REGULATION IN SWITCHED BIMODAL LINEAR SYSTEMS
by
Zhizheng Wu
A thesis submitted in conformity with the requirements
for the degree of Doctor of Philosophy
Department of Mechanical and Industrial Engineering
University of Toronto
© Copyright by Zhizheng Wu 2009
ii
REGULATION IN SWITCHED BIMODAL LINEAR SYSTEMS
Zhizheng Wu
Ph.D. Thesis, 2009
Department of Mechanical and Industrial Engineering
University of Toronto
ABSTRACT
In the past few decades, significant progress has been made in addressing control
problems for a variety of engineering systems having smooth dynamics. In practice, one often
encounters also non-smooth systems in various branches of science and engineering, such as
for example mechanical systems subject to impact. Motivated by the read/write head flying
height regulation problem in hard disk drives, where the close proximity of the read/write
head to the disk surface results in intermittent contact between the two and a bimodal system
behavior, this thesis studies the output regulation problem in switched bimodal linear systems
against known and unknown exogenous input signals.
The regulation problems in bimodal systems presented in this thesis are solved within sets
of Q -parameterized controllers, in which the Q parameters are designed to yield internal
stability and exact output regulation in the closed loop switched system. The proposed
parameterized controllers are constructed mainly in two steps. The first step is based on
constructing a switched observer-based state feedback central controller for the switched
iii
linear system. The second step involves augmenting the switched central controller with
additional dynamics (i.e. Q parameter) to construct a parameterized set of switched
controllers. Based on the proposed sets of Q -parameterized controllers, four main regulation
problems are addressed and corresponding regulator synthesis algorithms are proposed. The
first problem concerns regulation against known deterministic exogenous inputs, where no
stability or structural constraints are imposed on the Q parameter. The second problem is
similar to the first, except that the Q parameter is constrained to be a linear combination of
basis functions. This structure of the Q parameter is considered in the rest of the thesis. The
third problem involves regulation against exogenous inputs involving known deterministic
components and unknown random components, and where the regulator is designed subject to
an 2H performance constraint. The last problem involves the development of adaptive
regulators against unknown sinusoidal exogenous inputs. The different regulator synthesis
algorithms are developed based on solving sets of linear matrix inequalities or bilinear matrix
inequalities. The last two proposed regulation methods are successfully evaluated on an
experimental setup motivated by the flying height regulation problem in hard disk drives, and
involving a mechanical system with switched dynamics.
iv
ACKNOWLEDGEMENTS
I would like to thank my academic supervisor, Professor Foued Ben Amara, for his guidance,
encouragement, helpful discussions and support throughout the course of the work. My thanks
also go to Professor James K. Mills, Professor Beno Benhabib and Professor Yu Sun, who have
served as members of my thesis examination committee.
Furthermore, I would also like to express my sincere gratitude to all my colleagues in the
Mechanical and Industrial Engineering Department at the University of Toronto, including
Mr. Maurizio Ficocelli, Mr. Azhar Iqbal, Mr. Jason Li and Mr. Edward Lee who provided
invaluable help and advice in completing this thesis. I gratefully acknowledge the financial
support from the Ontario Graduate Scholarship (OGS), the Ontario Graduate Scholarship in
Science and Technology (OGSST), and the University of Toronto Fellowship. Finally, I
would like to thank my family for all their support and encouragement throughout my studies.
v
TABLE OF CONTENTS
ABSTRACT ........................................................................................................................................................II
ACKNOWLEDGEMENTS............................................................................................................................ IV
LIST OF FIGURES...................................................................................................................................... VIII
LIST OF TABLES..........................................................................................................................................XII
CHAPTERS
1 INTRODUCTION.................................................................................................................................... 1
1.1 MOTIVATION................................................................................................................. 1 1.2 LITERATURE REVIEW ON SWITCHED SYSTEMS ............................................................... 3
1.2.1 Stability of switched systems ........................................................................................... 4 1.2.2 Output regulation in switched systems .......................................................................... 9 1.2.3 Adaptive regulation in linear and nonlinear systems...............................................12
1.3 PROBLEM STATEMENT................................................................................................. 14 1.4 CONTRIBUTIONS OF THE THESIS .................................................................................. 16 1.5 ORGANIZATION OF THE THESIS.................................................................................... 19
2 PARAMETERIZED REGULATOR SYNTHESIS FOR SWITCHED BIMODAL SYSTEMS BASED ON BILINEAR MATRIX INEQUALITIES.........................................22
2.1 INTRODUCTION............................................................................................................ 22 2.2 REGULATION PROBLEM FOR SWITCHED BIMODAL SYSTEMS ........................................ 24 2.3 PARAMETERIZATION OF A SET OF CONTROLLERS......................................................... 26
2.3.1 Observer-based state feedback controller ...................................................................27 2.3.2 Parameterized output feedback controller...................................................................27 2.3.3 Stability of the parameterized switched closed loop system .................................30
2.4 REGULATION CONDITIONS FOR THE SWITCHED SYSTEM .............................................. 32 2.4.1 Regulation conditions for 1
clΣ and 2clΣ .....................................................................33
2.4.2 Equivalent impulsive switched closed loop system model....................................36 2.4.3 Regulation conditions for the switched system.........................................................38
2.5 REGULATOR SYNTHESIS FOR THE SWITCHED SYSTEM.................................................. 43 2.6 NUMERICAL EXAMPLE................................................................................................. 48 2.7 CONCLUSION............................................................................................................... 53
3 REGULATION IN SWITCHED BIMODAL SYSTEMS BASED ON LINEAR MATRIX INEQUALITIES....................................................................................................................................56
3.1 INTRODUCTION............................................................................................................ 56 3.2 THE REGULATION PROBLEM FOR SWITCHED BIMODAL SYSTEMS ................................. 58
vi
3.3 PARAMETERIZATION OF A SET OF STABILIZING CONTROLLERS .................................... 60 3.3.1 Q -parameterized controller ...........................................................................................60 3.3.2 Internal stability of the Q -parameterized switched closed loop system ..........62
3.4 REGULATION OF BIMODAL SYSTEMS: A COMMON LYAPUNOV FUNCTION APPROACH... 67 3.4.1 Input-output stability.........................................................................................................67 3.4.2 State space regulation conditions for 1
clΣ and 2clΣ ................................................73
3.4.3 Regulation conditions for the switched system.........................................................76 3.4.4 Regulator synthesis for the switched system .............................................................88 3.4.5 Numerical example............................................................................................................91
3.5 REGULATION OF BIMODAL SYSTEMS: A MULTIPLE LYAPUNOV FUNCTION APPROACH . 95 3.5.1 Regulation conditions for 1
clΣ and 2clΣ in the frequency domain ....................98
3.5.2 Regulation condition and regulator synthesis for the switched system............104 3.5.3 Numerical example..........................................................................................................112
3.6 CONCLUSION............................................................................................................. 114
