by the D/A channel to drive the PSA
through the high-voltage amplifier. The position responses of the nanopositioning
stage is recorded using a sampling rate of 5 kHz.
The input–output data sets are used to identify the plant transfer function by esti-
mating a model from the frequency response data. The frequency responses obtained
from the experimental data and the identified model are compared in Fig. 4.3. It is
observed that the nanopositioning system exhibits two main resonant modes. The
first resonant mode occurs around 522 Hz, and the identified second-order model
matches the system dynamics well in the frequencies up to 600 Hz. The identified
transfer function of the plant model is:

5.876 × 107
G(s) = . (4.76)
s 2 + 183.7s + 1.076 × 107
96 4 Model Predictive Discrete-Time Sliding-Mode Control

Fig. 4.3 Frequency

Magnitude (dB)
40
responses of system plant
obtained by experiment and 20
identified second-order
model 0
Modeled dynamics Residual modes
−20 2 3
10 10

Phase (degree)
0

−100

−200
Experiment
−300 Plant model
2 3
10 10
Frequency (Hz)

It is notable that in order to capture the high-frequency dynamics accurately, a

much higher order model is required to be identified. Here, a simple second-order
model is employed to demonstrate the effectiveness of the control scheme.

4.5.4 Controller Parameter Design

Comparing the transfer function (4.1) with Eq. (4.76) yields the mass–spring–damper
system parameters, which allows the calculation of the discrete-time system model
(4.5) with a sampling time of 0.5 ms. That is,
   
−0.0388 0.0003 5.6729
A= , B= , C = [1 0]. (4.77)
−3126.6906 −0.0921 17075.5580

First, the state observer is constructed by the pole placement technique for the
closed-loop observer system. Simulation study shows that small pole values lead to
rapid convergence in that a high bandwidth is achieved by the observer. However,
experiments reveal that too small pole produces instability of the system, which
may arise from the sensor noises and spillover effects of the neglected dynamics.
Thus, the poles are assigned as (−0.15, −0.15), which are selected to guarantee the
stability of the observer in the presence of unmodeled high-frequency dynamics. The
corresponding gains are obtained as L = [0.1691, 0.0003]T .
Next, the sliding function is designed by selecting appropriate proportional and
integral gain vectors G = [g1 , g2 ] and K I = [k1 , k2 ]. It is found that the matrix W
has two constant eigenvalues of 0 and 1. Each of the other two matrices U and V
has an eigenvalue fixed at 0. Thus, the gains are chosen to guarantee that the free
4.5 Experimental Investigation 97

eigenvalues λu and λv of U and V, respectively, locate inside the unit circle. Noting
that Ga = [ρ1 , ρ2 ] = [g1 + k1 , g2 + k2 ], the eigenvalues can be expressed by:

u 1 ρρ21 − u 2 v1 kk21 + v2
λu = , λv = (4.78)
u 1 ρρ21 + u 2 v1 kk21 + v2

where u 1 , u 2 , v1 , and v2 are all constants. It is seen that the ratios ρρ21 and kk21 are
related to the two eigenvalues λu and λv , respectively. The eigenvalues are assigned
as λu = 0.2 and λv = 0.4 to ensure a quick response of the closed-loop control
system. Then, by selecting k2 = 1, extensive simulations show that g2 < k2 does
not improve the performance. It is found that g2 = 10000 results in a better tracking
result, whereas further increase of g2 does not make more contribution. Thus, the gain
vectors can be calculated as G = [344418.5, 10000] and K I = [20293358.5, 1].
After some trials, the weighting parameter λ = 1019 is chosen for the MPDTSMC,
and parameters Ks = [1 50] and δ = 108 are selected for the DTSMC controller to
generate better results. In addition, the control action is constrained between −2 to
10 V, which will be produced by the D/A channel and then amplified by the high-
voltage amplifier with a gain of 10 to provide the driving voltage range of −20 to
100 V.

4.5.5 Simulation Studies

First, the performance of the implemented controllers are tested by conducting sim-
ulation studies. By defining the tracking error e = xd − x, the root-mean-square
(RMS) error is defined as follows:


1 
N
erms = ei2 . (4.79)
N
i=1

For a sinusoidal reference input as shown in Fig. 4.4a, the DTSMC tracking errors
and control actions with different parameter δ are shown in Fig. 4.4b, c, respectively.
With a larger value of δ = 1010 , it is observed from Eq. (4.47) that sat(Sk ) =
Sk /δ always holds. As a result, a smooth control action is produced as shown in
Fig. 4.4c. On the contrary, using a smaller value of δ = 106 , sat(Sk ) = sgn(Sk ) when
|Sk | > δ. Hence, slight oscillations are produced in the tracking results as shown
in Fig. 4.4b. Correspondingly, slight oscillation phenomenon occurs in the control
action, as revealed in Fig. 4.4c.
The RMS error versus the parameter δ is depicted in Fig. 4.4d. To generate a better
tracking result for the DTSMC scheme, δ = 108 is selected, which produces a RMS
error of 0.043 µm.
98 4 Model Predictive Discrete-Time Sliding-Mode Control

(a) (b)
20 1

Displacement error (μm)

Displacement (μm)

15 0.5

10 0

5 −0.5

0 −1
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time (s) Time (s)

(c) (d)
4 0.4

3
RMS error x (μm)
Control action (V)

0.3
2
0.2
1
0.1
0

−1 0 6 7 8 9 10 11
0 0.5 1 1.5 2 10 10 10 10 10 10
Time (s) Parameter δ

Fig. 4.4 Simulation results of DTSMC controller. a Desired position input. b Tracking errors of
DTSMC with δ = 106 (solid line), 108 (dash-dot line), and 1010 (dashed line). c Control actions
of DTSMC with δ = 106 (solid line), 108 (dash-dot line), and 1010 (dashed line). d RMS error of
DTSMC versus the parameter δ

Regarding the MPDTSMC, it is necessary to choose the suitable prediction hori-

zon N . Simulations are carried out to disclose the relationship between the steady-
state tracking error and the prediction horizon N . For a sinusoidal reference input
as shown in Fig. 4.5a, the RMS error versus the prediction horizon N is obtained as
shown in Fig. 4.5d. It is found that the larger the prediction horizon N , the smaller
the tracking error. However, the larger the prediction horizon N , the heavier the
calculation burden.
It is further observed from Fig. 4.5d that no significant improvement on the track-
ing result is achieved when N > 12. Hence, N = 12 is selected to make a com-
promise between the tracking accuracy and the computation load. With N = 12,
the tracking error and control action of the MPDTSMC are shown in Fig. 4.5b, c,
respectively. Obviously, no chattering phenomenon exists in the control action. Sim-
ulation result shows that the MPDTSMC achieves a RMS error of 0.044 µm, which
is comparable to the DTSMC simulation result.
The performances of the MPDTSMC and DTSMC controllers are validated
through experimental studies as conducted in the following discussion.
4.5 Experimental Investigation 99

(a) (b)
20 0.1

Displacement error (μm)

Displacement (μm)

15 0.05

10 0

5 −0.05

0 −0.1
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time (s) Time (s)

(c) (d)
4 0.05
Control action (V)

RMS error x (μm)

3
0.048
2
0.046
1
0.044
0

−1 0.042
0 0.5 1 1.5 2 5 10 15 20
Time (s) Prediction horizon N

Fig. 4.5 Simulation results of MPDTSMC controller. a Desired position input. b Tracking errors
of MPDTSMC with N = 12. c Control actions of MPDTSMC with prediction horizon N . d RMS
error of MPDTSMC versus the prediction horizon N

4.5.6 Experimental Testing Results

The performance testing of the designed controllers is carried out through several
experimental studies. For comparisons, the maximum tracking error and percent
RMS error are defined as:
max(|e|)
emax (%) = × 100 % (4.80)
max(xd ) − min(xd )
⎛ ! " ⎞
1 N 2
e
i=1 i
erms (%) = ⎝ ⎠ × 100 %.
N
(4.81)
max(xd ) − min(xd )

First, in order to conduct a comparison with the simulation results, the 2-Hz
reference signal with 20-µm amplitude as shown in Fig. 4.6a is used. The tracking
errors of the DTSMC and MPDTSMC schemes are illustrated in Fig. 4.6b. It is
observed that the maximum error emax and RMS error erms of the DTSMC controller
are 0.81 and 0.52 %, respectively. In contrast, the MPDTSMC scheme further reduces
100 4 Model Predictive Discrete-Time Sliding-Mode Control

Fig. 4.6 Experimental (a)

results of DTSMC and 20

Displacement (μm)
MPDTSMC controller.
a Reference position input. 15
b Experimental results of the
10
DTSMC and MPDTSMC
tracking errors, and the state 5
observer estimation errors
0
0 0.5 1 1.5 2
Time (s)
(b)
Displacement error (μm) 0.2 MPDTSMC DTSMC Observer

0.1

−0.1

−0.2
0 0.5 1 1.5 2
Time (s)
(c)
3
Control action (V)

0
MPDTSMC DTSMC
−1
0 0.5 1 1.5 2
Time (s)

the errors to emax = 0.77 % and erms = 0.39 %, respectively. In comparison with
the DTSMC results, the MPDTSMC reduces the maximum error by 5 % and RMS
error by 25 %, respectively. In addition, Fig. 4.6b also indicates that the state observer
estimates the position value with the maximum error below 0.67 %.
Additionally, by comparing the experimental results as shown in Fig. 4.6b with
simulation results as shown in Fig. 4.5b, it can be observed that the actual tracking
errors are worse than the simulation results of both DTSMC and MPDTSMC con-
trollers. The differences mainly arise from the noises and high-frequency dynamics,
which are not captured by the second-order model as employed in the simulation
study. The presence of high-frequency dynamics degrades the performance of the
controllers as indicated by the experimental results.
Second, the dynamic resolution of the nanopositioning stage is tested. It is known
that the resolution can be well presented using the sensitivity and noise variance of
the sensing system [2, 5]. However, such approach is more suitable for evaluating
the static resolution of the system. In order to exhibit the dynamic resolution and
demonstrate the effectiveness of the state observer, sinusoidal signals with small
4.5 Experimental Investigation 101

Fig. 4.7 Dynamic resolution (a)

test results of the 30
nanopositioning stage using Actual Estimated Reference

Displacement (nm)
a 2-Hz (a) and a 20-Hz 20
(b) sinusoidal wave inputs
with 10 nm peak-to-peak 10
amplitude
0

−10

−20
0 0.5 1 1.5 2
Time (s)

(b)
30
Displacement (nm)

20

10

−10
Actual Estimated Reference
−20
0 0.05 0.1 0.15 0.2
Time (s)

amplitudes are used as inputs in the experimental testing. For instance, Fig. 4.7a,
b depict the results which are obtained using reference inputs of 10-nm amplitude
along with the frequencies of 2 and 20 Hz, respectively. The sine waves can be clearly
identified even with a 20-Hz higher frequency input. It indicates that the resolution
of the positioning system is better than 10 nm. Furthermore, the results also illustrate
that the implemented state observer acts as a low-pass filter with the extra functions
of filtering the noisy sensor readings and suppressing the residual modes of high-
frequency dynamics.
Next, with the proposed MPDTSMC controller, the hysteresis effects are tested
by applying a 0.5-Hz reference input with the amplitude of 80 µm. Scaling the output
displacement with the maximum input voltage value yields the results as illustrated
in Fig. 4.2. It is observed that the closed-loop hysteresis is significantly suppressed
to 0.56 %. As compared with the open-loop hysteresis width (14.4 %), it has been
substantially reduced by 96.1 % owing to the effectiveness of the MPDTSMC strat-
egy. Actually, the proposed scheme significantly mitigates the tracking errors to
emax = 0.33 % and erms = 0.19 %, respectively.
Moreover, the bandwidth of the closed-loop system with the MPDTSMC control
is tested by applying a sinusoidal reference input with the amplitude of 0.2-µm and
varying frequencies of 1–900 Hz. The Bode diagrams are plotted in Fig. 4.8. It is
observed that the −3 dB bandwidth is 413 Hz, which is equivalent to 79 % of the first
102 4 Model Predictive Discrete-Time Sliding-Mode Control

Fig. 4.8 Frequency 20

Magnitude (dB)
responses of the closed-loop
0
system with the MPDTSMC (413 Hz, −3 dB)
controller −20

−40
1 2 3
10 10 10

Phase (degree)
0

−200

−400

−600
1 2 3
10 10 10
Frequency (Hz)

resonant frequency (522 Hz). It is notable that the above bandwidth is tested using a
small amplitude of the input command where the hysteresis can be neglected. As the
motion range increases, the bandwidth decreases due to the effect of the hysteresis.
Even so, it demonstrates the effectiveness of the implemented controller as compared
with the open-loop testing result as depicted in Fig. 4.3, which is also obtained using
a low-amplitude input.

4.5.7 Discussion on System Performance

The foregoing experimental studies confirm the feasibility and effectiveness of the
proposed control strategy. Moreover, it is observed that the resonant mode is sup-
pressed by the implemented MPDTSMC control. The observation spillover is suc-
cessfully eliminated by the state observer at the cost of a lower response speed as
reflected by the assigned pole values. On the other hand, the noise in sensor readings
requires a further treatment to enhance the positioning resolution of the system.
By adopting a closed-loop sampling rate of 2 kHz, which is sufficient for the
investigated plant, the above simulations and experiments demonstrate the effective-
ness of the reported control approaches. The controller parameters are not optimally
designed, and the current amplitudes of steady-state errors as shown in Fig. 4.6 indi-
cate that the performance of the MPDTSMC may be improved to further reduce the
tracking error down to the positioning resolution. For instance, an iterative control
[22] may be employed to compensate for the periodic errors.
In addition, the conducted investigations demonstrate an enhancement of the posi-
tioning accuracy for the nanopositioning system with the MPDTSMC over the con-
ventional DTSMC control. The results validate the effectiveness of the MPDTSMC
control scheme with PI sliding function and displays a great potential for the future
research.
4.6 Chapter Summary 103

4.6 Chapter Summary

This chapter shows that a nanopositioning stage with piezoelectric actuation can
be precisely controlled by using only a second-order mass–spring–damper sys-
tem model without modeling the hysteresis effects. Such a simple framework is
very attractive for its easy implementation characteristics, which is enabled by the
MPDTSMC with PI sliding function. It is found that the MPC strategy forces the
system state to slide in the vicinity of the sliding surface with an O(T 3 ) bound-
ary layer in an optimal manner, which eliminates the chattering phenomenon. The
PI action in the MPDTSMC drives the state tracking error to a boundary layer of
O(T 2 ) in the sliding mode, which indicates a much lower tracking error in compari-
son with a commonly designed DTSMC controller. The experimental results confirm
that the MPDTSMC controller improves the tracking performance as compared with
the DTSMC approach. In addition, the hysteresis has been substantially reduced to
0.56 % and a 10-nm dynamic resolution is achieved by the nanopositioning stage.
Because the implementation of the controller does not require any bounds on the
uncertainties and unknown parameters of the system, the MPDTSMC scheme can
be easily extended to the control of micro-/nanopositioning stages driven by other
types of actuators as well.

References

1. Abidi, K., Xu, J.X., Yu, X.: On the discrete-time integral sliding mode control. IEEE Trans.
Automat. Control 52(4), 709–715 (2007)
2. Aphale, S.S., Bhikkaji, B., Moheimani, S.O.R.: Minimizing scanning errors in piezoelectric
stack-actuated nanopositioning platforms. IEEE Trans. Nanotechnol. 7(1), 79–90 (2008)
3. Bartoszewicz, A.: Discrete-time quasi-sliding-mode control strategies. IEEE Trans. Ind. Elec-
tron. 45(5), 633–637 (1998)
4. Dannehl, J., Fuchs, F.W.: Discrete sliding mode current control of three-phase grid-connected
pwm converters. In: Proceeding of the 13th European Conference on Power Electronics and
Applications, pp. 1–10 (2009)
5. Dong, J., Salapaka, S.M., Ferreira, P.M.: Robust control of a parallel-kinematic nanopositioner.
J. Dyn. Syst. Meas. Control 130(4), 041007 (2008)
6. Furuta, K.: Sliding mode control of a discrete system. Syst. Control Lett. 14(2), 145–152 (1990)
7. Gao, W., Wang, Y., Homaifa, A.: Discrete-time variable structure control systems. IEEE Trans.
Ind. Electron. 42(2), 117–122 (1995)
8. Garcia-Gabin, W., Zambrano, D., Camacho, E.F.: Sliding mode predictive control of a solar
air conditioning plant. Control Eng. Pract. 17(6), 652–663 (2009)
9. Kim, B., Washington, G.N.: Nonlinear position control of smart actuators using model predic-
tive sliding mode control. In: Proceeding of ASME Conference on Smart Materials, Adaptive
Structures and Intelligent Systems, pp. 511–522 (2008)
10. Koshkouei, A.J., Zinober, A.S.I.: Sliding mode state observers for discrete-time linear systems.
Int. J. Syst. Sci. 33(9), 751–758 (2002)
11. Liaw, H.C., Shirinzadeh, B.: Enhanced adaptive motion tracking control of piezo-actuated
flexure-based four-bar mechanisms for micro/nano manipulation. Sens. Actuator A-Phys. 147,
254–262 (2008)
104 4 Model Predictive Discrete-Time Sliding-Mode Control

12. Liaw, H.C., Shirinzadeh, B.: Robust generalised impedance control of piezo-actuated flexure-
based four-bar mechanisms for micro/nano manipulation. Sens. Actuator A-Phys. 148, 443–453
(2008)
13. Liaw, H.C., Shirinzadeh, B., Smith, J.: Enhanced sliding mode motion tracking control of
piezoelectric actuators. Sens. Actuator A-Phys. 138, 194–202 (2007)
14. Liaw, H.C., Shirinzadeh, B., Smith, J.: Robust motion tracking control of piezo-driven flexure-
based four-bar mechanism for micro-nano manipulation. Mechatronics 18, 111–120 (2008)
15. Liaw, H.C., Shirinzadeh, B., Smith, J.: Sliding-mode enhanced adaptive motion tracking control
of piezoelectric actuation systems for micro/nano manipulation. IEEE Trans. Control Syst.
Technol. 16(4), 826–833 (2008)
16. Neelakantan, V.A., Washington, G.N., Bucknor, N.K.: Model predictive control of a two stage
actuation system using piezoelectric actuators for controllable industrial and automotive brakes
and clutches. J. Intell. Mater. Syst. Struct. 19(7), 845–857 (2008)
17. Pai, M.C., Sinha, A.: Sliding mode control of vibration in a flexible structure via estimated
states and H∞ /μ techniques. In: Proceeeding of American Control Conference, pp. 1118–1123
(2000)
18. Sarpturk, S., Istefanopulos, Y., Kaynak, O.: On the stability of discrete-time sliding mode
control systems. IEEE Trans. Automat. Control 32(10), 930–932 (1987)
19. Su, W.C., Drakunov, S.V., Ozguner, U.: An O(T 2 ) boundary layer in sliding mode for sampled-
data systems. IEEE Trans. Automat. Control 45(3), 482–485 (2000)
20. Wang, D.A., Huang, Y.M.: Application of discrete-time variable structure control in the vibra-
tion reduction of a flexible structure. J. Sound Vibr. 261(3), 483–501 (2003)
21. Wills, A.G., Bates, D., Fleming, A.J., Ninness, B., Moheimani, S.O.R.: Model predictive control
applied to constraint handling in active noise and vibration control. IEEE Trans. Control Syst.
Technol. 16(1), 3–12 (2008)
22. Wu, Y., Zou, Q.: Iterative control approach to compensate for both the hysteresis and the
dynamics effects of piezo actuators. IEEE Trans. Control Syst. Technol. 15(5), 936–944 (2007)
23. Xiao, L., Su, H., Chu, J.: Sliding mode prediction tracking control design for uncertain systems.
Asian J. Control 9(3), 317–325 (2007)
24. Xu, J.X., Abidi, K.: A discrete-time integral sliding mode control approach for output tracking
with state estimation. In: Proceeding of the 17th IFAC World Congress, pp. 14,199–14,204
(2008)
25. Xu, J.X., Abidi, K.: Discrete-time output integral sliding-mode control for a piezomotor-driven
linear motion stage. IEEE Trans. Ind. Electron. 55(11), 3917–3926 (2008)
26. Yong, Y.K., Aphale, S., Moheimani, S.O.R.: Design, identification and control of a flexure-
based XY stage for fast nanoscale positioning. IEEE Trans. Nanotechnol. 8(1), 46–54 (2009)
Chapter 5
Model Predictive Output Integral
Discrete-Time Sliding-Mode Control

This chapter presents a control scheme termed model predictive output integral
discrete-time sliding-mode control (MPOIDSMC) to achieve a precise positioning
for a piezoelectric actuation stage. The controller features an integral type of slid-
ing function based on the output error along with a sliding-mode state observer.
The model predictive control methodology and the integral item are introduced to
eliminate the chattering phenomenon and to achieve a low tracking error. The feasi-
bility and effectiveness of the MPOIDSMC scheme is verified by conducting several
experimental studies on a piezoelectric nanopositioning stage.

5.1 Introduction

In the last Chap. 4, a DTSMC scheme has been developed based on the full state
feedback approach. In a typical micro-/nano-positioning system, only the output is
available from the displacement sensor. Therefore, it is more attractive to design an
output-based DTSMC strategy [4, 8]. In addition, the sliding function is commonly
designed as a proportional action based on the output or output error. Recent works
[6, 15, 16] show that an integral action in the sliding function is capable of improving
the performance of the conventional DTSMC. At the same time, the strategy of model
predictive control (MPC) can be employed to enhance the performance of a discrete-
time system [12]. By combining the SMC and MPC concepts, model predictive
sliding-mode control has been recently proposed to achieve the advantages of both
methods [3, 14].
In this chapter, a scheme of MPOIDSMC with a sliding-mode state observer is pre-
sented for micro-/nano-positioning applications. The presented strategy is capable of
improving the system performance by integrating MPC, output integral discrete-time
sliding-mode control (OIDSMC), and state observer techniques together. Specifi-
cally, the MPC generates an optimal control to drive the system output to the sliding

© Springer International Publishing Switzerland 2016 105

Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2_5
106 5 Model Predictive Output Integral Discrete-Time Sliding-Mode Control

surface, and the OIDSMC maintains the output trajectory on the sliding surface for
the subsequent time. Moreover, a sliding mode observer is adopted to estimate the
system state and to suppress the spillover effects. It is demonstrated that the inte-
grated scheme eliminates the chattering effect by forcing the system trajectory to a
vicinity of the sliding surface with a boundary layer of thickness O(T 3 ) in an optimal
manner, and achieves an output tracking error of the order O(T 2 ). The produced ulti-
mate error has a much lower magnitude than those reported in most of the previous
works [15]. The theoretical analysis and the effectiveness of the proposed strategy
are validated by experimental studies conducted on a piezo-driven nanopositioning
system.

5.2 Problem Formulation

The dynamics model of a piezo-driven micro-/nano-positioning system can be

described by a linear time-invariant (LTI) model accompanied by disturbance:

Ẋ(t) = Ac X(t) + Bc u(t) + Dc f(t) (5.1)

Y(t) = CX(t) (5.2)

where the state vector X ∈ R n , the output Y ∈ R p , the control input u ∈ R m , the
disturbance vector f ∈ R q is supposed to be smooth and bounded, and m ≤ p < n.
The system matrices Ac , Bc , Dc , and C are constant with approximate dimensions.
The disturbance f describes the combined effects of unmodeled nonlinear
hysteresis, higher order dynamics, and external perturbations. It is assumed that
f satisfies the matching condition, i.e., Dc = Bc δd where δd is a scalar.
Using a sampling time T , the continuous-time system model (5.1) can be dis-
cretized as:

Xk+1 = AXk + Buk + dk (5.3)

Yk = CXk (5.4)

where Xk = X(kT ) and

A = e Ac T (5.5)
 T
B= eAc τ dτ Bc (5.6)
0
 T
dk = eAc τ Bc δd f((k + 1)T − τ )dτ. (5.7)
0

It is observed that both B and dk are of the order O(T ) with respect to the sampling
time T . Additionally, the Properties 4.1 and 4.2 are employed to facilitate the control
design. In addition, it is assumed that the triplet (A, B, C) of the nominal system is
5.2 Problem Formulation 107

both controllable and observable with the matrices B and C being of full rank. In
addition, the matrix CB is supposed to be invertible.
For a single-input and single-output (SISO) piezoelectric actuation system (i.e.,
m = p = 1), the concerned motion control problem is how to make the output
(Yk ) precisely track a command reference input. A discrete-time sliding mode-based
controller is presented in the next section.

5.3 MPOIDSMC Design

The design procedure of a MPOIDSMC control scheme is outlined in this section.

First, the OIDSMC is designed based on an integral sliding function. Then, an inte-
grated MPOIDSMC scheme is presented. The stability and tracking accuracy of the
control algorithm are evaluated in detail.

5.3.1 OIDSMC Controller Design

The output tracking error is defined as:

E k = Yk − Rk (5.8)

where Rk denotes the desired system output.

Based on the output error (5.8), a PI-type sliding function is defined as follows:

Sk = G E k + K I ξk (5.9)

where

S = {Sk |Sk = 0, for k = 0, 1, . . . } (5.10)

represents the sliding surface or sliding hyperplane, and the integral error is

ξk = E k + ξk−1 . (5.11)

In addition, G and K I are constant positive proportional and integral gains, which
are specified to assign appropriate eigenvalues for the system.
eq
Considering that the equivalent control u k is the solution of ΔS = Sk+1 − Sk = 0
[2], it can be deduced that

G E k+1 + K I ξk+1 = Sk
⇒ G E k+1 + K I (E k+1 + ξk ) = Sk
⇒ (G + K I )E k+1 + K I ξk = Sk
⇒ (G + K I )[C(AXk + Bu k + dk ) − Rk+1 ] + K I ξk = Sk (5.12)
108 5 Model Predictive Output Integral Discrete-Time Sliding-Mode Control

which allows the derivation of the equivalent controller:

u k = −(ΩCB)−1 (−Sk + ΩCAXk + ΩCdk + K I ξk − Ω Rk+1 )

eq
(5.13)

where
Ω = G + KI . (5.14)

However, the controller (5.13) is not realizable because it contains the unknown
disturbance term dk at the current time step. To solve this problem, the disturbance
dk is obtained by its one-step delayed estimation:

d̂k = dk−1 = Xk − AXk−1 − Bu k−1 (5.15)

which is derived by noting (5.3).

Thus, Eq. (5.13) becomes

u k = −(ΩCB)−1 (−Sk + ΩCAXk + ΩCdk−1 + K I ξk − Ω Rk+1 ). (5.16)

eq

5.3.1.1 Stability Analysis

Substituting Eq. (5.16) into Eq. (5.3) along with a necessary calculation, yields the
closed-loop system state dynamics:

Xk+1 = UXk + Vdk−1 + dk − dk−1 + B(ΩCB)−1 (Ω Rk+1 − G Rk ) (5.17)

with

U = A − B(ΩCB)−1 (ΩCA − GC) (5.18)

V = I − B(ΩCB)−1 ΩC (5.19)

where I denotes an identity matrix.

The eigenvalues of the matrices U and V are the solutions of det(zI − U) = 0
and det(zI − V) = 0, respectively, which determine the closed-loop state dynamics
behavior. To ensure the stability of the closed-loop system, the eigenvalues of the
two matrices should locate inside the unit circle in the complex z-plane.

5.3.1.2 Tracking Error Analysis

In view of Eqs. (5.12) and (5.14), the following expression can be obtained:

Sk+1 = Ω[C(AXk + Bu k + dk ) − Rk+1 ] + K I ξk . (5.20)

5.3 MPOIDSMC Design 109

Substituting Eq. (5.16) into Eq. (5.20), yields

Sk+1 = Sk + ΩCε k (5.21)

where

εk = dk − dk−1
 T  (k+1)T −τ
Ac τ
= e Bc δd ḟ(σ )dσ dτ = O(T 2 ). (5.22)
0 kT −τ

Thus, in consideration of Property 4.2, it can be deduced from Eq. (5.21) that Sk
has a magnitude of the order O(T ), i.e., the sliding surface has an ultimate bound of
O(T ).
By noting Eqs. (5.8) and (5.9), the one-step forward error equation can be
expressed as:

E k+1 = Yk+1 − Rk+1

= C(AXk + Bu k + dk ) − Rk+1 . (5.23)

Substituting Eq. (5.16) into Eq. (5.23), yields

E k+1 = C(dk − dk−1 ) − Ω −1 (K I ξk − Sk )

= Cε k + Ω −1 G E k . (5.24)

Then, in view of Eq. (5.24) and Property 4.2, it is derived that the ultimate bound
of the output tracking error E k is of the order O(T ).
The error dynamics equation (5.24) also indicates that the gains should be designed
to satisfy the condition:
|Ω −1 G| < 1 (5.25)

so that the closed-loop system is stable.

The equivalent controller as designed above takes effect in the sliding mode when
the system trajectory is kept on the sliding surface. However, if the initial state of the
system does not lie on the sliding surface S or external disturbances occur during the
sliding-mode motion, the equivalent control alone cannot drive the trajectory toward
S. Thus, an augmented controller is designed in the following discussion.

5.3.2 MPOIDSMC Controller Design

A common method to design a robust controller is to augment the equivalent con-

trol with a discontinuous switching control term, whose role is to force the system
trajectory to reach the sliding mode. However, the common switching control may
induce unwanted chattering phenomenon [17].
110 5 Model Predictive Output Integral Discrete-Time Sliding-Mode Control

Here, an integrated MPC and OIDSMC control scheme is established to handle

the nonlinearities and disturbances existed in the system. Specifically, the MPC is
mp
used to generate an optimal control (u k ) to drive the system output to the sliding
eq
surface, and the OIDSMC (u k ) is employed to maintain the output trajectory on
the sliding surface for the subsequent time. In addition, a sliding-mode observer is
designed to estimate the system state and to eliminate the spillover effects.

5.3.2.1 Controller Design

Substituting the augmented control

eq mp
uk = uk + uk (5.26)

into Eq. (5.20) and noting Eq. (5.16), gives

mp
Sk+1 = Sk + ΩCBu k + ΩCε k (5.27)

which describes the dynamics of the sliding mode and also represents a one-step
future prediction of the sliding-mode dynamics.
Based on Eq. (5.27), an N -step prediction of the sliding-mode dynamics can be
obtained as follows:
mp mp mp
Sk+N = Sk + ΩCB(u k + u k+1 + · · · + u k+N −1 )
+(ΩCε k+N −1 + · · · + ΩCε k+1 + ΩCε k ) (5.28)

where the positive integer N is known as the prediction horizon.

The N prediction functions can be stacked together into the form:

Sak = ΛSk + Φuak−1 + Γ ζ k−1 (5.29)

where the vectors for the future sliding function, control effort, and disturbance
estimation errors are:
⎡ ⎤ ⎡ mp ⎤ ⎡ ⎤
Sk+1 uk εk
⎢ Sk+2 ⎥ ⎢ u mp ⎥ ⎢ ε k+1 ⎥
⎢ ⎥ a ⎢ k+1 ⎥ ⎢ ⎥
Sk = ⎢ . ⎥ , uk−1 = ⎢
a
. ⎥ , ζ k−1 = ⎢ .. ⎥. (5.30)
⎣ .. ⎦ ⎣ .. ⎦ ⎣ . ⎦
mp
Sk+N u k+N −1 ε k+N −1

Additionally, the three matrices take on the following forms:

5.3 MPOIDSMC Design 111
⎡ ⎤
I
⎢I⎥
⎢ ⎥
Λ = ⎢.⎥ (5.31)
⎣ .. ⎦
I
⎡ ⎤
ΩCB 0 . . . 0
⎢ ΩCB ΩCB . . . 0 ⎥
⎢ ⎥
Φ=⎢ . .. . . .. ⎥ (5.32)
⎣ .. . . . ⎦
ΩCB ΩCB . . . ΩCB
⎡ ⎤
ΩC 0 . . . 0
⎢ ΩC ΩC . . . 0 ⎥
⎢ ⎥
Γ =⎢ . .. . . .. ⎥ . (5.33)
⎣ .. . . . ⎦
ΩC ΩC . . . ΩC

Using the above notations, a cost function for minimization can be expressed by:

T
J = (Sak )T Sak + w uak−1 uak−1 (5.34)

where w is the weighting to limit the partial control effort. It is notable that uak−1
approaches to zero as Sak tends to zero. Thus, the cost function J will vanish at the
steady state.
Substituting Eq. (5.29) into the cost function (5.34) and applying the optimization
criterion by setting ∂u∂aJ = 0, yields
k−1

uak−1 = −(Φ T Φ + wI)−1 Φ T (ΛSk + Γ ζ k−1 ) (5.35)

where the future disturbance values in ζ k−1 are unknown. Thus, they are estimated
by
ζ̂ k−1 = [εk−1
T
ε k−1
T
· · · ε k−1
T
]T . (5.36)

Because only the first value of the predicted control sequence is used, the optimal
control effort can be obtained as:

= −e(Φ T Φ + wI )−1 Φ T (ΛSk + Γ ζ̂ k−1 )

mp
uk (5.37)

where the selection vector e = [1, 0, 0, . . . , 0].

Therefore, the total control action can be expressed as:
eq mp
uk = uk + uk
= −(ΩCB)−1 [−Sk + ΩCAX k + ΩCdk−1 + K I ξk − Ω Rk+1 ]
−e(Φ T Φ + wI)−1 Φ T (ΛSk + Γ ζ̂ k−1 ). (5.38)
112 5 Model Predictive Output Integral Discrete-Time Sliding-Mode Control

5.3.2.2 Stability Analysis

Substituting Eq. (5.35) into the stacked equation (5.29) yields

Sak = ΛSk − Φ(Φ T Φ + wI)−1 Φ T (ΛSk + Γ ζ̂ k−1 ) + Γ ζ k−1 (5.39)

where w is defined in the cost function (5.34) to limit the partial control input. For the
convenience of analysis, it is assumed that there is no penalty for the partial control
mp
effort u k , i.e., w = 0.
Then, Eq. (5.39) becomes

Sak = ΛSk − (ΛSk + Γ ζ̂ k−1 ) + Γ ζ k−1

= Γ (ζ k−1 − ζ̂ k−1 ). (5.40)

Considering only the first value of the predicted sliding hyperplane vector (5.40)
and taking into account Eqs. (5.30) and (5.36), yields

Sk+1 = ΩC(ε k − ε k−1 )

= ΩC(dk − 2dk−1 + dk−2 ) = O(T 3 ) (5.41)

which indicates that the magnitude of Sk+1 is of the order O(T 3 ) with the consider-
ation of Property 4.1.
It is assumed that the change rate of εk is limited:

|Sk+1 | = |ΩC(ε k − ε k−1 )| ≤ Δ S (5.42)

where 2Δ S is defined as the quasi-sliding mode bandwidth.

According to Definition 1 and Definition 2 in [1], it can be concluded that the
system (5.3) satisfies the reaching condition of the quasi-sliding mode in the Δ S
vicinity of the sliding surface S in a finite number of steps. Therefore, the closed-
loop control system is stable.
Remark 5.1 The essence of the presented MPOIDSMC is to drive the system tra-
jectory to a quasi-sliding mode as shown in Eq. (5.42). One advantage of this quasi-
sliding model [1] lies in that it does not require the system state to cross the sliding
hyperplane S in each successive control step. Therefore, the chattering phenomenon
is eliminated. This further results in a reduction of the control effort and improved
control performance [1].
Remark 5.2 It is observed that Sk+1 has the magnitude of order O(T 3 ) as shown in
Eq. (5.41). Therefore, the MPOIDSMC controller (5.38) drives the system trajectory
to slide in the vicinity of S with an ultimate boundary layer of thickness O(T 3 ),
whose magnitude is much lower than the O(T ) thickness achieved by the OIDSMC
controller. Therefore, a much lower control error is expected from the MPOIDSMC
scheme, which will be proved later.
5.3 MPOIDSMC Design 113

5.3.2.3 Sliding-Mode Observer Design

An insight into the controller (5.38) reveals that its implementation needs the state
feedback Xk = [xk , ẋk ]T of the system. However, only the displacement xk is avail-
able in practice. In addition, a low-order linear model is usually employed as a nom-
inal system model. Thus, the higher order dynamics is neglected in the controller
design. However, the neglect of residual modes may result in both control spillover
and observation spillover [9], which are undesirable because they may induce sys-
tem instability and performance degradation [11]. It has been shown [5] that the
sliding-mode observer is capable of eliminating the observation spillover effects in
the vibration control of flexible structures.
Although various types of state observers are available [13], it is found that a
second-order sliding-mode observer [7] is better than the ordinary first-order sliding-
mode observer [10] in terms of chattering reduction. Hence, a second-order discrete
sliding-mode state observer [7] is designed in this work to suppress the spillover
effects and to estimate the full state by making use of the only measurable position
information of the nanopositioning system.
The second-order discrete-time sliding-mode observer takes on the form:

Ỹk
X̂k+1 = AX̂k + Bu k + HỸk + dk−1 + Ko sat( ) (5.43)
δo
Ŷk = CX̂k (5.44)

where Ỹk = Yk − Ŷk is the sliding function, X̂k represents the estimate of Xk , H
is the observer gain vector, and δo denotes the boundary layer. The switching gain
vector Ko is designed as:

Ko ≥ |dk − dk−1 |. (5.45)

Subtracting Eq. (5.43) from Eq. (5.3) allows the derivation of

 Ỹk
Xk+1 = (A − HC)
Xk + dk − dk−1 − Ko sat( ) (5.46)
δo

where Xk = Xk − X̂k denotes the estimation error of the state Xk . It is observed that
the state observer is independent of the controller. To guarantee the stability of the
observation unit, the observer gain H should be designed to make all the eigenvalues
of the matrix A − HC locate inside the unit circle.

Remark 5.3 The convergence speed of the observer relies on the location of the
poles of the matrix A − HC. The poles can be arbitrarily placed by the gain vector
H. Generally, small pole value will give rapid convergence, and the convergence
speed of the observer should be faster than the system response so that the observer
dynamics is insignificant compared to the system dynamics. However, too small pole
114 5 Model Predictive Output Integral Discrete-Time Sliding-Mode Control

produces instability of the system, which comes from the noise and spillover effects
of the neglected system dynamics. Hence, a trial-and-error approach is needed for
an appropriate pole placement.

5.3.2.4 Tracking Accuracy Analysis

In view of the convergence of the second-order sliding-mode observer, the following

equation can be deduced at the ultimate steady state:

Xk+1 ≈ (A − HC)
 Xk + dk − dk−1 . (5.47)

Considering that dk − dk−1 is of the order O(T 2 ), it can be observed from

Eq. (5.47) that the ultimate bound of the estimation error  Xk is of the order O(T ).
To evaluate the tracking error bound, Eq. (5.38) is substituted into Eq. (5.23). Then,
a necessary calculation leads to

E k+1 = Cε k + Ω −1 G E k − CBe(Φ T Φ + wI)−1 Φ T (ΛSk + Γ ζ̂ k−1 ). (5.48)

To facilitate the analysis, it is assumed that there is no penalty for the partial
mp
control effort u k , i.e., w = 0. Then, Eq. (5.48) reduces to

E k+1 = Cε k + Ω −1 G E k − CBeΦ −1 (ΛSk + Γ ζ̂ k−1 )

= Cε k + Ω −1 G E k − CB(ΩCB)−1 (Sk+1 + ΩCε k )
= Ω −1 G E k − Ω −1 Sk+1 . (5.49)

Taking into accout that Sk+1 is of the order O(T 3 ) as shown in Eq. (5.41), it can
be deduced from Eq. (5.49) that E k = O(T 2 ). That is, the ultimate bound of the
output tracking error is O(T 2 ), which is much lower than the O(T ) tracking error
obtained with the OIDSMC designed earlier.

5.4 Experimental Investigations

In this section, the designed controller is verified by a series of experimental studies

conducted on a piezo-driven nanopositioning system.

5.4.1 Experimental Setup

The experimental setup employed in this work is depicted in Fig. 4.1. The piezostage
is composed of a flexure-based compliant mechanism driven by a PSA with the
stroke of 14.5 µm (model: TS18-H5-202, from Piezo Systems, Inc.). The stage output
5.4 Experimental Investigations 115

PC + MATLAB System plant

Voltage
dSPACE D/A amplifier PSA Nano-
DS1103 positioning
board Signal Capacitive stage
A/D
conditioner sensor

Fig. 5.1 Experimental setup for a nanopositioning stage

platform delivers a one-axis translational motion by making use of elastic deforma-

tions of the flexure hinges, and the output displacement is measured by a capacitive
sensor (model: D-510.050, from Physik Instrumente Co., Ltd.). A dSPACE DS1103
(from dSPACE GmbH) rapid prototyping system equipped with 16-bit A/D and D/A
converters are employed to implement the controller. The D/A channel produces an
analog voltage output which is then amplified by a high-voltage amplifier (model:
EPA-104, from Piezo Systems, Inc.) to provide a voltage ranging from −100 to 100 V
for driving the PSA. In addition, the sensor output voltage signal is passed through
a signal conditioner (model: E-852, from Physik Instrumente Co., Ltd.) and then
acquired by the A/D channel. Control algorithms are developed with MATLAB
software and downloaded to DS1103 board via the ControlDesk interface to realize
the real-time control. A block diagram of the hardware is shown in Fig. 5.1.

5.4.2 Plant Model Identification

The nanopositioning system is viewed as a linear system with bounded disturbances

as described in (5.1). The linear plant model is identified by the frequency response
approach via experiments. Specifically, swept sine waves with the amplitude of 0.1 V
and frequency range of 1–2000 Hz are applied to the voltage amplifier to drive the
PSA. The displacement responses of the nanopositioning stage are recorded using a
sampling rate of 5 kHz.
The input–output data sets are used to identify the plant model by estimating a
transfer function from the frequency response data. The frequency responses obtained
by the experiment and the identified model are compared in Fig. 5.2. The first resonant
mode occurs around 522 Hz, and the identified second-order model matches the
system dynamics well in the frequencies below 600 Hz. The identified second-order
transfer function is described by Eq. (5.50)

5.876 × 107
G(s) = . (5.50)
s 2 + 183.7s + 1.076 × 107
116 5 Model Predictive Output Integral Discrete-Time Sliding-Mode Control

Magnitude (dB)
40

20

0
Unmodeled dyn.
−20
2 3
10 10
Phase (degree)

−100

−200
Experiment
−300 Plant model
2 3
10 10
Frequency (Hz)

Fig. 5.2 Plant frequency responses obtained by experiments and the identified second-order model

In order to capture the high-frequency dynamics accurately, the identification of a

much higher order model is required. Here, a simple second-order model is employed
to demonstrate the effectiveness of the control scheme.

5.4.3 Controller Parameter Design

By comparing the state-space model (5.1) with (5.50) allows the generation of the
system parameters, which can be converted into the discrete-time system model (5.3)
with a sampling time T . In order to select a suitable T , the influences of the sampling
rate selection on the controllability and observability of the discrete-time system are
investigated.
As quantitative measures, the minimum singular values of the controllability
matrix (C ) and observability matrix (O):

C
C = , O = [B AB] (5.51)
CA

are employed as the controllability index and observability index, respectively. They
serve as the indicators of how far the system is from being uncontrollable and unob-
servable. As the sampling frequency increases from 1000 to 5000 Hz, the two indices
are generated as shown in Fig. 5.3.
It is observed that the magnitude of the observability index is about 104 times lower
than that of the controllability index. Hence, the system is prone to be unobservable
5.4 Experimental Investigations 117

−4
x 10

7.5 Controllability 3
Observability
Controllability index

Observability index
5 2

2.5 1

0 0
1000 1500 2000 2500 3000 3500 4000 4500 5000
Sampling frequency (Hz)

Fig. 5.3 The controllability index and observability index versus the sampling frequency

rather than uncontrollable. Both of the two indices arrive at the maximum value
under the sampling frequency around 2000 Hz, and they decay as the increasing of
the sampling rate. Additionally, in order to capture the system behavior precisely, a
sampling frequency larger than the first resonant frequency (522 Hz) of the system
is preferred. Considering the above issues, a sampling rate of 2000 Hz is adopted.
With T = 0.0005 s, the discrete-time system is represented by:
 
−0.0388 0.0003 5.6729
A= , B= , C = [1 0]. (5.52)
−3126.6906 −0.0921 17075.5580

Based on the obtained linear model, a sliding-mode observer is then constructed.

The observer gain vector H is designed by the pole placement technique. Simulations
suggest that small pole values produce rapid convergence in that a high bandwidth is
achieved by the observer. However, experiments reveal that too small pole leads to
instability of the system, which arises from the sensor noises and spillover effects of
the neglected dynamics. To make a tradeoff, the two poles are assigned as (−0.15,
−0.15), which are chosen to guarantee a relatively quicker convergence as well as
robustness of the observer in the presence of unmodeled dynamics. The correspond-
ing gains are calculated as H = [0.1691, 0.0003]T . In addition, the switching gain
vector is selected as Ko = [0.1, 0.1]T and the boundary layer is chosen as δo = 50
for the observer.
Next, the sliding function is designed by assigning appropriate proportional and
integral gains G and K I . It is found that the matrix U has a fixed eigenvalue of
λu1 = −0.9669 and the matrix V has two constant eigenvalues of 0 and 1. Thus, the
gains are chosen to ensure that the free eigenvalue λu2 of the matrix U locates inside
118 5 Model Predictive Output Integral Discrete-Time Sliding-Mode Control

the unit circle. Moreover, it is found that as the increasing of sampling rate, the fixed
eigenvalue moves more closely to the boundary of the unit circle, which drives the
closed-loop system to unstable as a consequence. Noting that Ω = G + K I , the
eigenvalue can be obtained as:
G
λu2 = (5.53)
Ω
which allows the calculation of
G
KI = − G. (5.54)
λu2

By selecting G = 1, K I can be considered as a function of the eigenvalue λu2 .

The influences of the selection of λu2 on the controller performances are investigated
in the subsequent section. In addition, to obtain better results, the prediction horizon
N = 10 and weighting parameter w = 2 × 105 are chosen for the MPOIDSMC after
some trials.

5.4.4 Experimental Studies

The performance of the designed controller is verified by conducting several exper-

imental studies.
First, a 10-μm step positioning is carried out to discover the effects of the assign-
ment of the free eigenvalue λu2 . For instance, with three different λu2 of 0.80, 0.90 and
0.95, the corresponding K I values are calculated as 0.25, 0.11 and 0.05, respectively.
The positioning results are shown in Fig. 5.4. It is observed that as the increasing of
the eigenvalue λu2 , the integral action in the sliding function is weakened as the gain
K I is decreased. As a result, the speed of the system transient response becomes slow.
The experimental result also demonstrates the effectiveness of the integral effect in
the controller. To make a compromise between the overshoot and the response speed,
λu2 = 0.90 corresponding to K I = 0.11 is selected for the set-point positioning.
Next, a consecutive step positioning with a step size of 10 nm is performed and
the results are shown in Fig. 5.5. The steps can be identified from the results, which
indicates that the positioning resolution of the system is about 10 nm. Additionally,
the observer output indicates that the sliding-mode observer is capable of estimating
the state from the noisy sensor readings.
Then, the motion tracking performance of the implemented controller is tested
using a 5-Hz sinusoidal reference input as shown in Fig. 5.6a. In order to obtain a
better result, the integral gain is selected as K I = 0.6667 which corresponds to
λu2 = 0.60. The influence of the weighting parameter w selection is investigated by
carrying out a sinusoidal motion tracking task. The 5-Hz reference input is shown in
Fig. 5.6a. The tracking errors of the MPOIDSMC with different w values are shown
in Fig. 5.6b. It is found that a larger weighting parameter w = 2.5 × 105 results in
5.4 Experimental Investigations 119

11

10

9
λu2=0.80 (KI=0.25)
8
Displacement (μm)

7 λ =0.90 (K =0.11)
u2 I
6
λ =0.95 (K =0.05)
5 u2 I

4
Reference
3

0
0.95 1 1.05 1.1 1.15 1.2
Time (s)

Fig. 5.4 The 10-µm step positioning results with different integral gains K I

60
Actual
50 Estimated
Reference
40
Displacement (nm)

30

20

10

−10

−20

−30
0 0.5 1 1.5 2 2.5 3 3.5
Time (s)

Fig. 5.5 The 10-nm consecutive step positioning results

a larger RMS tracking error of 0.0327 µm. By reducing the weighting parameter to
w = 2.0 × 105 , a smaller RMS error of 0.0288 µm is obtained. However, further
reduction of the weighting parameter to w = 1.5 × 105 leads to the oscillation in
the control result, which causes a larger RMS error of 0.1660 µm. The reason lies
in that a smaller weighting parameter imposes less suppression on the MPC control
action. To generate a better control result, w = 2.0 × 105 is selected in this work.
120 5 Model Predictive Output Integral Discrete-Time Sliding-Mode Control

(a) 10

Displacement (μm)
8

0
0 0.2 0.4 0.6 0.8 1
Time (s)

(b) 0.1
Displacement error (μm)

0.05

−0.05

w=2.5e5
−0.1
w=2.0e5
(b) w=1.5e5
−0.15
0 0.2 0.4 0.6 0.8 1
Time (s)

Fig. 5.6 Sinusoidal tracking results of MPOIDSMC with different parameters w

For a comparison study, the OIDSMC controller is also implemented. The control
results of the OIDSMC and MPOIDSMC controllers are illustrated in Fig. 5.7. Due
to the ummodeled hysteresis effect, the OIDSMC control produces the maximum
peak-to-peak (PP) tracking error of 0.3117 µm and RMS error of 0.0731 µm, i.e., 3.12
and 0.73 % of the concerned motion range, respectively. In contrast, the MPOIDSMC
scheme leads to the maximum PP error of 0.1586 µm and RMS error of 0.0288 µm,
which accounts of 1.59 and 0.29 % of the motion range, respectively. As compared
to the OIDSMC, the MPOIDSMC controller substantially reduces the PP and RMS
errors by 49 and 60 %, respectively. The effectiveness of the proposed control is
clearly revealed.
In addition, the position estimated by the sliding-mode observer is also illustrated
in Fig. 5.7a, and the error of the observer output with respect to the reference input is
depicted in Fig. 5.7b. The error has PP and RMS values of 1.27 and 0.44 % relative to
the reference motion range, which demonstrates the efficiency of the sliding-mode
observer.
5.4 Experimental Investigations 121

(a) 10

Displacement (μm)
8

2 Reference
Observer
0
0 0.2 0.4 0.6 0.8 1
Time (s)

(b) 0.2
Displacement error (μm)

0.1

−0.1

OIDSMC MPOIDSMC Observer

−0.2
0 0.2 0.4 0.6 0.8 1
Time (s)

Fig. 5.7 Comparison of sinusoidal tracking results. a The reference position input and the position
estimated by the sliding mode observer. b OIDSMC and MPOIDSMC control errors and the observer
output error

5.4.5 Discussion on Controller Performance

The foregoing experimental results show that the observation spillover is successfully
eliminated by the designed sliding-mode observer. A precise positioning with a res-
olution around 10 nm is achieved, and a fine motion tracking superior to OIDSMC
result is obtained, which indicates that the hysteresis effects have been alleviated
effectively. It is notable that the positioning resolution of the system can be enhanced
by filtering the noises in the sensor readings.
The controller parameters are not optimally designed, and the current ampli-
tudes of steady-state errors indicate that the performance of the MPOIDSMC may
be improved to further reduce the control error down to the positioning resolution.
Although there is plenty of room for performance improvement of the proposed
controller, the enhancement of positioning precision for the micro-/nano-positioning
system over OIDSMC control as elaborated by the conducted investigations demon-
strates the effectiveness of the MPOIDSMC control scheme with integral sliding
function and displays great potential for the future research.
122 5 Model Predictive Output Integral Discrete-Time Sliding-Mode Control

As an output-based DTSMC, the major restriction of the proposed control scheme

lies in that it is only applicable to minimum phase systems. In addition, it is found
that the selection of the sampling frequency is limited by the observability of the
discretized system. How to extend the controller to nonminimum phase systems and
how to improve the system observability remain to be explored in the future work.

5.5 Chapter Summary

This chapter shows that a piezoelectrically driven micro-/nano-positioning system

can be precisely controlled with the MPOIDSMC controller. The OIDSMC, MPC,
and state observer are integrated together to improve the system performance, and
the accuracy and stability of the control algorithm are analyzed in detail. Experimen-
tal results confirm that the unmodeled hysteresis has been effectively suppressed.
A 10-nm resolution is achieved by the nanopositioning stage, and the MPOIDSMC
scheme substantially improves the tracking performance as compared to the tra-
ditional OIDSMC control. Because the implementation of the controller does not
require any bounds on uncertainties and unknown parameters of the system, the
reported controller can be easily extended to other types of micro-/nano-positioning
systems. Future works may be conducted in terms of the extension of the controller to
nonminimum phase systems and the enhancement of the observability for discretized
systems.

References

1. Bartoszewicz, A.: Discrete-time quasi-sliding-mode control strategies. IEEE Trans. Ind.

Electron. 45(5), 633–637 (1998)
2. Furuta, K.: Sliding mode control of a discrete system. Syst. Control Lett. 14(2), 145–152 (1990)
3. Garcia-Gabin, W., Zambrano, D., Camacho, E.F.: Sliding mode predictive control of a solar
air conditioning plant. Control Eng. Pract. 17(6), 652–663 (2009)
4. Ha, Q.P., Trinh, H., Nguyen, H.T., Tuan, H.D.: Dynamic output feedback sliding-mode control
using pole placement and linear functional observers. IEEE Trans. Ind. Electron. 50(5), 1030–
1037 (2003)
5. Kim, M.H., Inman, D.J.: Reduction of observation spillover in vibration suppression using a
sliding mode observer. J. Vib. Control 7(7), 1087–1105 (2001)
6. Li, Y.F., Wikander, J.: Model reference discrete-time sliding mode control of linear motor
precision servo systems. Mechatronics 14(7), 835–851 (2004)
7. Mihoub, M., Nouri, A.S., Ben Abdennour, R.: A chattering free second order discrete slid-
ing mode observer: an experimentation on a chemical reactor. In: Proceedings of the 18th
Mediterranean Conference on Control and Automation, pp. 739–744 (2010)
8. Nguyen, T., Su, W.C., Gajic, Z.: Singular perturbation analysis of discrete-time output feedback
sliding mode control with disturbance attenuation. In: Proceedings of the American Control
Conference, pp. 757–762 (2009)
9. Pai, M.C., Sinha, A.: Sliding mode control of vibration in a flexible structure via estimated states
and H∞ /μ techniques. In: Proceedings of the American Control Conference, pp. 1118–1123
(2000)
References 123

10. Thein, M.W.L.: A discrete time variable structure observer with overlapping boundary layers.
In: Proceedings of the American Control Conference, pp. 2633–2638 (2002)
11. Wang, D.A., Huang, Y.M.: Application of discrete-time variable structure control in the vibra-
tion reduction of a flexible structure. J. Sound Vib. 261(3), 483–501 (2003)
12. Wills, A.G., Bates, D., Fleming, A.J., Ninness, B., Moheimani, S.O.R.: Model predictive control
applied to constraint handling in active noise and vibration control. IEEE Trans. Control Syst.
Technol. 16(1), 3–12 (2008)
13. Wu, L., Gao, H., Wang, C.: Quasi sliding mode control of differential linear repetitive processes
with unknown input disturbance. IEEE Trans. Ind. Electron. 58(7), 3059–3068 (2011)
14. Xiao, L., Su, H., Chu, J.: Sliding mode prediction tracking control design for uncertain systems.
Asian J. Control 9(3), 317–325 (2007)
15. Xu, J.X., Abidi, K.: A discrete-time integral sliding mode control approach for output tracking
with state estimation. In: Proceedings of the 17th IFAC World Congress, pp. 14,199–14,204
(2008)
16. Xu, J.X., Abidi, K.: Discrete-time output integral sliding-mode control for a piezomotor-driven
linear motion stage. IEEE Trans. Ind. Electron. 55(11), 3917–3926 (2008)
17. Zhu, Z., Xia, Y., Fu, M.: Adaptive sliding mode control for attitude stabilization with actuator
saturation. IEEE Trans. Ind. Electron. 58(10), 4898–4907 (2011)
 Part III
Hysteresis-Model-Free,
State-Observer-Free Feedback Control
Chapter 6
Digital Sliding-Mode Control
of Second-Order Systems

This chapter presents the precision motion control of a piezoelectric bimorph actuator
without using a hysteresis model and a state observer. A digital sliding-mode control
(DSMC) scheme is developed based on a second-order plant model of the system.
The nonlinearity effect is considered as a lumped perturbation, which is estimated
by resorting to a one-step delayed estimation technique. The effectiveness of the
approach without using state observer and hysteresis model has been validated by
experimental investigations.

6.1 Introduction

Generally, discrete-time sliding-mode control (DTSMC) can be categorized into

state-based and output-based methods. The former is developed based on the system
state or state error [1, 17, 21], while the latter is implemented on the base of the
system output or output error [18, 20]. Usually, the implementation of both methods
requires the state feedback of the system. However, in majority of practical situa-
tions, only the position information of a piezoelectric actuator system is provided
by the displacement sensor. Hence, a state observer is indispensable for the practical
realization of DSMC [1, 18, 20, 21], which complicates the control design proce-
dure. Furthermore, an improperly designed state observer may cause instability of
the system. In this sense, it is desirable to eliminate the use of state observer. A
DTSMC whose implementation does not need a state observer is termed a DSMC
scheme.
In the literature, a few efforts have been made toward the DSMC strategy. For
example, an input–output-based adaptive DSMC has been developed in [13], which
is based solely on input and output data. However, the controller is dedicated to a first-
order model with long dead time. It is not suitable for a piezoelectric actuation system
which typically features a plant model of higher order preceded by complicated
nonlinearity. To relax the restriction of the first-order plant model as required in
© Springer International Publishing Switzerland 2016 127
Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2_6
128 6 Digital Sliding-Mode Control of Second-Order Systems

[13], a new input–output-based DSMC control has been figured out for a velocity
tracking system [14]. Adaptive coefficients of the switching hyperplane are employed
to tolerate the modeling error, and the effectiveness is verified by simulation studies.
Nonetheless, due to a relatively long parameter adaptation process, the adaptive
control normally leads to a slow transient response before attaining the steady state
[4, 19]. On the other hand, the generalized minimum variance control (GMVC)
scheme can be employed to develop a DSMC control without using a state observer.
To improve the tracking accuracy, a DSMC scheme is presented in [10] to achieve the
O(T 2 ) and O(T 3 ) accuracy of the sliding hypersurface. This scheme is developed
based on input/output measurements and the GMVC technique. Nevertheless, the
output tracking accuracy is unknown and filtering is required to suppress the induced
chattering phenomenon. To eliminate the chattering completely, a fuzzy sliding mode
control is demanded [9]. In addition, the multirate output feedback technique has been
developed to release the need of the entire state vector for the realization of DSMC
control scheme [2].
This chapter presents the design of a simple DSMC scheme without using the
complicated hysteresis model and state observer for precision motion control of a
piezoelectric actuator. Specifically, the unmodeled nonlinearity effects are treated as
a lumped perturbation and the perturbation is estimated by resorting to a one-step
delayed estimation technique. Furthermore, the avoidance of state observer is real-
ized by developing a new DSMC based on a discrete-time second-order dynamics
model of the system. The local stability of the closed-loop system is proved theoreti-
cally, and the effectiveness of the proposed scheme is validated through experimental
investigations.

6.2 Dynamics Model and Problem Formulation

Piezoelectric bimorph actuators are popularly employed to construct microgrippers

for performing delicate manipulation tasks. A multi-layer piezoelectric bimorph actu-
ator is illustrated in Fig. 6.1, which consists of a central layer of metal and four lay-
ers of piezoelectric materials. Driven by an excitation voltage u, the free end of the

L W
+
u H x
−
(a) (b)
Fig. 6.1 Illustration of a four-layer piezoelectric bimorph with a brass-reinforced inner layer.
a Front view. b Side view
6.2 Dynamics Model and Problem Formulation 129

cantilever undergoes an output displacement x. In order to achieve a precise posi-

tioning, both piezoelectric nonlinearities and external disturbances call for a suitable
control technique.
The dynamics model of a piezoelectric actuator driven by an input voltage can be
established as follows:

M ẍ(t) + B ẋ(t) + Kx(t) = Du(t) + P(t) (6.1)

where t is the time variable; parameters M, B, K , and x represent the mass, damping
coefficient, stiffness, and output displacement of the actuator, respectively; D is the
piezoelectric coefficient; and u denotes the input voltage. In addition, it is assumed
that the parameters M, B, K , and D are known by the approach of system iden-
tification. The perturbation term P(t) describes the lumped effect of piezoelectric
hysteresis, creep, external force, parameter uncertainties, and other disturbances. It
is notable that when the driving voltage is u = 0, P does not include the piezoelectric
hysteresis and creep effects.
To devise a DSMC control scheme, a discrete-time dynamics model is developed.
Specifically, dividing both sides of Eq. (6.1) by M results in

m ẍ(t) + b ẋ(t) + kx(t) = du(t) + p(t) (6.2)

where m = 1, b = M B
,k = MK
,d = M D
, and p(t) = P(t)
M .
Then, the continuous-time model (6.2) is discretized by adopting a small sampling
time T . For the purpose of discretization, several approaches are available, e.g., the
zero-order hold. In particular, the Euler backward difference is employed owing to
its simplicity [7, 15, 16]:

1
ẋ(t) ≈ [x(kT) − x(kT − T )] (6.3)
T
1
ẍ(t) ≈ 2 [x(kT) − 2x(kT − T ) + x(kT − 2T )] (6.4)
T
where k denotes the kth time step.
Thus, the continuous-time dynamics model (6.2) is converted into an equivalent
discrete-time form:

m̄xk−2 + b̄xk−1 + k̄xk = d̄u k + pk (6.5)

where xk−2 = x(kT − 2T ) and

1 b 2 b 1
m̄ = 2
, b̄ = − − 2 , k̄ = k + + 2 , d̄ = d. (6.6)
T T T T T
It is notable that the backward difference equations (6.3) and (6.4) are employed
to discretize the continuous-time model (6.2), and the discretization noises are not
considered here. The discrete-time model (6.5) includes the delayed versions xk−2
130 6 Digital Sliding-Mode Control of Second-Order Systems

and xk−1 of the plant output xk . Similar to other approaches such as zero-order hold,
the main problem of the discretization is the generated time delay about T /2. The time
delay causes slow response in transient behavior of the closed-loop control system
[3]. Here, the time delay is neglected because a small sampling time T is selected.
Based on the perturbation estimation technique [5], the perturbation term pk can
be generated by its one-step delayed estimation:

p̂k = pk−1 = −d̄u k−1 + m̄xk−3 + b̄xk−2 + k̄xk−1 . (6.7)

Hence, the dynamics model (6.5) can be rewritten as

m̄xk−2 + b̄xk−1 + k̄xk = d̄u k + p̂k − p̃k (6.8)

where p̃k = p̂k − pk is the perturbation estimation error, which can be further
expressed as
T
p̃k = pk−1 − pk ≈ − ṗ(t)T = − Ṗ(t). (6.9)
M
Assumption 6.1 The first derivative of the lumped perturbation P(t) is bounded,
i.e., | Ṗ(t)| ≤ δ.

In view of Eq. (6.9) and Assumption 6.1, it can be deduced that p̃k is also
bounded, i.e.,
Tδ
| p̃k | ≤ . (6.10)
M

In order to overcome the error term p̃k and to achieve a precise position control,
a DSMC scheme is devised in the next section.

6.3 DSMC Design

A DSMC with integral action is developed in this section. First, substituting the
estimated perturbation term (6.7) into the dynamics model (6.8) allows the calculation
of the position:

1
xk = [d̄(u k − u k−1 ) + m̄xk−3 + (b̄ − m̄)xk−2 + (k̄ − b̄)xk−1 − p̃k ]. (6.11)
k̄

Based on the position error ek = xk − xd,k where xd,k is the desired position
trajectory, a proportional-integral (PI) type of sliding function is defined as follows:

sk = λ P ek + λ I εk (6.12)

where λ P and λ I are the proportional and integral gains, respectively.

6.3 DSMC Design 131

In addition, the integral error is defined as


k
εk = ei = ek + εk−1 . (6.13)
i=1

Concerning the reaching law, there are two different definitions as summarized
in [11, 22]. Here, the following definition is adopted [6]:

Δsk = sk − sk−1 = 0. (6.14)

eq
Considering that the equivalent control u k is the solution to Eq. (6.14), the fol-
lowing deductions hold:

λ P ek + λ I εk = sk−1 (6.15)
⇒ (λ P + λ I )ek + λ I εk−1 = sk−1 (6.16)
⇒ (λ P + λ I )(xk − xd,k ) + λ I εk−1 = sk−1 . (6.17)

Then, substituting Eq. (6.11) into Eq. (6.17) and ignoring the estimation error p̃k
lead to the equivalent control:

eq 1 k̄ k̄
u k = u k−1 + [ sk−1 − λ I εk−1 + k̄xd,k ]
d̄ λ A λA
1
− [m̄xk−3 + (b̄ − m̄)xk−2 + (k̄ − b̄)xk−1 ] (6.18)
d̄
where the gain λ A = λ P + λ I .
The equivalent control (6.18) represents the control action for the case of perfect
disturbance estimation, i.e., p̃k = 0. It takes effect in the sliding phase when the
position trajectory is kept on the sliding surface (sk = 0). However, if a large error p̃k
occurs during the sliding phase, the stand-alone equivalent control cannot drive the
position toward the sliding surface. Thus, the equivalent control is augmented by a
switching control u sw
k to give the total control action:

eq
u k = u k + u sw
k (6.19)

i.e.,

1 k̄ k̄
u k = u k−1 + [ sk−1 − λ I εk−1 + k̄xd,k ]
d̄ λ A λA
1 λS
− [m̄xk−3 + (b̄ − m̄)xk−2 + (k̄ − b̄)xk−1 ] − sgn(sk−1 ) (6.20)
d̄ d̄

where λ S is a positive control gain and sgn(·) denotes the signum function.
132 6 Digital Sliding-Mode Control of Second-Order Systems

Theorem 6.1 For the system (6.8) with Assumption 6.1 and sliding function (6.12),
if the controller (6.20) with gain λ S ≥ | p̃k | is employed, then the discrete sliding
mode will occur with a quasi-sliding domain width of QSD = λ Ak̄λ S + λk̄A | p̃k | after
a finite number of steps.

Proof Substituting Eq. (6.20) into the expression of the sliding function (6.12), a
necessary algebra operation gives

sk = λ A (xk − xd,k ) + λ I εk−1

1
= λ A { [d̄(u k − u k−1 ) + m̄xk−3 + (b̄ − m̄)xk−2
k̄
+(k̄ − b̄)xk−1 − p̃k ] − xd,k } + λ I εk−1
λA
= sk−1 − [λ S sgn(sk−1 ) − p̃k ]. (6.21)
k̄

It is notable that the parameters k̄, λ A , and λ S are all positive. In the case of
sk−1 ≥ 0, it can be derived that

sk ≤ sk−1 if λ S ≥ | p̃k |. (6.22)

Otherwise, if sk−1 < 0, then

sk ≥ sk−1 if λ S ≥ | p̃k |. (6.23)

Hence, in view of Eqs. (6.22) and (6.23), the following conclusion is derived:

|sk | ≤ |sk−1 | if λ S ≥ | p̃k |. (6.24)

Considering Eq. (6.10), it is deduced from Eq. (6.24) that sk decreases monoto-
nously, and the discrete sliding mode is reached after a finite number of steps.
According to Eq. (6.21), the change value of the sliding function can be expressed
into the form:
λA
sk − sk−1 = − [λ S sgn(sk−1 ) − p̃k ]
k̄
λ A λS λA
=∓ + p̃k = 0. (6.25)
k̄ k̄
Hence, the quasi-sliding domain, which represents the variation of the sliding
function change, can be determined as

λ A λS λA
|sk − sk−1 | ≤ + | p̃k | = QSD. (6.26)
k̄ k̄
6.3 DSMC Design 133

Remark 6.1 It is notable that only the local convergence is obtained in Theorem 6.1
due to the lack of knowledge of the bound value for p̃k . In addition, the relationship
(6.24) represents a sufficient condition for the existence of discrete sliding mode
[12]. The selection of parameter λ S has no direct relation to the initial conditions.
Generally, λ S as well as λ P and λ I can be assigned in consideration of the quasi-
sliding domain width as represented by Eq. (6.26).

Remark 6.2 It has been shown that the relationship of |sk | ≤ |sk−1 | can be decom-
posed into two inequalities [12]:

(sk − sk−1 ) sgn(sk−1 ) ≤ 0 (6.27)

(sk + sk−1 ) sgn(sk−1 ) > 0 (6.28)

which are called sliding condition and convergence condition, respectively. Because
the condition (6.27) itself may cause instability and divergence, the convergence on
the sliding surface is assured by the condition (6.28).

Remark 6.3 Due to the discontinuity of the signum function sgn(·), chattering may
occur in the control input. To alleviate the chattering phenomenon, the boundary
layer technique is adopted by replacing the signum function in Eq. (6.20) with the
saturation function [21]:

sgn(sk ) if |sk | > ε
sat(sk ) = (6.29)
sk /ε if |sk | ≤ ε

where the boundary layer thickness ε ensures that sk is bounded by ±ε. In practice, a
tradeoff between the chattering and tracking error is needed to assign the parameter ε.

6.4 Experimental Studies

In this section, the DSMC controller is validated by conducting a series of experi-

mental studies on a prototype system.

6.4.1 Experimental Setup

The experimental setup of a piezoelectric microgripper prototype is depicted in

Fig. 6.2. The gripper is composed of two four-layer piezoelectric bimorph actua-
tors with dimension of L×H ×W = 26 × 5 × 0.86 mm3 , as shown in Fig. 6.1. In
this work, only one actuator is used which is driven by a high-voltage amplifier
(model: EPA-104, from Piezo Systems, Inc.). The output displacement at the free
end of the actuator is measured by a laser displacement sensor (model: LK-H055,
from Keyence Corp.), which has a resolution of 25 nm within a measuring range
134 6 Digital Sliding-Mode Control of Second-Order Systems

Fig. 6.2 Experimental setup of a piezoelectric bimorph actuator system

of 20 mm. In addition, a National Instruments (NI) cRIO-9075 real-time controller

(from National Instruments Corp.) equipped with NI-9263 D/A module and NI-9870
A/D module is adopted to produce excitation voltage signals and acquire the sen-
sor readings, respectively. The NI cRIO-9075 combines a real-time processor and a
reconfigurable field-programmable gate array (FPGA) within the same chassis. The
chassis is connected to a computer via Ethernet port for communication. Moreover,
TM
LabVIEW software is employed to realize a deterministic real-time control of the
piezoelectric actuator system.

6.4.2 Plant Model Identification

By applying an input sine wave with the amplitude of 0.5 V and varying frequency
of 1–1000 Hz to the actuator, the output position responses are recorded. The input–
output data sets are then used to identify the plant model by estimating a transfer
function G from the frequency response data. The identified second-order model is
shown below:

1.141 × 108
G(s) = . (6.30)
s 2 + 126.8 s + 5.943 × 106

The frequency responses obtained from the experimental data and the identified
model G in Eq. (6.30) are compared in Fig. 6.3.
Here, the simple second-order model is employed to demonstrate the effectiveness
of the control scheme. By comparing Eq. (6.2) and inverse Laplace transform of
Eq. (6.30), the dynamics model parameters can be derived as b = 126.8 N s/µm,
k = 5.943 × 106 N/µm, and d = 1.141 × 108 µm/V.
The dominant time constant of the plant is calculated as τ = 0.0158 s. Generally,
for the digital control implementation, it has been suggested to select a sampling time
less than τ /3 (i.e., 0.0053 s) of the plant [23]. In this work, a sampling time is chosen
as T = 0.004 s. With such a sampling time, the discrete-time model parameters are
obtained by resorting to Eq. (6.6).
The influence of the sampling time T on the performance of sliding-mode control
systems has been investigated in the literature [7]. It has been shown that the smaller
6.4 Experimental Studies 135

60

Magnitude (dB)
40

20

0 Experimental data
Identified model G
−20
0 1 2 3
10 10 10 10

0
Phase (degree)

−200

−400

−600
0 1 2 3
10 10 10 10
Frequency (Hz)

Fig. 6.3 Frequency responses of the system obtained by experiment and identified second-order
model

the sampling time interval, the better the control performance. Therefore, a smaller
sampling time can be adopted to further improve the control results.