4 REGULATION IN SWITCHED BIMODAL SYSTEMS WITH AN 2H
PERFORMANCE CONSTRAINT................................................................................................117
4.1 INTRODUCTION.......................................................................................................... 117 4.2 THE REGULATION PROBLEM FOR THE SWITCHED BIMODAL SYSTEM .......................... 119 4.3 PARAMETERIZATION OF A SET OF STABILIZING CONTROLLERS .................................. 120
4.3.1 Q -parameterized controller .........................................................................................120 4.3.2 Stability of the Q -parameterized switched closed loop system........................123
4.4 REGULATION CONDITIONS FOR SWITCHED SYSTEM ................................................... 127 4.4.1 Regulation conditions for 1Σ and 2Σ .....................................................................128
4.4.2 Regulation condition for the switched system.........................................................134 4.4.3 Regulation condition for the switched closed loop system cl
rΣ with an 2H performance constraint ..................................................................................................137
4.5 REGULATOR SYNTHESIS ............................................................................................ 139 4.6 NUMERICAL EXAMPLE............................................................................................... 144 4.7 EXPERIMENTAL EVALUATION.................................................................................... 149
4.7.1 Description of the experimental setup .......................................................................150 4.7.2 Simulation of the external disturbances and of the contact force ......................151 4.7.3 Identification of the switched system model............................................................155 4.7.4 Experimental results........................................................................................................158
4.8 CONCLUSION............................................................................................................. 162
5 ADAPTIVE REGULATION IN SWITCHED BIMODAL LINEAR SYSTEMS ...........167
5.1 INTRODUCTION.......................................................................................................... 167 5.2 THE ADAPTIVE REGULATION PROBLEM FOR BIMODAL SYSTEMS................................ 169 5.3 PARAMETERIZATION OF A SET OF STABILIZING CONTROLLERS .................................. 170
5.3.1 Q -parameterized controller .........................................................................................170 5.3.2 Stability of the Q -parameterized switched closed loop system........................173
vii
5.4 REGULATION CONDITIONS FOR THE SWITCHED SYSTEM ............................................ 174 5.5 ADAPTIVE REGULATION IN THE SWITCHED CLOSED LOOP SYSTEM ............................ 185 5.6 NUMERICAL EXAMPLE............................................................................................... 193 5.7 EXPERIMENTAL EVALUATION.................................................................................... 197 5.8 CONCLUSION............................................................................................................. 200
6 CONCLUSION .....................................................................................................................................207
6.1 CONTRIBUTIONS OF THE THESIS ................................................................................ 207 6.1.1 Development of a regulator synthesis procedure in bimodal linear systems
against known deterministic exogenous inputs......................................................208 6.1.2 Development of a regulator synthesis procedure in bimodal linear systems
with an 2H performance constraint. .......................................................................209
6.1.3 Development of an adaptive regulation approach in bimodal linear systems against unknown sinusoidal exogenous inputs.......................................................210
6.2 FUTURE WORK .......................................................................................................... 211
APPENDIX ...................................................................................................................................................214
BIBLIOGRAPHY .......................................................................................................................................228
viii
LIST OF FIGURES
Figure 2.1. Closed loop system with Q -parameterized controller. ....................................... 27 Figure 2.2. State trajectories of 1
clΣ with initial conditions on the switching surface inside the set Bγ for case 1. ................................................................................................ 43
Figure 2.3. State trajectories of 2
clΣ with initial conditions on the switching surface inside the set Bγ for case 2. ................................................................................................ 43
Figure 2.4. Diagram of a mechanical system with switched dynamics. ................................. 48 Figure 2.5. Simulation results for the case of 30ev = − micrometers showing the
performance variable e and the switching component (1,1)rd in the disturbance
rd . ........................ ………………………………………………………………54 Figure 2.6. Simulation results for the case of 30ev = − micrometers showing the regulated
mass height v and the contact surface displacement sv . .................................. 54 Figure 3.1. State trajectories of 1
clΣ with initial conditions on the switching surface inside the set xB for case 1…....………………… ……………………………………… 86
Figure 3.2. State trajectories of 2
clΣ with initial conditions on the switching surface inside the set xB for case.. ................................................................................................. 86
Figure 3.3. Simulation results for the case of 30ev = − micrometers showing the
performance variable e and the switching component (1,1)rd in the disturbance rd ...………………………………………………………………..93
Figure 3.4. Simulation results for the case of 30ev = − micrometers showing the regulated
mass height v and the contact surface displacement sv . ................................. 93
Figure 3.5. State space partitions where the vector Tex exC C= . ............................................ 105
ix
Figure 3.6. Simulation results for the case of 30ev = − micrometers showing the performance variable e and the switching component (1,1)rd in the disturbance rd . . ………………………………………………………………115
Figure 3.7. Simulation results for the case of 30ev = − micrometers showing the regulated
mass height v and the contact surface displacement sv . ............................... 115 Figure 4.1. Closed loop system with a Q -parameterized controller. ……………………121 Figure 4.2. Simulation results showing the performance variable e obtained using the
controller designed without accounting for the 2H performance constraint. .. 146 Figure 4.3. Simulation results showing the switching component (1,1)rd in the known input
rd and the random disturbance wd . ................................................................. 146 Figure 4.4. Simulation results showing the displacement of the mass obtained using the
controller designed without accounting for the 2H performance constraint. .. 147 Figure 4.5. Simulation results showing the performance variable e obtained using the
controller designed based on the 2H performance constraint.......................... 147 Figure 4.6. Simulation results showing the displacement of the mass obtained using the
controller designed based on the 2H performance constraint.......................... 148 Figure 4.7. Schematic diagram of the experimental setup.................................................... 151 Figure 4.8. The experimental setup: (a) General view, (b) Close-up view of the suspension
beam with the actuators...................................................................................... 153 Figure 4.9. The frequency response of the identified analytical model (solid line) and the real
system based on experimental data (dashed line) in the non-contact mode: (a) transfer function from u to vy , (b) transfer function from d to vy . .......... 158
Figure 4.10. The frequency response of the identified analytical model (solid line) and the real
system based on experimental data (dashed line) in the contact mode: (a) Transfer function from u to vy , (b) transfer function from d to vy ........ 159
Figure 4.11. Experimental results showing the performance variable e and the position y
for the case of 0dw = and obtained using the controller designed without accounting for the 2H performance constraint. ............................................. 163
x
Figure 4.12. Experimental results showing the random disturbance dw and the total current
in the coil for the case of 0dw = and obtained using the controller designed without accounting for the 2H performance constraint. ................................ 163
Figure 4.13. Experimental results showing the performance variable e and the position y
in the presence of dw and obtained using the controller designed without accounting for the 2H performance constraint. ............................................. 164
Figure 4.14. Experimental results showing the random disturbance dw and the total current
in the coil obtained using the controller designed without accounting for the 2H performance constraint. .................................................................................... 164
Figure 4.15. Experimental results showing the performance variable e and the position y
in the presence of dw and obtained using the controller designed by accounting for the 2H performance constraint................................................................. 165
Figure 4.16. Experimental results showing the random disturbance dw and the total current
in the coil obtained using the controller designed by accounting for the 2H performance constraint………………………………………………………...165
Figure 5.1. Diagram of the switching sequence for the closed loop system. ……………….186 Figure 5.2. Simulation results showing the performance variable e and the switching
component (1,1)id in the input id ................................................................... 195
Figure 5.3. Simulation results showing estimated parameter vector 2θ and nominal parameter vector 0θ ........................................................................................... 195
Figure 5.4. Experimental results showing the performance variable e for the case where the
nominal parameter vector 0θ is used. .............................................................. 201 Figure 5.5. Experimental results showing the tip position ty of the suspension beam and the
contact surface cs for the case where the nominal parameter vector 0θ is used............................................................................................................................. 201
Figure 5.6. Experimental results showing the current i in the coil used to simulate the
disturbance force and the contact force for the case where the nominal parameter vector 0θ is used............................................................................................... 202
xi
Figure 5.7. Experimental results showing the control input signal u driving the piezoelectric actuator for the case where the nominal parameter vector 0θ is used.............. 202
Figure 5.8. Experimental results showing the performance variable e for the case where the
adaptive regulator isused.………………………………………………………203 Figure 5.9. Experimental results showing the tip position ty of the suspension beam and the
contact surface cs for the case where the adaptive regulator is used............... 203 Figure 5.10. Experimental results showing the current i in the coil used to simulate the
disturbance force and the contact force for the case where the adaptive regulator is used. .............................................................................................................. 204
Figure 5. 11. Experimental results showing the control input signal u driving the
piezoelectric actuator for the case where the adaptive regulator is used. ........ 204 Figure 5.12. Experimental results showing the desired parameter vector 0θ (dashed line) and
the estimated parameter vector θ (solid line) for the case where the adaptive regulator is used. .............................................................................................. 205
xii
LIST OF TABLES
Table 1.1. Summary of the contributions................................................................................ 21 Table 6.1. Summary of advantages and limitations of the proposed algorithms .................. 211
1
CHAPTER 1
Introduction
The main theme of the research presented in this dissertation is the development and analysis
of regulation techniques against exogenous input signals in switched bimodal linear systems.
The exogenous input signals could represent reference signals to be tracked or disturbance
signals to be rejected. Two kinds of regulator design problems are addressed in this thesis.
The first problem is the design of regulators for switched bimodal systems to reject known
exogenous input signals. The second problem is the development of an adaptive regulation
method in switched bimodal systems to reject unknown exogenous sinusoidal input signals.
The rest of this chapter is organized as follows. A motivation for the regulation problem
treated in this thesis is presented in Section 1.1. A literature review on the stability analysis
and regulation of switched systems is then presented in section 1.2. Section 1.3 outlines the
problems addressed in this thesis, followed by a summary of contributions in Section 1.4.
The organization of the rest of the thesis is presented in Section 1.5.
1.1 Motivation
To this day, Hard disk drives (HDDs) remain the major archival mass data storage medium in
computing devices and are also playing an increasingly significant role in consumer
2
electronics devices. Over the past decade, HDD data storage capacity has been increasing at a
rate of 60% to 100% per year. To accommodate the demand for more storage capacity in the
same volume, the magnetic medium has to store the information at ever higher areal densities,
which requires that the read/write (R/W) head be held at a very small constant flying height
above the moving magnetic medium (disk surface). Most of the existing modern hard-disk
drives operate in the flying mode [1-4], where the R/W head attached on the slider is lifted by
the air flow beneath and flies above the disk surface when the hard disk rotates. The flying
height will vary with time due to a number of unknown factors, including spindle vibrations,
hard disk surface microwaviness, and also airflow disturbances generated by the disk rotation.
Therefore, it is highly desirable to design a closed-loop feedback system to actively suppress
the flying height variations.
One of the major issues associated with the design of a flying height regulation system is
the fact that, for small flying heights, the slider/disk interface becomes a system with
switched dynamics due to the intermittent contact between the R/W head and the disk surface.
Moreover, the various causes of flying height variation (microwaviness of the disk surface,
aerodynamic disturbances, spindle vibrations, etc.) are not well characterized and are not
known a priori. It is well known that the techniques developed to address traditional
regulation problems in linear and smooth nonlinear systems cannot be applied to switched
systems directly. At present time, there are only limited results on the regulation problem for
switched systems. In this thesis, motivated directly by the flying height regulation problem in
hard disk drives, some original results related to the exact output regulation problem in
switched bimodal systems against both known and unknown exogenous input signals is
presented.
3
1.2 Literature Review on Switched Systems
The purpose of this section is to summarize the main results presented in the literature related
to the control of switched systems. Recently, switched control systems have attracted much
attention in the control community since they present problems that are not only
academically challenging, but also of practical importance. Switched systems are hybrid
systems that combine various types of dynamic behavior, including continuous-time
dynamics, discrete-time dynamics, switching and state variable jumps [5-6]. There are three
main factors that motivate the study of switched systems. First, many of the natural non-
smooth dynamic systems can be modeled as switched systems. Examples of such systems
include mechanical systems with impact, hopping robots, linear systems with static
nonlinearities such as relays and saturations, network communication systems with delay, etc.
[2,7,8,9]. Second, switched systems offer much better approximations of nonlinear or hybrid
systems than linear systems. This is due to the fact that linear systems only approximate the
true dynamics in small regions around equilibrium points and do not capture important
properties of nonlinear systems such as limit cycle oscillations [10]. Lastly, the study of the
switched systems is also motivated by the widespread applications of switched controllers in
systems with increasing performance requirements, especially in the presence of large
uncertainties or disturbances. A good example of such systems is the so-called supervised
control design that involves switching between different controllers. It is well understood
that switched controller systems, as a special type of switched systems, can provide control
solutions that guarantee a level of stability and good performance that no single-controller
system can provide [11-14]. For example, many complex nonlinear systems that are not
4
stabilizable by a single controller can be stabilized by switching between a finite number of
controllers [15-17].