6.4.3 Experimental Results

For the purpose of comparison, a traditional PID controller is also implemented:


k
u PID
k = K p ek + K i e j + K d (ek − ek−1 ) (6.31)
j=0

where the position error ek = xd,k − xk with xd,k and xk representing the desired
and actual system output at the kth time step, respectively. K p , K i , and K d denote
the proportional, integral, and derivative gains, respectively. In this work, the control
gains are tuned by the trial-and-error approach to obtain a small tracking error.

6.4.3.1 Set-Point Positioning Results

First, the set-point positioning performance of the designed controller is examined.

The parameters of the DSMC and PID controllers are tuned by several trials to
generate a rapid response with the same magnitude of overshoot. By assigning
136 6 Digital Sliding-Mode Control of Second-Order Systems

Table 6.1 Control parameters of the PID and DSMC controllers

Controller Parameters
Kp Ki Kd
PID #1 0.017 0.534 2×10−4
PID #2 0.032 1.011 8×10−5
λP λI λS ε
DSMC #1 30 1 1×104 100
DSMC #2 0.5 1 8×105 10

Fig. 6.4 Set-point (a)

positioning results of PID #1
and DSMC #1. a Positioning 20
results. b Control actions

15
Position (μm)

10

5
Reference
PID 1
0 DSMC 1
1 1.1 1.2 1.3 1.4
Time (s)

(b) 1.2

1
Control action (V)

0.8

0.6

0.4

0.2

PID 1
0 DSMC 1
1 1.1 1.2 1.3 1.4
Time (s)

λ P = 30, λ I = 1, λ S = 1×104 , and ε = 100 as shown in Table 6.1, the position-

ing result of the DSMC #1 is depicted in Fig. 6.4. In addition, the parameters of
PID controller #1 are tuned as shown in Table 6.1. The comparison of the two
6.4 Experimental Studies 137

Table 6.2 Positioning results and performance improvement of DSMC with respect to PID
controller
Performance PID DSMC Improvement (%)
Set-point positioning
5 % settling time (s) 0.095 0.048 49.5
Steady-state RMSE (µm) 0.042 0.031 26.2
Sinusoidal motion tracking
MAXE% 7.26 1.48 79.6
RMSE% 4.00 0.62 84.5
Arbitrary motion tracking
MAXE% 18.86 2.86 84.8
RMSE% 9.99 1.29 87.1

controller results is tabulated in Table 6.2, where the steady-state root-mean-square

error (RMSE) is calculated from the data of the last one second.
From the positioning results as shown in Fig. 6.4, it is observed that both con-
trollers produce no explicit overshoot. Moreover, the DSMC achieves a shorter set-
tling time and a lower steady-state error, which indicates the improvements of 49.5
and 26.2 %, respectively, over the PID control. The control actions are depicted in
Fig. 6.4b, which shows that no chattering effect exists in DSMC. However, a further
decrease of the control gain λ P or increase of λ S will produce a quicker transient
response with clear overshoot effect. Therefore, a compromise between the transient
speed and overshoot is required to adjust the control gains in practice.
Concerning the DSMC scheme, the time histories of the sliding function sk and
its variation Δsk are depicted in Fig. 6.5a. The perturbation estimation error is shown
in Fig. 6.5b, which reveals an initial value around 1.2×108 . Fig. 6.5a exhibits that
Δsk is well constrained within the quasi-sliding domain, which confirms the QSD
as predicted by (6.26) for the given parameters and the perturbation estimation error.
Additionally, in the controller DSMC #1, λ S = 1×104 is selected, which is less than
the initial value of | p̃k |. This means that it is not necessary to assign λ S as the initial
value of | p̃k |. Actually, the initial value of | p̃k | is not always available in practice.
Thus, λ S can be assigned by the trial-and-error approach.
In addition, Fig. 6.5a indicates that when the sliding mode is arrived, the sliding
function sk is kept within the interval between −15 and −4. Hence, sk is well con-
strained by the boundary thickness parameter ε = 100. On the other hand, the sliding
function variation Δsk varies between −3 and 3, but it does not arrive at zero due to
the inherent property of discrete sliding mode. Anyway, Δsk is more close to zero
than sk . This phenomenon confirms the effectiveness of the selected reaching law
(6.14).
Moreover, the data points that do not satisfy the condition of λ S ≥ | p̃k | in
Eq. (6.24) are marked by circles in Fig. 6.5b. The time histories of the sliding condition
(6.27) and the convergence condition (6.28) are shown in Fig. 6.5c, d, respectively.
Fig. 6.5d illustrates that all of the circle points meet the convergence condition except
138 6 Digital Sliding-Mode Control of Second-Order Systems

(a) 300
(b) 1.5 x 108
s
k
200 Δs
k 1
100
sk and Δsk

0.5
0

−100 5
QSD
QSD 0
100
0
−200
0 −QSD
−5 −0.5
−300 −100 −QSD
1 1.05 1.9 1.95 2
−400 −1
1 1.2 1.4 1.6 1.8 2 1 1.1 1.2 1.3 1.4
Time (s) Time (s)

(c) 100 (d) 800

0 600

−100 400

−200 200

−300 0

−400 −200

−500 −400

−600 −600
1 1.1 1.2 1.3 1.4 1 1.1 1.2 1.3 1.4
Time (s) Time (s)

Fig. 6.5 Set-point positioning results of DSMC #1. a Time evolutions of sk and Δsk . b Perturbation
estimation error p̃k . c Evolution of sliding condition. d Evolution of convergence condition

for the initial one at the time of 1 s. On the other hand, this initial point satisfies the
sliding condition as revealed in Fig. 6.5c. It means that the convergence on the sliding
manifold is reached in one sampling period. Hence, all of the points meet either the
sliding condition or the convergence condition. This explain the reason why the con-
trol system still converges even though the relationship of λ S ≥ | p̃k | is not satisfied
(Fig. 6.5).

6.4.3.2 Sinusoidal Tracking Results

Next, the tracking performance of the designed controller for a 1.25-Hz sinusoidal
motion, as shown in Fig. 6.6a, is verified. Although the motion tracking can be imple-
mented by employing the foregoing PID #1 and DSMC #1 controllers directly, rela-
tively large tracking errors are produced. In order to achieve better tracking accuracy
for the sinusoidal input, PID #2 and DSMC #2 controllers are adopted by finely tun-
ing the parameters through several trials. Specifically, the control parameters of the
two controllers are shown in Table 6.1. The positioning errors of PID #2 and DSMC
#2 are depicted in Fig. 6.6b.
6.4 Experimental Studies 139

Fig. 6.6 Sinusoidal tracking (a) 4

results of PID #2 and DSMC
#2. a Reference input. 3
b Position errors

Reference position (μm)

2

−1

−2

−3

−4
0.5 1 1.5 2 2.5 3 3.5 4
Time (s)

(b) 0.6

0.4
Position error (μm)

0.2

−0.2

−0.4

−0.6 PID 2
DSMC 2
−0.8
0.5 1 1.5 2 2.5 3 3.5 4
Time (s)

For comparison purpose, one-cycle (2.4–3.2 s) positioning results of both con-

trollers are calculated as shown in Table 6.2, where the percent maximum error
(MAXE) and RMSE are calculated as follows:

max(|ek |)
MAXE% = × 100 % (6.32)
max(xd,k ) − min(xd,k )
 
1 N 2
N k=1 ek
RMSE% = × 100 %. (6.33)
max(xd,k ) − min(xd,k )

It is observed that the DSMC scheme produces smaller tracking errors than the
PID controller. Specifically, the DSMC approach achieves the MAXE and RMSE
of 0.118 and 0.049 µm, respectively, which are almost the noise level of the sensor.
140 6 Digital Sliding-Mode Control of Second-Order Systems

These results indicate significant enhancements of 79.6 and 84.5 % in comparison

with the PID results, respectively.
In addition, Fig. 6.7a, b depicts the time histories of sk , Δsk , and the perturbation
estimation error p̃k of DSMC #2. It is observed that Δsk is well restricted within
the QSD band. By selecting a larger value of λ S = 8×105 , | p̃k | ≤ λ S holds for this
motion trajectory except for one point at the time of 0.8 s, i.e., the initial point of
the sinusoidal trajectory. In addition, both sk and Δsk vary in the vicinity of zero as
shown in Fig. 6.7a. Moreover, Δsk is one order of magnitude lower than sk , which
also demonstrates the effectiveness of the adopted reaching law (6.14).
The sliding condition and convergence condition of the discrete sliding mode are
depicted in Fig. 6.7c, d, respectively. It is seen that the point at 0.8 s meets the sliding
condition as denoted by a circle in Fig. 6.7c. In addition, almost all the data points
satisfy the convergence condition except for the four points marked by diamonds
as shown in Fig. 6.7d. Figure 6.7c reveals that these four points meet the sliding
condition instead, that is, the controller converges because all of the points in the
trajectory satisfy either the sliding or the convergence condition.

6
(a) 1.5 (b) x 10
s 1
k
1 Δs
k
QSD
0.5
0.5
k
s and Δs

0 0
k

−0.5
−0.5
−QSD
−1
−1
−1.5
0 1 2 3 4 0.5 1 1.5 2 2.5 3 3.5 4
Time (s) Time (s)

(c) 0.2
(d) 3

2.5
0.1
2
0 1.5

−0.1 1

0.5
−0.2
0

−0.3 −0.5
0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4
Time (s) Time (s)

Fig. 6.7 Sinusoidal tracking results of DSMC #2. a Time evolutions of sk and Δsk . b Perturbation
estimation error p̃k . c Evolution of sliding condition. d Evolution of convergence condition
6.4 Experimental Studies 141

6.4.3.3 Arbitrary Motion Tracking Results

Then, the tracking performance of the proposed control scheme for an arbitrary refer-
ence input is tested. Particularly, by applying an arbitrary input as shown in Fig. 6.8a,
the tracking errors of the aforementioned PID #2 and DSMC #2 controllers are shown
in Fig. 6.8b. The errors are calculated as shown in Table 6.2. As compared with PID,
the DSMC has reduced the MAXE and RMSE by 84.8 and 87.1 %, respectively.
Hence, the superiority of DSMC over PID is evident from the experimental results.
For this arbitrary trajectory, the sliding function and perturbation estimation error
of the DSMC are shown in Fig. 6.9a, b, respectively. Fig. 6.9a reveals that Δsk is
well constrained by the width of QSD. It is seen that a large estimation error of

Fig. 6.8 Arbitrary motion (a) 5

tracking results of PID #2
and DSMC #2. a Reference
input. b Position errors
Reference position (μm)

−5
0.5 1 1.5 2 2.5 3 3.5 4
Time (s)

(b) 2

1.5

1
Position error (μm)

0.5

−0.5

−1

−1.5 PID 2
DSMC 2
−2
0.5 1 1.5 2 2.5 3 3.5 4
Time (s)
142 6 Digital Sliding-Mode Control of Second-Order Systems

6
(a) (b) 4 x 10
2
1.5 3
1 QSD
2
0.5
k
s and Δs

0 1
−0.5
k

−1 −QSD 0
−1.5
sk −1
−2 Δsk
−2.5 −2
0 1 2 3 4 0.5 1 1.5 2 2.5 3 3.5 4
Time (s) Time (s)

(c) 0.8 (d) 4

0.6
3
0.4
0.2 2
0
−0.2 1

−0.4
0
−0.6
−0.8 −1
0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4
Time (s) Time (s)

Fig. 6.9 Arbitrary motion tracking results of DSMC #2. a Time evolutions of sk and Δsk .
b Perturbation estimation error p̃k . c Evolution of sliding condition. d Evolution of convergence
condition

p̃k = 3.2×106 occurs at the time of 1 s. This indicates a sudden change of ṗ(t),
which is mainly caused by the sharp transition of ẋ(t) at 1 s.
In addition, Fig. 6.9b shows that the assigned λ S = 8×105 is smaller than the
maximum value of | p̃k |. The data points that violate the condition of | p̃k | ≤ λ S
are denoted by circles in Fig. 6.9b. For these points, the corresponding sliding and
convergence conditions are marked by circles in Fig. 6.9c, d, respectively. Fig. 6.9d
illustrates that almost all the points satisfy the convergence condition expect for
a few points which are marked by diamonds. Alternatively, these points meet the
sliding condition as represented in Fig. 6.9c. Therefore, the closed-loop system still
converges because the data points that violate the condition of | p̃k | ≤ λ S satisfy
either the sliding or the convergence condition of the discrete sliding mode. The
experimental results reveal that the relationship (6.24) is not a necessary condition
for the existence of discrete sliding mode.
6.4 Experimental Studies 143

6.4.3.4 Robustness Resting Results

The forgoing experiments verify the robustness of the DSMC control with respect
to internal disturbances in terms of model uncertainties and parameter perturbations.
The robustness against external disturbance is examined by applying an external
force on the piezoelectric actuator during the motion tracking task.
Specifically, an external force is applied by hanging a weight of 25 mN on the
piezoelectric bimorph during the sinusoidal motion tracking. By exerting the force
as shown in Fig. 6.10a, the tracking result of DSMC #2 is described in Fig. 6.10b. In
addition, the time histories of sk and Δsk are plotted in Fig. 6.10c, and the tracking
errors are depicted in Fig. 6.10d. In addition, by applying the same external force,
the tracking result of the PID #2 controller is depicted in Fig. 6.11a, and the tracking
error is shown in Fig. 6.11b.
It is observed that the external force causes an increase of the tracking error at
its moment of occurrence around 1.7 s for both PID and DSMC control schemes.

(a) 30 (b) 8
Reference
25 6 Actual
Applied force (mN)

20 4
Position (μm)

15 2

10 0

5 −2

0 −4

−5 −6
0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4
Time (s) Time (s)

(c) 25 (d) 5
s
k
4
20 Δsk
Position error (μm)

3
15 2
k
s and Δs

10 1
0
k

5
−1
0
−2
−5 −3
0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4
Time (s) Time (s)

Fig. 6.10 Robustness testing results of DSMC #2 for a sinusoidal motion tracking. a Time history
of the applied external force. b Tracking results. c Sliding function sk and its variation Δsk . d
Tracking errors
144 6 Digital Sliding-Mode Control of Second-Order Systems

Fig. 6.11 Robustness (a) 30

testing results of PID #2 for Reference
a sinusoidal motion tracking. Actual
a The tracking results. b 20
Tracking errors

Position (μm)
10

−10

−20
0.5 1 1.5 2 2.5 3 3.5 4
Time (s)

(b) 20

10
Position error (μm)

−10

−20

−30
0.5 1 1.5 2 2.5 3 3.5 4
Time (s)

The PID and DSMC controllers produce the maximum errors of 27.67 and 4.75 µm,
respectively. Therefore, as compared with the PID result, the DSMC has reduced the
maximum error by 83 %. Moreover, after the occurrence of the external disturbance,
the tracking accuracy of DSMC is recovered quickly within 0.18 s, while the result
of PID is restored after 0.30 s. Thus, in comparison with PID, the DSMC shortens
the recovery time by 40 %. The results demonstrate that the DSMC possesses a much
better robustness property than PID against external disturbances.
6.4 Experimental Studies 145

6.4.4 Discussion

The foregoing experimental investigations validate the effectiveness of the DSMC

scheme with Assumption 6.1. Actually, the perturbation estimation errors p̃k as plot-
ted in Figs. 6.5b, 6.7b and 6.9b are obtained by Eq. (6.9), i.e., p̃k = pk−1 − pk .
Experimental results also demonstrate the superiority of the DSMC over PID strat-
egy in terms of transient-state response time and steady-state positioning error. The
reason why the finely tuned PID controller is still not able to achieve satisfactory posi-
tioning results mainly attributes to the nonlinear hysteresis effect of the piezoelectric
actuator. In contrast, the fact that a rapid and precise positioning is accomplished by
the DSMC controller demonstrates the effectiveness of the DSMC scheme without
modeling the hysteresis effect.
It is observed from Fig. 6.3 that the second-order model matches the magnitude
response of system well at low frequencies up to 500 Hz. To capture the dynamics
behavior accurately, a higher-order model is required to be identified. Here, a simple
second-order model is employed and the higher-frequency dynamics is involved
in the lumped disturbance pk . The achieved experimental results demonstrate the
effectiveness of the reported control scheme using a lower-order plant model.
The obtained experimental results illustrate that the relationship (6.24) is a suf-
ficient but not necessary condition for the existence of discrete sliding mode. A
necessary and sufficient condition is expected to be deduced in the next step. In
addition, only a local stability of the closed-loop system is proved in this work. The
issue of deriving a simple control scheme with global asymptotic stability is a topic
of further work. Additionally, the adaptive mechanism [4, 8, 19] can be employed
to adjust the controller parameters online in the future work.
In addition, the quantization errors are not considered in the discretization
approach in the current work. It has been shown in [16] that the quantization errors
will increase the width of quasi-sliding domain. Hence, the influence of the quanti-
zation errors on the performance of the presented control deserves an investigation
for potential applications using 8- or 16-bit microcontrollers.

6.5 Chapter Summary

Instead of using a state observer, a DSMC scheme has been developed based on a
second-order plant model of the system in this chapter. The unmodeled nonlinearity
effect is considered as a lumped perturbation, which is estimated by resorting to the
one-step delayed estimation technique. The effectiveness of the presented approach
without using state observer and hysteresis model has been validated by experimental
investigations. Both set-point positioning and sinusoidal tracking results confirm the
superiority of the DSMC over PID control in terms of response speed, positioning
accuracy, and robustness property. The presented idea can be extended to precision
control of other systems that can be described using a discrete-time plant model
preceded by disturbances.
146 6 Digital Sliding-Mode Control of Second-Order Systems

References

1. Abidi, K., Xu, J.X., Yu, X.: On the discrete-time integral sliding mode control. IEEE Trans.
Autom. Control 52(4), 709–715 (2007)
2. Bandyopadhyay, B., Fulwani, D.: High-performance tracking controller for discrete plant using
nonlinear sliding surface. IEEE Trans. Ind. Electron. 56(9), 3628–3637 (2009)
3. Bibian, S., Jin, H.: Time delay compensation of digital control for DC switchmode power
supplies using prediction techniques. IEEE Trans. Power Electron. 15(5), 835–842 (2000)
4. Chen, X., Hisayama, T.: Adaptive sliding-mode position control for piezo-actuated stage. IEEE
Trans. Ind. Electron. 55(11), 3927–3934 (2008)
5. Elmali, H., Olgac, N.: Implementation of sliding mode control with perturbation estimation
(SMCPE). IEEE Trans. Control Syst. Technol. 4(1), 79–85 (1996)
6. Furuta, K.: Sliding mode control of a discrete system. Syst. Control Lett. 14(2), 145–152 (1990)
7. Galias, Z., Yu, X.: Euler’s discretization of single input sliding-mode control systems. IEEE
Trans. Autom. Control 52(9), 1726–1730 (2007)
8. Huang, S., Tan, K.K., Lee, T.H.: Adaptive sliding-mode control of piezoelectric actuators.
IEEE Trans. Ind. Electron. 56(9), 3514–3522 (2009)
9. Mitic, D., Milojkovic, M., Antic, D.: Tracking system design based on digital minimum vari-
ance control with fuzzy sliding mode. In: Proceedings of the 8th International Conference on
Telecommunications in Modern Satellite, Cable and Broadcasting Services (TELSIKS 2007),
pp. 494–497. Nis, Serbia (2007)
10. Mitic, D., Milosavljevic, C.: Sliding mode-based minimum variance and generalized minimum
variance controls with O(T 2 ) and O(T 3 ) accuracy. Electr. Eng. 86(4), 229–237 (2004)
11. Monsees, G.: Discrete-time sliding mode control. Ph.D. thesis, Delft University of Technology
(2002)
12. Sarpturk, S., Istefanopulos, Y., Kaynak, O.: On the stability of discrete-time sliding mode
control systems. IEEE Trans. Autom. Control 32(10), 930–932 (1987)
13. Sha, D., Bajic, V.B.: Robust discrete adaptive input-output-based sliding mode controller. Int.
J. Syst. Sci. 31(12), 1601–1614 (2000)
14. Sha, D., Bajic, V.B., Yang, H.: New model and sliding mode control of hydraulic elevator
velocity tracking system. Simul. Pr. Theory 9(6), 365–385 (2002)
15. Tarokh, M.: A discrete-time adaptive control scheme for robot manipulators. J. Robot. Syst.
7(2), 145–166 (1990)
16. Veselic, B., Perunicic-Drazenovic, B., Milosavljevic, C.: Improved discrete-time sliding-mode
position control using Euler velocity estimation. IEEE Trans. Ind. Electron. 57(11), 3840–3847
(2010)
17. Xi, Z., Hesketh, T.: Discrete time integral sliding mode control for overhead crane with uncer-
tainties. IET Control Theory Appl. 4(10), 2071–2081 (2010)
18. Xu, J.X., Abidi, K.: Discrete-time output integral sliding-mode control for a piezomotor-driven
linear motion stage. IEEE Trans. Ind. Electron. 55(11), 3917–3926 (2008)
19. Xu, Q., Jia, M.: Model reference adaptive control with perturbation estimation for a micropo-
sitioning system. IEEE Trans. Control Syst. Technol. 22(1), 352–359 (2014)
20. Xu, Q., Li, Y.: Micro-/nanopositioning using model predictive output integral discrete sliding
mode control. IEEE Trans. Ind. Electron. 59(2), 1161–1170 (2012)
21. Xu, Q., Li, Y.: Model predictive discrete-time sliding mode control of a nanopositioning
piezostage without modeling hysteresis. IEEE Trans. Control Syst. Technol. 20(4), 983–994
(2012)
22. Young, K.D., Utkin, V.I., Ozguner, U.: A control engineer’s guide to sliding mode control.
IEEE Trans. Control Syst. Technol. 7(3), 328–342 (1999)
23. Zhu, Y.: Multivariable System Identification for Process Control. Elsevier Science Inc., New
York (2001)
Chapter 7
Digital Sliding-Mode Control of High-Order
Systems

This chapter presents the design and implementation of an input–output-based digital

sliding-mode control (IODSMC) algorithm to suppress the nonlinearity and distur-
bance in piezoelectric micro-/nano-positioning systems. It is applicable to preci-
sion motion tracking of a class of micro-/nano-positioning systems, which can be
described by a high-order linear model preceded by disturbances. The stability of
the control system is proved and its effectiveness is validated through experimental
investigations on a piezo-driven micropositioning system.

7.1 Introduction

The DSMC scheme as presented in Chap. 6 is applicable to a second-order system. In

practice, majority of piezoelectric micro-/naon-positioning systems exhibit a plant
mode of higher order. Hence, it is necessary to extend the control strategy to higher
order systems. To this end, an input–output-based DSMC (IODSMC) is developed
in this chapter. Unlike the DSMC algorithms as devised in the previous chapters, the
presented IODSMC exhibits the following three advantages simultaneously, which
are particularly attractive for practical applications.
1. Its implementation requires an input–output data-based model only, whereas the
hysteresis model and the state observer are not needed.
2. It is capable of suppressing the model disturbance and achieving an O(T 2 ) accu-
racy of the output position tracking.
3. It eliminates the chattering phenomenon by means of boundary layer technique.

© Springer International Publishing Switzerland 2016 147

Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2_7
148 7 Digital Sliding-Mode Control of High-Order Systems

7.2 Problem Formulation

7.2.1 System Modeling

A piezo-driven micro-/nano-positioning system can be described by a linear time-

invariant (LTI) model preceded by disturbance:

Ẋ(t) = Ac X(t) + Bc u(t) + Dc f(t) (7.1)

Y(t) = CX(t) (7.2)

where the state vector X ∈ R n 0 , the output Y ∈ R p0 , the control input u ∈ R m 0 ,

the disturbance vector f ∈ R q0 is supposed to be smooth and bounded, and m 0 ≤
p0 < n 0 . The system matrices Ac , Bc , Dc , and C are constants with approximate
dimensions.
The disturbance f represents the lumped effect of the unmodeled nonlinear hys-
teresis, residual dynamics, and external perturbations. It is assumed that f satisfies
the matching condition, i.e., Dc = Bc δd where δd is a scalar.
Adopting a sampling time T , the continuous-time system model (7.1) can be
discretized by the zero-order-hold (ZOH) sampling:

Xk+1 = AXk + Buk + dk (7.3)

Yk = CXk (7.4)

where the notation Xk = X(kT ), and

A = e Ac T (7.5)
 T
B= eAc τ dτ Bc (7.6)
0
 T
dk = eAc τ Bc δd f((k + 1)T − τ )dτ. (7.7)
0

It is observed that both B and dk are of the order O(T ) with respect to the sampling
time T .
Majority of the existing DSMC schemes have been developed based on the
discrete-time model as represented by Eqs. (7.3) and (7.4). Therefore, the full state
information is required and various state observers are usually needed. To eliminate
the use of state observer, a control scheme is devised in the following discussion.
First, for a SISO system (i.e., m 0 = p0 = 1), the discrete-time model is expressed
into the z-domain form [1]:

z −1 M(z −1 ) z −1 L(z −1 )
Y (z) = U (z) + D(z) (7.8)
N (z −1 ) N (z −1 )
7.2 Problem Formulation 149

with

M(z −1 ) = z −n+1 C adj(zI − A)B (7.9)

L(z −1 ) = z −n+1 C adj(zI − A) (7.10)
N (z −1 ) = z −n det(zI − A) (7.11)

where Y (z), U (z), and D(z) are the z-transforms of Yk , u k , and dk , respectively. In
addition, z −1 represents a unit delay.
Assume that the linear part of the plant model (7.8) can be represented as follows:

Y (z) z −1 M(z −1 ) b1 z −1 + b2 z −2 + · · · + bm z −m
= −1
= (7.12)
U (z) N (z ) 1 + a1 z −1 + a2 z −2 + · · · + an z −n

where ai and bi are the nominal model parameters, and n ≥ m holds. Then, the
discrete-time model (7.8) can be expanded into the following form:


n 
m
xk = ai xk−i + bi u k−i + pk (7.13)
i=1 i=1

where xk = Yk is the output position, pk denotes the nonlinear disturbance term in

the system model (7.8), and Property 4.1 applies to the disturbance pk .
Subsequently, a controllable canonical form of the plant model (7.13) can be
derived as follows:

X̄k+1 = GX̄k + Hu k + QVk + Pk (7.14)

Ȳk = C̄X̄k (7.15)

where
⎡ ⎤
xk−n+1
⎢ .. ⎥
⎢ ⎥
X̄k = ⎢ . ⎥ ∈ R n×1 (7.16)
⎣ xk−1 ⎦
xk
⎡ ⎤
u k−m+1
⎢ .. ⎥
⎢ . ⎥
Vk = ⎢ ⎥ ∈ R (m−1)×1 (7.17)
⎣ u k−2 ⎦
u k−1
⎡ ⎤
0
⎢ .. ⎥
⎢ ⎥
Pk = ⎢ . ⎥ ∈ R n×1 (7.18)
⎣ 0 ⎦
pk
150 7 Digital Sliding-Mode Control of High-Order Systems
⎡ ⎤
0 1 ··· 0
⎢ .. .. . . .. ⎥
⎢ . . ⎥
G=⎢ . . ⎥ ∈ R n×n (7.19)
⎣ 0 0 ··· 1 ⎦
−an −an−1 · · · −a1
⎡ ⎤
0
⎢ .. ⎥
⎢ ⎥
H = ⎢ . ⎥ ∈ R n×1 (7.20)
⎣0⎦
b1
⎡ ⎤
0 0 ··· 0
⎢ .. .. . . .. ⎥
⎢ . . ⎥
Q=⎢ . . ⎥ ∈ R n×(m−1) (7.21)
⎣ 0 0 ··· 0 ⎦
bm bm−1 · · · b2


C̄ = 0 · · · 0 1 ∈ R 1×n . (7.22)

It is observed that the models (7.14) and (7.15) are constructed using the history
data of the system input and output only. Hence, this state-space model is called an
input–output model.

7.2.2 Disturbance Estimation

Based on the perturbation estimation technique, the disturbance term Pk in Eq. (7.14)
can be estimated by its one-step delayed value as follows:

P̂k = Pk−1 = X̄k − G X̄k−1 − Hu k−1 − QVk−1 . (7.23)

Then, the system models (7.14) and (7.15) become

X̄k+1 = GX̄k + Hu k + QVk + P̂k − P̃k (7.24)

Ȳk = C̄X̄k (7.25)

where P̃k = P̂k − Pk is the disturbance estimation error, which can be further
expressed as:

P̃k = Pk−1 − Pk
= G(X̄k − X̄k−1 ) + H(u k − u k−1 ) + Q(Vk − Vk−1 ) − (X̄k+1 − X̄k ) (7.26)

where G, H, and Q are constant matrices.

7.2 Problem Formulation 151

For a piezo-driven micro-/nano-positioning system, the input voltage u k and out-

put position xk are limited in practice. Hence, X̄k and Vk are also limited. Then,
it is reasonable to assume that P̃k as expressed by Eq. (7.26) is bounded. In view
of Eq. (7.18), it can be deduced that only the last element p̃k of P̃k is nonzero. To
facilitate the control design, the following assumption is proposed:

Assumption 7.1 The estimation error p̃k of the lumped disturbance pk is bounded,
i.e., | p̃k | ≤ σ .

The objective of the motion tracking control is to force the system output (Ȳk =
xk ) to precisely follow a desired reference input under the influence of the lumped
disturbance Pk . To accomplish this goal, a control algorithm is developed in the next
section.

7.3 IODSMC Design

The design procedure of an IODSMC control scheme is outlined in this section. First,
the equivalent controller is devised based on an incremental proportional-integral-
derivative (PID)-type sliding function. Then, an augmented IODSMC scheme is
constructed. The stability and tracking error bound of the control system are also
evaluated.

7.3.1 Sliding Function Definition

First, the output tracking error is defined as follows:

ek = Ȳk − rk (7.27)

where rk is the desired output position of the system.

Based on the output error in Eq. (7.27), an incremental PID type of sliding function
is defined:

sk = sk−1 + K p (ek − ek−1 ) + K i ek + K d (ek − 2ek−1 + ek−2 ) (7.28)

where K p , K i , and K d represent the proportional, integral, and derivative gains,

respectively.
After a fundamental algebra operation, the sliding function (7.28) is simplified
into the form:
sk = sk−1 + K 3 ek−2 + K 2 ek−1 + K 1 ek (7.29)
152 7 Digital Sliding-Mode Control of High-Order Systems

where

K1 = K p + Ki + Kd (7.30)
K 2 = −K p − 2K d (7.31)
K3 = Kd . (7.32)

7.3.2 Design of IODSMC

In the literature, there are different definitions of the reaching law. In particular, the
IODSMC algorithm is developed on the basis of the following reaching law:

sk+1 = 0 (7.33)

so that the sliding surface is reached in one sampling instant.

According to Eq. (7.33), the following deduction is generated

sk + K 3 ek−1 + K 2 ek + K 1 (Ȳk+1 − rk+1 ) = 0. (7.34)

Then, solving the expression of Ȳk+1 from Eq. (7.25) and substituting it into
Eq. (7.34) leads to

sk + K 3 ek−1 + K 2 ek + K 1 [C̄(GX̄k + Hu k + QVk + P̂k − P̃k ) − rk+1 ] = 0. (7.35)

eq eq
Considering that the equivalent control u k is the solution to Eq. (7.33), u k is
derived from Eq. (7.35) by neglecting the disturbance estimation error P̃k :

u k = (K 1 C̄H)−1 (−sk − K 3 ek−1 − K 2 ek + K 1 rk+1 − K 1 C̄GX̄k − K 1 C̄QVk − K 1 C̄P̂k )

eq

(7.36)

where the one-step forward reference position rk+1 is required to realize the con-
troller. In micro-/nano-positioning applications, the trajectory to be tracked is usually
predefined. Hence, it is reasonable to assume that rk+1 is known.

7.3.2.1 Stability Analysis

Substituting Eq. (7.36) into Eq. (7.14), the following expression is produced after a
necessary calculation

X̄k+1 = RX̄k + WPk−1 + (Pk − Pk−1 ) + E (7.37)

7.3 IODSMC Design 153

where

R = G − H(K 1 C̄H)−1 [K 2 C̄ + K 1 C̄G] (7.38)

W = I − H(K 1 C̄H)−1 K 1 C̄ (7.39)
E = H(K 1 C̄H)−1 (−K 3 ek−1 + K 2 rk + K 1rk+1 − K 1 C̄QVk ) (7.40)

with I denoting an identity matrix.

The eigenvalues of the matrices R and W are the solutions of det(zI − R) = 0
and det(zI − W) = 0, respectively, which govern the closed-loop state dynamics
behavior. To ensure the stability of the closed-loop system, the eigenvalues of the
two matrices are designed to locate inside the unit circle in the z-plane.

7.3.2.2 Tracking Error Bound Analysis

Once given Eqs. (7.14), (7.15), and (7.29), the expression of sk+1 can be derived as
follows

sk+1 = sk + K 3 ek−1 + K 2 ek + K 1 ek+1

= sk + K 3 ek−1 + K 2 ek + K 1 (Ȳk+1 − rk+1 )
= sk + K 3 ek−1 + K 2 ek + K 1 [C̄(GX̄k + Hu k + QVk + Pk ) − rk+1 ]. (7.41)

Substituting Eq. (7.36) into Eq. (7.41), a fundamental computation leads to

sk+1 = K 1 C̄(Pk − Pk−1 ). (7.42)

Then, in view of Property 4.1, the following deduction is generated from Eq. (7.42):

sk = O(T 2 ). (7.43)

Hence, sk has a magnitude of the order O(T 2 ). That is, the sliding surface has an
ultimate bound of O(T 2 ).
In addition, by noting Eqs. (7.14), (7.15), and (7.27), the one-step forward error
equation can be expressed as:

ek+1 = Ȳk+1 − rk+1

= C̄(GX̄k + Hu k + QVk + Pk ) − rk+1 . (7.44)

Then, substituting Eq. (7.36) into Eq. (7.44) gives

ek+1 = C̄(Pk − 2Pk−1 + Pk−2 ) − K 1−1 (K 2 ek + K 3 ek−1 ). (7.45)

154 7 Digital Sliding-Mode Control of High-Order Systems

According to Property 4.1, the relation Pk − 2Pk−1 + Pk−2 = O(T 3 ) can be

derived. Then, in consideration of Eq. (7.45) and Property 4.2, it is derived that

ek = O(T 2 ) (7.46)

which indicates that the ultimate bound of the output tracking error ek is of the order
O(T 2 ).
The error dynamics equation (7.45) suggests that the control gains should be
designed to satisfy the following conditions so as to make the closed-loop system
stable

|K 1−1 K 2 | < 1 (7.47)

|K 1−1 K 3 | < 1. (7.48)

The equivalent controller (7.36) is designed based on the assumption of zero

disturbance estimation error, i.e., P̃k = 0. The controller takes effect in the sliding
phase when the system trajectory is kept on the sliding surface. Actually, P̃k = 0.
Furthermore, if the initial state of the system does not lie on the sliding surface sk = 0
or external disturbances emerge during the sliding phase motion, the equivalent
control alone cannot drive the trajectory toward the sliding surface. In this sense, a
robust IODSMC controller is designed in the following section.

7.3.3 Robust IODSMC Design

To design a robust IODSMC controller, the equivalent control is augmented with a

discontinuous switching control action. The role of the switching control is to force
the system trajectory to reach the sliding mode by overcoming external disturbances.
Here, the switching control is designed as follows:
−1
k = −b1 K s sign{sk }
u sw (7.49)

where K s is a positive control gain and sign{·} denotes the signum function.
eq
The equivalent control u k is then augmented by the switching control u sw
k to give
the overall control action:
eq
u k = u k + u sw
k . (7.50)

That is,

u k = (K 1 C̄H)−1 (−sk − K 3 ek−1 − K 2 ek + K 1 rk+1 − K 1 C̄GX̄k − K 1 C̄QVk − K 1 C̄P̂k )

−b1−1 K s sign{sk }. (7.51)
7.3 IODSMC Design 155

Theorem 7.1 For the system described by Eqs. (7.24) and (7.25) along with the
sliding function (7.52) and Assumption 7.1, if the controller (7.51) is employed, then
the quasi-sliding mode will occur in a finite number of steps.

Proof Substituting Eq. (7.51) into the expression of the sliding function (7.52), a
necessary algebra operation gives:

sk+1 = −K 1 C̄Hb1−1 K s sign{sk } + K 1 C̄(Pk − Pk−1 )

= −K 1 [K s sign{sk } + ( pk−1 − pk )]
= −K 1 (K s sign{sk } + p̃k ) (7.52)

where both K 1 and K s are positive gains.