1.2.1 Stability of switched systems
The stability issues in switched systems, especially switched linear systems, have been of
increasing interest for researchers in the past decade, and include several interesting
phenomena. For example, even when all the subsystems in a switched system are
exponentially stable, the switched system may have divergent trajectories for certain
switching signals [5,18,19]. Another remarkable fact is that one may carefully switch
between unstable subsystems to make the switched system exponentially stable [20,21]. As
these examples suggest, the stability of switched systems depends not only on the dynamics
of each subsystem in the switched system, but also on the properties of switching signals.
The stability analysis of switched systems involves two main types of problems, namely, the
stability analysis for switched systems under arbitrary switching and the stability analysis for
switched systems under constrained switching.
Stability of switched systems under arbitrary switching
Among the large variety of stability analysis problems encountered in switched systems,
one can assume that the switching sequence is not known a priori and look for stability
results under arbitrary switching sequences. A well known approach to deal with this
problem is based on the existence of common quadratic Lyapunov functions (CQLF) to
check the asymptotic stability for the switched system under arbitrary switching. The
existence of a common quadratic Lyapunov function for all subsystems in the switched
system assures the quadratic stability of the switched system. However, the conditions for
the existence of a common quadratic Lyapunov function are not easy to determine. In fact,
5
there have been various attempts to derive conditions for the existence of a common
quadratic Lyapunov function. Ooba, et al. [22] and Shorten, et al. [23] presented some
easily verifiable algebraic conditions for proving the existence of the common quadratic
Lyapunov function for general switched linear systems. Dayawansa and Martin [24]
proved that uniform exponential stability is equivalent to the existence of a smooth
positive definite function, which is a common Lyapunov function for the switched system,
but cannot guarantee the existence of a common quadratic Lyapunov function. If a
common quadratic Lyapunov function does not exist or cannot be found, then the multiple
Lyapunov function method can be used for stability analysis. Branicky [25,26] proposed
conditions for multiple Lyapunov functions to ensure the stability of the switched system.
Numerical algorithms to compute multiple Lyapunov functions were developed by
Brayton and Tong [27] and Polanski [28]. However, the computation of the multiple
Lyapunov functions is not a tractable problem in practice. For discrete-time switched
systems, Daafouz, et al. [29] developed a more practical scheme to check the stability of
the system using multiple quadratic Lyapunov functions. The main idea is to associate a
quadratic Lyapunov function with each subsystem in the switched system, and then check
the asymptotic stability using the switched quadratic Lyapunov functions which are
designed to decrease at each of the switching instances. The existence of such Lyapunov
functions can be reduced to the compatibility of a set of linear matrix inequalities (LMIs).
There also have been other results in the literature that propose constructive synthesis and
computation methods to find multiple Lyapunov functions using linear matrix inequalities
or bilinear matrix inequalities [30,31,32].
6
Stability of switched systems under constrained switching
In this class of problems, one important case of stability studies involves a finite number
of linear systems together with a set of switching rules defined by switching surfaces
(state dependent switching). In this case, the state space is partitioned into cells, and the
dynamics of the system are described using different linear models corresponding to
operation in the different cells. Johansson and Rantzer [33] presented the first rigorous
result on the analysis of general switched systems where switching is based on space
bounds (switching surfaces). This method relaxes the problem to the search for piecewise
quadratic Lyapunov functions for the switched system, which can be formulated by
solving a set of linear matrix inequalities. If the piecewise quadratic Lyapunov functions
cannot be found with the given natural partitioning of the state space, then a refinement of
the partitions can be applied to improve the flexibility of finding piecewise quadratic
Lyapunov functions. For high order systems, although it is very hard to obtain a
refinement of the proper partitions in the state space to efficiently analyze the stability
using piecewise quadratic Lyapunov functions, the idea of this approach still appears to be
very powerful and can be furthered in a large number of directions, including performance
analysis, global linearization, controller optimization, and model reduction. Hassibi and
Boyd [34] addressed the question of stability and control of piecewise linear time-
invariant systems using linear matrix inequalities. Using Lyapunov theory, they derived
sufficient conditions for stability and performance that can be checked by solving convex
optimization problems with LMI constraints along the switching surface. The approach is
to search among special classes of functions for a Lyapunov function that proves stability
or performance for the piecewise linear system. They demonstrated applications of the
7
proposed methods by considering controller synthesis of a simple mechanical system
subject to input saturation, and stability analysis of an electrical circuit with multiple
equilibrium points. Goncalves, Mgretski, and Dahleh [35] developed a stability analysis
methodology for switched linear systems using impact maps and surface Lyapunov
functions. This methodology solely studies the system behaviour at the switching surface
and constructs a surface Lyapunov function to show that impact maps associated with the
switched linear system are contracting. This representation of impact maps allows the
search for surface Lyapunov functions to be done by simply solving a semidefinite
program, allowing global asymptotic stability, robustness, and performance of limit cycles
and equilibrium points of piecewise linear system to be efficiently checked. This analysis
methodology has been successfully applied to relay feedback, on/off and saturation
systems. Rodrigues and How [36] proposed a synthesis method for both state and dynamic
output feedback control of a class of piecewise affine systems, where the state space
related to each subsystem is partitioned as polytopic cells. The synthesis procedure
delivers stabilizing controllers that do not necessarily have measurements of the switching
parameters and enables switching in the controller based on the estimated states rather
than on measured parameters. The proposed technique formulates the search for a
piecewise quadratic control Lyapunov function and a piecewise affine control law as a bi-
convex optimization program subject to some linear constraints and a bilinear matrix
inequality, which can be solved effectively using iterative algorithms.
Another type of switched systems with constrained switching corresponds to switched
systems where the switching is time dependent. One well known approach to guarantee
stability in this case is to find a lower bound on the dwell time or average dwell time in
8
the switched system in any given mode of operation of the switched system [5]. The basic
idea is that a switched system is stable if all individual subsystems are stable and the
switching is sufficiently slow as to allow the transient effects to dissipate after each switch.
Recently, the stability of switched systems under random switching, in which the
switching times are unknown but can characterized by stochastic processes, has also been
studied in the literature. For example, the stochastic stabilizability for discrete-time linear
switching systems is investigated by Boukas and Benzaouia [37], where the switching
signal is a finite-state Markovian jump process. A sufficient condition for stochastic
stability of the system is deduced by using a non-quadratic Lyapunov function, and the
necessary and sufficient conditions allowing the constrained control law to always be
admissible despite the stochastic character of the system are also presented. Similar
problems are also addressed in the papers of Fang and Loparo [38], and Costa and Tuesta
[39]. In other cases, where the switching signal can be controlled and designed, one then
can study the existence of specific switching rules that ensure stability of the switched
system. In these cases, both the switching signal and the control input are assumed to be
design variables. As in the case of linear systems, the concepts of controllability play a
fundamental role in the design and synthesis for switching systems if the switching signal
can be designed. Loparo, Aslanis, and Hajek [40] addressed the issues of controllability
and reachability for continuous-time planar switched systems. Sufficient conditions and
necessary conditions for controllability and reachability were reported by Ezzine and
Haddad [41], Szigeti [42], Sun and Ge [43], Xie, et al. [44], Yang [45], and Cheng, et al.