Assigning K s = | p̃k | + ε where ε denotes an arbitrary positive constant, the
following conclusion can be derived from Eq. (7.52):

K 1 ε ≤ |sk+1 | ≤ K 1 (2| p̃k | + ε). (7.53)

Recalling Assumption 7.1, i.e., p̃(k) is bounded, it can be obtained that

K 1 ε ≤ |sk+1 | ≤ K 1 (2σ + ε). (7.54)

Hence, it is deduced from Eq. (7.54) that the quasi-sliding mode in the vicinity
of the sliding plane sk = 0 is reached within a finite number of steps.

Remark 7.1 The essence of the switching control action u sw k of the IODSMC is to
overcome the estimation error P̃k by driving the system trajectory to a quasi-sliding
eq
domain as indicated by Eq. (7.54). Afterwards, the equivalent control action u k takes
effect to keep the trajectory sliding in the O(T ) vicinity of the sliding surface as
2

revealed in Eq. (7.43).

Remark 7.2 Chattering may be induced due to the discontinuity of the signum func-
tion sign{·}. To alleviate the chattering phenomenon, the boundary layer technique
is adopted to replace the signum function in Eq. (7.51) with the saturation function:

sign{sk } if |sk | > ε
sat{sk } = (7.55)
sk / if |sk | ≤ ε.

In practice, a compromise between the chattering effect and tracking accuracy is

made to assign the parameter .

By substituting Eq. (7.23) into Eq. (7.51) and replacing the sign{·} function with
the saturation function (7.55) results in the overall robust controller:

u k = u k−1 − (K 1 C̄H)−1 [sk + K 3 ek−1 + K 2 ek − K 1rk+1 + K 1 C̄G(X̄k − X̄k−1 )

+K 1 C̄Q(Vk − Vk−1 ) + K 1 Ȳk ] − b1−1 K s sat{sk }. (7.56)
156 7 Digital Sliding-Mode Control of High-Order Systems

Remark 7.3 It is notable that only a linear input–output model of the system is
needed, whereas neither a state observer nor a hysteresis model is required for the
realization of the controller (7.56). Hence, as compared with the existing DSMC
algorithms [4–6], one advantage of the reported IODSMC control is that it is easy to
implement.

In the following sections, the performance of the IODSMC controller is verified by

conducting a collection of experimental studies on a piezo-driven micropositioning
system.

7.4 Experimental Setup and Controller Setup

7.4.1 Experimental Setup

The experimental setup is depicted in Fig. 7.1. A custom-built piezoelectric actu-

ation compliant microgripper is fabricated using Al-7075 alloy and actuated by a
piezo-stack actuator (PSA) (model: TS18-H5-202, from Piezo Systems, Inc.). To
measure the gripper tip position, four strain gages (model: SGD-3/350-DY13, from
Omega Engineering Ltd.) are employed to construct a full-bridge circuit. In addition,
for measuring the grasping force, one strain gage (model: SGD-3/350-LY13, from
Omega Engineering Ltd.) is adopted to form a quarter-bridge circuit. The control
hardware is realized using a National Instruments (NI) cRIO-9022 real-time con-
troller combined with cRIO-9118 reconfigurable chassis (from National Instruments
Corp.). The chassis is equipped with NI-9263 D/A module and NI-9237 A/D module
for producing analog voltage signals and acquiring the strain gage sensor readings,
respectively. The analog excitation signal is amplified by a high-voltage amplifier
(model: EPA-104 from Piezo Systems, Inc.) with an adjustable gain of then before
TM
being fed into the PSA. LabVIEW software is employed to realize a deterministic
real-time control for the micropositioning system.

Fig. 7.1 Experimental setup of a piezo-actuated micropositioning system

7.4 Experimental Setup and Controller Setup 157

For the calibration of the position strain gages, the tip position of the gripper is
measured by a laser displacement sensor (model: LK-H055, from Keyence Corp.)
with a resolution of 25 nm. After calibration, the position and force strain gage sensors
offer the three standard deviation resolutions of 0.34 µm and 1.50 mN, respectively.
In addition, a digital microscope with a magnification ratio of 200 is employed to
monitor the gripper tips and the grasped object during the micropositioning process.

7.4.2 Plant Model Identification

The dynamics model parameters of the micropositioning system are identified by

resorting to the frequency response approach. Specifically, swept sine waves with
the amplitude of 0.1 V and frequency range of 1–500 Hz are produced by the NI-9263
module to drive the PSA. Without an contact force exerted, the position response is
measured by the position strain gage sensor. Then, a spectral analysis is conducted
to obtain the frequency responses as depicted in Fig. 7.2.
It is observed from Fig. 7.2 that the system exhibits a resonant peak around 244 Hz.
Based on the experimental data, a fourth-order plant model is identified. Figure 7.2
reveals that the identified model well approximates the system behavior up to 300 Hz.
Using a sampling time of T = 0.0005 s, the model is discretized as follows:

−0.9373z −1 + 5.474z −2 − 6.101z −3 + 5.513z −4

G x (z) = . (7.57)
1 − 2.126z −1 + 2.613z −2 − 1.591z −3 + 0.621z −4

Fig. 7.2 Open-loop 60

frequency response of the
Magnitude (dB)

micropositioning system
40
generated by experiment and
the identified model
20

Modeled dynamics
0
1 2
10 10
Phase (degree)

−200

−400
Experiment
Identified model
−600
1 2
10 10
Frequency (Hz)
158 7 Digital Sliding-Mode Control of High-Order Systems

Then, the parameters of the nominal dynamics models (7.14) and (7.15) are
obtained by comparing Eq. (7.12) with Eq. (7.57).
It is notable that a model of higher order can be identified to better describe the
high-frequency dynamics. In the present work, a fourth-order model is employed to
make a tradeoff between the model accuracy and complexity, and to demonstrate the
effectiveness of the control scheme.

7.4.3 Controller Parameter Design

As a model-free controller, PID solves the control command using the tracking error
only. For the purpose of comparison, a PID control is employed due to its popularity.
An incremental-type PID algorithm is expressed as follows

u PID
k = u k−1 + K P (ek − ek−1 ) + K I ek + K D (ek − 2ek−1 + ek−2 ) (7.58)

where u k−1 is the control command in the previous time step. K P , K I , and K D
represent the proportional, integral, and derivative gains, respectively. Here, the PID
gains are tuned by the trial-and-error approach to achieve a quicker response without
overshoot phenomenon.
In addition, the tuned PID gains are adopted to determine the control gains K 1 , K 2
and K 3 of the IODSMC using Eqs. (7.30)–(7.32), respectively. As compared with
PID algorithm, IODSMC has an additional switching gain K s , which is introduced
to suppress the disturbance estimation error p̃k . In practice, the switching gain K s
and the boundary layer thickness  are tuned by several trials through experimental
testings. The controller parameters are summarized in Table 7.1, which satisfy the
conditions as specified by Eqs. (7.38), (7.39), (7.47) and (7.48).

Table 7.1 Parameters of the PID and IODSMC controllers

Controller Parameter Value
PID KP 0.0040
KI 0.0060
KD 0.0005
IODSMC K1 0.0105
K2 −0.0050
K3 0.0005
Ks 5
 1
7.5 Experimental Results and Discussion 159

7.5 Experimental Results and Discussion

Using the developed controllers, several experimental studies are conducted to verify
the performance of the micropositioning system.

7.5.1 Set-Point Positioning Results

First, a 40-µm set-point positioning is carried out. The experimental results of the
PID and IODSMC controllers are shown in Fig. 7.3. The positioning results exhibit
that similar steady-state RMSE of 0.018 and 0.020 µm are generated by the PID and
IODSMC controllers, respectively.
In addition, the PID and IODSMC schemes produce the 5 % settling time of 0.0638
and 0.0141 s, respectively. As compared with PID, the IODSMC algorithm enables
a fourfold more rapid transient response while at the expense of a 8 % overshoot.
It is notable that the control gains K 1 , K 2 , and K 3 of IODSMC are directly
calculated from the PID gains, as shown in Table 7.1. The overshoot magnitude can
be reduced by finely tuning the control gains of the IODSMC controller.

7.5.2 Sinusoidal Positioning Results

Next, the sinusoidal motion tracking performance of the controllers is tested using a
5-Hz, 40 µm-amplitude sinusoidal reference input as shown in Fig. 7.4a. The tracking
errors of the PID and IODSMC controllers are illustrated in Fig. 7.4b.

Fig. 7.3 The 40-µm 50

set-point positioning results
45
of PID and IODSMC
controllers 40
Displacement (µm)

35
5% error bounds
30

25

20

15

10
Reference
5 PID
IODSMC
0
0 0.05 0.1 0.15 0.2 0.25 0.3
Time (s)
160 7 Digital Sliding-Mode Control of High-Order Systems

Fig. 7.4 Sinusoidal tracking (a)

results of PID and IODSMC 50

Displacement (µm)
controllers. a The tracking 40
results. b PID and IODSMC 30
tracking errors. c IODSMC 20
controller output
10 Reference
PID
0
IODSMC
−10
0 0.2 0.4 0.6 0.8 1

(b) Time (s)

Displacement error (µm) 10

−5
PID IODSMC
−10
0 0.2 0.4 0.6 0.8 1
Time (s)
(c)
4
Controller output (V)

−1
0 0.2 0.4 0.6 0.8 1
Time (s)

It is observed that the PID control produces the maximum peak-to-peak error
(PPE) and RMSE of 17.773 and 5.882 µm, i.e., 44.43 and 14.71 % of the motion
range, respectively. By contrast, the IODSMC scheme induces the maximum PPE
of 3.782 µm and RMSE of 0.802 µm, which are equivalent to 9.46 and 2.01 %
of the total range, respectively. As compared with PID, the IODSMC controller
substantially reduces the PPE and RMSE by 78.7 and 86.3 %, respectively. The
reason why the PID produces a worse result is mainly attributed to its bandwidth
limit and the presence of piezoelectric hysteresis effect.
The control action u k of the IODSMC scheme is depicted in Fig. 7.4c, which
exhibits no chattering phenomenon. The time histories of the perturbation estima-
tion error p̃k and the sliding function sk are shown in Fig. 7.5a, b, respectively. It is
observed from Fig. 7.5a that the error p̃k is constrained by | p̃k | ≤ σ = 2.65, which
confirms the effectiveness of Assumption 7.1. In addition, the magnitude of the slid-
ing function sk is well constrained by the assigned boundary layer thickness  = 1 as
depicted in Fig. 7.5b.
7.5 Experimental Results and Discussion 161

Fig. 7.5 Experimental (a) 4

results of IODSMC
controller. a The
perturbation estimation error 2
p̃k . b Time history of
IODSMC sliding function 0

−2

−4
0 0.2 0.4 0.6 0.8 1
Time (s)

(b) 1.5

0.5

−0.5

−1

−1.5
0 0.2 0.4 0.6 0.8 1
Time (s)

7.5.3 Bandwidth Testing Results

In order to examine the control bandwidth of the piezoelectric micropositioning

system, the motion tracking testing of the 40-µm sinusoidal signal with the frequency
varying from 1 to 80 Hz is undertaken. The closed-loop frequency responses of the
system are shown in Fig. 7.6.
It is found that the ordinary −3 dB bandwidth corresponds to a large phase lag.
Hence, the closed-loop control bandwidth is defined as the frequency at which the
phase is lagged 30◦ . Using the PID and IODSMC controllers, the 30◦ -lag bandwidths
of 6.1 and 40.4 Hz are achieved, which equal to 2.5 and 16.6 % of the resonant
frequency, respectively. These cutoff frequencies correspond to magnitude errors of
−0.7 and 2.8 dB of the PID and IODSMC responses, respectively. As compared
with PID algorithm, the IODSMC scheme creates an over sixfold improvement on
the control bandwidth for the micropositioning system.
162 7 Digital Sliding-Mode Control of High-Order Systems

Fig. 7.6 Control bandwidth (a) 10

results of the

Magnitude (dB)
micropositioning system
with PID and IODSMC 0
controllers
−10
PID
IODSMC
−20
0 1 2
10 10 10

(b)
Phase (degree) 0

−30

−60

−90

−120
0 1 2
10 10 10
Frequency (Hz)

7.5.4 Robustness Testing Results

In order to verify the robustness property of the IODSMC controller against internal
and external disturbances, two case studies are conducted.
First, the motion tracking of a 1-Hz, 110 µm-amplitude sinusoidal trajectory as
shown in Fig. 7.7a is evaluated. The tracking error is plotted in Fig. 7.7c without
force exerted. It is observed that a PPE of 4.230 µm and a RMSE of 0.693 µm are
produced, which account for 3.85 and 0.63 % of the motion range, respectively. It is
notable that such a low tracking error is achieved under the influence of piezoelectric
hysteresis effect.
The closed-loop and open-loop hysteresis curves are compared in Fig. 7.8. As
compared with the significant open-loop hysteresis of 14.1 %, the IODSMC scheme
greatly mitigates the hysteresis to a low level of 1.2 %. Hence, the IODSMC con-
troller has reduced the open-loop hysteresis by over tenfold. The experimental results
demonstrate the robustness of the IODSMC control in front of internal disturbance,
i.e., unmodeled piezoelectric nonlinearity effect.
Second, the robustness of the IODSMC controller in the presence of external dis-
turbance is evaluated for the micropositioning system. Specifically, the microgripper
is commanded to grasp a copper microwire of 300 µm diameter. To realize this task,
the gripper arms are controlled to track a 1-Hz, 110 µm-amplitude sinusoidal tra-
jectory, as shown in Fig. 7.7a. Figure 7.7b illustrates the signal of the grasping force,
where the peak force occurs when the gripper is closed to grasp the object and zero
force means that the gripper is opened. The motion tracking experiments have been
carried out under two situations, i.e., without and with grasping the copper wire. The
tracking errors for the two situations are shown in Fig. 7.7c.
7.5 Experimental Results and Discussion 163

Fig. 7.7 a The reference (a)

displacement trajectory.

Displacement (µm)
b Signal of the grasping 100
force. c IODSMC tracking
errors without and with
grasping the copper wire 50

0 1 2 3 4 5
Time (s)

(b)
30
Force (mN)

20

10

−10
0 1 2 3 4 5
Time (s)
(c)
Displacement error (µm)

3
2
1
0
−1
−2
Without force With force
−3
0 1 2 3 4 5
Time (s)

Fig. 7.8 Open-loop and 120

closed-loop hysteresis
curves. The open-loop curve 100
Actual displacement (µm)

is generated by scaling the

driving voltage with the
motion range 80

60

40

20
Open−loop
Closed−loop
0
0 20 40 60 80 100 120
Desired displacement (s)
164 7 Digital Sliding-Mode Control of High-Order Systems

It is found that by exerting an external force with the peak value of 28-mN, the
IODSMC scheme is still capable of tracking the predefined trajectory precisely. As a
consequence, the experimental results without and with the grasping force applied are
very similar in that the RMSEs are calculated as 0.693 and 0.689 µm, respectively.
The small discrepancy between the two tracking errors demonstrates the robustness
property of the IODSMC control scheme under the influence of external disturbances.

7.5.5 Further Discussion

The generated experimental results validate the effectiveness of the IODSMC control
scheme for a piezo-actuated micropositioning system. Because the scheme has been
implemented without modeling the hysteresis nonlinearity, the results reveal the
efficiency of the reported idea for the model disturbance suppression. Considering
that the nominal input–output model has been employed to construct the disturbance
estimator for the estimate of the unmodeled dynamics and uncertainties in the control
design, the results also demonstrate the feasibility of the identified lower order model
for precise positioning applications.
As the frequency of the reference input increases, Fig. 7.6 indicates that the perfor-
mance of PID control drops more quickly than that of the IODSMC control. Indeed,
it is one of the reasons why a 5-Hz sinusoidal reference input is adopted in the com-
parative study. Actually, the IODSMC achieves a less than 10 % RMSE for a higher
speed tracking of 20-Hz reference input trajectory.
It is notable that the control gains of IODSMC have not been optimally tuned. An
optimal tuning of the four control gains is a possible solution to further improve the
positioning results. In addition, similar to existing work [2, 3], the IODSMC scheme
is targeted at minimum-phase plants. The issue of extending the control scheme to
nonminimum-phase systems deserves a further investigation.
As compared with the existing DSMC schemes, the reported IODSMC only
requires a linear input–output model for its implementation without using a state
observer. It is able to accomplish an O(T 2 ) output tracking accuracy using a
chattering-free control action. Moreover, the motion tracking ability is robust against
both internal and external disturbances. The design process of the control parameters
is as simple as that of a PID control. Comparatively, the IODSMC performs much
better than the PID as demonstrated by the aforementioned experimental results.

7.6 Chapter Summary

This chapter presents an IODSMC control algorithm dedicated to precision position-

ing control of a piezoelectric actuation system. Its novelty lies in the achievement
of three advantages when applied to the piezoelectric positioning system control.
The tracking accuracy of the controller has been analyzed and the stability in the
7.6 Chapter Summary 165

presence of model disturbances has been proved. The effectiveness and robustness
of the presented control have been validated on a piezo-driven microgripper system.
Results show that the IODSMC scheme is superior to the popular PID control in terms
of both transient response speed and motion tracking accuracy, which is enabled by
an enlarged control bandwidth. Because the implementation of the reported input–
output model based control does not require any hysteresis model and state observer,
it can be easily extended to other types of micro-/nano-positioning systems. One
possible future work is to extend the control scheme to nonminimum-phase systems.

References

1. Mitic, D., Milosavljevic, C.: Sliding mode-based minimum variance and generalized minimum
variance controls with O(T 2 ) and O(T 3 ) accuracy. Electr. Eng. 86(4), 229–237 (2004)
2. Sha, D., Bajic, V.B.: Robust discrete adaptive input-output-based sliding mode controller. Int.
J. Syst. Sci. 31(12), 1601–1614 (2000)
3. Sha, D., Bajic, V.B., Yang, H.: New model and sliding mode control of hydraulic elevator velocity
tracking system. Simul. Pract. Theory 9(6), 365–385 (2002)
4. Xu, J.X., Abidi, K.: Discrete-time output integral sliding-mode control for a piezomotor-driven
linear motion stage. IEEE Trans. Ind. Electron. 55(11), 3917–3926 (2008)
5. Xu, Q., Li, Y.: Micro-/nanopositioning using model predictive output integral discrete sliding
mode control. IEEE Trans. Ind. Electron. 59(2), 1161–1170 (2012)
6. Xu, Q., Li, Y.: Model predictive discrete-time sliding mode control of a nanopositioning
piezostage without modeling hysteresis. IEEE Trans. Control Syst. Technol. 20(4), 983–994
(2012)
Chapter 8
Digital Sliding-Mode Prediction Control

This chapter presents the design and verification of a digital sliding-mode predic-
tion control (DSMPC) scheme for precise position control of piezoelectric micro-
/nanopositioning systems. Its implementation only needs input/output measure-
ments, whereas the burdens on hysteresis modeling and state observer design are
released. As compared with DSMC, the DSMPC is capable of further attenuating
the positioning error through an optimal control, which is provided by the predictive
control strategy. Its stability is proved, and the ultimate tracking error bounds are
evaluated analytically. The feasibility of the control scheme is validated by experi-
mental investigations on a piezo-driven micropositioning device.

8.1 Introduction

This chapter presents the implementation of a scheme of DSMPC for the precision
motion tracking of a class of piezoelectric micro-/nanopositioning systems, which
can be represented by a linear model preceded by disturbances. As compared with
the controllers developed in Chaps. 4 and 5, the control scheme exhibits the following
characteristics:
1. It is established based on an input/output data-based model. Hence, its implemen-
tation does not require the knowledge on the system states.
2. The design of a state observer is not required. Instead, only the past and current
values of the input and output of the system are needed.
3. An incremental PID-type sliding function is employed, which is more flexible
than the PI-type sliding function used before.
4. The overall control action of the DSMPC scheme involves the equivalent control,
switching control, and predictive control, while the robust switching control action
is not involved in the previous work. Hence, the presented control is more robust
than the previous algorithms against the disturbances.

© Springer International Publishing Switzerland 2016 167

Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2_8
168 8 Digital Sliding-Mode Prediction Control

5. The sampling frequency is limited by the controllability and observability indices

in previous works. In contrast, the presented DSMPC does not impose such a
limitation, and the sampling rate is constrained by the hardware limit only.

8.2 Problem Formulation

8.2.1 System Modeling

A piezo-driven micro-/nanopositioning system can be described by a discrete-time

dynamics model preceded by disturbance:


n 
m
xk = ai xk−i + bi u k−i + pk (8.1)
i=1 i=1

where xk is the output position, u k is the input voltage, and pk represents the dis-
turbance term. The disturbance pk describes the combined effect of the unmodeled
nonlinear hysteresis, residual dynamics, and external perturbations.
Then, a controllable canonical form of SISO plant model (8.1) can be derived as
follows:

Xk+1 = GXk + Hu k + QVk + Pk (8.2)

Yk = CXk (8.3)

where
 T
Xk = xk−n+1 , . . . , xk−1 , xk (8.4)
 T
Vk = u k−m+1 , . . . , u k−2 , u k−1 (8.5)
Pk = [0, . . . , 0, pk ] T
(8.6)
H = [0, . . . , 0, b1 ]T (8.7)
⎡ ⎤
0 1 ··· 0
⎢ .. .. . . .. ⎥
⎢ . . ⎥
G=⎢ . . ⎥ (8.8)
⎣ 0 0 ··· 1 ⎦
−an −an−1 · · · −a1
⎡ ⎤
0 0 ··· 0
⎢ .. .. . . .. ⎥
⎢ . . ⎥
Q=⎢ . . ⎥ (8.9)
⎣ 0 0 ··· 0 ⎦
bm bm−1 · · · b2
8.2 Problem Formulation 169

C = [0, . . . , 0, 1]T . (8.10)

It is notable that the state-space models (8.2) and (8.3) are constructed using the
history data of the system input and output only. This chapter develops an efficient
control scheme based on this input/output data-based model.

8.2.2 Disturbance Estimation

According to the perturbation estimation technique [2], the disturbance term Pk in

Eq. (8.2) can be estimated by its one-step delayed value as follows:

P̂k = Pk−1 = Xk − GXk−1 − Hu k−1 − QVk−1 . (8.11)

Then, the system models (8.2) and (8.3) become

Xk+1 = GXk + Hu k + QVk + P̂k − P̃k (8.12)

Yk = CXk (8.13)

where P̃k = P̂k − Pk is the disturbance estimation error, which can be further
expressed as
P̃k = Pk−1 − Pk ≈ −Ṗ(t)T (8.14)

where T is the sampling time.

Assumption 8.1 The first-order derivative of the lumped disturbance P(t) is bounded,
i.e., |Ṗ(t)| ≤ Δ, which means that the norm of Ṗ(t) is not larger than a positive con-
stant Δ.

In view of Assumption 8.1 as well as (8.14), it can be deduced that P̃k is also
bounded, i.e.,
|P̃k | = |Pk−1 − Pk | ≤ ΔT. (8.15)

The goal of the motion tracking control is to make the system output follow a
desired reference input precisely under the influence of the lumped disturbance Pk ,
more specifically, the bounded disturbance estimation error P̃k .

8.3 DSMC Design

First, the design procedure of a DSMC control scheme is outlined in this section.
170 8 Digital Sliding-Mode Prediction Control

8.3.1 Sliding Function Definition

The output tracking error is defined as

ek = Yk − rk (8.16)

where rk is the desired position output of the system.

Based on the output error (8.16), an incremental PID type of sliding function is
defined as follows:

sk = sk−1 + K p (ek − ek−1 ) + K i ek + K d (ek − 2ek−1 + ek−2 ) (8.17)

where K p , K i , and K d represent the proportional, integral, and derivative gains,

respectively.
After a fundamental algebraic operation, the sliding function (8.17) is simplified
as
sk = sk−1 + K 3 ek−2 + K 2 ek−1 + K 1 ek (8.18)

where
K1 = K p + Ki + Kd (8.19)
K 2 = −K p − 2K d (8.20)
K3 = Kd . (8.21)

8.3.2 Design of DSMC

Theorem 8.1 For the system described by Eqs. (8.12) and (8.13) with the sliding
function (8.18), if the following controller (8.22) is employed, then the discrete sliding
mode will occur within a finite number of steps:
−1
k = (K 1 CH) (−K 3 ek−1 − K 2 ek + K 1 rk+1
u sm
−K 1 CGX k − K 1 CQVk − K 1 CPk−1 ) − (CH)−1 K s sign(sk ) (8.22)

where K s is a positive switching gain and sign(·) denotes the signum function.

Proof Taking into account the definition (8.18), the one-step forward value of the
sliding function is derived as follows

sk+1 = sk + K 3 ek−1 + K 2 ek + K 1 ek+1

= sk + K 3 ek−1 + K 2 ek + K 1 (Yk+1 − rk+1 )
= sk + K 3 ek−1 + K 2 ek + K 1 [C(GXk + Hu k + QVk + Pk ) − rk+1 ] (8.23)

which is deduced using Eqs. (8.12) and (8.13).

8.3 DSMC Design 171

Substituting Eq. (8.22) into Eq. (8.23), a fundamental calculation yields

sk+1 = sk − K 1 [K s sign(sk ) − C(Pk − Pk−1 )]. (8.24)

In consideration of Eq. (8.15), it is obtained that

|C(Pk − Pk−1 )| ≤ CΔT. (8.25)

In addition, it is assumed that the switching gain is designed to meet the condition

K s ≥ CΔT + δ (8.26)

where δ is an arbitrary positive constant.

In view of the attribute of the discrete sliding-mode system [3], i.e., starting from
the initial state, the trajectory will move monotonically toward the switching plane,
it is supposed that sk is a positive sequence or negative sequence in the reaching
phase [4]. Actually, if sk and sk+1 have opposite signs, it means that the system state
crosses the sliding hyperplane.
If sk is a positive sequence, i.e., sk ≥ 0, then the following deduction holds:

sk+1 ≤ sk − K 1 [CΔT + δ − C(Pk − Pk−1 )]

≤ sk . (8.27)

Otherwise, if sk is a negative sequence, i.e., sk < 0, then

sk+1 > sk + K 1 [CΔT + δ + C(Pk − Pk−1 )]

> sk . (8.28)

Taking into account Eqs. (8.27) and (8.28) together yields

|sk+1 | ≤ |sk | for k ≥ k0 (8.29)

which implies that sk decreases monotonously and the discrete sliding mode is
reached within a finite number (k0 ) of steps. It has been shown that the relation-
ship (8.29) is a sufficient condition for the existence of discrete sliding mode [6].

Remark 8.1 Due to the discontinuity of the signum function sign(·), chattering may
occur in the control action. To alleviate the chattering phenomenon, the signum
function in (8.22) is replaced with the saturation function:

sign(sk ) if |sk | > ε
sat(sk ) = . (8.30)
sk /ε if |sk | ≤ ε

Actually, a tradeoff between the chattering effect and tracking error is needed to
assign the parameter ε.
172 8 Digital Sliding-Mode Prediction Control

8.3.3 Error Bound Analysis

The DSMC controller (8.22) consists of the equivalent and switching control actions.
The essence of the switching control u sw , i.e., the last term in Eq. (8.22), is to
overcome the disturbance P̃k and to drive the system trajectory toward the sliding
surface. The trajectory is then kept on the sliding surface by the equivalent control
action. Hence, the ultimate tracking error is attributed to the equivalent control u eq ,
i.e., controller (8.22) without the last term, which is employed in the following error
bound analysis.
Substituting the equivalent controller into Eq. (8.23), a fundamental algebraic
operation gives

sk+1 = sk + K 1 C(Pk − Pk−1 ). (8.31)

According to Property 4.1, Pk − Pk−1 = O(T 2 ) can be derived. Then, in view of

Property 4.2, it is deduced from Eq. (8.31):

sk = O(T ). (8.32)

Thus, sk has a magnitude of the order O(T ), which indicates that the sliding surface
has an ultimate bound of O(T ).
In view of Eqs. (8.2), (8.3), and (8.16), the one-step forward error equation can
be expressed as

ek+1 = Yk+1 − rk+1

= C(GXk + Hu k + QVk + Pk ) − rk+1 . (8.33)

Then, substituting the controller expression (8.22) into Eq. (8.33) yields

ek+1 = C(Pk − Pk−1 ) − K 1−1 (K 2 ek + K 3 ek−1 ). (8.34)

Taking into account Property 4.2 and Pk − Pk−1 = O(T 2 ), it can be derived from
Eq. (8.34) that
ek = O(T ). (8.35)

The error dynamics equation (8.34) also indicates that the control gains should be
designed to satisfy the following conditions so that the closed-loop system is stable:

|K 1−1 K 2 | ≤ 1 (8.36)
|K 1−1 K 3 | ≤ 1. (8.37)

In addition, substituting the controller equation (8.22) into Eq. (8.2), a necessary
calculation gives the closed-loop system state equation:
8.3 DSMC Design 173

Xk+1 = RXk + WPk−1 + (Pk − Pk−1 ) + E (8.38)

in which

R = G − H(K 1 CH)−1 [K 2 C + K 1 CG] (8.39)

−1
W = I − H(K 1 CH) K 1 C (8.40)
−1
E = H (K 1 CH) (−sk − K 3 ek−1 + K 2 rk + K 1rk+1 − K 1 CQVk ) (8.41)

with I denoting an identity matrix.

The eigenvalues of the matrices R and W are the solutions of det(zI − R) = 0
and det(zI − W) = 0, respectively, which determine the closed-loop state dynamics
behavior. To ensure the stability of the closed-loop system, the eigenvalues of the
two matrices (R and W) should locate inside the unit circle in the complex plane.

8.4 DSMPC Design

The DSMC controller developed in the previous section is able to produce a tracking
error bound of O(T ) order. To further reduce the error bound and to attenuate the
chattering effect, a DSMPC control scheme is developed in this section.

8.4.1 DSMPC Design

pc
The MPC is used to generate an optimal control (u k ) to drive the system output to
the sliding surface. Substituting the augmented control
pc
u k = u sm
k + uk (8.42)

into Eq. (8.23) and noting Eqs. (8.14) and (8.22) gives
pc
sk+1 = sk + K 1 CHu k − K 1 CP̃k − K 1 K s sat(sk ) (8.43)

which describes the sliding-mode dynamics and also represents a one-step forward
prediction of the sliding function.
Based on Eq. (8.43), an Nh -step forward prediction of the sliding function can be
obtained as follows:
pc pc pc
sk+Nh = sk + K 1 CH(u k + u k+1 + · · · + u k+Nh −1 ) − K 1 C(P̃k + P̃k+1 + · · · + P̃k+Nh −1 )
−K 1 K s [sat(sk ) + sat(sk+1 ) + · · · + sat(sk+Nh −1 )] (8.44)
174 8 Digital Sliding-Mode Prediction Control

where the positive integer Nh is called the prediction horizon as well as control
horizon in this research.
The Nh equations can be stacked together into the following form:

Sk = Λsk + ΦUk−1 − Γ ζk−1 − LJk−1 (8.45)

where the vectors for the future sliding functions, control actions, disturbance esti-
mation errors, and saturation functions are
 T
Sk = sk+1 , sk+2 , . . . , sk+Nh (8.46)
 T
pc pc pc
Uk−1 = u k , u k+1 , . . . , u k+Nh −1 (8.47)
 T T T
T
ζk−1 = P̃k , P̃k+1 , . . . , P̃k+Nh −1 (8.48)
 T
Jk−1 = sat(sk ), sat(sk+1 ), . . . , sat(sk+Nh −1 ) . (8.49)

Additionally, the four matrices exhibit the following forms:

Λ = [I, I, . . . , I]T (8.50)

⎡ ⎤
K 1 CH 0 ... 0
⎢ K 1 CH K 1 CH ... 0 ⎥
⎢ ⎥
Φ=⎢ . .. .. .. ⎥ (8.51)
⎣ .. . . . ⎦
K 1 CH K 1 CH . . . K 1 CH
⎡ ⎤
K1C 0 . . . 0
⎢ K1C K1C . . . 0 ⎥
⎢ ⎥
Γ =⎢ . .. . . .. ⎥ (8.52)
⎣ .. . . . ⎦
K1C K1C . . . K1C
⎡ ⎤
K1 Ks 0 ... 0
⎢ K1 Ks K1 Ks ... 0 ⎥
⎢ ⎥
L=⎢ . .. .. .. ⎥. (8.53)
⎣ .. . . . ⎦
K1 Ks K1 Ks . . . K1 Ks

Using the aforementioned notations, a cost function for minimization can be

expressed by

f = SkT Sk + wUk−1
T
Uk−1 (8.54)

where w is the weight value to tune the magnitude of the control action.
The optimal control sequence is obtained by minimizing the cost function f .
Substituting Eq. (8.45) into Eq. (8.54) and applying the optimality condition ∂U∂k−1
f
=
0 [5] yield
8.4 DSMPC Design 175

Uk−1 = −(Φ T Φ + wI)−1 Φ T (Λsk − Γ ζk−1 − LJk−1 ) (8.55)

where the future disturbance values in ζk−1 and saturation function values in Jk−1
are unknown. Hence, they are estimated as follows:

T T T
ζ̂k−1 = [P̃k−1 , P̃k−1 , . . . , P̃k−1 ]T (8.56)
Ĵk−1 = [sat(sk ), sat(sk ), . . . , sat(sk )] .
T
(8.57)

Because only the first value of the predicted control sequence is used, the optimal
control action can be obtained:

u k = −v(Φ T Φ + wI)−1 Φ T (Λsk − Γ ζ̂k−1 − LĴk−1 )

pc
(8.58)

where the selection vector v = [1, 0, 0, . . . , 0].

Therefore, in view of Eq. (8.42) along with Eqs. (8.22) and (8.58), the global
control action can be derived as follows

u k = (K 1 CH)−1 (−K 3 ek−1 − K 2 ek + K 1 rk+1 − K 1 CGXk − K 1 CQVk − K 1 CPk−1 )

−(CH)−1 K s sat(sk ) − v(Φ T Φ + wI)−1 Φ T (Λsk − Γ ζ̂k−1 − LĴk−1 ). (8.59)

It is observed from Eq. (8.59) that the one-step forward reference position rk+1 is
required to realize the controller. In micro-/nanopositioning applications, the trajec-
tory to be tracked is usually predefined. Hence, it is reasonable to assume that rk+1
is known. The block diagram of the DSMPC control scheme is depicted in Fig. 8.1.

8.4.2 Stability Analysis

Theorem 8.2 For the system described by Eqs. (8.12) and (8.13) along with the
sliding function (8.18), if the controller (8.59) is employed, then the quasi-sliding
mode will occur within a finite number of steps.

uksm
rk DSMC
+
Reference uk Piezoelectric xk
micropositioning
+ system
MPC
ukpc
DSMPC

Fig. 8.1 Block diagram of the DSMPC control scheme

176 8 Digital Sliding-Mode Prediction Control

Proof Substituting Eq. (8.55) along with Eqs. (8.56) and (8.57) into the stacked
equation (8.45) yields

Sk = Λsk − Φ(Φ T Φ + wI)−1 Φ T (Λsk − Γ ζ̂k−1 − L Ĵk−1 ) − Γ ζk−1 − LJk−1 (8.60)

where the switching control term LJk−1 takes effect initially in the reaching phase.
For a steady-state analysis, the transient switching control action is not considered.
For the convenience of analysis, it is assumed that there is no penalty for the
pc
partial control action u k , i.e., w = 0. Therefore, Eq. (8.60) further reduces to

Sk = Γ (ζ̂k−1 − ζk−1 ). (8.61)

Next, considering only the first value of the predicted sliding hyperplane vector
(8.61) and taking into account Eqs. (8.46) and (8.56), it is derived that

sk+1 = K 1 C(P̃k−1 − P̃k )

= K 1 C(Pk − 2Pk−1 + Pk−2 ) = O(T 3 ) (8.62)

which indicates that the magnitude of sk+1 is of the order O(T 3 ) in consideration of
Property 4.1.
Therefore, the ultimate magnitude of sk+1 is limited:

|sk+1 | ≤ λs (8.63)

where λs is defined as the quasi-sliding mode band width.

According to Definition 1 and Definition 2 in [1], it can be concluded that the
system described by Eqs. (8.12) and (8.13) satisfies the reaching condition of the
quasi-sliding mode in λs vicinity of the sliding surface in a finite number of steps.
Therefore, the closed-loop control system is stable.