[46].
9
In addition to the results mentioned before, numerous other results are published in recent
years and are summarized in books [47-49] and the references therein.
1.2.2 Output regulation in switched systems
Till now, most of the literature on switched systems focused on the stability analysis for such
systems. However, as mentioned in section 1.1, in practice, in additional to stability
requirements, there is a need to find controllers that would achieve output regulation. Output
regulation is the problem of finding a control law by which the output of the system of
interest asymptotically tracks a reference signal and rejects undesired disturbance inputs,
both of which are deterministic signals generated by an exosystem. The control law must also
asymptotically stabilize the system whenever the exosignal is absent. The output regulation
problem for linear time-invariant systems has been solved by Francis, Wonham and Davison
in [50-52]. The main approach is based on the Internal Model Principle, which states that
asymptotic disturbance rejection can be achieved if a model of the exosystem generating the
disturbance is included in the stable closed loop system. For output regulation in nonlinear
systems, Isidori and Byrnes [53] and Byrnes et al. [54] presented necessary and sufficient
conditions to solve this problem. In general, the solvability of the problem is based on the
center manifold theory. Isidori and Byrnes found that it is possible to use a set of mixed
nonlinear partial differential and algebraic equations, called regulator equations, to
characterize the steady state of the system. The solution of the regulator equations provided a
feedward control to cancel the steady state tracking error. Based on the solution of the
regulator equations, both state feedback and error feedback control laws can be readily
synthesized to achieve asymptotic tracking and disturbance rejection for an exactly known
plant. Additional results have also been reported in the application of the theory to practical
10
design problems [55-60].
Although significant progress has been made in addressing the regulation problem for
both linear and nonlinear systems, the existing results cannot be applied to switched systems
since the vector fields of switched systems are discontinuous and non-smooth. The literature
on regulation for switched systems covers two main problems: exact output regulation and
optimal output regulation.
1) Exact output regulation: In this problem, it is desired to design a controller such that the
output of the resulting closed loop switched system asymptotically tracks a reference
signal and/or rejects disturbance inputs. The work of Devasia [61,62] and Sakurama [63]
are closely related to this problem. Devasia et al. studied the exact-output tracking
problem for linear switching systems, and necessary and sufficient conditions for the
existence of exact output-tracking bounded state trajectories are established. It is
assumed that the reference trajectories are generated by an exosystem, where either the
dynamics and initial conditions of the exosystem are completely known, or the dynamics
of the exosystem is known but the initial conditions are not known. Moreover, the
switching times are assumed known a priori, so the necessary and sufficient regulation
conditions are derived only for the elimination of switching induced output transients.
Therefore, the developed results cannot be applied to the case when switching is signal
driven and the switching times cannot be set a priori. Sakurama et al. recently discussed
the trajectory tracking problem for bimodal switched linear systems. It is assumed that
the reference trajectory for the switched system is known and the switching times are
measurable. An error variable and an error system are introduced, based on which the
tracking controller is designed using a Lyapunov-like function. However, the conditions
11
for existence of the desired regulator are sometimes very difficult to check. van de
Wouw and Pavlov [64] also studied the tracking problem for a class of piecewise affine
systems, where the desired system state trajectory is assumed to be known and
predefined. An observer-based putput-feedback control design method is proposed,
which consists of a feedforward, a piecewise affine feedback law and a model-based
observer. The synthesis conditions are formulated in terms of bilinear matrix inequalities.
2) Optimal output regulation: In this case, the controller is designed so that the resulting
switched closed loop system achieves a pre-specified optimal performance. To date, only
few results related to this topic have been reported in the literature based on 2l gain
analysis and H∞ closed loop system performance. Hespanha [65] studied the root-mean-
square (RMS) gain of switched linear systems and a bisection algorithm is proposed to
compute the slow switching RMS gain of the switched system in terms of the solutions
to a collection of Riccati equations. Lee and Dullerud [66] studied the disturbance
attenuation problem for discrete-time switched and markovian jump linear systems. An
exact condition for uniform stabilization and disturbance attenuation is given via the
union of an increasing family of linear matrix inequalities. A semidefinite programming
based controller synthesis method was proposed. Lee and Khargonekar [67] studied the
infinite-horizon suboptimal LQG control problem in discrete-time switched systems,
which aims at maintaining the average output variance below a given level subject to the
uniform exponential stability of the closed-loop system. The solution is given in the form
of a dynamic linear output feedback controller which not only observes the present mode
of the system but also recalls a finite number of past modes. These results are also
extended to Markovian jump linear systems. Rodrigues and Boukas [68] studied the
12
optimal inventory control problem in switched production systems, where the inventory
control problem is formulated as a switched H∞ control problem and the synthesis of a
state feedback controller that quadratically stabilizes the production dynamics and at the
same time rejects the external demand fluctuations is cast as a set of linear matrix
inequalities. Lin and Antsaklis [69] studied the asymptotic stability with a finite 2l
induced gain for a class of discrete-time switched linear control systems, where
sufficient conditions for the co-design of continuous variable controllers and discrete
event switching logic are proposed. Geromel et al. [70] investigated the stability
conditions and the output feedback regulator design for switched linear systems in both
the continuous and discrete-time domains, where optimal switching control strategy are
derived in terms of Lyapunov-Metzler inequalities. Some other results concerning 2l
gain analysis and H∞ control for certain classes of hybrid systems or hybrid control are
also available in [71-75].
1.2.3 Adaptive regulation in linear and nonlinear systems
When the disturbances or reference signals are deterministic signal with unknown properties,
adaptive regulation approaches are usually used. Most of the proposed approaches are based
on the internal model principle, where an internal model-based controller is designed online
to accommodate uncertainties in the exogenous signals. Marino and Tomei [76] designed an
adaptive internal model using an adaptive observer which provides estimates of the
observable states and of the exosystem unknown parameters. Ben Amara et al. [77,78]
proposed an adaptive regulation approach within a set of parameterized stabilizing controllers,
where the parameter in the stabilizing controller is tuned online to asymptotically reject
13
sinusoidal disturbances. A similar approach has been considered by Landau, et al. [79] to
adaptively reject narrow-band unknown disturbances in active suspension systems. Another
approach, based on the external model principle, involves the use of a model of the
disturbance outside the feedback loop to estimate the values of the disturbance input. Bodson
and Douglas [80] developed an adaptive feedforward compensator, where the feedforward
controller outside the feedback loop is designed to offset the steady state error induced by the
exogenous disturbance. The parameters in the feedforward controller are updated online
using the regulated error. Tomizuka et al. [81] developed an adaptive regulator, where a
stabilizing controller is computed first, then a model of the disturbance outside the feedback
loop is used to estimate the values of the disturbance input. The latter are then used in a
feedforward disturbance rejection algorithm to provide cancellation of the disturbance.