Remark 8.2 The essence of the presented DSMPC scheme is to force the system
trajectory to reach a quasi-sliding mode in an optimal manner which is enabled by
pc
the predictive control action u k as well as u sm
k . One advantage of this quasi-sliding
model lies in that it does not require the system state to cross the sliding hyperplane
in each successive control step. Therefore, the chattering phenomenon is eliminated.
As a consequence, the control action is reduced and the control performance is
improved [1].
8.4 DSMPC Design 177

8.4.3 Error Bound Analysis

Substituting Eq. (8.59) into the one-step forward error equation (8.33) yields

ek+1 = C(Pk − Pk−1 ) − K 1−1 (K 2 ek + K 3 ek−1 )

−vCH(Φ T Φ + wI )−1 Φ T (Λsk − Γ ζ̂k−1 − LĴk−1 ). (8.64)

As mentioned earlier, the ultimate tracking error is analyzed without considering

the transient switching control action LĴk−1 . Additionally, for the sake of simplicity,
it is assumed that there is no penalty for the predictive control action, i.e., w = 0.
Thus, Eq. (8.64) reduces to

ek+1 = −K 1−1 sk − K 1−1 (K 2 ek + K 3 ek−1 ). (8.65)

In view of Property 4.2 as well as sk = O(T 3 ), it can be derived from (8.65) that

ek = O(T 2 ). (8.66)

Hence, the ultimate bound of the output tracking error ek is of the order O(T 2 ),
which is much lower than the O(T ) tracking error obtained by DSMC as shown in
Eq. (8.35).

Remark 8.3 It is observed from Eq. (8.62) that sk+1 has the magnitude of order
O(T 3 ). Therefore, the DSMPC controller (8.59) drives the system trajectory to slide
in the vicinity of the sliding surface with an ultimate bound of O(T 3 ) thickness,
whose magnitude is much lower than the O(T ) thickness achieved by DSMC con-
troller as shown in Eq. (8.32). As a consequence, a much lower control error of order
O(T 2 ) is produced by the DSMPC as indicated in Eq. (8.66).

8.5 Experimental Studies and Discussion

In this section, the performance of the developed controllers is verified by carrying

out a series of experimental studies.

8.5.1 Experimental Setup

The experimental setup of a piezo-actuated micropositioning system is shown in

Fig. 7.1. A custom-built flexure-based microgripper is actuated by a piezoelectric
stack actuator (PSA). For the measurement of the gripper tip position, four strain
gages (model: SGD-3/350-DY13, from Omega Engineering Ltd.) are adopted to
178 8 Digital Sliding-Mode Prediction Control

construct a full Wheatstone bridge circuit. The control hardware is realized using a
National Instruments (NI) cRIO-9022 real-time (RT) controller combined with cRIO-
9118 reconfigurable chassis which contains a field-programmable gate array (FPGA)
core. Additionally, the chassis is equipped with NI-9263 D/A module and NI-9237
A/D module. The analog excitation signal is amplified by a high-voltage amplifier
TM
before being fed into the PSA. LabVIEW software is employed to implement a real-
time control. After calibrating using a laser displacement sensor (model: LK-H055,
from Keyence Corporation), the strain gage sensor offers a three-standard-deviation
resolution of 0.34 µm.

8.5.2 Plant Model Identification

The plant model of the system is identified by means of frequency response approach.
Specifically, swept sine waves with the amplitude of 0.1 V and frequency range of
1–500 Hz are used to drive the PSA. The position response is measured by the strain
gage sensor using a sampling rate of 4 kHz. Then, a spectral analysis is conducted to
obtain the frequency responses as depicted in Fig. 8.2. The system exhibits a resonant
peak around 244 Hz. Based on the experimental data, a forth-order plant model G p
is identified with MATLAB . Using a sampling time of T = 0.00025 s, the model
G p is discretized and the parameters of models (8.2) and (8.3) are obtained.
It is observed that the forth-order model is able to capture the system dynamics
behavior up to 300 Hz. To describe the dynamics at frequencies higher than 300 Hz,
a higher-order model is needed. In this work, a lower-order model is used to make
a compromise between model complexity and accuracy, and to reveal the efficiency
of the developed control scheme.

Fig. 8.2 Bode diagrams of 60

Magnitude (dB)

the open-loop system

obtained by experiment and 40
the identified model
20

0
0 1 2
10 10 10
Phase (degrees)

−200

−400 Experiment
Identified model
−600
0 1 2
10 10 10
Frequency (Hz)
8.5 Experimental Studies and Discussion 179

8.5.3 Controller Parameter Design

For comparison purpose, an incremental PID control is implemented due to its pop-
pid
ularity. The controller can be expressed by Eq. (8.17) with sk replaced by u k . The
PID gains K p , K i , and K d are optimally tuned through intensive tests. The gains
are K p = 0.001, K i = 0.0055, and K d = 0.001 in the present work. The suitability
of these gains have been confirmed because other PID gains either produce larger
tracking errors or result in oscillatory responses.
The tuned PID gains are adopted to calculate the DSMC and DSMPC gains K 1 ,
K 2 , and K 3 in Eq. (8.18). In addition, K s = 1.5 and ε = 5.0 are selected for the DSMC
and DSMPC. The DSMPC has two extra parameters, i.e., the prediction horizon Nh
and weighting parameter w. In this work, Nh = 10 and w = 5×105 are chosen by
the trial-and-error approach to achieve better result. It is notable that the selected
controller parameters satisfy the conditions in Eqs. (8.36)–(8.37) and (8.39)–(8.40).

8.5.4 Experimental Studies

8.5.4.1 Sinusoidal Positioning Testing

First, the motion tracking of a 10 Hz, 40 µm amplitude sinusoidal reference input as

shown in Fig. 8.3a is examined. The tracking errors of the PID, DSMC, and DSMPC
controllers are illustrated in Fig. 8.3b.
It is observed that the PID control produces the maximum error (MAXE) and
root-mean-square error (RMSE) of 7.327 and 4.523 µm, i.e., 18.3 % and 11.3 % of
the motion range, respectively. In contrast, the DSMC scheme leads to the MAXE
of 1.748 µm and RMSE of 0.668 µm, which account for 4.4 % and 1.7 % of the
total range, respectively. As compared with PID, the DSMC controller reduces the
MAXE and RMSE by 76.1 % and 85.2 %, respectively. Furthermore, the DSMPC
mitigates the MAXE and RMSE to 0.868 and 0.243 µm, i.e., 2.2 % and 0.6 % of
the positioning range, respectively. Thus, as compared with DSMC, the DSMPC
substantially reduces the MAXE and RMSE by 50.3 % and 63.6 %, respectively. The
experimental results agree well with the ultimate tracking error bounds as predicted
earlier.
Moreover, the sliding function values of the DSMC and DSMPC controllers are
illustrated in Fig. 8.3c. It is seen that the sliding functions of both control schemes
are well constrained by the thickness parameter of ε = 5.0. Specifically, the DSMC
drives the sliding function to the range of ±2.52 µm. In contrast, the sliding function
has been further suppressed within ±0.90 µm by the DSMPC scheme. In compari-
son with DSMC, the DSMPC mitigates the sliding surface boundedness by 64.3 %
substantially. The experimental results are consistent with the ultimate bound analy-
ses of the sliding surfaces for DSMC and DSMPC schemes as conducted earlier,
respectively.
180 8 Digital Sliding-Mode Prediction Control

Fig. 8.3 Comparison of (a)

PID, DSMC, and DSMPC 50

Displacement (µm)
sinusoidal tracking results. 40
a The reference input. b PID,
30
DSMC, and DSMPC
tracking errors. c Sliding 20
functions of DSMC and 10
DSMPC controllers 0
−10
0 0.1 0.2 0.3 0.4 0.5 0.6

(b) Time (s)

Displacement error (µm)
10

0
PID
−5 DSMC
DSMPC
−10
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (s)
(c)
3
Sliding function (µm)

2
1
0
−1
DSMC
−2
DSMPC
−3
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (s)

For a better understanding of the DSMPC scheme, the three components of the
control action in terms of equivalent control, switching control, and predictive con-
trol are depicted in Fig. 8.4a, b, and c, respectively. It is seen that no chattering
phenomenon exists in the DSMPC control action. Furthermore, it is found that the
equivalent control is the dominant control action of the DSMPC and DSMC con-
trollers. As compared with DSMC scheme, the DSMPC has an extra control action
of predictive control as shown in Fig. 8.4c, whose magnitude is similar to that of the
switching control as plotted in Fig. 8.4b. This extra predictive control action explains
the reason why the DSMPC performs better than the conventional DSMC control
strategy.
However, as compared with the DSMC tracking results, there is a significant
amount of noise in the DSMPC results as shown in Fig. 8.3b. The spectra of the
switching control, predictive control, and tracking errors of the DSMPC scheme are
depicted in Fig. 8.5a, b, and c, respectively. Figure 8.5c shows that the high-frequency
noise occurs at 230 Hz, which is close to the resonant frequency of the system. The
8.5 Experimental Studies and Discussion 181

Fig. 8.4 Control action (a)

components of the DSMPC

Equivalent control (V)

4
controller. a Equivalent
control action. b Switching 3
control action. c Predictive 2
control action
1

−1
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (s)
(b)
0.02
Switching control (V)

0.01

−0.01

−0.02
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (s)
(c)
Predictive control (V)

0.02

0.01

−0.01

−0.02
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (s)

noise level can be alleviated by reducing the magnitude of predictive control, which
leads to a larger tracking error as the cost. In practice, a tradeoff between the noise
and tracking error is required to adjust the predictive control action.
In addition, the motion tracking experiments for the 40-µm amplitude sinusoidal
reference input with higher frequencies of 20 and 40 Hz have also been carried out.
The tracking results of the three controllers are compared in Table 8.1. Evidently, the
DSMPC outperforms both PID and DSMC schemes.

8.5.4.2 Control Bandwidth Testing

To further explore the performance of the control schemes, the control bandwidth
of the piezoelectric micropositioning system is examined. Specifically, the motion
tracking of the sinusoidal signal with the amplitude of 10 µm and frequency varying
182 8 Digital Sliding-Mode Prediction Control

Fig. 8.5 Amplitude spectra (a) 0.015

of a switching control u sw ,
b predictive control u pc , and
0.01

|Usw(f)|
c tracking error of DSMPC

0.005
230Hz

0
0 100 200 300 400 500
Frequency (Hz)
−3
(b) 8
x 10

6
|Upc(f)|

2 230Hz

0
0 100 200 300 400 500
Frequency (Hz)

(c) 0.2

0.15
|E(f)|

0.1 230Hz

0.05

0
0 100 200 300 400 500
Frequency (Hz)

Table 8.1 Sinusoidal positioning results achieved by PID, DSMC, and DSMPC controllers
Frequency Performance PID DSMC DSMPC
10 Hz MAXE (µm) 7.327 1.748 0.868
RMSE (µm) 4.523 0.668 0.243
20 Hz MAXE (µm) 8.344 3.463 1.647
RMSE (µm) 13.486 1.392 0.494
40 Hz MAXE (µm) 12.988 5.323 3.843
RMSE (µm) 18.942 3.164 2.142

from 1 to 500 Hz has been undertaken. The closed-loop frequency responses of the
system are shown in Fig. 8.6.
In this work, the closed-loop control bandwidth is defined as the frequency at
which the phase is lagged 30◦ [8]. With the PID, DSMC, and DSMPC controllers, the
30◦ -lag bandwidths of 17.1, 47.4, and 53.9 Hz are achieved, which are equivalent to
7.0, 19.4, and 22.1 % of the resonant frequency, respectively. These cutoff frequencies
8.5 Experimental Studies and Discussion 183

Fig. 8.6 Closed-loop 20

Bode diagrams of the

Magnitude (dB)
micropositioning system 10
obtained with PID, DSMC,
and DSMPC controllers
0

−10

−20
0 1 2
10 10 10

17.1 47.4 53.9

−30
Phase (degree)

−200

−400
PID
DSMC
−600
DSMPC
0 1 2
10 10 10
Frequency (Hz)

correspond to small magnitude errors of −0.27, 0.48, and 0.92 dB of the closed-loop
responses of PID, DSMC, and DSMPC, respectively. Hence, as compared with PID
and DSMC, the DSMPC enables an improvement of 215 and 14 % on the control
bandwidth, respectively.

8.5.5 Further Discussion

Figure 8.6 indicates a peak value of about 10 dB in the magnitude response around the
resonant frequency of the system. It means that the closed-loop system does not damp
the system significantly. This behavior is similar for the DSMC and DSMPC schemes.
The resonant peak induces high-frequency noise as depicted in Fig. 8.5c. In the future
work, the control scheme can be further enhanced to address the deficiencies.
As the frequency of the reference input increases, the performances of PID and
DSMC controllers drop more quickly than that of DSMPC as indicated by Fig. 8.6.
Actually, the reason why the PID and DSMC achieve worse positioning results is
mainly attributed to their bandwidth limits and the inherent piezoelectric hysteresis
effect. In contrast, the fact that a precise high-speed positioning is achieved by the
reported DSMPC indicates the robustness of the DSMPC scheme under the influence
of piezoelectric nonlinearities and other disturbances. As compared with DSMC, the
184 8 Digital Sliding-Mode Prediction Control

DSMPC enables a wider control bandwidth which is dominantly contributed by the

predictive control action.
Unlike the existing observer-based DSMC algorithms [7, 9, 10], the DSMPC does
not require a state observer because it is developed on the basis of an input/output
data-based model. It is known that a state observer is capable of estimating the full
states from noisy output signals. Instead, the presented observer-free approach uses
the output directly, and it is not able to suppress the noise. Hence, the DSMPC scheme
is preferable to be adopted for the situations where the sensor noise is not a critical
issue. Otherwise, the implementation of an observer or a filter is desirable.
Although the MPC control action of the DSMPC scheme has been optimized,
other control gains have not been optimally tuned. An optimum tuning of the con-
troller parameters is a potential solution to further enhance the positioning perfor-
mance. In addition, the motion tracking of sinusoidal trajectories has been carried
out in this work to show the superiority of DSMPC over PID and DSMC control
algorithms. Other type of trajectories (e.g., square and triangle waves) can also be
employed for such purpose. Even so, the improvement of positioning accuracy for
the micropositioning system over the conventional PID and DSMC as demonstrated
by the conducted experiments validates the effectiveness of the DSMPC scheme and
exhibits a potential for the future work.

8.6 Chapter Summary

A DSMPC control algorithm has been developed in this chapter dedicated to pre-
cision positioning control of a piezoelectric micro-/nanopositioning system without
hysteresis modeling and state observer design. Theoretical analyses of the tracking
accuracy and stability under the influence of disturbances have been carried out. Ex-
perimental results show that the DSMPC scheme produces a smaller tracking error
bound than DSMC, thanks to an optimal predictive control action. Moreover, the
DSMPC affords an enlarged control bandwidth than PID and DSMC, which is the
reason why DSMPC performs better in terms of high-speed tracking accuracy. In the
future, further work deserves to extend the control scheme to the nonminimum-phase
system control.

References

1. Bartoszewicz, A.: Discrete-time quasi-sliding-mode control strategies. IEEE Trans. Ind.

Electron. 45(5), 633–637 (1998)
2. Elmali, H., Olgac, N.: Implementation of sliding mode control with perturbation estimation
(SMCPE). IEEE Trans. Control Syst. Technol. 4(1), 79–85 (1996)
3. Gao, W., Wang, Y., Homaifa, A.: Discrete-time variable structure control systems. IEEE Trans.
Ind. Electron. 42(2), 117–122 (1995)
References 185

4. Mitic, D., Antic, D., Milojkovic, M., Nikolic, S., Peric, S.: Input-output based quasi-sliding
mode control of DC-DC converter. Facta Univ. Ser.: Elec. Energ. 25(1), 69–80 (2012)
5. Neelakantan, V.A., Washington, G.N., Bucknor, N.K.: Model predictive control of a two stage
actuation system using piezoelectric actuators for controllable industrial and automotive brakes
and clutches. J. Intell. Mater. Syst. Struct. 19(7), 845–857 (2008)
6. Sarpturk, S., Istefanopulos, Y., Kaynak, O.: On the stability of discrete-time sliding mode
control systems. IEEE Trans. Automat. Contr. 32(10), 930–932 (1987)
7. Xu, J.X., Abidi, K.: Discrete-time output integral sliding-mode control for a piezomotor-driven
linear motion stage. IEEE Trans. Ind. Electron. 55(11), 3917–3926 (2008)
8. Xu, Q., Jia, M.: Model reference adaptive control with perturbation estimation for a micropo-
sitioning system. IEEE Trans. Contr. Syst. Technol. 22(1), 352–359 (2014)
9. Xu, Q., Li, Y.: Micro-/nanopositioning using model predictive output integral discrete sliding
mode control. IEEE Trans. Ind. Electron. 59(2), 1161–1170 (2012)
10. Xu, Q., Li, Y.: Model predictive discrete-time sliding mode control of a nanopositioning
piezostage without modeling hysteresis. IEEE Trans. Contr. Syst. Technol. 20(4), 983–994
(2012)
Chapter 9
Model-Reference Adaptive Control
with Perturbation Estimation

This chapter presents the design and testing of a model-reference adaptive control
(MRAC) scheme to compensate for the hysteresis effect of a class of piezo-actuated
systems, which possess a second-order nominal model. Specifically, by treating the
uncertainties as a lumped perturbation to the nominal system model, a scheme of
MRAC with perturbation estimation (MRACPE) is developed and validated on a
micropositioning system. As compared with existing works, the reported scheme
allows the predesign of the maximum tracking error. It is capable of estimating the
unmodeled perturbation of the system online.

9.1 Introduction

The idea of adaptive control is to adjust the system response by using a closed-
loop controller whose parameters can be updated online. Unlike robust control
approach, the adaptive control does not require a prior knowledge about the bounds
on uncertain or time-varying items. Hence, adaptive control paves a more straightfor-
ward way to the precision control of micro-/nanopositioning systems. Nevertheless,
only limited works have been made toward the extension of adaptive controllers to
micro-/nanopositioning system control. In the previous work [6], a MRAC strategy
is reported to compensate for the hysteresis effect of a micropositioning stage. Even
though the adaptive controller has been realized without modeling the hysteresis
effect nor acquiring the uncertainty bounds, a Prandtl–Ishlinskii hysteresis model is
required to convert the desired motion trajectory into a voltage input. More recently,
a MRAC scheme based on the hyperstability theory is presented for a piezo-actuated
system [7]. Nonetheless, a Bouc–Wen hysteresis model is still employed to identify
the dynamics equation of the system.
From a practical point of view, it is preferable to develop a MRAC scheme without
modeling the complicated nonlinear effects. By considering the nonlinearity as per-
turbations to the system, several perturbation estimation methods have been reported,
© Springer International Publishing Switzerland 2016 187
Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2_9
188 9 Model-Reference Adaptive Control with Perturbation Estimation

which can be integrated with MRAC schemes. To name a few, a scheme of MRAC
with disturbance rejection strategy is presented for the systems, which can be rep-
resented by parabolic or hyperbolic partial differential equations along with known
disturbance model or constant disturbance [5]. However, these scenarios are differ-
ent from the situation in a piezo-actuated system, where the disturbance involving
hysteresis nonlinearity is not constant. In addition, an adaptive perturbation approx-
imator is outlined to estimate the structured perturbations of second-order systems
online [2], and a perturbation estimation mechanism is proposed to design a MRAC
controller for a class of MIMO dynamic systems [1]. The aforementioned two meth-
ods have been implemented based on the assumption that the system states are all
measurable. Nevertheless, this assumption does not always hold for microposition-
ing systems which typically only offer the displacement feedback. Moreover, the
reference signals as used in the preceding literature are all given in terms of voltage.
For practical applications, a desired motion trajectory instead of voltage signal is
predefined.
This chapter presents a scheme of MRACPE for the precision motion control of
a piezoelectric actuation micropositioning system. One advantage of the proposed
scheme lies in that the size of the tracking error can be predesigned, which is desirable
in practice. A second-order nominal system is assumed and the unmodeled dynamics
and nonlinearity effect are treated as a lumped perturbation, which is approximated
by a perturbation estimation technique. A dead-zone modification of the adaptive
rules is introduced to mitigate the parameter drifts and to speed up the parameter
convergence process. Moreover, the MRACPE scheme employs a desired displace-
ment trajectory rather than a voltage signal as the reference input. The stability of
the closed-loop control system is proved through Lyapunov stability analysis. Exper-
imental studies show that the MRACPE is superior to conventional PID control in
terms of positioning accuracy for both set-point and sinusoidal positioning tasks,
which is enabled by a significantly enlarged control bandwidth.

9.2 Dynamics Modeling and Perturbation Estimation

9.2.1 Dynamics Modeling

Considering the micropositioning stage as a mass–spring–damper system, the dynam-

ics model, which integrates the components of the stage and PSA, can be developed
as follows:
M ẍ(t) + B ẋ(t) + K x(t) = Du(t) + P(t) (9.1)

where t is the time variable, x denotes the output displacement, M, B, K , and D rep-
resent the equivalent mass, damping parameter, output stiffness, and displacement-
voltage coefficient, respectively. In addition, the perturbation term P(t) describes
9.2 Dynamics Modeling and Perturbation Estimation 189

the lumped effect of hysteresis, model parameter uncertainties, unmodeled high-

frequency dynamics, and other disturbances. It is assumed that P(t) is bounded.
The system model (9.1) can be rewritten into the form:

ẍ(t) + α1 ẋ(t) + α0 x(t) = β0 u(t) + f (t) (9.2)

where α1 = M B
, α0 = MK
, β0 = M
D
, and f (t) = P(t)
M .
The dynamics model (9.2) indicates that the piezo-actuated system is represented
by a second-order linear model, and the hysteresis effect is contained in the unknown
perturbations f (t) to the nominal system. The unmodeled nonlinearity and distur-
bances need to be compensated for in order to achieve a precise control of the posi-
tion x. In the subsequent section, a perturbation estimation technique is employed
for the online estimation of the unmodeled perturbation involving the hysteresis
nonlinearity.

9.2.2 Perturbation Estimation

A brief review of the perturbation estimation method is outlined here. More details
can be found in the literature [3].
Consider a general nonlinear system given by:

x(n) = f(X) + Δf(X) + [B(X) + ΔB(X)]u(t) + d(t) (9.3)

(n ) (n ) (n ) (n )
where the vector x(n) = [x1 1 , x2 2 , . . . , xm m ]T ∈ m with xi i ∈ , the global
m
state vector X = [X1T , . . . , Xm ] ∈ r (r = i=1
T T
n i ) with the state sub-vector
(n i −1) T
Xi = [xi , ẋi , . . . , xi ] ∈  and xi (i = 1, . . . , m) denoting m independent
n i
(n )
coordinates. In addition, the superscript in xi i refers to the n i th order of time
derivative.
The bounded perturbations in the system (9.3) can be combined together to form
a perturbation vector:

Ψ (t) = Δf + ΔBu(t) + d(t)

= x(n) − f − Bu(t). (9.4)

The estimation for the perturbation Ψ is approximated as:

(n)
Ψ est (t) = xcal − f − Bu(t − T ) (9.5)

where x(n)
cal denotes a calculated state vector because the measurement values of the
higher order states of the system are not always available. Additionally, T is the
sampling time interval and u(t − T ) represents the control input in the previous
190 9 Model-Reference Adaptive Control with Perturbation Estimation

time step. In practice, a high enough sampling frequency is selected to ensure that
u(t) ≈ u(t − T ).
(n)
In order to obtain the state vector xcal , various types of state observers can be
employed, e.g., Luenberger observer. In this work, the state vector is computed
based on a backward difference equation:

(n−1) (n−1)
(n) xcal (t) − xcal (t − T )
xcal (t) = . (9.6)
T
The above approach is employed due to its advantage in terms of computational
efficiency.
Given the system model (9.2), the perturbation f can be expressed by:

f (t) = ẍ(t) + α1 ẋ(t) + α0 x(t) − β0 u(t). (9.7)

Based on the perturbation estimation strategy, f is estimated as follows:

fˆ(t) = ẍ(t) + α1 ẋ(t) + α0 x(t) − β0 u(t − T ). (9.8)

It is observed from (9.8) that the full states (x, ẋ, and ẍ) of the system are required
to implement the perturbation estimation. In a typical positioning system, only the
position information x is provided by displacement sensors. Hence, the other states
ẋ and ẍ are estimated by the aforementioned state calculator (9.6).
In practice, the dynamics model parameters α1 , α0 , and β0 can only be estimated
with some degree of uncertainty. In the subsequent section, a MRACPE control
scheme is devised to achieve a precise position control in the presence of model
uncertainties.

9.3 MRACPE Control Design

The basic idea of MRACPE lies in the establishment of a closed-loop controller with
parameters updated online to change the response of the system. The control error is
defined as the deviation of the system output from the desired response of a reference
model. Based on this error, the controller parameters are adaptively adjusted. The
control objective is to drive the parameters to converge to ideal values for matching
the reference model response.

9.3.1 MRACPE Controller Design

Using the estimated perturbation (9.8), the system model (9.2) becomes

ẍ(t) + α1 ẋ(t) + α0 x(t) = β0 u(t) + fˆ(t). (9.9)

9.3 MRACPE Control Design 191

For the purpose of motion tracking control, a reference model is defined as follows:

ẍm (t) + a1 ẋm (t) + a0 xm (t) = b0 u d (t) (9.10)

where xm is the reference model output, u d is the input to the model, and a1 , a0 , and
b0 are positive parameters. For brevity, the time variable t is omitted hereafter in this
section.
Defining the tracking error as

e = x − xm . (9.11)

Then, subtracting Eq. (9.10) from Eq. (9.9) leads to an error dynamics equation:

ë + a1 ė + a0 e = β0 u − b0 u d + (a1 − α1 )ẋ + (a0 − α0 )x + fˆ. (9.12)

Assigning a vector of state error E = [e ė]T , Eq. (9.12) can be expressed into the
state-space form:
Ė = AE + β0 Bu + Δ (9.13)

where
     
0 1 0 0
A= , B= , Δ= (9.14)
−a0 −a1 1 δ

with δ = (a1 − α1 )ẋ + (a0 − α0 )x − b0 u d + fˆ.

It has been shown that if all the eigenvalues of the matrix A have negative real
parts, then all solutions of E to Eq. (9.13) will tend to zero as t → ∞ [4]. Moreover,
the system (9.13) is asymptotically stable. Because given any symmetric positive
definite matrix Q, there exists a symmetric positive definite matrix P, which is the
unique solution to the linear equation:

AT P + PA = −Q. (9.15)

In addition, an auxiliary control item is defined as follows:

ê = ET PB. (9.16)

Theorem 9.1 For the system described by Eq. (9.9) with a reference model (9.10),
the tracking error e defined by Eq. (9.11) satisfies lim e(t) = 0 if the MRACPE
t→∞
control law is given by
u = k0 u d + k1 x + k2 ẋ + k3 fˆ (9.17)
192 9 Model-Reference Adaptive Control with Perturbation Estimation

where the adaptive laws for the control parameters k0 , k1 , k2 , and k3 are chosen as:

k̇0 = −η0 ê u d (9.18)

k̇1 = −η1 ê x (9.19)
k̇2 = −η2 ê ẋ (9.20)
k̇3 = −η3 ê fˆ (9.21)

where η0 , η1 , η2 , and η3 are all positive constants.

Proof To demonstrate the stability of the controller, a Lyapunov candidate function

is defined as follows:
1 T 1 1
V = E PE + (β0 k0 − b0 )2 + (β0 k1 + a0 − α0 )2
2 2η0 β0 2η1 β0
1 1
+ (β0 k2 + a1 − α1 )2 + (β0 k3 + 1)2 . (9.22)
2η2 β0 2η3 β0

In view of Eqs. (9.13) and (9.16), the time derivative of the first item in Eq. (9.22)
can be expressed as

V̇1 = ET PĖ
= ET PAE + ET P(β0 Bu + Δ)
1
= ET (AT P + PA)E + ET PB(β0 u + δ)
2
1
= − ET QE + ê(β0 u + δ). (9.23)
2
Hence, the time derivative of the Lyapunov function is derived:

1 k̇0
V̇ = − ET QE + ê(β0 u + δ) + (β0 k0 − b0 )
2 η0
k̇1 k̇2 k̇3
+ (β0 k1 + a0 − α0 ) + (β0 k2 + a1 − α1 ) + (β0 k3 + 1). (9.24)
η1 η2 η3

Next, substituting Eq. (9.17) into Eq. (9.24) and rearranging the items leads to

1 k̇0 k̇1
V̇ = − ET QE + (êu d + )(β0 k0 − b0 ) + (êx + )(β0 k1 + a0 − α0 )
2 η0 η1
k̇2 k̇3
+(ê ẋ + )(β0 k2 + a1 − α1 ) + (ê fˆ + )(β0 k3 + 1). (9.25)
η2 η3

Then, substituting the adaptive laws (9.18)–(9.21) into Eq. (9.25) and considering
Eq. (9.15), allows the derivation:
9.3 MRACPE Control Design 193

1
V̇ = − ET QE ≤ 0. (9.26)
2
Therefore, the stability of the closed-loop control system is proved.

9.3.2 Dead-Zone Modification of Adaptive Laws

In practice, due to the noises in the tracking error, the control parameters (k0 , k1 ,
k2 , and k3 ) drift as demonstrated by the experiments as conducted later. In order to
alleviate the drift issue, a dead-zone modification scheme [8] is adopted to turn off
the adaptation processes once the tracking error attains the predefined small value.
Specifically, the adaptive laws (9.18)–(9.21) are revised as follows:

−η0 ê u d , if |e| > ε
k̇0 = (9.27)
0, if |e| ≤ ε

−η1 ê x, if |e| > ε
k̇1 = (9.28)
0, if |e| ≤ ε

−η2 ê ẋ, if |e| > ε
k̇2 = (9.29)
0, if |e| ≤ ε

−η3 ê fˆ, if |e| > ε
k̇3 = (9.30)
0, if |e| ≤ ε

where ε is the assigned dead-zone size, which enables the restriction of the tracking
error e. Generally, the smaller the ε, the lower the tracking errors. The ε value can
be tuned by the trial-and-error approach through experimental studies.

9.3.3 Overview of Control Scheme

Recalling the expression for the estimated perturbation in Eq. (9.8), the control law
(9.17) can be further written into the form:

u(t) = k0 u d (t)+(k1 +k3 α0 )x(t)+(k2 +k3 α1 )ẋ(t)+k3 ẍ(t)−k3 β0 u(t −T ). (9.31)

The block diagram of the control scheme is illustrated in Fig. 9.1. It is observed that
the controller has two inputs, i.e., the desired control voltage u d and the displacement
x of the system. The output of the controller is the voltage u that will be used to drive
the piezoelectric actuator.
By using a voltage signal as the reference input, several MRAC schemes have been
realized in the previous work [1, 2]. However, for practical applications, a desired
displacement trajectory instead of the voltage is predefined for the stage. Thus, once
194 9 Model-Reference Adaptive Control with Perturbation Estimation

MRACPE Control
Reference xm , xm
model

Adaptive
laws

Desired xd ud
u
Gm−1 MRACPE
trajectory controller Piezo-driven
positioning
x, x , x State x system
calculator

Fig. 9.1 Block diagram of the MRACPE control scheme

a displacement is given, the corresponding desired voltage needs to be determined.

Here, this mapping is implemented by resorting to an inverse model of the reference
model (9.10), which is expressed by the following transfer function:

Ud (s) 1
G −1
m (s) = = (s 2 + a1 s + a0 ) (9.32)
X m (s) b0

where Ud and X m represent the Laplace transform of u d and xm , respectively.

9.4 Experimental Setup and Controller Setup

In this work, the effectiveness of the MRACPE control scheme is verified by con-
ducting a series of experimental studies on a piezo-driven micropositioning system.

9.4.1 Experimental Setup

As illustrated in Fig. 9.2a, the employed micropositioning stage is composed of flex-

ure hinges. The stage is actuated by a piezoelectric stack actuator (PSA) through
a bridge-type displacement amplifier, which employs right circular flexure hinges.
The linear displacement of the output platform is guided by four leaf springs which
are connected to the mounting frame. When the PSA is driven by a voltage amplifier,
the stage delivers an amplified displacement along the x-axis direction without para-
sitic motions, which has been validated by the simulation result of the finite-element
analysis (FEA) as shown in Fig. 9.2b. More details about the working principle and
analytical modeling of the displacement amplifier can be found in [12].
9.4 Experimental Setup and Controller Setup 195

(a) (b)
Leaf spring Mounting frame PSA

y
x

Output platform Displacement-amplifier

Fig. 9.2 a A micropositioning stage driven by a piezoelectric stack actuator through displacement
amplifier. b Deformed shape obtained by FEA simulation

Fig. 9.3 Experimental setup of a piezo-driven micropositioning system

The experimental setup of the custom-built micropositioning system is depicted

in Fig. 9.3. Actually, a PSA and a voice coil motor are used to construct a dual-
stage micropositioning system as detailed in [10]. In this work, the PSA-driven
inner flexure stage is employed to verify the proposed control scheme. The PSA is
connected to a high-voltage amplifier (model: EPA-104, from Piezo Systems, Inc.)
with an adjustable gain value of then. The output position of the stage is measured by
a laser displacement sensor (model: LK-H055, from Keyence Corp.). In addition, a
Natural Instruments (NI) cRIO-9022 real-time controller combined with cRIO-9118
reconfigurable chassis is adopted to implement the control algorithm. The chassis is
equipped with NI-9215 A/D and NI-9263 D/A modules for the acquisition of sensor
readings and production of excitation voltage signals, respectively.
The NI cRIO-9118 chassis contains a field-programmable gate array (FPGA)
core, and the associated cRIO-9022 real-time controller communicates with a per-
sonal computer (PC) via Ethernet port. The control algorithm is programmed using
TM
the LabVIEW software. With the MRACPE control algorithm, the control hard-
196 9 Model-Reference Adaptive Control with Perturbation Estimation

ware enables a maximal closed-loop sampling frequency of 100 kHz. In this work, a
sufficient sampling frequency of 5 kHz is adopted.

9.4.2 Statics Testing and Dynamics Model Identification

First, the motion range of the micropositioning system is tested by applying a sinu-
soidal voltage signal with the amplitude of 10 V and frequency of 0.2 Hz. The results
are illustrated in Fig. 9.4, which shows that a motion range over 94 µm is obtained.
In addition, the open-loop output versus input curve exhibits a hysteresis width of
17.4 %. The presence of the significant hysteresis nonlinearity necessitates the control
scheme development in order to achieve a precise positioning.
Next, the dynamics model parameters α1 , α0 , and β0 of the nominal plant model
(9.2) are identified by means of frequency response approach. Specifically, swept sine
waves with the amplitude of 0.5 V and frequency range of 1–650 Hz are produced
to drive the PSA via the high-voltage amplifier. Both the excitation voltage and the
displacement output signals are acquired to generate the frequency response of the
system, as shown in Fig. 9.5. From the experimental data, a second-order model is
identified and the model parameters are tabulated in Table 9.1.
It is observed from Fig. 9.5 that the second-order model is capable of approxi-
mating the frequency response up to 600 Hz, which covers the resonant frequency of
470 Hz. In order to capture the system dynamics at frequencies higher than 600 Hz,
a model of higher order is needed. In the current work, a simple second-order model
is employed to demonstrate the effectiveness of the control scheme, and the residual
modes at higher frequencies are considered as disturbances, which will be suppressed
by the MRACPE controller.