Serrani [82] developed a hybrid adaptive external model-based controller to reject harmonic
disturbances occurring at the input of a linear system. The adaptive unit is outside the stable
loop and possesses a self-tuning mechanism to achieve regulation even in the absence of
persistency of excitation.
Significant progress has also been reported in the literature in addressing the adaptive
output regulation problem for nonlinear systems. The internal model principle is still best
suited to this problem. Serrani et al. [83] studied the problem of output regulation for
nonlinear systems driven by a linear and neutrally stable exosystem whose frequencies are
not known a priori. The solution is realized by the parallel connection of a robust stabilizer
and a parameterized family of internal models which is adaptively tuned to reproduce the
signals generated by the exosytem and the nonlinearities of the plant. This adaptive design
methodology was also recognized as a useful way to cope with parameter uncertainties in the
14
exosystem. For instance, it was used in [84] for the solution of a global decentralized
regulation problem for a class of nonlinear systems with a neutrally stable uncertain
exosystem. Priscoli et al. [85] extended the existing theory of adaptive output regulation for
linear systems to nonlinear systems by allowing nonlinear internal models based on the
theory of adaptive observers for nonlinear systems [86,87] and the general nonequilibrium
theory [88]. Xi and Ding [89] discussed the adaptive output regulation problem for a class of
nonlinear systems with nonlinear exosystems. The nonlinear internal model is constructed
based on the circle criterion to produce the desired feedforward control term and
accommodate the nonlinearities in the exosystems. Ding [90] also studied the global
disturbance rejection problem in nonlinear systems, where the disturbance is assumed to be
sinusoidal with completely unknown phases, amplitude and frequencies, but the number of
distinct frequencies or the order of the corresponding unknown linear exosystem is known.
Despite the progress mentioned above, to the best of the author’s knowledge, no results
exist in the literature on the exact output regulation for switched systems in the presence of
unknown exogenous input signals.
1.3 Problem Statement
A formal statement of the main focus of this dissertation is given as follows. Consider a
switched bimodal linear system given by
:
,,
, r
xr r r ry yr r re er r r
x A x B u D wy C x D w
e C x D w
Σ
⎧Δ = + +⎪
= +⎨⎪ = +⎩
(1.1)
15
where xΔ denotes dxdt
in the continuous-time case and ( 1)x k + in the discrete-time case,
nx∈R is the state vector, u∈R is the control input, y∈R is the measurement signal to be
fed to the controller, hrw = R is the exogenous input signal, e∈R is the performance
variable to be regulated and is assumed to be measurable, and { }1, 2r∈ is the index of the
system rΣ under consideration at time t or k . The switching between the systems 1Σ and
2Σ is performed according to the value of the performance variable e as
1 if ,2 if ,
e er r re er r r
e C x D wr
e C x D wδδ
⎧ = + ≤= ⎨
= + >⎩ (1.2)
where δ is constant and 0δ ≠ . For the plant (1.1) subject to the exogenous input rw , it is
desired to design a controller that yields internal stability and asymptotic exact output
regulation, namely as time goes to infinity, the value of the performance variable e should
go to zero. This thesis concerns the development of regulator synthesis methods for the
above regulation problem in switched systems (1.1) under the switching law defined by (1.2).
More specifically, this thesis aims mainly at:
1) The development of a regulator design method to reject known deterministic exogenous
inputs in the switched system (1.1).
2) The development of adaptive regulator design method to reject sinusoidal exogenous
inputs with unknown properties in the switched system (1.1).
16
1.4 Contributions of the Thesis
The regulation problems in bimodal linear systems presented in this thesis are solved within
sets of Q -parameterized controllers, where the parameterizing dynamics are designed not
only to make the switched closed loop system stable, but also to achieve exact output
regulation in the switched system. The proposed sets of parameterized controllers are
constructed in two steps. The first step is based on constructing a switched observer-based
state feedback central controller for the switched linear system. The second step involves
augmenting the switched central controller with additional dynamics, referred to as the Q
parameter, to construct a parameterized set of switched controllers. Then, based on the sets of
Q -parameterized controllers, four regulator synthesis methods are developed to deal with
different cases. The original contributions presented in this thesis are summarized as follows:
1) Development of a regulator synthesis procedure for bimodal systems based on bilinear
matrix inequalities. A parameterized regulator design method based on solving bilinear
matrix inequalities is proposed for continuous-time switched bimodal linear systems
subject to known deterministic exogenous inputs. A set of Q -parameterized controllers
is considered, where, unlike the case of controller parameterization in linear systems, no
stability or structural constraints are imposed on the Q parameter. Sufficient regulation
conditions are derived for the resulting parameterized switched closed loop system and a
corresponding regulator synthesis approach is proposed based on solving properly
formulated bilinear matrix inequalities. The advantage of considering Q -parameterized
controllers where no stability or structural constraints are imposed on the Q parameter is
that it offers more flexibility in the design of the desired regulators. However, the
17
synthesis of the regulators based on solving bilinear matrix inequalities results only in
local optimal solutions in the search algorithm. Moreover, a general structure for the Q
parameter is not convenient to work with when it comes to tuning the Q parameter in the
case of adaptive regulation against unknown exogenous sinusoidal inputs.
2) Development of a regulator synthesis procedure in bimodal systems based on linear
matrix inequalities. Parameterized regulator design methods based on solving properly
formulated linear matrix inequalities are proposed for continuous-time switched bimodal
linear systems subject to known deterministic sinusoidal exogenous inputs. The design
of the proposed regulators involves two main steps. First, a set of observer-based Q -
parameterized stabilizing controllers for the switched system is constructed. The Q
parameters are represented as linear combination of stable basis functions. The stability
of the resulting closed loop switched system with the stabilizing Q -parameterized
controllers is analyzed. Second, regulation conditions for each of the two subsystems in
the resulting bimodal switched closed loop system are presented in both the state space
domain and the frequency domain. Then, regulation conditions and regulator synthesis
methods for the switched closed loop system are proposed using a common Lyapunov
function approach and a multiple Lyapunov function approach. There are two main
advantages in considering Q parameters that are expressed as linear combinations of
stable basis functions. First, the regulator synthesis procedure for the case of known
exogenous inputs can be developed based on solving a set of linear matrix inequalities.