Fig. 9.4 Open-loop 100

hysteresis loop obtained with
h/H ×100% = 17.4%
a 0.2-Hz sinusoidal reference
80
Output displacement (µm)

input
H = 94.9 µm
h = 16.5 µm

60

40

20

0
0 2 4 6 8 10
Input voltage (V)
9.4 Experimental Setup and Controller Setup 197

Fig. 9.5 Frequency 40

responses obtained by
experiment and the identified 30
second-order model
20

Magnitude (dB)
10

−10

−20 Experimental data

The 2nd−order model
−30
0 1 2 3
10 10 10 10
Frequency (Hz)

Table 9.1 Parameters of the Parameter Value

plant model and designed
controller α0 8.721 × 106
α1 2.953 × 102
β0 4.404 × 107
a0 8.721 × 106
a1 1.477 × 103
b0 4.404 × 107
η0 1 × 10−2
η1 1 × 10−4
η2 1 × 10−9
η3 1 × 10−17
ε 0.7
Q diag{3 × 105 , 1 × 10−3 }

9.4.3 Controller Setup

For experimental investigations, the reference model is chosen with parameters a0 , a1 ,

and b0 as described in Table 9.1. The assigned parameters of the MRACPE controller
are also shown in Table 9.1. As depicted in Fig. 9.1, by passing the desired position
trajectory xd through the reference model inverse G −1
m , a reference voltage signal u d
is generated. Excited by u d , an output xm = xd is then produced by the reference
model.
For a comparative study, the popular PID controller is also implemented. A digital
PID control strategy is employed as follows:
198 9 Model-Reference Adaptive Control with Perturbation Estimation


k
u k = K p ek + K i ek + K d (ek − ek−1 ) (9.33)
j=0

with the positioning error ek = x dk − xk , where x dk and xk represent the desired and
actual system output at the kth time step, and K p , K i , and K d denote the proportional,
integral, and derivative gains, respectively. The PID gains are initially tuned by
simulation studies using the Ziegler–Nichols (Z–N) method. The gains are then finely
adjusted through experimental studies to eliminate the overshoot, i.e., K p = 0.01,
K i = 0.0023, and K d = 0.029. A sampling time of T = 0.0002 s is adopted for both
MRACPE and PID controllers.

9.5 Experimental Results and Discussion

9.5.1 Set-Point Positioning Testing

First, the set-point positioning capability of the micropositioning system is examined.

By commanding the output platform from home position to 20 µm, the positioning
results using PID and MRACPE controllers are depicted in Fig. 9.6. Both controllers
are tuned to achieve a quick response without overshoot. For a quantitative compar-
ison, the performances are shown in Table 9.2.
It is observed that the PID scheme produces a slightly quicker response because
its 5 % settling time is 0.006 s smaller than that of SMCPE. Even so, the MRACPE
delivers a more accurate positioning as it creates a steady-state root-mean-square
(RMS) error which is 7 % lower than the PID result.

Fig. 9.6 Experimental 25

results of set-point
positioning using PID and
20
MRACPE controllers
Displacement (µm)

15
21

10 20

19
5
0.8 0.9 1

0 Reference
PID
MRACPE
−5
0 0.2 0.4 0.6 0.8 1
Time (s)
9.5 Experimental Results and Discussion 199

Table 9.2 Positioning results of PID and MRACPE controllers

Performance PID MRACPE
Set-point positioning (20 μm)
5 % settling time (s) 0.025 0.031
RMS steady-state error (µm) 0.275 0.256
Sinusoidal positioning (10 Hz)
Maximum steady-state error (%) 39.20 6.51
RMS steady-state error (%) 25.60 2.96

0.8 0.012

0.6 0.01

0.008
0.4
k0

k1
0.006
0.2
0.004
0 0.002

−0.2 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

−5 −10
x 10 x 10

1 0

−2
k2

−4
k

0.5
−6

−8

0 −10
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)

Fig. 9.7 Adaptation processes of the MRACPE control parameters with ε = 0.7 for the set-point
positioning

Additionally, the adaptation processes of the four control parameters of the

MRACPE scheme are depicted in Fig. 9.7. The dead-zone parameter is assigned
as ε = 0.7 to limit the positioning errors within ±0.7 µm, which is the noise level of
the displacement sensor. It is seen that as the response approaches the steady state,
the controller parameters converge to the corresponding ideal values, respectively. It
is notable that a quicker response speed of MRACPE can be obtained by optimally
tuning the control gains η0 , η1 , η2 , and η3 .
200 9 Model-Reference Adaptive Control with Perturbation Estimation

9.5.2 Sinusoidal Positioning Testing

Next, a 10-Hz, 20 µm-amplitude sinusoidal positioning control is performed by

employing the two controllers, and the results are shown in Fig. 9.8. It is found
that the PID algorithm leads to identical responses in each cycle. Concerning the
MRACPE, after the transient progress within the first few cycles, it gives identical
responses afterwards. Additionally, the PID control produces the maximum and RMS
steady-state positioning errors of 7.840 and 5.119 µm, which account for 39.20 and
25.60 % of the positioning range, respectively. In contrast, the MRACPE leads to the
maximum and RMS steady-state errors of 1.302 and 0.591 µm, i.e., 6.51 and 2.96 %
of the positioning range, respectively. The sinusoidal positioning results are summa-
rized in Table 9.2. As compared with PID control, the MRACPE scheme improves
the positioning accuracy by 83 and 88 % in terms of the maximum and RMS errors,
respectively.
The sinusoidal positioning results are generated by the same controllers as used
in the aforementioned set-point positioning tests, except that ε = 1.3 is reassigned
in MRACPE. By optimally tuning the PID gains, the PID 2 controller is generated,
which creates the results as shown in Fig. 9.8. It is observed that the maximum and
RMS errors of 25.96 and 14.44 % are produced, respectively, which are about four
times worse than the MRACPE results. Hence, the superiority of MRACPE over PID
control is evident from the experimental results.
Furthermore, to discover the influence of the dead-zone selection on the MRACPE
performance, the positioning results with three different dead-zones, i.e., 2.3, 1.6 and
1.3 µm, are examined. The MRACPE positioning errors and the parameter adaptation
procedures are shown in Figs. 9.9 and 9.10, respectively. Figure 9.9 indicates that the
maximum steady-state positioning errors are restricted by the assigned dead-zone
values. The smaller the dead-zone, the better the positioning accuracy.
Figure 9.10 reveals that the controller parameters k0 , k1 , k2 , and k3 arrive at the
steady states quickly in two of the three case studies (ε = 2.3 and 1.6). Concerning the
case with a smaller dead-zone value (ε = 1.3), the convergence speed of the controller
parameters is slower due to the parameter drift. Further experimental testings show
that the parameters get saturated finally. Hence, a tradeoff between the tracking
accuracy and parameter convergence speed is needed to assign the dead-zone value.

9.5.3 Control Bandwidth Testing

In addition, the control bandwidth of the PSA-actuated system is tested by applying

the 20-µm sinusoidal signal with the frequency varying from 1 to 100 Hz. The closed-
loop frequency responses are shown in Fig. 9.11. It is observed that there are large
phase lags (over 90◦ ) within the ordinary −3 dB bandwidth, which lead to large
9.5 Experimental Results and Discussion 201

Fig. 9.8 a The 10-Hz (a)

sinusoidal motion tracking 25
results of PID and MRACPE
(ε = 1.6) controllers.
20
b Positioning errors

Displacement (µm)
15

10

Reference PID PID 2 MRACPE

−5
1.6 1.65 1.7 1.75 1.8 1.85 1.9 1.95 2
Time (s)

(b) 10
PID PID 2 MRACPE
8

6
Displacement error (µm)

−2

−4

−6

−8

−10
0 0.5 1 1.5 2
Time (s)

tracking errors. Hence, the closed-loop control bandwidth is defined as the frequency
at which the phase is lagged 30◦ [11]. With the PID and MRACPE controllers, the
30◦ -lag bandwidths of 12.8 and 79.3 Hz are achieved, which are equivalent to 2.7 and
16.9 % of the resonant frequency, respectively. These cutoff frequencies correspond
to small errors of 0.43 and 1.72 dB for the magnitude responses of the PID and
MRACPE controllers, respectively. As compared with PID algorithm, the MRAPCE
scheme improves the control bandwidth by over six times.
202 9 Model-Reference Adaptive Control with Perturbation Estimation

Fig. 9.9 a The sinusoidal (a) 25

positioning results of
MRACPE with different
dead-zones ε. b Positioning 20
errors

Displacement (µm)
15

10

−5
1.6 1.65 1.7 1.75 1.8 1.85 1.9 1.95 2
Time (s)

(b) 10

8
Displacement error (µm)

−2

−4
0 0.5 1 1.5 2
Time (s)

9.5.4 Discussion

The obtained experimental results confirm the effectiveness of the MRACPE con-
trol scheme for a PSA-actuated micropositioning system. Because the scheme is
implemented without modeling the hysteresis nonlinearity, the results demonstrate
the efficiency of the reported idea for the disturbance suppression. Considering that
the nominal plant model is used to develop the perturbation estimator for the estima-
tion of the unmodeled dynamics and uncertainties in the control design, the results
also reveal the feasibility of the identified second-order model for the positioning
application.
Figure 9.11 indicates that the performance of PID control drops quicker than
that of the MRACPE control as the increasing of the reference input frequency.
9.5 Experimental Results and Discussion 203

−3
x 10
1 10

0.8
5
0.6
0

1
k

k
0.4
0
0.2

0 −5
0 0.5 1 1.5 2 0 0.5 1 1.5 2

−5 −9
x 10 x 10
6 5

4 0
2

2 −5
k

0 −10

−2 −15
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time (s) Time (s)

Fig. 9.10 Adaptation processes of the control parameters of MRACPE with different dead-zones
ε for sinusoidal positioning

Fig. 9.11 Control 5

bandwidth test results of the
Magnitude (dB)

micropositioning system 0
with PID and MRACPE
controllers −5

−10

−15 0 1 2
10 10 10

0
Phase (degrees)

−50

−100

−150
PID
−200 MRACPE
0 1 2
10 10 10
Frequency (Hz)
204 9 Model-Reference Adaptive Control with Perturbation Estimation

This is one of the reasons why a 10-Hz sinusoidal reference input is adopted in the
aforementioned comparative study. Actually, the MRACPE achieves a RMSE less
than 10 % of positioning range for a higher speed tracking of reference input with
the frequency of 40 Hz.
It is notable that although the MRACPE produces a slightly lower transient
response than PID control, it achieves more accurate positioning in both set-point
and sinusoidal tracking tasks. Another advantage of the MRACPE is that its posi-
tioning error size can be predefined by assigning the dead-zone (ε) value. Exper-
imental results reveal that the smaller the dead-zone, the slower the convergence
speed of controller parameters. Even so, the convergence speed of the parameters
does not affect the steady-state tracking error because the latter is restricted by the
assigned ε value. The smaller the ε, the more accurate the positioning result. How-
ever, too small ε produces the maximum positioning error exceeding the ε value, or
even causes instability of the system due to the drift of the controller parameters,
which are mainly induced by the noises of the displacement sensor. Hence, a tradeoff
between the tracking accuracy and the stability is required to choose the dead-zone
value. The tracking errors can be reduced by employing displacement sensors with
a lower noise level. Alternatively, a state observer may be adopted to estimate the
state values. Unavoidably, the employment of state observer complicates the control
design process at the same time.
In addition, the control gains of the MRACPE are not optimally tuned. An optimal
tuning of the four control gains is a possible solution to further improve the position-
ing results. Moreover, the robust controller design procedure [9] can be employed to
achieve high positioning accuracy at higher speeds.

9.6 Chapter Summary

This chapter provides the design procedure of the MRACPE control scheme for a
PSA-actuated micropositioning system without modeling the hysteretic nonlinear-
ity effect. The controller is designed based on Lyapunov stability analysis, and its
performance is verified by a series of experimental studies. Results show that the
MRACPE with dead-zone modification of the adaptive parameters is capable of fur-
ther suppressing the disturbances and substantially mitigating the positioning errors
as compared with conventional PID control. The MRACPE enables over six times
increase of control bandwidth versus the PID algorithm. Moreover, the dead-zone
allows the specification of the maximum tracking error for the MRACPE scheme,
which is attractive for practical applications. Owing to a not complicated structure,
the proposed control scheme can also be extended to position control of other types
of micro-/nanopositioning systems.
References 205

References

1. Cheng, C.C., Chang, C.C., Su, T.M.: Design of model reference adaptive tracking controllers
for mismatch perturbed nonlinear systems with input nonlinearity. In: Proceedings of 17th
IFAC World Congress, pp. 5974–5979. Seoul, Korea (2008)
2. Demetriou, M.A., Fahroo, F.: Model reference adaptive control of structurally perturbed
second-order distributed parameter systems. Int. J. Robust Nonlinear Control 16(16), 773–
799 (2006)
3. Elmali, H., Olgac, N.: Implementation of sliding mode control with perturbation estimation
(SMCPE). IEEE Trans. Control Syst. Technol. 4(1), 79–85 (1996)
4. Ioannou, P.A., Sun, J.: Robust Adaptive Control. Prentice Hall, New Jersey (1996)
5. Kim, J.Y., Bentsman, J.: Disturbance rejection in robust model reference adaptive control of
parabolic and hyperbolic systems. In: Proceedings of 45th IEEE Conference on Decision and
Control, pp. 3083–3088. San Diego, CA, USA (2006)
6. Li, Y., Xu, Q.: Hysteresis modeling and compensation for an XY micropositioning stage with
model reference adaptive control. In: Proceedings of 48th IEEE Conference on Decision and
Control, pp. 5580–5585. Shanghai, China (2009)
7. Liu, Y.T., Chang, K.M., Li, W.Z.: Model reference adaptive control for a piezo-positioning
system. Precis. Eng. 34(1), 62–69 (2010)
8. Narendra, K.S., Annaswamy, A.M.: Stable Adaptive Systems. Prentice Hall, New Jersey (1989)
9. Tuma, T., Haeberle, W., Rothuizen, H., Lygeros, J., Pantazi, A., Sebastian, A.: A dual-stage
nanopositioning approach to high-speed scanning probe microscopy. In: Proceedings of 51st
IEEE Conference on Decision and Control, pp. 5079–5084. Maui, HI, USA (2012)
10. Xu, Q.: Design and development of a flexure-based dual-stage nanopositioning system with
minimum interference behavior. IEEE Trans. Autom. Sci. Eng. 9(3), 554–563 (2012)
11. Xu, Q., Jia, M.: Model reference adaptive control with perturbation estimation for a micropo-
sitioning system. IEEE Trans. Control Syst. Technol. 22(1), 352–359 (2014)
12. Xu, Q., Li, Y.: Analytical modeling, optimization and testing of a compound bridge-type
compliant displacement amplifier. Mech. Mach. Theory 46(2), 183–200 (2011)
 Part IV
Applications to Micromanipulation
Chapter 10
Adaptive Impedance Control of Piezoelectric
Microgripper

Delicate interaction control is a crucial issue for automated microsystems dedicated

to micromanipulation of microobjects. This chapter presents a framework of dig-
ital sliding mode generalized impedance control with adaptive switching gain to
regulating both the position and contact force of a piezoelectric-bimorph microgrip-
per for micromanipulation and microassembly applications. Based on a second-order
dynamics model, its implementation does not require a state observer and a hysteresis/
creep model. The stability of the control system is proved in theory, which ensures
the tracking performance in the presence of model uncertainties and disturbances.
The effectiveness of the scheme is validated by experimental investigations on the
grasp operation of a microgear.

10.1 Introduction

Microassembly is an important approach to producing complex micro-/nanosystems

which are composed of microparts fabricated from different materials [2, 26]. As a
crucial device to realize the grasp-hold-release operation in the aforementioned appli-
cation, microgripper has gained intensive attention in the literature. Various types
of microgrippers have been proposed in terms of electrostatic [3], electrothermal
[15], and piezoelectric microgrippers [24], etc. In particular, smart material-based
piezoelectric actuators offer the merits of subnanometer positioning resolution and
rapid response speed. Hence, constructed by a pair of piezoelectric unimorph or
bimorph actuators, the piezoelectric microgrippers have been popularly applied in
micromanipulation and microassembly tasks [6, 22].
To avoid damaging the fragile microobjects and microgripper arms which contact
each other during microassembly process, the interaction control is critical to regu-
late the desired position and contact force simultaneously [20]. Generally, there are
two basic approaches to interaction control, i.e., hybrid position/force control and
impedance control. The first method generates a compliant motion by controlling
© Springer International Publishing Switzerland 2016 209
Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2_10
210 10 Adaptive Impedance Control of Piezoelectric Microgripper

the end-effector position in position subspace and the interaction force in force
subspace, whose implementation relies on a switching of the control law. However,
this approach suffers from the robustness problem during the transition between
unconstrained and constrained motions [30]. Alternatively, the impedance control
enables a compliant motion of the end-effector in a unified framework for both
constrained and unconstrained directions [13]. The goal of impedance control is to
regulate a desired dynamics relation between the position and contact force. That
is, both position and force are indirectly controlled by regulating the predefined
dynamics response, i.e., the desired impedance.
Impedance control emerges as a promising approach to interaction control in
micro-/nanoscales [11, 18]. For example, a position-based impedance control of a
piezoelectric-bimorph microgripper is reported in [29], where a state-feedback posi-
tion controller and a low-pass filtered force feedback controller are implemented.
However, the impedance control usually requires an exact dynamics model of the sys-
tem. Due to the uncertainties in dynamics model, it is practically difficult to achieve
the desired impedance. To remedy this shortcoming, adaptive [17] and robust [19]
impedance control algorithms have been developed to tolerate the dynamics parame-
ter uncertainties. Nevertheless, the conventional impedance control framework has
a drawback of inability of tracking the reference force input. Several schemes have
been exploited to overcome this disadvantage. Typically, a generalized impedance
control (GIC) scheme is established based on a generalized desired impedance [5].
It integrates both the position and force errors, and enables the force tracking ability.
Recently, intensive efforts have been made toward force tracking impedance con-
trol, which is robust to parameter uncertainties in both dynamics and environment
models [14]. Although this robustness property appears attractive, it is at the expense
of implementation complication. Considering that the impedance control possesses
some inherent robustness to environment model errors [1, 28], a GIC with robustness
to dynamic model uncertainties is desirable for a precise interaction control.
Regarding piezoelectric microgrippers, in addition to the model uncertainties,
further challenges to achieve a delicate interaction control come from the nonlinear
hysteresis and creep effects of the piezoelectric actuators. Thus, control strategies
need to be well developed to achieve a precise position and force control. Usually, a
feedforward control based on an inverse hysteresis model and an inverse creep model
is adopted to compensate for the piezoelectric nonlinearities. Nevertheless, modeling
the hysteresis and creep effects is a time-consuming work and the tracking result is
very sensitive to the model accuracy [34]. Hence, by treating the nonlinearities as
disturbances, the robust feedback control is more attractive owing to its merit of easy
implementation [32, 35]. In the literature, a state-of-the-art robust GIC algorithm
is presented in [16], where the desired motion and force trajectories are tracked
simultaneously. It has demonstrated the potential of sliding-mode control (SMC) in
precise interaction control in the presence of model uncertainties and disturbances.
However, the reported SMC-based GIC is realized in the continuous-time form. Its
implementation needs a state observer and prior knowledge on the uncertainty bounds
of the dynamics model parameters.
10.1 Introduction 211

In practice, the control scheme of a microassembly system is usually realized on a

computer or a DSP apparatus, i.e., a sampled-data system. As it is known that applying
a continuous-time controller directly to a digital system can induce instability due to
the computational delays [4]. This is true for a microassembly system, e.g., a visual
servoing system [25], which typically has a low sampling rate, e.g., in the order of
10 Hz. Hence, to deploy a reliable interaction control to a sampled-data system, a
discrete-time impedance controller is more preferable.
In this chapter, an adaptive digital sliding-mode generalized impedance control
(ADSMGIC) is reported to realize an interaction control of a piezoelectric microgrip-
per dedicated to micromanipulation and microassembly. Unlike the continuous-time
and discrete-time SMC algorithms which are developed based on the system state
[16, 35], the presented one is established by using a discrete-time second-order
dynamics model of the system. Hence, it eliminates the use of a state observer and
enables a substantial simplification of practical implementation procedures. More-
over, the unmodeled piezoelectric nonlinearities and other uncertainties as well as
disturbances are treated as a lumped perturbation. This perturbation term is estimated
by resorting to a perturbation estimation technique, which allows the relief of bur-
den on hysteresis and creep modeling. Furthermore, the chattering phenomenon is
suppressed by employing an adaptive law for the switching gain. The stability of
the closed-loop system is proved theoretically and the effectiveness of the reported
interaction control scheme is validated by conducting several experimental studies.

10.2 Problem Formulation

A piezoelectric microgripper is illustrated in Fig. 10.1a. It is constructed by a pair of

piezoelectric bimorphs for the microobject grasping. Owing to the symmetry of the
microgripper, one gripper arm, i.e., a piezoelectric cantilever as shown in Fig. 10.1b
is selected for a detailed analysis. The cantilever suffers from an excitation voltage
and a contact force which is exerted by the microobject. In order to guarantee a
successful assembly of the microobject, both the position x and the contact force F
need to be controlled.

(a) (b) x

−F
Microobject

Piezo-bimorph
−−
u
−

Fig. 10.1 a Schematic depiction of grasping a microobject by a piezoelectric microgripper con-

structed by two piezoelectric cantilevers. b Illustration of output displacement x of a piezoelectric
cantilever under an excitation of input voltage u and contact force F
212 10 Adaptive Impedance Control of Piezoelectric Microgripper

10.2.1 Dynamics Modeling and Perturbation Estimation

The dynamics model of a piezoelectric cantilever under the actions of an excitation

voltage and an external force (−F) can be established as follows:

m̂ ẍ(t) + b̂ ẋ(t) + k̂x(t) = d̂u(t) + P(t) − F(t) (10.1)

where t is the time variable, parameters m̂, b̂, k̂, and d̂ represent the estimated values
of mass, damping coefficient, stiffness, and piezoelectric coefficient of the gripper
arm, respectively. In addition, u is the input voltage, x is the output position, and
F denotes the interaction force exerted on the microobject. The bounded lumped
perturbation term P describes the combined effect of the unmolded piezoelectric
nonlinearities (hysteresis and drift), the uncertainties of dynamic model parameters
(m̂, b̂, k̂, and d̂), and other disturbances.
By adopting a small sampling time T , the following approximations are valid
[31]:

1
ẋ(t) ≈ (xk − xk−1 ) (10.2)
T
1
ẍ(t) ≈ 2 (xk − 2xk−1 + xk−2 ) (10.3)
T
where xk = x(kT ), xk−1 = x(kT − T ), and xk−2 = x(kT − 2T ) with k representing
the kth time step.
Then, the continuous-time dynamics model (10.1) is converted into an equivalent
discrete-time form:

m̄xk−2 + b̄xk−1 + k̄xk = d̄u k + Pk − Fk (10.4)

where the parameters are

m̂ 2m̂ b̂ m̂ b̂
m̄ = 2
, b̄ = − 2 − , k̄ = 2 + + k̂, d̄ = d̂. (10.5)
T T T T T
By employing the perturbation estimation technique [7], the lumped perturbation
term Pk can be derived by its one-step delayed estimation:

P̂k = −d̄u k−1 + m̄xk−3 + b̄xk−2 + k̄xk−1 + Fk−1 . (10.6)

Hence, the dynamics model (10.4) can be rewritten into:

m̄xk−2 + b̄xk−1 + k̄xk = d̄u k + P̂k − P̃k − Fk (10.7)

where P̃k = P̂k − Pk is the perturbation estimation error.

10.2 Problem Formulation 213

10.2.2 Impedance Control Problem

The objective of impedance control is to regulate the position and contact force
simultaneously by achieving a desired impedance behavior.
Assume that the output position and predefined reference position of the gripper
arm are x and xr , respectively. By adopting the impedance control algorithm [10],
the desired impedance behavior between the output position error ex = x − xr and
the contact force F can be specified by a second-order dynamics equation:

M ëx (t) + B ėx (t) + K ex (t) = −F(t) (10.8)

where M, B, and K represent the target mass, damping, and stiffness parameters,
respectively. However, the major limitation of this kind of impedance control is the
inability of achieving explicit force tracking control.
To formulate an impedance behavior to track the desired force Fr as exerted on
the object, the GIC scheme can be adopted. The GIC combines the position and force
errors by using a generalized impedance equation [5]:

M ëx (t) + B ėx (t) + K ex (t) = −K f e f (t) (10.9)

where e f = F − Fr is the force error, and M, B, K , and K f are the target parameters
for the generalized impedance.
It is observed from Eq. (10.9) that by introducing the reference force Fr to the
desired impedance behavior, the interaction force can be explicitly controlled by
using the GIC approach.
Taking into account the approximations (10.2) and (10.3), the differential equation
(10.9) is equivalently converted into a difference equation:

M̄ex,k−2 + B̄ex,k−1 + K̄ ex,k = − K̄ f e f,k (10.10)

where ex,k−2 = ex (kT − 2T ) and e f,k = e f (kT ).

In addition, the relationships between the target parameters of the continuous-time
impedance (M, B, K , and K f ) and discrete-time impedance ( M̄, B̄, K̄ , and K̄ f ) are
derived below
M 2M B M B
M̄ = , B̄ = − 2 − , K̄ = 2 + + K , K̄ f = K f . (10.11)
T2 T T T T
However, it is difficult to obtain precise values for the parameters of dynamics
model (10.7) in practice. Additionally, the perturbation estimation error P̃k always
exists, especially when the sampling time T is large and the variation rate of the pertur-
bation Pk is high. Hence, the presence of model uncertainties and disturbances poses
a great challenge for the achievement of the desired impedance behavior described
by Eq. (10.10). In the following section, a control scheme is developed to overcome
this issue.
214 10 Adaptive Impedance Control of Piezoelectric Microgripper

10.3 ADSMGIC Scheme Design

An ADSMGIC scheme is designed to guarantee the achievement of the desired

impedance behavior (10.10) so that both the position and contact force of the micro-
gripper can be regulated. In particular, a digital sliding mode control (DSMC) scheme
is established as follows.

10.3.1 Sliding Function Definition

First, a sliding function is defined and then converted into the discrete-time form.
Based on the position error ex,k = xk − xr,k and the force error e f,k = Fk − Fr,k ,
an incremental PID type of sliding function is defined below

sk = K P (ex,k − ex,k−1 ) + K I ex,k + K D (ex,k − 2ex,k−1 + ex,k−2 ) + K 4 e f,k + K 5 sk−1

(10.12)

where K P , K I , and K D denote the proportional, integral, and derivative gains, respec-
tively. In addition, K 4 and K 5 are also positive control gains. Actually, the last term
in Eq. (10.12) represents a momentum term which is defined by the momentum
constant K 5 (0 < K 5 < 1).
Next, through a fundamental algebra operation, the sliding function (10.12) is
expressed into the following form

sk = K 1 ex,k−2 + K 2 ex,k−1 + K 3 ex,k + K 4 e f,k + K 5 sk−1 (10.13)

where

K1 = K D (10.14)
K 2 = −K P − 2K D (10.15)
K3 = K P + K I + K D . (10.16)

Then, by setting the following parameters

K 1 = M̄ (10.17)
K 2 = B̄ (10.18)
K 3 = K̄ (10.19)
K 4 = K̄ f (10.20)

the sliding function (10.13) can be further written into:

sk = M̄ex,k−2 + B̄ex,k−1 + K̄ ex,k + K̄ f e f,k + K 5 sk−1 . (10.21)

10.3 ADSMGIC Scheme Design 215

It is observed that the sliding function (10.21) becomes the desired impedance
behavior (10.10) if sk = 0 and sk−1 = 0 are assigned. Motivated by this obser-
vation, the sliding-mode control technique is applied here. That is, if the system
impedance trajectory is maintained on the sliding surface sk = 0 and sk−1 = 0, the
desired impedance behavior (10.10) will be achieved. In the following discussions, an
ADSMGIC scheme with adaptive gain is developed to ensure that the discrete sliding
mode is reached and maintained, i.e., the desired impedance behavior is achieved.

10.3.2 ADSMGIC Design

First, substituting Eq. (10.6) into Eq. (10.7) allows the calculation of the position:

1
xk = [d̄(u k − u k−1 ) + m̄xk−3 + (b̄ − m̄)xk−2 + (k̄ − b̄)xk−1 − (Fk − Fk−1 ) − P̃k ].
k̄
(10.22)
eq
Then, taking into account that the equivalent control u k is the solution to Δsk =
sk − sk−1 = 0 [8], the following deduction is derived:

M̄ex,k−2 + B̄ex,k−1 + K̄ ex,k + K̄ f e f,k + K 5 sk−1 = sk−1 . (10.23)

In view of the definition ex,k = xk − xr,k , Eq. (10.23) becomes

M̄ex,k−2 + B̄ex,k−1 + K̄ (xk − xr,k ) + K̄ f e f,k + K 5 sk−1 = sk−1 . (10.24)

eq
It is notable that the equivalent control u k is the control action without consider-
ing the perturbation estimation error P̃k . The error P̃k will be compensated for by
the switching control action u swk derived later. Thus, substituting the x k expression
(10.22) into (10.24), gives

eq k̄  
u k = u k−1 + (1 − K 5 )sk−1 − M̄ex,k−2 − B̄ex,k−1 − K̄ f e f,k
K̄ d̄
1 
− m̄xk−3 + (b̄ − m̄)xk−2 + (k̄ − b̄)xk−1 − k̄xr,k − (Fk − Fk−1 ) . (10.25)
d̄

The equivalent controller (10.25) takes effect in the sliding phase when the
impedance trajectory is kept on the sliding surface (i.e., sk = 0). However, if the
initial state of the trajectory does not lie on the sliding surface or a large perturbation
estimation error P̃k emerges owing to large uncertainties and disturbances during the
sliding phase, the stand-alone equivalent control cannot drive the impedance trajec-
tory towards the sliding surface. Under such situation, a switching control action u sw k
is necessary.
216 10 Adaptive Impedance Control of Piezoelectric Microgripper

F
xr u −
ADSMGIC Microgripper x
Fr +
+
1 G −1 Σ F̂
z − sp − +
− dp
+ sp
Force observer

Fig. 10.2 Block diagram of the ADSMGIC control scheme

Hence, the equivalent control is augmented by the switching control to create the
total control action:
eq
u k = u k + u sw
k . (10.26)

That is

k̄  
u k = u k−1 + (1 − K 5 )sk−1 − M̄ex,k−2 − B̄ex,k−1 − K̄ f e f,k
K̄ d̄
1 
− m̄xk−3 + (b̄ − m̄)xk−2 + (k̄ − b̄)xk−1 − k̄xr,k − (Fk − Fk−1 )
d̄
KS
− sign(sk−1 ) (10.27)
d̄

where the last term is the switching control action. In addition, sign(·) denotes the
signum function and K S is a positive control gain.
A block diagram of the ADSMGIC control scheme is shown in Fig. 10.2. It is
observed that both the desired position and force are controlled at the same time. A
force observer is employed to estimate the contact force during the grasp operation.
Alternatively, a force sensor can be used to provide the force signal.

10.3.3 Stability Analysis

Theorem 10.1 For the system (10.7) with the sliding function (10.12), if the con-
troller (10.27) is employed, then the discrete sliding mode will occur in a finite
number of steps.

Proof Substituting Eq. (10.27) along with the control gains (10.14)–(10.20) into the
sliding function (10.12), a necessary algebra operation gives
10.3 ADSMGIC Scheme Design 217

sk = K̄ (xk − xr,k ) + M̄ex,k−2 + B̄ex,k−1 + K̄ f e f,k + K 5 sk−1


1 
= K̄ d̄(u k − u k−1 ) + m̄xk−3 + (b̄ − m̄)xk−2 + (k̄ − b̄)xk−1 − (Fk − Fk−1 ) − P̃k
k̄

−xr,k + M̄ex,k−2 + B̄ex,k−1 + K̄ f e f,k + K 5 sk−1

K̄ K S K̄
= sk−1 − sign(sk−1 ) − P̃k
k̄ k̄
K̄  
= sk−1 − K S sign(sk−1 ) + P̃k (10.28)
k̄

where the parameters of k̄ and K̄ are positive.

If sk−1 ≥ 0, it is derived that

sk ≤ sk−1 if K S ≥ | P̃k |. (10.29)

Otherwise, in the case of sk−1 < 0, then

sk ≥ sk−1 if K S ≥ | P̃k |. (10.30)

Hence, combining Eqs. (10.29) and (10.30), the following deduction can be derived

|sk | ≤ |sk−1 | if K S ≥ | P̃k | (10.31)

which indicates that sk decreases monotonously, and the discrete sliding mode is
reached within a finite number of steps.

Remark 10.1 It has been shown that the relation (10.31) describes a sufficient con-
dition for the existence of discrete sliding mode [27]. Owing to the discontinuity
of the signum function sign(·), chattering effects may occur in the control action.
To reduce the chattering phenomenon, the boundary layer technique is employed by
replacing the signum function in (10.27) with the saturation function:

sign(sk ) if |sk | > δ
sat(sk ) = sk (10.32)
δ if |sk | ≤ δ

where a tradeoff between the chattering effect and tracking error is needed to assign
the parameter δ.

Remark 10.2 Even though the boundary layer technique is capable of alleviating the
chattering effect, the fixed gain K S is not feasible for some scenarios in practice. In
this work, an adaptive rule for the switching gain K S is employed to adjust the gain
value of the saturation function (10.32) online [21]:

K S,k = |K S,k−1 + γ sign(sk sk−1 )| (10.33)

218 10 Adaptive Impedance Control of Piezoelectric Microgripper

where the positive adaptation parameter γ determines the rate of adaptation of the
control gain.
By resorting to the adaptive rule (10.33), it is deduced that if the impedance tra-
jectory has not yet reached the sliding surface, then sign(sk sk−1 ) = +1. Hence, the
control gain is increased. Once the impedance trajectory crosses the sliding surface,
sign(sk sk−1 ) = −1 holds. Consequently, the control gain is decreased accordingly.
Therefore, the purpose of the adaptation law (10.33) is to produce an optimal switch-
ing gain at each time step so that the discrete sliding mode is reached and maintained,
i.e., the desired impedance behavior is achieved.

10.3.4 Evaluation of Steady-State Errors

The developed control scheme is based on the framework of GIC in conjunction with
adaptive DSMC. In order to evaluate the steady-state position and force errors, the
contact environment of the microgripper is modeled as follows:

F = K e (x − xe ) (10.34)

where K e describes the stiffness of the environment and xe represents the equilibrium
position of the environment in the absence of contact force.
Combing Eqs. (10.9) and (10.34) allows the generation of an equation of the
closed-loop impedance control system interacting with the environment:

M ẍ + B ẋ + (K + K f K e )x = M ẍr + B ẋr + K xr + K f (Fr + K e xe ) (10.35)

which is stable if positive target parameters are chosen for the desired impedance [5].
In order to achieve a satisfactory manipulation with desired position and force
behavior, the relationships between the desired position (xr ) and force (Fr ) trajec-
tories and the target impedance parameters (M, B, K , and K f ) demand a careful
design in compliance with the environment. Based on the assembly/manipulation
strategy as given in [5, 16], a tradeoff between the position control and force control
can be generated by varying the impedance parameters {M, B, K } and {K f }, which
are associated with the position and force, respectively.
For instance, in the case of a soft environment with a low stiffness K e , e.g., in the
air, the contact force F is negligible and the reference force trajectory can be selected
as Fr = 0. It follows that the desired generalized impedance (10.9) is simplified into a
position trajectory (xr ) defined by parameters M, B, and K . Thus, a position control
framework is established. On the contrary, in the presence of a rigid environment
with a large stiffness K e , the motion terms in the generalized impedance (10.9) are
neglected and the impedance control becomes a force control problem. Otherwise,
if a simultaneous position and force tracking is required, a tradeoff between the
position control and force control can be realized by adjusting the target impedance
parameters as detailed below.
10.3 ADSMGIC Scheme Design 219

At the steady state, the response of the closed-loop system (10.35) can be derived
as follows:

(K + K f K e )x = K xr + K f (Fr + K e xe ) (10.36)
⇒ K (x − xr ) = K f [Fr + K e (xe − x)] (10.37)
⇒ K (x − xr ) = K f (Fr − F) (10.38)
⇒ K (x − xr ) = K f {Fr + K e [(xe − xr ) − (x − xr )]}. (10.39)

In view of Eqs. (10.38) and (10.39), the following steady-state position and force
errors are calculated:
Fr + K e (xe − xr )
x =
ess (10.40)
Ke + K
Kf
K ss
essf = − e . (10.41)
Kf x

It is deduced that if K e is relatively small and K >> K f is assigned, then the

steady-state position error ess ss
x is relatively small and the steady-state force error e f
is relatively large. Thus, the position control is emphasized. In contrast, if K e is
relatively large and K << K f is assigned, then the force control is emphasized and
the desired force is tracked more precisely than the position. Therefore, a compromise
between the position and force tracking control can be realized by adjusting the
relative values of K and K f .
Additionally, other target impedance parameters (M and B) can be chosen to
produce a required transient response for the closed-loop system (10.35).

10.4 Experimental Setup and Controller Setup

In this section, the experimental setup is presented and the controller parameters are
selected for the subsequent experimental investigations.

10.4.1 Experimental Setup

The experimental setup of a prototype piezoelectric microgripper system is depicted

in Fig. 10.3. The gripper is composed of two four-layer piezoelectric bimorphs with
brass reinforced inner layers. Each actuator has the dimension of 26 × 5 × 0.86 mm3
and is driven by a high-voltage amplifier (model: EPA-104, from Piezo Systems,
Inc.). The end-point position of each gripper arm is measured by a laser displace-
ment sensor (model: LK-H055, from Keyence Corp.). The laser sensor provides a
220 10 Adaptive Impedance Control of Piezoelectric Microgripper

Fig. 10.3 Experimental

setup of a piezoelectric
microgripper system for a
microgear grasping. The
close-up view shows a
snapshot of the
microgripping experiment

resolution of 25 nm within a measuring range of 20 mm around the reference distance

of 50 mm, which endows a large free space for micromanipulation/microassembly
tasks. A digital microscope (magnification rate: 200x) is mounted above the micro-
gripper for the assembly process monitoring. In addition, a National Instruments
(NI) cRIO-9075 real-time controller (from National Instruments Corp.) equipped
with NI-9263 D/A module and NI-9870 A/D module is adopted to produce the
excitation signals and acquire the sensor readings, respectively. The NI cRIO-9075
combines a real-time processor and a reconfigurable field-programmable gate array
(FPGA) within the same chassis, which is connected to a computer via the Ethernet
TM
port for communication. Moreover, LabVIEW software is employed to implement
a real-time control of the microgripper system.
The microobject to be grasped is a plastic microgear with the diameter of 1450 µm,
as shown in the close-up view in Fig. 10.3. Due to the symmetry of the microgripper,
only the right gripper arm is employed to demonstrate the effectiveness of the reported
control scheme, whereas the left one acts as a passive arm in the experiments.

10.4.2 Force Observer Design

Regarding the force sensing issue, diverse methods have been presented to generate
the force signal [36, 37], and various types of force sensors have been used to detect
the contact force in piezoelectric cantilevers. For example, the strain-gage sensors
are widely adopted in the piezoelectric microgripper [12, 29]. Nevertheless, such
sensors are inclined to produce noisy signal output, and it is difficult to integrate a
suitable sensor into the microgripper.
10.4 Experimental Setup and Controller Setup 221

In view of the limitation of force sensors, the force estimation approach has been
studied extensively. For instance, a linear force estimator is reported in [9] based
on a linear dynamics model of the system. However, this linear estimator offers
a low accuracy due to the influences of piezoelectric hysteresis and creep effects.
Later, a nonlinear force estimator was presented in [23, 24], which was constructed
by modeling the nonlinear hysteresis and creep effects. Nevertheless, the modeling
processes are time-consuming and model errors always exist. Hence, a more effective
force estimation method for a piezoelectric microgripper has been developed in recent
work [33]. It is implemented without modeling the complicated hysteresis and drift
effects. This force observer is employed here to reconstruct the force signal. A brief
description is given below, and more details can be found in [33], where a fourth-
order model is identified and used in the force estimation. Alternatively, a simple
second-order dynamic model is employed in this work.
Basically, two position signals x(t) and x1 (t), i.e., with and without the force
applied, are required to estimate the force signal. Specifically, the force that is applied
on the microobject can be estimated in the time domain:

dp 1
F̂(t) = [u(t) − u(t − T )] − L −1 {G −1 (s)}[x(t) − x1 (t)] (10.42)
sp sp

where u(t) and u(t − T ) denote the excitation voltages at the time instances t and
t −T , respectively. x is the position output under the actions of both excitation voltage
u and applied force F, whereas x1 denotes the position output excited by the voltage
u alone. Additionally, L −1 represents the inverse Laplace transform operator, and
G(s) with unity DC gain is the plant model under the excitation of input voltage.
In addition, d p and s p are the piezoelectric constant and elastic constant of the
piezoelectric bimorph, respectively.
When an excitation voltage u is applied to the high-voltage amplifier with F = 0,
the plant model G p is identified below by the frequency response method.

1.1292 × 108
G p (s) = . (10.43)
s 2 + 126.8 s + 5.9430 × 106

The piezoelectric constant d p is calculated as the DC gain of G p , i.e., d p =

19.0 µm/V. Then, the plant model G is derived as G = G p /d p . In addition, by
hanging a known weight and recording the induced displacement, the elastic con-
stant of the piezo-bimorph is determined as s p = 0.46 µm/mN.
To verify the performance of the force observer, an input voltage signal as shown
in Fig. 10.4a is applied to the voltage amplifier to drive the gripper arm while without
the external force applied. The position response is illustrated in Fig. 10.4b, and the
force observer output is shown in Fig. 10.4d. It is seen that the force is accurately
estimated with a RMSE of 0.25 mN. In addition, Fig. 10.4c illustrates an open-loop
hysteresis loop of the piezoelectric gripper arm, which poses a challenge to the precise
interaction control.
222 10 Adaptive Impedance Control of Piezoelectric Microgripper

(a) (b)
1 20
Input voltage (V)

0.8

Position (μm)
15
0.6
10
0.4

0.2 5

0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)
(c) (d) 1
20

0.5
Position (μm)

15
Force (mN)
10 0

5 −0.5
Estimated Desired
0
−1
0 0.2 0.4 0.6 0.8 1 0 2 4 6 8 10
Input voltage (V) Time (s)

Fig. 10.4 Force estimation result of the force observer without applied force. a Input voltage.
b Output position. c Position-voltage hysteresis loop. d Force signal

Using the force signal provided by the force observer, experiments on interaction
control of the microgripper for micromanipulation/microassembly are conducted by
grasping a microgear in the subsequent section.

10.4.3 Controller Setup

10.4.3.1 Trajectory Planning

Initially, the microgear to be grasped is in contact with the gripper arms. To realize
an interaction control, a position trajectory xr and a corresponding force trajectory
Fr need to be planned concurrently.
By applying the excitation voltage as given in Fig. 10.5a accompanied with an
interaction force to the gripper arm, the resulted position and force signals are
acquired as shown in Fig. 10.5b, c (solid lines). Based on the actual position and
force outputs, the trajectories are planned as shown in Fig. 10.5b, c (dashed lines).
Figure 10.5d depicts the errors between the planned position and force trajectories
and the actual ones, respectively. The trajectories are planned to realize a typical
grasp-hold-release operation in microassembly tasks.
10.4 Experimental Setup and Controller Setup 223

(a) (b)
1 10
Input voltage (V)

0.8 8

Position (μm)
0.6 6

0.4 4

0.2 2
0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)

(c) 25 (d) 2
20

Planning errors
1
Force (mN)

15

10 0

5
−1 Position (μm)
0 Experiment
Reference Force (mN)
−5 −2
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)

Fig. 10.5 Force estimation result of the force observer with applied force. a Input voltage. b Output
position. c Signal of the force applied to microobject. d Errors between the planned position/force
trajectories and the actual ones

10.4.3.2 Controller Parameters Setup

Concerning the dynamics model parameters, the mass of one piezoelectric bimorph
(m̂) is chosen as the nominal value. Moreover, by comparing model (10.1) with the
inverse Laplace transform of the plant model G p as shown in Eq. (10.43), other
parameters (b̂, k̂, and d̂) are obtained as shown in Table 10.1.
Because the left gripper arm is passive, this arm together with the microgear to
be grasped act as the environment in this work. In view of the compliance of the
bimorph indicated by s p , the stiffness is calculated as 2.17 × 103 N/m. Considering
that the rigidity of the plastic microgear is much higher than that of the gripper arm,
the environment stiffness of K e ≈ 2 × 103 N/m is assumed. Other target parameters
in the desired impedance (10.9) are selected as: M = 1, B = 2ζ ωn , and K = ωn2 ,
where ζ = 1 is set to achieve a critical damping. In addition, ωn = 1000π rad/s is
chosen to give the target parameters as shown in Table 10.1.
In view of the relationship (10.41) of the steady-state errors, K f should be chosen
to constrain the maximum force to ensure that the gripper arms and the microobject
will not be damaged. For illustration, K f is selected as K f = K5 . Thus, K > K f ,
which indicates that the position control is emphasized in the impedance control.
224 10 Adaptive Impedance Control of Piezoelectric Microgripper

Table 10.1 Parameters of Parameter Value

dynamics model, target
impedance, and ADSMGIC m̂ 1.8 × 10−3 kg
controller b̂ 126.8 Ns/µm
k̂ 5.943 × 106 N/µm
d̂ 1.129 × 108 µm/V
M 1 kg
B 6.2832 × 103 Ns/m
K 9.8696 × 106 N/m
Kf 2000 mN/µm
KP 3.1416 × 105
KI 9.8696 × 106
KD 2500
K4 2000
K5 0.5
KS 3000
γ 5
δ 105

In addition, with a sampling time T = 0.02 s, the discrete-time dynamics model

and the target impedance parameters are calculated by Eqs. (10.5) and (10.11),
respectively.

10.5 Experimental Results and Discussion

Experimental studies on the prototype microassembly system is carried out to verify

the effectiveness of the developed interaction control scheme.

10.5.1 Interaction Control Results

Regarding the developed ADSMGIC controller, the control gains K P , K I , K D , and

K 4 are shown in Table 10.1, which are calculated by Eqs. (10.14)–(10.20). Other
controller parameters are tuned by several trials as given in Table 10.1. To investigate
the effects of the parameter tuning, three case studies are carried out below.
10.5 Experimental Results and Discussion 225

(a) 10
(b) 1

Position error (μm)

Position (μm)
0.5
5 0
Reference
−0.5
Actual
0
−1
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)
(c) 30 (d) 2

Force error (mN)

1
Force (mN)

20
0
10
Reference −1
Actual
0
−2
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)
5
(e) (f) 1
x 10
Control input (V)

1
Sliding function

0.5

0.5 0

−0.5
0
−1
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)
(g) 5000 (h) 0.98
Switching gain Ks

Control input (V)

4500 0.96

4000 0.94

3500 0.92

3000 0.9
0 2 4 6 8 10 6 6.5 7
Time (s) Time (s)

Fig. 10.6 Tracking results of ADSMGIC with fixed control gain K S = 3800. a Position trajectory.
b Position error. c Force trajectory. d Force error. e Control input. f Sliding function. g Switching
gain. h Close-up view of control input

10.5.1.1 Case Study I

First, the effectiveness of the adaptive switching gain is examined. For the planned
trajectories, the tracking results of the control algorithm with a fixed switching gain
K S = 3800 and an adaptive switching gain K S,k are shown in Figs. 10.6 and 10.7,
respectively. The position and force RMSEs of the two controllers are summarized
in Table 10.2. The position and force relative RMSEs are calculated as the ratios
of RMSEs to the peak-to-peak magnitudes of the position and force trajectories,
226 10 Adaptive Impedance Control of Piezoelectric Microgripper

(a) 10
(b) 1

Position error (μm)

Position (μm)
0.5
5 0
Reference −0.5
Actual
0
−1
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)
(c) 30 (d) 2

Force error (mN)

Force (mN)

1
20
0
10
Reference −1
Actual
0
−2
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)
5
(e) (f) 1
x 10
Control input (V)

1
Sliding function

0.5

0.5 0

−0.5
0
−1
0 2 4 6 8 10 0 2 4 6 8 10
Time (s) Time (s)
(g) 5000 (h) 0.98
s

Control input (V)

Switching gain K

4500 0.96

4000 0.94

3500 0.92

3000 0.9
0 2 4 6 8 10 6 6.5 7
Time (s) Time (s)

Fig. 10.7 Tracking results of ADSMGIC with adaptive control gain K S,k . a Position trajectory.
b Position error. c Force trajectory. d Force error. e Control input. f Sliding function. g Switching
gain. h Close-up view of control input

respectively. In addition, Figs. 10.6f and 10.7f indicate that the sliding function is
well constrained by the boundary layer parameter δ = 105 in both cases.
As compared with the fixed gain, the adaptive gain produces an improvement on
the position and force tracking accuracy. The close-up view of the control input as
shown in Fig. 10.6h indicates that the fixed switching gain in Fig. 10.6g causes oscil-
lation phenomenon. On the contrary, the oscillation is eliminated by the employment
of adaptive gain in Fig. 10.7g as implied by Fig. 10.7h. Hence, the dominant role of
the adaptive law for the switching gain is to alleviate the chattering phenomenon of
ADSMGIC control action.
10.5 Experimental Results and Discussion 227

Table 10.2 Interaction control results of ADSMGIC for three case studies
Position error Force error
Case study Control option RMSE (µm) Relative (%) RMSE (µm) Relative (%)
I Fixed gain K S 0.252 2.80 0.88 3.75
Adaptive gain K S 0.250 2.77 0.86 3.66
II Control gain K 5 = 0.43 0.30 3.33 0.77 3.28
III Impedance parameter 0.29 3.22 0.78 3.32
Kf =0

Fig. 10.8 Error bars of 6.5

position and force tracking Position error
6 Force error
relative RMSEs versus the
control gain K 5 5.5
Relative RMSE (%)

4.5

3.5

2.5

1.5
0 0.2 0.4 0.6 0.8 1
Control gain K
5

10.5.1.2 Case Study II

Next, to demonstrate the influence of the control gain on the interaction control, the
tracking performance is tested as the gain K 5 varies from 0.1 to 0.9. For each K 5
value, the experiment is conducted five times. The error bars of the position and force
tracking results are depicted in Fig. 10.8.
It is found that the position and force errors vary in opposite tendencies as the
control gain K 5 changes. Specifically, as K 5 increases from 0.1 to 0.9, the position
error is reduced whereas the force error is increased gradually. Hence, a good position
tracking is obtained at the sacrifice of force error, and vice versa. It is further observed
that a balance between the position and force tracking is achieved with the gain K 5 =
0.43. It leads to almost identical relative RMSEs (3.3 %) for both trajectories, as
shown in Table 10.2.
The experimental results indicate that the position and force tracking accuracy
can be compromised by adjusting the control gain K 5 for the proposed ADSMGIC
scheme.
228 10 Adaptive Impedance Control of Piezoelectric Microgripper

Fig. 10.9 Error bars of 10

position and force tracking Position error
relative RMSEs versus the 9 Force error
target impedance parameter
Kf 8

Relative RMSE (%)

7

2
0 2000 4000 6000 8000 10000
Kf

10.5.1.3 Case Study III

To illustrate the effect of selection of the target impedance parameters K and K f on

the tracking performance, the control scheme with K f varying from 0 to 104 is tested.
For each K f value, the experiment has been repeated by five times and experimental
data are obtained. The error bars of the relative RMSEs for both position and force
tracking results versus K f are plotted in Fig. 10.9.
It is observed that the magnitude of the relative RMSE of the position is lower
than that of the force. The reason arises from the fact that the relationship of K > K f
holds. Thus, the position control is more emphasized in the impedance control as
predicted by Eqs. (10.40) and (10.41). As a result, a relative small position tracking
error is achieved. Moreover, it is found that as K f increases from 0 to 6000, the
position error is reduced gradually. Whereas the force error is increased as the rising
of K f , which is consistent with the tendency as evaluated by Eq. (10.41).
In addition, Fig. 10.9 also illustrates how to select a balance point of the position
and force control by adjusting the target impedance parameter. It is evident that in the
case of K f = 0, a better balance between the position and force tracking accuracy is
achieved in that the ADSMGIC produces similar position and force relative RMSEs,
as shown in Table 10.2.
Furthermore, it is found that a too large K f induces oscillation of the control
action. For instance, as K f increases greater than 8000, a moderate oscillation in
the interaction control is produced. The oscillation deteriorates both the position and
force tracking accuracy as depicted in Fig. 10.9.
In the aforementioned case studies, the effects of the parameters are intuitively
illustrated to have a better understanding of the controller performance under the tun-
ing of control parameters and target parameters. These processes provide a guideline
for the extension of the control scheme to other applications.
10.5 Experimental Results and Discussion 229

10.5.2 Discussion on Control Performance

The fact that the controller works well for different values of the target impedance
parameter reveals that the control scheme is robust to the target stiffness. In addition,
the impedance parameters B and K are determined by the parameter ωn . Regarding
the selection of ωn , additional experimental testings reveal that, the larger the ωn
value, the smaller the achievable position and force tracking errors for the control
scheme.
Generally, in order to overcome the disturbances and uncertainties of the system
model, the upper bounds of the perturbation terms need to be acquired in SMC robust
control [16], which complicates the control design process. Alternatively, only an
upper bound K S [see Eq. (10.31)], i.e., the switching control gain, of perturbation
estimation error is required by the reported robust control framework. Moreover, by
adopting the adaptive rule (10.33), the switching gain can be set as zero initially
because it is automatically adjusted online. Thus, the burden of specifying the per-
turbation bounds is relieved in the presented approach. On the other hand, because
the perturbation estimation error is usually much lower than the actual perturbation
and the switching gain is adjusted to adapt to the sliding function, a much smaller
control gain is resulted as compared with conventional SMC algorithms. As a con-
sequence, the reported control framework achieves an improvement on the tracking
performance by suppressing the chattering phenomenon.
It is notable that the lumped perturbation term Pk is calculated by the one-step
delayed estimation (10.6). Thus, a good calculation accuracy can be obtained by
employing a high sampling rate T and a low variation rate of Pk . Due to the controller
hardware limit, a sampling time of T = 0.02 s is implemented for the prototype
microgripper in this work. Even so, a reasonable accuracy for both the position and
force tracking has been achieved as illustrated by the experimental results. With
such a sampling rate, the conventional continuous-time impedance control may even
not work if directly deployed to the sampled-data system. Hence, the experimental
results confirm the feasibility of the digital control scheme. The tracking errors can
be further reduced by adopting a smaller sampling time.
Additionally, the gain parameters of the ADSMGIC controller are not optimally
designed in the preceding experiments. The performance of the control scheme may
be improved to further enhance the tracking accuracy. Moreover, the precise micro-
manipulation is affected by the external disturbances including external vibrations,
temperature, etc. The fact that a precise position and force control can be simultane-
ously achieved demonstrates the robustness of the control scheme in the presence of
noise and temperature disturbances. Further work includes the investigation of grasp-
ing more microobjects to accomplish complex microassembly tasks. The adhesive
forces will also be taken into account during the assembly of smaller microscopic
objects.
230 10 Adaptive Impedance Control of Piezoelectric Microgripper

10.6 Chapter Summary

This chapter reports on a control scheme for position/force control of a piezoelectric-

bimorph microgripper dedicated to micromanipulation and microassembly. The
presented digital sliding mode-based scheme is easy to implement in that only a
second-order dynamic model is used, whereas neither the state estimation nor hys-
teresis modeling is needed. The stability of the control system has been proved
theoretically and the effectiveness of the scheme has been validated by a series
of experimental studies on typical grasp operation. The superiority of the adaptive
switching gain over fixed one in terms of chattering suppression has been confirmed.
Moreover, results indicate that the scheme is capable of achieving satisfactory inter-
action control results. A tradeoff between the position and force tracking accuracy
can be realized by varying the relative values of control gains or target impedance
parameters. As a future work, adaptive rules for the model parameters may be devised
to eliminate the procedure of dynamic model identification. Moreover, the extension
of the proposed methodology to interaction control of microgrippers driven by other
types of actuators will also be explored for micromanipulation and microassem-
bly tasks.

References

1. Almeida, F., Lopes, A., Abreu, P.: Force-impedance control: A new control strategy of robotic
manipulators. In: Kaynak, O., Tosunoglu, S., Ang, M. (eds.) Recent Advances in Mechatronics,
pp. 126–137. Springer, Singapore (1999)
2. Bargiel, S., Rabenorosoa, K., Clévy, C., Gorecki, C., Lutz, P.: Towards micro-assembly of
hybrid MOEMS components on a reconfigurable silicon free-space micro-optical bench. J.
Micromech. Microeng. 20(4), 045012 (2010)
3. Beyeler, F., Neild, A., Oberti, S., Bell, D.J., Sun, Y., Dual, J., Nelson, B.J.: Monolithically fab-
ricated microgripper with integrated force sensor for manipulating microobjects and biological
cells aligned in an ultrasonic field. J. Microelectromech. Syst. 16(1), 7–15 (2007)
4. Bonitz, R.G., Hsia, T.C.: Internal force-based impedance control for cooperating manipulators.
IEEE Trans. Robot. Autom. 12(1), 78–89 (1996)
5. Chan, S.P., Liaw, H.C.: Generalized impedance control of robot for assembly tasks requiring
compliant manipulation. IEEE Trans. Ind. Electron. 43(4), 453–461 (1996)
6. Chonan, S., Jiang, Z.W., Koseki, M.: Soft-handling gripper driven by piezoceramic bimorph
strips. Smart Mater. Struct. 5, 407–414 (1996)
7. Elmali, H., Olgac, N.: Implementation of sliding mode control with perturbation estimation
(SMCPE). IEEE Trans. Control Syst. Technol. 4(1), 79–85 (1996)
8. Furuta, K.: Sliding mode control of a discrete system. Syst. Control Lett. 14(2), 145–152 (1990)
9. Haddab, Y., Chaillet, N., Bourjault, A.: A microgripper using smart piezoelectric actuators.
In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems,
pp. 659–664. Takamatsu, Japan (2000)
10. Hogan, N.: Stable execution of contact tasks using impedance control. In: Proceedings of the
IEEE International Conference on Robotics and Automation, pp. 1047–1054. Raleigh, North
Carolina, USA (1987)
11. Huang, H.B., Sun, D., Mills, J.K., Cheng, S.H.: Robotic cell injection system with position and
force control: toward automatic batch biomanipulation. IEEE Trans. Robot. 25(3), 727–737
(2009)
References 231

12. Huang, X., Cai, J., Wang, M., Lv, X.: A piezoelectric bimorph micro-gripper with micro-force
sensing. In: IEEE International Conference on Information Acquisition, pp. 145–149. Hong
Kong and Macau, China (2005)
13. Jung, S., Hsia, T.C., Bonitz, R.G.: Force tracking impedance control for robot manipulators
with an unknown environment: theory, simulation, and experiment. Int. J. Robot. Res. 20(9),
765–774 (2001)
14. Jung, S., Hsia, T.C., Bonitz, R.G.: Force tracking impedance control of robot manipulators
under unknown environment. IEEE Trans. Control Syst. Technol. 12(3), 474–483 (2004)
15. Kim, K., Liu, X., Zhang, Y., Sun, Y.: Nanonewton force-controlled manipulation of biological
cells using a monolithic MEMS microgripper with two-axis force feedback. J. Micromech.
Microeng. 18(5), 055013 (2008)
16. Liaw, H.C., Shirinzadeh, B.: Robust generalised impedance control of piezo-actuated flexure-
based four-bar mechanisms for micro/nano manipulation. Sens. Actuator A-Phys. 148(2),
443–453 (2008)
17. Lu, W.S., Meng, Q.H.: Impedance control with adaptation for robotic manipulations. IEEE
Trans. Robot. Autom. 7(3), 408–415 (1991)
18. Lu, Z., Chen, P.C.Y., Lin, W.: Force sensing and control in micromanipulation. IEEE Trans.
Syst. Man Cybern. Part C, Appl. Rev. 36(6), 713–724 (2006)
19. Lu, Z., Kawamura, S., Goldenberg, A.A.: An approach to sliding-mode based control. IEEE
Trans. Robot. Autom. 11(5), 754–759 (1995)
20. Menciassi, A., Eisinberg, A., Carrozza, M.C., Dario, P.: Force sensing microinstrument for
measuring tissue properties and pulse in microsurgery. IEEE/ASME Trans. Mechatron. 8(1),
10–17 (2003)
21. Monsees, G., Scherpen, J.M.A.: Adaptive switching gain for a discrete-time sliding mode
controller. Int. J. Control 75(4), 242–251 (2002)
22. Qiu, L., Cui, Y., Feng, F.: Design of a new micro-gripper based on piezoelectric bimorphs.
Appl. Mech. Mater. 101–102, 173–177 (2011)
23. Rakotondrabe, M., Haddab, Y., Lutz, P.: Nonlinear modeling and estimation of force in a piezo-
electric cantilever. In: Proceedings of the IEEE/ASME International Conference on Advanced
Intelligent Mechatronics, pp. 1–6. Zurich, Switzerland (2007)
24. Rakotondrabe, M., Ivan, I.A.: Development and force/position control of a new hybrid thermo-
piezoelectric microgripper dedicated to micromanipulation tasks. IEEE Trans. Autom. Sci.
Eng. 8(4), 824–834 (2011)
25. Reddy, A.N., Maheshwari, N., Sahu, D.K., Ananthasuresh, G.K.: Miniature compliant grippers
with vision-based force sensing. IEEE Trans. Robot. 26(5), 867–877 (2010)
26. Sariola, V., Jaaskelainen, M., Zhou, Q.: Hybrid microassembly combining robotics and water
droplet self-alignment. IEEE Trans. Robot. 26(6), 965–977 (2010)
27. Sarpturk, S., Istefanopulos, Y., Kaynak, O.: On the stability of discrete-time sliding mode
control systems. IEEE Trans. Autom. Control 32(10), 930–932 (1987)
28. Schutter, J.D., Bruyninckx, H., Zhu, W.H., Spong, M.W.: Force control: A bird’s eye view. In:
Siciliano, B. (ed.) Control Problems in Robotics and Automation: Future Directions, pp. 1–17.
Springer, Berlin (1998)
29. Seki, H.: Modeling and impedance control of a piezoelectric bimorph microgripper. In:
Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems,
vol. 2, pp. 958–965. Raleigh, USA (1992)
30. Seraji, H., Colbaugh, R.: Force tracking in impedance control. Int. J. Robot. Res. 16(1), 97–117
(1997)
31. Tarokh, M.: A discrete-time adaptive control scheme for robot manipulators. J. Robot. Syst.
7(2), 145–166 (1990)
32. Wu, J., Shieh, L.S., Zhang, Y., Song, G.: Digital controller design for Bouc-Wen model with
high-order hysteretic nonlinearities through approximated scalar sign function. Int. J. Sys. Sci.
42(10), 1581–1599 (2011)
33. Xu, Q.: A new method of force estimation in piezoelectric cantilever-based microgrip-
per. In: Proceedings of the IEEE/ASME International Conference on Advanced Intelligent
Mechatronics, pp. 574–579. Kaohsiung, Taiwan (2012)
232 10 Adaptive Impedance Control of Piezoelectric Microgripper

34. Xu, Q.: Identification and compensation of piezoelectric hysteresis without modeling hysteresis
inverse. IEEE Trans. Ind. Electron. 60(9), 3927–3937 (2013)
35. Xu, Q., Li, Y.: Model predictive discrete-time sliding mode control of a nanopositioning
piezostage without modeling hysteresis. IEEE Trans. Control Syst. Technol. 20(4), 983–994
(2012)
36. Zhang, L., Dong, J.: High-rate tunable ultrasonic force regulated nanomachining lithography
with an atomic force microscope. Nanotechnology 23(8), 085303 (2012)
37. Zhang, Y., Zou, Q.: High-speed force load in force measurement in liquid using scanning probe
microscope. Rev. Sci. Instrum. 83(1), 013707 (2012)
Chapter 11
Position/Force Switching Control
of a Miniature Gripper

This chapter presents the design and implementation of combined position and force
control of a piezoelectrically actuated compliant gripper. The challenge of achieving
a smooth transition between the position and force switching control is addressed
by means of an incremental control scheme. Precision control under the influence
of hysteretic nonlinearity is guaranteed by a digital sliding-mode control (DSMC)
algorithm. The scheme is implemented with an FPGA platform. Experimental inves-
tigations are undertaken to verify the effectiveness of the gripper system by executing
grasp-hold-release operations of a micro copper wire.

11.1 Introduction

As a key device in microhandling system, miniature gripper plays an important role

dedicated to micromanipulation and microassembly tasks. Typically, the gripper is
employed to execute an automated grasp-hold-release operation for such applications
as material characterization, biological sample pick-and-place, and micro-object
assembly. Diverse miniature grippers have been reported using different actuation
principles [15]. In particular, piezoelectric actuators are prevailing in many indus-
trial applications owing to their ultrahigh position resolution and quick response
capabilities [13, 23].
To grasp an object gently without impairing the object and the gripper, it is impor-
tant to regulate both the relative position and touch force between the gripper and
object [14, 19]. In addition, a gripper with integrated position and force sensing is
able to characterize the stiffness and viscosity properties of various materials, such
as biological tissues [2]. However, it is challenging to implement a position and force
control for a PSA-driven gripper.
One of the major challenges stems from the combination of the position and force
control. Intuitively, a position/force switching control can be adopted to adjust the
gripper tip position and gripping force in an alternate manner [17]. The concerned
© Springer International Publishing Switzerland 2016 233
Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2_11
234 11 Position/Force Switching Control of a Miniature Gripper

dominant issue lies in how to achieve a smooth transition when different control
variables are switched [12]. Similar problem resides in the hybrid position/force
control scheme [4, 16], which calls for a stable switch to accomplish the position
and force control in different subspaces. Even though the switching control problem
can be avoided by employing the compliance [10] or impedance control framework
[8, 20], care must be taken for the position and force trajectory planning to ensure
a desired performance. To achieve a smooth transition of the position and force
control, an intermediate brake mode can be added between the two control modes [6,
18]. Furthermore, to accomplish a rapid intermediate control and to reduce the force
overshoot in the initial contact, an optimal velocity profile can be used [1]. However,
an extra control mode is required, whose execution increases the overall operation
time of the gripper. More recently, fuzzy control has been proposed to smooth the
transition [5, 11]. However, the membership functions have to be readjusted to adapt
to different objects, which complicates the practical implementation process. Hence,
a simple yet efficient approach is desirable to realize a smooth transition between
the position and force control. Another challenge is organized from the piezoelectric
actuator which exhibits strong nonlinearities in terms of hysteresis and drift effects
once driven by a voltage source [9].
To this end, a scheme of incremental control is reported in this chapter to achieve
a smooth transition of the position/force switching control for the gripper system.
Specifically, in the closing and opening phases, an incremental DSMC is devised to
cater for the position control. During the contact phase, an incremental-type digital
PID force control is adopted. The incremental controller solves the change value
of the control action in each time step. At the moment of switching between two
consecutive control phases, the controller adopts the control action in the previous
phase as the base value for the next one. In this way, a smooth transition between two
adjacent phases is easily achieved. The proposed control algorithm is deployed to an
FPGA-based digital control platform. The feasibility of the scheme is examined by
a typical grasp-hold-release operation of a micro copper wire through experimental
studies. Moreover, its superiority over the conventional approach is demonstrated
via comparison studies.

11.2 Experimental Setup and Calibration

11.2.1 Working Principle of the Gripper

The mechanical structure of the employed gripper is shown in Fig. 11.1. The com-
pliant gripper consists of a bridge-type displacement amplifier, a rotary bearing
mechanism, and two gripper arms/tips. The stroke (2S) of PSA is magnified by
the bridge-type amplifier, which provides an amplification ratio of A1 . Figure 11.1b
exhibits a pseudo-rigid-body (PRB) model of the amplifier. The amplified displace-
ment (A1 S) at the output end of the amplifier is transmitted to the two gripper arms
11.2 Experimental Setup and Calibration 235

Fig. 11.1 Schematic diagram of a miniature gripper. a The gripper structure. b Bridge-type
displacement amplifier

through two sets of rotary guiding flexures, respectively. With a distance a away from
the bearing mechanism, the displacement of each gripper tip is further magnified by
a factor of A2 = 1 + 2a/b, where b is the diameter of the rotary bearing mechanism.
Hence, the overall displacement at each tip of the gripper can be expressed as A1 A2 S,
and the gripping range of the gripper is 2 A1 A2 S. In contrast to a parallelogram-based
gripper, the employed one reduces the number of guiding flexures and simplifies the
gripper structure, which indicates a reduction of the fabrication cost.
To measure the displacement of the gripper tips and the contact force during the
gripping operation, a vision-based sensing approach can be employed [14]. Never-
theless, the relative low sensing resolution and sampling rate are the major disad-
vantages of such approach. Recently, it has been shown that strain gages provide
a straightforward way to measure the position and gripping force of a PSA-driven
monolithic gripper [19]. In this work, two groups of strain gages are adopted and
surface-bonded on the flexures as the position and force sensors, respectively. Due to
a symmetric architecture, the measurement for a single gripper arm is sufficient. To
generate a high sensitivity and high signal-to-noise ratio (SNR) for the measurement,
the strain-gage position sensor is glued on the bridge amplifier and the strain-gage
force sensor is attached on the base end of the gripper tip, as indicated in Fig. 11.1. A
preloading screw is used to adjust the amplification ratio A1 of the bridge amplifier
236 11 Position/Force Switching Control of a Miniature Gripper

Table 11.1 Main parameters of a compliant gripper

Parameter Symbol Value
Length of arm a 62.0 mm
Diameter of rotary bearing b 60.0 mm
Length of guiding flexure R 22.0 mm
Width of guiding flexure t1 0.4 mm
Length of amplifier arm l1 11.5 mm
Length of amplifier flexure l2 3.0 mm
Width of amplifier flexure t2 0.4 mm
Incline angle of amplifier θ 5.0◦
Thickness of material h 5.0 mm

and the initial gap between the two gripper tips. The entire gripper can be mounted
on a support base through the central fixing hole of the rotary bearing.
The structural parameters are designed as shown in Table 11.1 to enable a gripping
range of 200 µm using a PSA of 14.5-µm stroke.

11.2.2 Experimental Setup

A prototype gripper is shown in Fig. 11.2, which is fabricated from a piece of Al-
7075 material by the wire-EDM process. A PSA (model: TS18-H5-202, from Piezo

Fig. 11.2 Experimental setup of a prototype gripper system

11.2 Experimental Setup and Calibration 237

Systems, Inc.) is chosen to actuate the gripper. The PSA is driven through a
commercial high-voltage amplifier to deliver a maximal stroke of 14.5 µm. For
the measurement of the gripper tip position, four strain gages (model: SGD-3/350-
DY13, from Omega Engineering Ltd.) are employed to construct a full-bridge circuit.
Due to the space limit, another strain gage (model: SGD-3/350-LY13, from Omega
Engineering Ltd.) is adopted to form a quarter bridge for the gripping force measure-
ment. In addition, a Natural Instruments (NI) cRIO-9022 real-time (RT) controller
combined with cRIO-9118 reconfigurable chassis (from National Instruments Corp.)
is adopted to implement the control algorithm. The NI-9263 D/A and NI-9237 A/D
modules are adopted for producing the analog excitation signals and acquiring the
strain-gage sensor signals, respectively. The NI cRIO-9118 chassis contains a FPGA
core, and the associated cRIO-9022 RT controller communicates with a personal
computer (PC) via the 10/100 Mbit/s Ethernet port.
For the calibration of strain-gage position sensor, the tip position of the gripper
is also measured by a laser displacement sensor (model: LK-H055, from Keyence
Corp.), which provides a 25-nm resolution within a 20-mm measuring range. Addi-
tionally, a digital microscope (model: AM2011 Dino-Lite Basic, from AnMo Elec-
tronics Corp.) with a magnification ratio of 200 is employed to monitor the gripper
tips and grasped objects during the gripping process.
TM
The control algorithms are programmed using LabVIEW software to realize a
deterministic RT control for the gripper system. Specifically, the algorithm is devel-
oped and debugged in PC with LabVIEW. Afterward, the program is compiled to
generate FPGA codes automatically. When the compilation is completed, the codes
are downloaded into cRIO platform through the Ethernet connection. Once down-
loaded completely, the cRIO system can work independently.