Second, an adaptive online tuning algorithm for the free Q parameters can be
conveniently developed in the case of unknown sinusoidal exogenous inputs.
18
3) Development of the regulator synthesis procedure in bimodal systems within a class of
2H controller. A parameterized regulator synthesis method is proposed for switched
discrete-time bimodal linear systems where it is desired to achieve regulation
simultaneously against known deterministic sinusoidal exogenous inputs and unknown
random exogenous inputs. First, a Q -parameterized set of stabilizing controllers for the
discrete-time switched system is constructed and that yield regulation against the known
deterministic sinusoidal exogenous inputs. A subset of 2H controllers is then identified
within the already constructed set of regulators to effectively deal with the random
exogenous input signals. A corresponding regulator synthesis method is developed by
solving properly formulated linear matrix inequalities. A switched mechanical system
experimental setup is used to successfully validate the proposed regulator design
approach.
4) Development of an adaptive regulation approach in bimodal systems. An adaptive
regulation approach is proposed for discrete-time switched bimodal linear systems,
where it is desired to achieve regulation against unknown deterministic sinusoidal
exogenous input signals. First, assuming the properties of the exogenous input signal are
known, a set of observer-based Q -parameterized stabilizing controllers for the switched
system is constructed, and a sufficient regulation condition for the resulting switched
closed loop system is presented. To deal with the case of exogenous sinusoidal inputs
with unknown properties, an adaptation algorithm is developed to tune the Q parameter
in the expression of the parameterized set of controllers. The tuning is such that the
19
tuned Q parameter converges to the desired Q parameter that guarantees regulation in
the switched system. The proposed adaptation approach is successfully validated on an
experimental setup involving a switched mechanical system.
The four contributions are each treated in Chapters 2, 3, 4 and 5, respectively, and are
summarized in Table 1.1.
1.5 Organization of the Thesis
The remainder of this thesis is organized as follows. Chapter 2 presents a parameterized
regulator synthesis method for continuous-time switched bimodal linear systems against
known sinusoidal exogenous inputs, where no stability or structural constraints are imposed
on the Q parameter. A regulator synthesis algorithm is developed based on solving bilinear
matrix inequalities. Chapter 3 presents a regulator synthesis approach along the same lines as
that outlined in chapter 2, but where the Q parameter is constrained to be a linear
combination of stable basis functions. A regulator synthesis algorithm is developed based on
solving properly formulated linear matrix inequalities. Two regulator synthesis approaches
are proposed. The first approach is based on considering a common Lyapunov function for
the closed loop switched system, whereas the second is based on considering multiple
Lyapunov functions for the switched closed loop system. In chapter 4, the problem of output
regulation in discrete-time switched bimodal systems against known deterministic sinusoidal
exogenous inputs and unknown random exogenous inputs is solved within a class of 2H
controllers. Chapter 5 presents an adaptive regulation method for discrete-time switched
bimodal linear systems against unknown sinusoidal exogenous inputs. Chapters 4 and 5 also
20
include experimental results that validate the regulation approaches presented in those two
chapters. The experimental results are obtained using a switched mechanical system
experimental setup that is motivated by the flying height regulation problem in hard disk
drives.
21
Contributions
1st 2nd 3rd 4th Chapter
2 3 4 5 Known √ √ √ √ Continuous-time √ √ Plant model Discrete-time √ √ Deterministic known √ √ √ Deterministic unknown √
PLA
NT
Exogenous inputs Random √ Unconstrained √
Q Parameter Linear combination of stable basis functions √ √ √ Common Lyapunov function √ √ √
Closed loop system Lyapunov function Multiple Lyapunov functions √ √ Derived based on frequency domain regulation conditions (interpolation conditions) for each of the subsystems in the switched closed loop system
√ √ √
Regulation conditions Derived based on state space domain regulation conditions for each of the subsystems in the switched closed loop system
√ √
BMIs √ √
CO
NT
RO
LL
ER
Regulator synthesis algorithm LMIs √ √ √
EXPERIMENTAL RESULTS √ √
Table 1.1 Summary of the contributions.
22
CHAPTER 2
Parameterized Regulator Synthesis for
Switched Bimodal Systems Based on Bilinear
Matrix Inequalities
2.1. Introduction
In this chapter, a controller design approach is proposed for switched bimodal linear systems
where it is desired to reject known disturbance signals and track known reference inputs
simultaneously. Switching in the bimodal system is defined by a time-varying switching
surface. A regulator synthesis method based on solving bilinear matrix inequalities (BMIs) is
presented. The proposed regulator design approach consists of three steps. In the first step, a
switched observer-based state-feedback central controller is constructed for the switched
linear system. The second step involves augmenting the switched central-controller with
additional dynamics to construct a parameterized set of switched controllers. Stability
analysis of the resulting switched closed loop system is then presented. In the third step, two
sufficient regulation conditions are derived for the switched closed loop system. The first
sufficient condition is derived based on the input-output stability property of the switched
23
closed loop system. The second sufficient condition for regulation is derived by transforming
the forced switched closed loop system into an unforced impulsive switched system. As such,
the regulation problem is transformed into a stability analysis problem for the impulsive
switched system. Based on the parameterized controller structure and the derived regulation
conditions, proper bilinear matrix inequalities are formulated and a regulator synthesis
method is proposed. The main advantage of the proposed regulator synthesis approach is that
it offers a numerical procedure that can be implemented and used to develop the desired
switched regulator.
The rest of the chapter is organized as follows. In section 2.2, the general regulation
problem for switched bimodal linear systems is presented. In section 2.3, the construction of
a parameterized set of switched controllers for the switched system is discussed and the
stability properties of the resulting closed loop switched system are analyzed. Regulation
conditions for the switched system are presented in section 2.4 and a regulator synthesis
algorithm for switched system is proposed in section 2.5. The controller design method is
illustrated in section 2.6 using a numerical example, followed by the conclusion in section
2.7.
Notation: In the following, the state space representations :x Ax Buy Cx Du= +⎧
Σ ⎨ = +⎩ and
:A BC D⎡ ⎤
Σ ⎢ ⎥⎣ ⎦
will be used interchangeably for the sake of brevity. For a given 1 N× matrix X ,
X denotes the 1N × vector with the same entries as those of TX .