11.2.3 Calibration of Position and Force Sensors

The purpose of calibration is to determine the position and force gains so as to

convert the position and force strain-gage outputs into the position and force values,
respectively. The position and force gains can be derived by comparing the strain-gage
outputs with the known position and force values under the assumption that there is a
constant linear relationship between the strain-gage outputs and the measured values.
It has been shown that the strain-gage gains can be calibrated by using different input
signals such as sinusoidal and square waves [7]. In this work, the step signals are
adopted for the calibration of the position and force sensors.
To calibrate the strain-gage position sensor, a step signal with a 3-V amplitude
is used to drive the PSA. By comparing the strain-gage output and the laser sensor
readings, a gain of ηd = 185.541 µm/mV is derived to calibrate the position sen-
sor. The output of the two sensors are plotted in Fig. 11.3a. The root-mean-square
error (RMSE) of the strain-gage position sensor output with respect to the laser
sensor output is calculated as 0.154 µm. With a zero input, the noise signal of the
238 11 Position/Force Switching Control of a Miniature Gripper

(a) 40 (d) 30

Displacement (µm) 30

Force (mN)
20
20
10
10
Laser sensor
0 Strain gage 0

0 10 20 30 40 0 10 20 30 40
Time (s) Time (s)
(b) 0.5 (e) 2
Displacement (µm)

Force (mN)
0.25 1

0 0

−0.25 −1

−0.5 −2
0 2.5 5 7.5 10 0 2.5 5 7.5 10
Time (s) Time (s)
(c) 3000 (f) 4000
−0.016±0.345 0.002±1.501
3000
2000
Counts

Counts

2000
1000
1000

0 0
−0.5 −0.25 0 0.25 0.5 −2 −1 0 1 2
Displacement (µm) Force (mN)

Fig. 11.3 Calibration results of strain-gage position and force sensors. a The output of laser sensor
and strain-gage position sensor. b Strain-gage position sensor noise. c Histogram of strain-gage
position sensor noise. d The output of strain-gage force sensor. e Strain-gage force sensor noise.
f Histogram of strain-gage force sensor noise

strain-gage position sensor is acquired as shown in Fig. 11.3b and the histogram is
given in Fig. 11.3c, which reveals that the white noise follows a normal distribution
(solid curve). The limit of three standard deviations (3σ ) from the mean (−0.016 µm)
is obtained as 0.345 µm, which indicates that 99.7 % of the sensor readings fall within
the interval of −0.016 ± 0.345 µm. In addition, a slight drift is observed in Fig. 11.3a,
which is caused by the PSA with a constant voltage input.
Additionally, the strain-gage force sensor is calibrated by hanging a known weight
(25 mN) and acquiring the caused sensor output voltage. The force sensor gain is then
calculated as η f = 658.345 mN/mV, which is employed to convert the voltage into
the force value as plotted in Fig. 11.3d. In addition, Fig. 11.3e, f display the noise
and its histogram of the force sensor without a force exerted. It indicates a 3σ limit
11.2 Experimental Setup and Calibration 239

of 1.501 mN. Hence, 99.7 % of force sensor readings lie within the range of 0.002
± 1.501 mN around the mean value.

11.2.4 Gripping Range Testing

Using the strain-gage position sensor, the gripping range of the gripper is exam-
ined experimentally. By applying a sinusoidal signal of 0.5-Hz frequency and 10-V
amplitude to drive the PSA through the high-voltage amplifier, the gripping range
is tested as shown in Fig. 11.4. It is found that displacement of one gripper tip is
138 µm, indicating a gripping range of 276 µm for the gripper. As compared with
the simulation result (200 µm), the experimental result of the gripping range is larger,
which is caused by adjusting the preloading screw.

(a) 150
Displacement (µm)

100

50

0
0 2.5 5 7.5 10
Time (s)

(b) 150
Displacement (µm)

100

50

Laser sensor
0 Strain gage
0 2.5 5 7.5 10
Input voltage (V)

Fig. 11.4 Test result of gripping range by applying a 0.5-Hz, 10-V amplitude sinusoidal input.
a Time history results. b Hysteresis curves
240 11 Position/Force Switching Control of a Miniature Gripper

Figure 11.4b exhibits a significant hysteresis effect which calls for a control
technique for the precision control of the gripper tips. It is also observed that there
is a slight discrepancy between the strain-gage position sensor and laser sensor out-
puts. The discrepancy is mainly caused by different dynamic properties of the two
types of sensors. In addition, the results indicate that the strain-gage gain, which is
determined by using a step signal input, applies to the sinusoidal input as well. This
fact verifies the assumption that there is a constant linear relationship between the
strain-gage outputs and the measured values.

11.2.5 Frequency Response Testing

The dynamics characteristics of the gripper system are examined by means of

frequency response. Particularly, a sweep sine wave with a 0.1-V amplitude and
frequency ranging from 1 to 500 Hz is used to drive the PSA. The position response
is measured by the strain-gage position sensor. Spectral analysis is then conducted
to derive the frequency responses of the gripper, which are shown in Fig. 11.5. A res-
onant peak at the frequency of 244 Hz is observed. The resonance limits the usable
bandwidth of the system [22]. The reason lies in that the input signals with high-
frequency components will excite the resonance vibration. Hence, a high resonant
frequency is desirable to obtain a large control bandwidth.
Based on the experimental data, a fourth-order model of the plant is obtained.
Figure 11.5 indicates that the identified model well approximates the system dynam-
ics behavior up to 300 Hz. It is notable that a higher order model can be generated

60
Magnitude (dB)

40

20

0
0 1 2
10 10 10
Phase (degree)

−200

−400 Experiment
Identified model
−600
0 1 2
10 10 10
Frequency (Hz)

Fig. 11.5 Open-loop frequency response of the gripper system

11.2 Experimental Setup and Calibration 241

to better describe the dynamics at higher frequencies. In this work, a model of lower
order is adopted to make a compromise between the accuracy and complexity of the
plant model.

11.3 Position/Force Switching Control Scheme Design

11.3.1 Event-Based Switching Control Framework

For a precision microhandling task, the grasp process of the gripper can be classified
into three stages in terms of closing phase (I), contact phase (II), and opening phase
(III). Here, an event-based switching control strategy between the three phases is
employed to regulate the position and force alternately, which is depicted in Fig. 11.6.
The closing phase occurs before the contact of gripper tips and object, and open-
ing phase runs after the said contact procedure. In the closing and opening phases,
a precise position control of the gripper tips is sufficient. Depending on different
treatments of the contact phase, two control approaches can be developed in terms
of position control with force monitoring and position/force switching control.
The first method realizes only the position control in the closing and opening
phases. It applies a constant control voltage during the contact phase once a desired
force magnitude is attained. In contrast, the second method implements a position
control in the closing and opening phases and a force control in the contact phase,
respectively. Considering that only the position control is needed in the first approach,
any precision position control strategy can be employed. Regarding the position/force
switching control scheme, the system exhibits one input variable, i.e., the driving
voltage, and two output variables, i.e., the position and force. Because the three
phases adopt the same control input, the switching between different controllers may
cause instability and unwanted response of the position and force outputs. Therefore,

Fig. 11.6 An event-based Home position

switching control framework
of the gripper system with
three phases of operation
Clo tion
con e
on phas

Po
l
tro

sin con
si
III

I hase
gp
Po ening
Op
s i ti

tro
l

Force monitoring
or control
Contact phase
Opened Closed
II
242 11 Position/Force Switching Control of a Miniature Gripper

NI 9237 Position strain-gage

x
xr +
DSM High- Compliant
−
Switch u voltage gripper
NI 9263
fr criterion amplifier PSA
+ − PID
f
NI 9237 Force strain-gage
NI cRIO-9022 RT+9118 FPGA I/O modules Gripper prototype

Fig. 11.7 Block diagram of position/force switching control for the gripper system

it is essential to guarantee a stable and smooth transition between two consecutive

phases so that a satisfactory response is accomplished.
To achieve such goal, an incremental control framework is presented in this work.
Specifically, in both the closing and opening phases, an incremental DSMC position
control is proposed. During the contact phase, an incremental-type digital PID force
control is used. The incremental control algorithm only solves the change value of
the control action in each time step. During the switching between two consecutive
control phases, the incremental controller adopts the control action in the previous
phase as the base value for the next one. In this way, a smooth transition between the
two phases is accomplished. Moreover, a switching criterion of the position/force
control is assigned based on a threshold of the contact force.
The framework of the position/force switching control is described in Fig. 11.7.
In the following discussions, an incremental DSMC and an incremental PID control
are designed to regulate the position and contact force, respectively.

11.3.2 Incremental DSMC Position Controller

It is known that the dynamics model of the piezo-actuated system can be represented
as follows [3]:


n 
m
x(k) = ai x(k − i) + bi u(k − i) + d(k) (11.1)
i=1 i=0

where x(k) and u(k) represent the output position and input voltage at the time step
k, respectively. ai and bi are the model coefficients. In addition, d(k) describes the
lumped effect of piezoelectric hysteresis, drift, external force, parameter uncertain-
ties, and other model disturbances.
By expressing e(k) = x(k)−xr (k) as the position error, where xr (k) is the desired
position trajectory, a PI-type sliding function is defined below
11.3 Position/Force Switching Control Scheme Design 243

s(k) = λ P e(k) + λ I ε(k) (11.2)

k λ I are the positive gains. Additionally, the integral error is calculated

where λ P and
by ε(k) = i=1 e(i). Hence,

ε(k) = ε(k − 1) + e(k). (11.3)

The equivalent control action u eq (k) is solved from the relationship of s(k) = 0,
which gives

n 
m 
u eq (k) = −(b0−1 ) ai x(k − i) + bi u(k − i) + d̂(k) − d̃(k) − xr (k)
i=1 i=1
−1
−(λb0 ) λ I ε(k − 1) (11.4)

where λ = λ P + λ I , d̃(k) = d̂(k) − d(k) represents the estimation error of the

disturbance, and d̂(k) = d(k − 1) denotes the estimated value of the disturbance
term d(k). That is,


n 
m
d̂(k) = x(k − 1) − ai x(k − i − 1) − bi u(k − i − 1). (11.5)
i=1 i=0

It is notable that the disturbance estimation error can be expressed into the form:

d̃(k) = d(k − 1) − d(k) ≈ −ḋ(t)T (11.6)

where T is the sampling time.

Assume that ḋ(t) is bounded, i.e., |ḋ(t)| ≤ Λ. It is then derived that d̃(k) is also
bounded, i.e., |d̃(k)| ≤ ΛT .
Substituting Eq. (11.5) into Eq. (11.4) and assigning d̃(k) = 0 results in
 
n
u eq (k) = u(k − 1) − (b0−1 ) x(k − 1) − xr (k) + ai [x(k − i) − x(k − i − 1)]
i=1

m 
+ bi [u(k − i) − u(k − i − 1)] − (λb0 )−1 λ I ε(k − 1). (11.7)
i=1

The equivalent control (11.7) represents the control action for the case of perfect
disturbance estimation, i.e., d̃(k) = 0. It acts during the sliding phase in which the
position trajectory is maintained on the sliding hyperplane. Otherwise, if the initial
position of the system is not resident on the sliding hyperplane or a large estimation
error d̃(k) is caused, the trajectory cannot be kept on sliding hyperplane by the
stand-alone equivalent control. Thus, a nonlinear control action u sw (k) is required.
244 11 Position/Force Switching Control of a Miniature Gripper

Here, a nonlinear switching control action is employed to enhance the equivalent

control. The overall control action is given by:

u DSM (k) = u eq (k) + u sw (k) (11.8)

i.e.,
 
n
u DSM (k) = u(k − 1) − (b0−1 ) x(k − 1) − xr (k) + ai [x(k − 1) − x(k − i − 1)]
i=1

m 
+ bi [u(k − i) − u(k − i − 1)]
i=1
−(λb0 )−1 λ I ε(k − 1) − λ S b0−1 sgn{(s(k − 1)} (11.9)

where λ S represents a positive control gain and sgn{·} denotes the signum function.
To evaluate the stability of the control system, the relation (11.9) is substituted
into dynamics model (11.1). Then, some fundamental algebra operation gives


n 
m
x(k) = ai x(k − i) + b0 u(k) + bi u(k − i) + d(k)
i=1 i=1
= xr (k) − d̃(k) − (λ)−1 λ I ε(k − 1) − λ S sgn{(s(k − 1)}. (11.10)

Substituting Eq. (11.10) into the equation of sliding function (11.2) yields

s(k) = λ[x(k) − xr (k)] + λ I ε(k − 1)

= −λ[λ S sgn{s(k − 1)} + d̃(k)]. (11.11)

It is notable that the parameters λ and λ S are selected as positive values.

Assigning λ S = |d̃(k)| + δ where δ represents an arbitrary positive constant, the
following deduction can be drawn:

λδ ≤ |s(k)| ≤ 2λ|d̃(k)| + λδ. (11.12)

Recalling that d̃(k) is bounded by |d̃(k)| ≤ ΛT , it is deduced from Eq. (11.12)

that the discrete sliding mode in the vicinity of the sliding hyperplane s(k) = 0 is
attained in finite steps.

Remark 11.1 In practice, chattering phenomenon may be induced by the discon-

tinuous signum function sgn{·}. To alleviate the chattering effect, the technique
of boundary layer is adopted by using the saturation function instead of sgn{·} in
Eq. (11.9): 
sgn{s(k)} for |s(k)| > ε
sat{s(k)} = (11.13)
s(k)/ε for |s(k)| ≤ ε
11.3 Position/Force Switching Control Scheme Design 245

where the thickness parameter ε of the boundary layer guarantees that the sliding
function is bounded within ±ε. In practice, a compromise between control error and
chattering is considered in the selection of parameter ε.

It is notable that the presented control scheme is easy to implement in that a single
linear plant model is needed. Whereas neither a state observer nor a hysteresis model
is required for the control realization.

11.3.3 Incremental PID Force Controller

In the contact phase, the touch force f between the gripper tips and the object is
controlled to follow a desired force trajectory fr by employing an incremental PID
controller. Defining e(k) = fr (k) − f (k) as the force error, the incremental PID
controller equation is expressed as follows:

u PID (k) = u(k − 1) + K P [e(k) − e(k − 1)] + K I e(k)

+K D [e(k) − 2e(k − 1) + e(k − 2)] (11.14)

where u(k − 1) represents the control command in the previous time step. K P , K I ,
and K D are the positive control gains.
Generally, the PID control parameters can be adjusted with Ziegler–Nichols (Z–N)
method by simulation studies, and then finely tuned through experimental investiga-
tions. Due to the lack of an accurate force response model, the PID gains are directly
determined by the trial-and-error approach via experimental studies in this work.

11.3.4 Switching Criterion

In the grasp process, the two gripper tips initially move from their home positions
toward the object at a velocity of vr (µm/s). During this closing phase, the DSMC
position control is employed. Once the gripper tips touch the object and the force
sensor indicates a contact force exceeding the threshold value f 0 , the control system
switches to force control. In the duration of the contact phase, the regulation of the
contact force f grip is implemented by the incremental PID force control. After a
cumulative number of time steps kcontact , the grasped object can be moved to the
destination in microhandling tasks. Then, the gripper tips start to move away from
the object at the velocity of vr . In the opening phase, no force profile is required
and the controller switches back to the DSMC position control again. The switching
process between the position and force control is given below
246 11 Position/Force Switching Control of a Miniature Gripper

if | f (k)| < f 0
xr (k) = vr k;
u(k) = u DSM (k);
u(k − 1) = u(k);
else if | f (k)| ≥ f 0 and k ≤ kcontact
fr (k) = f grip ;
u(k) = u PID (k); (11.15)
u(k − 1) = u(k);
else
xr (k) = x|k=kcontact − vr (k − kcontact );
u(k) = u DSM (k);
u(k − 1) = u(k);
end.

11.4 Experimental Investigations and Discussion

The developed control schemes have been validated on the gripper system by con-
ducting several experimental studies. In particular, the gripping and releasing testing
of a copper wire with the diameter of 300 µm has been implemented to verify the
effectiveness of the presented incremental control strategy.

11.4.1 Controller Setup

The scheme of the position control with force monitoring adopts the DSMC control
algorithm, and the position/force switching control involves both DSMC and PID
algorithms. Table 11.2 shows the control parameters of DSMC position controller and
PID force controller, which are tuned by the trial-and-error method via experimental
studies. Moreover, a force threshold of f 0 = 3.5 mN and a 2.7 s duration of the

Table 11.2 Parameters of the DSMC position controller and PID force controller
Controller Parameter Value
DSMC λP 4000
λI 1
λS 0.4
ε 7000
PID KP 0.0010
KI 0.0030
KD 0.0006
11.4 Experimental Investigations and Discussion 247

gripping force f grip = 20 mN are selected as the first and second switching criteria,
respectively. In addition, the sampling rate is chosen as 2 kHz.

11.4.2 Position/Force Switching Control Results

Concerning the scheme of the position control with force monitoring, the reference
trajectories for the position and force tracking are assigned in Fig. 11.8a, b, respec-
tively. Figure 11.8c depicts the control action, and the position and force errors are
shown in Fig. 11.9a, b, respectively. The operation procedures are described below:

(a) 100
Displacement (µm)

Actual
50
Reference

0 I II III

0 1 2 3 4 5 6
Time (s)
(b) 30

20
Force (mN)

10
Actual
0 Reference

−10
0 1 2 3 4 5 6
Time (s)
(c) 8
Control action (V)

−2
0 1 2 3 4 5 6
Time (s)

Fig. 11.8 The results of position control with force monitoring for gripping a copper wire.
a Position control results, b grasp force control results, c control action
248 11 Position/Force Switching Control of a Miniature Gripper

(a)

Displacement error (µm)

2

−1

−2

−3
0 1 2 3 4 5 6
Time (s)
(b) 4
Force error (µm)

−2

−4

−6
0 1 2 3 4 5 6
Time (s)

Fig. 11.9 Tracking errors of the position control with force monitoring approach. a Position errors.
b Force errors

1. First, the gripper tips move at a constant velocity of vr = 50 µm/s as commanded by

the DSMC position controller. The griper tip closes when the touch force arrives
at the desired force f grip and the corresponding driving voltage u 1 is recorded.
2. Second, the constant control voltage u 1 is applied aiming to maintain a constant
gripping force of 20 mN for a specified duration of time of 2.7 s.
3. Third, the DSMC position control is switched on again to open the gripper at the
constant velocity of vr to release the copper wire.
Regarding the scheme of the position/force switching control, the reference posi-
tion and force trajectories are planned as depicted in Fig. 11.10a, b, respectively. In
addition, Fig. 11.11a, b display the control errors. The operation processes are shown
as follows:
1. Initially, the gripper tip is controlled by the DSMC position controller to close at
a velocity of vr = 50 µm/s until the contact force arrives at the threshold f 0 .
2. Then, the DSMC position controller is switched off and PID force control is
switched on to hold a constant gripping force of 20 mN for a specified duration
of 2.7 s.
3. Afterward, the PID force control is switched off and DSMC position control is
turned on again to open the gripper at the velocity of vr to release the copper wire.
11.4 Experimental Investigations and Discussion 249

(a) 100

Displacement (µm)
50 Actual
Reference

0 I II III

0 1 2 3 4 5 6
Time (s)
(b) 30

20
Force (mN)

10
Actual
0 Reference

−10
0 1 2 3 4 5 6
Time (s)
(c) 8
Control action (V)

−2
0 1 2 3 4 5 6
Time (s)

Fig. 11.10 The results of position/force switching control for gripping a copper wire. a Position
control results. b Grasp force control results. c Control action

The controller output is exhibited in Fig. 11.10c. Moreover, the snapshots of

the grasp process are shown in Fig. 11.12. These sequences represent the closing
(Fig. 11.12a–c), contact (Fig. 11.12d–e), and opening (Fig. 11.12f– h) phases of the
grasp-hold-release operation, respectively.

11.4.3 Further Discussion

For comparison purpose, the experimental results of the two implemented control
schemes are tabulated in Table 11.3, where RMSE describes the steady-state posi-
tion/force tracking errors. It is observed that both of the two schemes offer a precision
positioning without chattering phenomenon. It benefits from the nice tracking capa-
250 11 Position/Force Switching Control of a Miniature Gripper

(a)

Displacement error (µm)

1

0.5

−0.5

−1

−1.5
0 1 2 3 4 5 6
Time (s)
(b) 20
Force error (µm)

15

10

−5
0 1 2 3 4 5 6
Time (s)

Fig. 11.11 Tracking errors of the position/force switching control scheme. a Position errors.
b Force errors

Fig. 11.12 Snapshots of a–c grasp, d–e hold, and f–h release operations for a copper wire of
300-µm diameter

Table 11.3 Position and force control results of the two schemes
Control scheme Position RMSE (µm) Force RMSE (mN) Overall time (s)
Position control with 0.214 1.135 5.96
force monitoring
Position/Force 0.202 0.747 5.29
switching control
11.4 Experimental Investigations and Discussion 251

bility of the DSMC control algorithm in front of piezoelectric nonlinearities. As

compared with the first scheme of the position control with force monitoring [21],
the second approach of the position/force switching control has reduced the position
and force tracking RMSE by 5.6 and 34.2 %, respectively. The accurate force control
is vital for some demanding applications such as biological sample handling.
The first method is relatively easy to implement because no force control is applied.
However, during the contact phase, an obvious drift phenomenon occurs in the grip-
ping force, as shown in Fig. 11.8b. The drift effect stems from the PSA with a constant
voltage applied. As a result, a constant gripping force cannot be maintained exactly.
On the contrary, the second control strategy is capable of regulating both the position
and gripping force precisely, as indicated in Fig. 11.10. Moreover, because a force
control is adopted in the contact phase, the step response as shown in Fig. 11.10b
is much quicker than the ramp tracking response as depicted in Fig. 11.8b. Conse-
quently, the second scheme has reduced the overall execution time of the three-phase
grasp process by 11 % as compared with the first approach.
In view of the position trajectories as shown in Figs. 11.8a and 11.10a, it is
observed that the opening and closing phases are implemented within a large motion
range due to a large initial gap between the gripper tips and the copper wire. In
addition, the trajectories consist of both rising and falling curves. Hence, the open-
ing and closing phases exhibit a significant hysteretic nonlinearity. By contrast, the
contact phase is realized within a smaller gripping range. In such a small motion
range, the hysteresis effect is not dominant while the drift phenomenon requires an
explicit treatment. Therefore, the advanced DSMC controller is adopted for the hys-
teresis suppression in the position control during the opening and closing phases.
Whereas the force control in the contact phase is realized using the PID control,
whose objective is to maintain a desired touch force by alleviating the drift phenom-
enon. Another reason of using PID controller for the force control lies in that a force
response model of the gripper is not needed in that PID is a model-free control tech-
nique. Applying DSMC control strategy in force control requires the identification of
a force response model, which complicates the practical implementation procedure.
On the other hand, the stand-alone PID control is not able to produce a satisfactory
suppression of the hysteresis effect in the opening and closing phases. Hence, DSMC
instead of PID is employed in the position control regime.
Additionally, the slopes of the position and force trajectories in the closing or open-
ing phases can be calculated to determine the combined stiffness of the gripper tips
and grasped object. For instance, similar stiffness values of 1.104 and 1.092 mN/µm
are obtained from the opening phase trajectories of the first and second control meth-
ods, respectively. The similarity arises from the reason that the same DSMC position
control is applied in the opening phase in both of the two methods, hence similar
performance is expected.
The fact that satisfactory tracking results in the closing and opening phases have
been achieved reveals the robustness property of the implemented DSMC control
strategy under the influence of contact force. Moreover, Figs. 11.8 and 11.10 indicate
that a smooth transition between the position and force control has been achieved.
252 11 Position/Force Switching Control of a Miniature Gripper

The experimental results demonstrate the effectiveness of the presented incremental

control scheme dedicated to smooth position/force switching control.
It is notable that a velocity of 50 µm/s is employed in both closing and opening
phases in the gripping operations. Actually, different velocities can be realized in the
two phases. Moreover, other trajectory types and velocity values can also be used.
Further experiments indicate that a faster positioning velocity of 300 µm/s can be
implemented with an RMSE less than 0.5 µm.

11.5 Chapter Summary

This chapter reports on the position and force control of a compliant gripper, which
possesses integrated strain-gage position and force sensors. The gripper provides a
gripping range of 276 µm along with the position and force sensing resolutions of
0.345 µm and 1.501 mN, respectively. In addition, the gripping range and initial gap
between the gripper tips can be adjusted to accommodate different sizes of objects. An
incremental-based smooth position/force switching control framework is proposed
to demonstrate a grasp-hold-release operation of a copper wire of 300-µm diameter.
Furthermore, the reported DSMC control algorithm simplifies the control design
and implementation procedures. Experimental results confirm that the presented
incremental control scheme makes it easy to realize a smooth transition during the
switching control. Moreover, the position/force switching control performs better
than the conventional approach of position control with force monitoring, as the
former allows an exact delivery of the desired position and contact force.
In view of the promising accuracy for both the position and gripping force reg-
ulation of the presented incremental switching controller, it will be employed for
handling deformable biological materials in the future. Moreover, an architectural
optimization of the gripper structure will be performed to adapt to the manipulation
and assembly of more micrometer-sized objects dedicated to industrial automation.

References

1. Alkkiomaki, O., Kyrki, V., Kalviainen, H., Liu, Y., Handroos, H.: Smooth transition from
motion to force control in robotic manipulation using vision. In: Proceeding 9th International
Conference on Control, Automation, Robotics and Vision, pp. 1–6 (2006)
2. Beyeler, F., Neild, A., Oberti, S., Bell, D.J., Sun, Y., Dual, J., Nelson, B.J.: Monolithically fab-
ricated microgripper with integrated force sensor for manipulating microobjects and biological
cells aligned in an ultrasonic field. J. Microelectromech. Syst. 16(1), 7–15 (2007)
3. Cao, Y., Cheng, L., Chen, X.B., Peng, J.Y.: An inversion-based model predictive control with
an integral-of-error state variable for piezoelectric actuators. IEEE/ASME Trans. Mechatron.
18(3), 895–904 (2013)
4. Capisani, L.M., Ferrara, A.: Trajectory planning and second-order sliding mode
motion/interaction control for robot manipulators in unknown environments. IEEE Trans. Ind.
Electron. 59(8), 3189–3198 (2012)
References 253

5. Gao, B., Shao, J., Han, G., Sun, G., Yang, X., Wu, D.: Using fuzzy switching to achieve the
smooth switching of force and position. Appl. Mech. Mater. 274, 638–641 (2013)
6. Hara, S., Yamada, Y.: A control method switching from servo automatic transfer to force
sensorless impedance control manual positioning. In: Proceeding IECON 2007—33rd Annual
Conference of the IEEE Industrial Electronics Society, pp. 292–298 (2007)
7. Leang, K.K., Shan, Y., Song, S., Kim, K.J.: Integrated sensing for IPMC actuators using strain
gages for underwater applications. IEEE/ASME Trans. Mechatron. 17(2), 345–355 (2012)
8. Liaw, H.C., Shirinzadeh, B.: Robust adaptive constrained motion tracking control of piezo-
actuated flexure-based mechanisms for micro/nano manipulation. IEEE Trans. Ind. Electron.
58(4), 1406–1415 (2011)
9. Minase, J., Lu, T.F., Cazzolato, B., Grainger, S.: A review, supported by experimental results,
of voltage, charge and capacitor insertion method for driving piezoelectric actuators. Precis.
Eng. 34(10), 692–700 (2010)
10. Motoi, N., Shimono, T., Kubo, R., Kawamura, A.: Task realization by a force-based variable
compliance controller for flexible motion control systems. IEEE Trans. Ind. Electron. 61(2),
1009–1021 (2014)
11. Navarro, C., Treesatayapun, C., Baltazar, A.: Determination of the instantaneous initial contact
point on a parallel gripper using a multi input fuzzy rules emulated network controller with
feedback from ultrasonic and force sensors. In: Batyrshin I., Gonzalez Mendoza M. (eds.)
Advances in Artificial Intelligence, Lecture Notes in Computer Science, vol. 7629, pp. 261–
272. Springer, Berlin (2013)
12. Niksefat, N., Wu, Q., Sepehri, N.: Stable control of an electro-hydraulic actuator during contact
tasks: Theory and experiments. In: Proceeding of American Control Conference, pp. 4114–
4118 (2000)
13. Rakotondrabe, M., Ivan, I.A.: Development and force/position control of a new hybrid thermo-
piezoelectric microgripper dedicated to micromanipulation tasks. IEEE Trans. Automat. Sci.
Eng. 8(4), 824–834 (2011)
14. Reddy, A.N., Maheshwari, N., Sahu, D.K., Ananthasuresh, G.K.: Miniature compliant grippers
with vision-based force sensing. IEEE Trans. Robot. 26(5), 867–877 (2010)
15. Sakaino, S., Sato, T., Ohnishi, K.: Multi-DOF micro-macro bilateral controller using oblique
coordinate control. IEEE Trans. Ind. Inform. 7(3), 446–454 (2011)
16. Sakaino, S., Sato, T., Ohnishi, K.: Precise position/force hybrid control with modal mass
decoupling and bilateral communication between different structures. IEEE Trans. Ind. Inform.
7(2), 266–276 (2011)
17. Shimada, N., Yoshioka, T., Ohishi, K., Miyazaki, T.: Novel force-sensor-less contact motion
control for quick and smooth industrial robot motion. In: Proc. IECON 2011 - 37th Annual
Conf. on IEEE Industrial Electronics Society, pp. 4238–4243 (2011)
18. Takahashi, T., Tsuboi, T., Kishida, T., Kawanami, Y., Shimizu, S., Iribe, M., Fukushima, T.,
Fujita, M.: Adaptive grasping by multi fingered hand with tactile sensor based on robust force
and position control. In: Proceeding International Conference on Robotics and Automation,
pp. 264–271 (2008)
19. Wang, D.H., Yang, Q., Dong, H.M.: A monolithic compliant piezoelectric-driven microgripper:
Design, modeling, and testing. IEEE/ASME Trans. Mechatron. 18(1), 138–147 (2013)
20. Xu, Q.: Adaptive discrete-time sliding mode impedance control of a piezoelectric microgripper.
IEEE Trans. Robot. 29(3), 663–673 (2013)
21. Xu, Q.: New robust position and force regulation for a compliant microgripper. In: Proceeding
9th IEEE International Conference on Automation Science and Engineering, pp. 801–806
(2013)
22. Yong, Y.K., Moheimani, S.O.R., Kenton, B.J., Leang, K.K.: High-speed flexure-guided nanopo-
sitioning: Mechanical design and control issues. Rev. Sci. Instrum. 83(12), 121101 (2012)
23. Zhou, Q., Korhonen, P., Laitinen, J., Sjovall, S.: Automatic dextrous microhandling based on
a 6-DOF microgripper. J. Micromechatronics 3(3–4), 359–387 (2006)
Index

A Digital PID control, 73, 197

Actuators, 2 Digital signal processor (DSP), 10, 25, 211
Adaptive control, 9 Digital sliding-mode control (DSMC), 15,
Adaptive digital sliding-mode generalized 127, 170, 214, 234
impedance control (ADSMGIC), 211 Digital sliding-mode prediction control
Adaptive laws, 192 (DSMPC), 167
Adaptive rule, 217 Digital systems, 14
Adhesive force, 229 Discrete sliding mode, 216
Artificial neural networks (ANN), 8, 23 Discrete-time model, 129, 148, 168, 212
Discrete-time sliding-mode control
(DTSMC), 10, 127
B Discretization, 129
Backlash operator, 30 Discretization noise, 129
Backward difference equation, 190 Displacement amplifier, 194, 234
Bayesian inference, 35 Displacement sensor, 3
Bouc–Wen hysteresis, 58 Disturbance, 148
Bouc–Wen model, 28 Disturbance estimation, 150, 169
Boundary layer technique, 88, 133 Disturbance observer, 13
Dual formulation, 34, 61
Dynamics model, 58, 81, 212
C
Calibration, 237
Capacitive sensors, 3 E
Chattering, 88, 147 E-beam lithography, 3
Compliant gripper, 234 Eigenvalue, 86
Compliant mechanisms, 1 Electrical discharge machining (EDM), 94,
Control bandwidth, 161, 181, 200 236
Controllability, 116 Environment interaction, 218
Controllable canonical form, 149, 168 Equivalent control, 84, 131, 152, 215, 243
Convergence condition, 133 Euler backward difference, 129
Event-based switching, 241

D
Dead-zone modification, 193 F
Dead-zone operator, 31 Feedback control, 9
Digital microscope, 237 Feedforward compensation, 71
© Springer International Publishing Switzerland 2016 255
Q. Xu and K.K. Tan, Advanced Control of Piezoelectric
Micro-/Nano-Positioning Systems, Advances in Industrial Control,
DOI 10.1007/978-3-319-21623-2
256 Index

Feedforward control, 5, 47 LSSVM model, 33

Feedforward plus feedback control, 8, 49, 73 Luenberger observer, 92
Field-programmable gate array (FPGA), Lyapunov candidate, 192
134, 195, 234
Finite-element analysis (FEA), 194
Flexure hinge, 2 M
Force control, 245 Matching condition, 81, 148
Force observer, 220 Metal strain gauges, 3
Frequency response, 63, 95, 240 Micro-/nano-positioning, 1
Fuzzy logic control, 9 Micro-/nano-positioning system, 1
Microassembly, 13, 209
Microgripper, 13, 209, 233
G Micromanipulation, 13, 209
Generalized impedance control (GIC), 14, Model generalization, 44, 66
210 Model predictive control (MPC), 11, 173
Generalized minimum variance control Model predictive discrete-time sliding-
(GMVC), 11 model control (MPDTSMC), 80
Genetic algorithm (GA), 36 Model predictive output integral
discrete-time sliding-mode control
(MPOIDSMC), 105
H Model-reference adaptive control (MRAC),
H∞ control, 9 12
Hammerstein model, 5 Modified Prandtl–Ishlinskii (MPI) model,
High-order model, 147 29
Hybrid position/force control, 14 Momentum term, 214
Hysteresis, 4 Motion tracking, 159, 179
Hysteresis models, 6, 28 MRAC with perturbation estimation
(MRACPE), 187, 190
Multivalued mapping, 58
I
Impedance control, 14, 210, 213
Incremental control, 15, 242 N
Incremental PID control, 49, 158, 234, 245 Neural network control, 9
Inductance sensors, 3 Noise, 48, 57, 88, 114, 180, 204
Input–output-based DSMC (IODSMC), 147 Nonlinear regression, 60
Integral error, 131 Normalization transformation, 41
Intelligent hysteresis model, 59
Interaction control, 13, 209, 224
Inverse hysteresis model, 8 O
Inverse MPI model, 32 Observability, 116
Iterative learning control, 13 Optical microscopes, 3
Optical sensors, 3
Optimization, 28, 33, 39, 58
K Output integral discrete-time sliding-mode
Kernel trick, 34, 61 control (OIDSMC), 105

L P
Lagrangian function, 33 Parameter optimization, 36
Laser displacement sensor, 3, 62, 195, 237 Particle swarm optimization (PSO), 24, 36,
Laser interferometer, 3 64
Least squares support vector machines Perturbation estimation, 189
(LSSVM), 8, 24, 57 PID control, 135
Linear time-invariant (LTI) model, 106, 148 PID sliding function, 151, 170
Index 257

Piezo benders, 2 Second-order model, 129, 188

Piezo diaphragms, 2 Sensitivity, 3, 235
Piezoelectric actuators, 2 Sensor calibration, 237
Piezoelectric bimorph, 62, 128, 209 Sensors, 3
Piezoelectricity, 4 Set-point positioning, 159, 198
Piezoelectric stack actuator (PSA), 194, 233 Signal-to-noise ratio (SNR), 235
Piezoresistive strain gauges, 3 Singular value, 116
Piezo stacks, 2 Sinusoidal tracking, 200
PI sliding function, 84, 107, 130, 242 SISO, 107
Pole placement, 93 Sliding condition, 133
Position and force control, 13 Sliding-mode control (SMC), 10
Position/force switching control, 15, 233 Sliding-mode observer, 113
Prandtl–Ishlinskii model, 30 Sliding surface, 133
Precision alignment, 3 Smart actuators, 2, 209
Proportional-integral-derivative (PID) con- Smooth transition, 15, 234
trol, 9 Spectral analysis, 157, 240
Pseudo-rigid-body (PRB) model, 234 Spillover, 79
PZT, 4 Stability, 91, 108, 112, 152, 175, 216
State observer, 11, 92
State-space model, 81
Q Strain gage, 156, 237
Quadratic programming (QP), 8, 24 Strain sensors, 3
Quasi-sliding domain, 132 Support vector, 34, 61
Support vector machines (SVM), 8, 23
Switching control, 86, 131, 154, 215
R System identification, 63, 134, 157
Rate dependency, 33, 40
Rate-dependent hysteresis, 7, 23
Rate-independent hysteresis, 7
T
Reaching law, 131, 152
Target impedance, 213
Regression, 32
Trajectory planning, 222
Regression model, 60
Tustin approximation, 83
Repetitive control, 13
Robust control, 154
Robustness, 143, 162
W
White noise, 238
S
Sampled-data system, 10
Sampling time, 134 Z
Scanning electron microscopes, 3 Zero-order-hold (ZOH), 148
Scanning probe microscopes, 3 Ziegler–Nichols (Z–N) method, 73, 198