24
2.2. Regulation problem for switched bimodal systems
Consider the continuous-time switched system given by the state space representation:
0
0
, (0) , , (0) ,
, :
,
1 if ,
2 if ,
xr r r r
r r r r ry yr r r
r e er r
e er r
e er r
x A x B u D w x xw H w w w
y C x D we C x D w
e C x D wr
e C x D wδδ
= + + == =
= +∑
= +
⎧ = + ≤= ⎨
= + >⎩
⎧⎪⎪⎪⎪⎨⎪⎪⎪⎪⎩
(2.1)
where nx∈R is the state vector, u∈R is the control input, y∈R is the measurement signal
to be fed to the controller, e∈R is the performance variable to be regulated and is assumed
to be measurable, hrw ∈R is the state vector of the exogenous systems generating the signal
rw , h hrH ×∈R have simple eigenvalues on the imaginary axis, { }1,2r∈ is the index of the
system rΣ under consideration at time t , and δ is a constant satisfying 0δ > . The
switching between the systems 1Σ and 2Σ is performed according to the value of the
performance variable e , and is determined based on the location of x with respect to a
switching surface S given by:
{ }: n e er rS x e C x D w δ= ∈ = + =R . (2.2)
The switching surface S is not fixed but changes with time given that the term er rD w is in
general a time-varying term. The switching between the two modes takes place as follows: if
1r = and e becomes strictly greater than δ , then the mode switches to 2r = ; and if 2r =
and e becomes less than or equal to δ , then the mode switches to 1r = . In the following, it
is assumed that for any given 0t ≥ , the system must operate in only one of the two modes
25
corresponding to { }1,2r∈ . This assumption is motivated by physical considerations in some
applications of interest, such as the example presented at the end of the chapter or the system
treated in [2,3]. For the switched system (2.1), it is desired to construct an output feedback
controller to regulate the performance variable e of the switched system against the external
input signal rw . Given the switching nature of the plant, the output feedback controller is
also chosen to be a switching feedback controller r , { }1, 2r∈ , with a state space
representation given by:
,:
,
c cc r c r
r cr c
x A x B y
u C x
⎧ = +⎪⎨
=⎪⎩ (2.3)
where cncx ∈R , and where the switching among controllers is to obey the same rule given in
(2.1) for switching between the two plant models. Therefore, the resulting closed loop system
is given by:
,
: 0 ,
1 if ,
2 if .
c xr r r r
rc y c c yc cr r r r r
cl e er r r
c
e er r
e er r
x xA B C Dw
x xB C A B D
xe C D w
x
e C x D wr
e C x D wδδ
⎧ ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤= +⎪ ⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦⎣ ⎦ ⎣ ⎦⎪⎪ ⎡ ⎤⎪ ⎡ ⎤∑ = +⎨ ⎢ ⎥⎣ ⎦
⎣ ⎦⎪⎪ ⎧ = + ≤⎪ = ⎨⎪ = + >⎩⎩
(2.4)
The output feedback regulation problem for the switched system (2.1) can be stated as
follows.
Output feedback regulation problem:
Given the switched system (2.1), where switching is performed according to the rule given in
(2.1), find a switched output feedback controller of the form (2.3) such that the resulting
closed loop system satisfies the following conditions:
26
1) Internal stability: With (0) 0rw = , the equilibrium point 0TT T
cx x⎡ ⎤ =⎣ ⎦ of the unforced
switched closed loop system is asymptotically stable.
2) Output regulation: For any (0) (0) cT nT T n
cx x⎡ ⎤ ∈ ×⎣ ⎦ R R and (0) hrw ∈R , the response ( )e t
of the closed loop system involving the switched system (2.1) and the switched controller
(2.3) satisfies:
lim ( ) 0t
e t→∞
= .
The following section presents a framework within which the design of an output
feedback regulator for the switched system (2.1), including the development of regulation
conditions and a regulator synthesis procedure, will be conducted.
2.3. Parameterization of a set of controllers
The controller design approach presented in this chapter relies on the construction of a
parameterized set of output feedback controllers for the switched system (2.1). In this section,
the construction of such a set is first discussed, followed by an analysis of the stability
properties of the resulting switched closed loop system. The construction of a parameterized
set of switched controllers involves two steps. The first step consists of designing a central
controller in the form of an observer-based state feedback controller. The second step
involves augmenting the central controller with additional dynamics to construct a
parameterized set of controllers.
27
2.3.1 Observer-based state feedback controller
Consider the following observer-based state feedback controller for the switched system (2.1):
( ) 0ˆ ˆ ˆ ˆ ˆ, (0) ,:
ˆ,o r r rr
r
x A x B u L y y x xu K x
⎧ = + + − =⎪⎨
=⎪⎩ (2.5)
where ˆ nx∈R is the estimate of the plant state vector x and ˆ ˆyry C x= is an estimate of the
plant output y . The mode { }1, 2r∈ is determined according to the rule given in (2.4).
Moreover, it is assumed that there are no impulsive changes in the controller states at the
switching times.
2.3.2 Parameterized output feedback controller
The construction of a parameterized set of controllers for the switched system is based on
considering, for each { }1, 2r∈ , a linear fractional transformation involving a fixed system
rJ and a proper system rQ as shown in Figure 2.1. The proposed controllers are similar to
the parameterized stabilizing controllers for linear systems [91-93], but where no stability or
structural assumptions are placed on the system rQ . The state space representation of the
system rJ is given by:
0
1
ˆ ˆ ˆ ˆ ( ) , (0) ,( )
1 0ˆ: ,
ˆ ˆ ,
y y xr r r r r r Qr r r r
qr r Q
yr
x A B K L C x L y y x xB L D D
J u K x y
y y C x y×
⎧ = + + − + =⎡ ⎤+⎣ ⎦⎪⎪ = + ⎡ ⎤⎨ ⎣ ⎦⎪
− = −⎪⎩
(2.6)
and the system rQ is given by:
( ) 0ˆ , (0) ,
:,
r r
r
Q Q Q Q Q Q
r QQ Q
q q
x A x B y y x x
Q Cy x
I ×
⎧ = + − =⎪⎪
⎡ ⎤⎨= ⎢ ⎥⎪⎢ ⎥⎪ ⎣ ⎦⎩
(2.7)
28
Figure 2.1. Closed loop system with Q -parameterized controller.
where qQx ∈ and 1q
Qy +∈ . In particular, throughout the rest of the chapter, the system rQ
is such that the matrices rQA ,
rQB and rQC also change according to r .
Remark 2.1
The state space representation of the system rJ given in (2.6) differs from that traditionally
used in the construction of parameterized sets of stabilizing controllers for linear systems in
that the state equation in (2.6) contains the additional term ( )y xr r r QL D D x+ . The presence of
this term makes it possible to derive the sufficient conditions for regulation presented in
Theorem 2.1.
By combining the systems rJ and rQ in (2.6) and (2.7), the state space representation of the
regulator is then given by (2.3), where
rQ
ˆy y− Qy
rJ
Plant r∑
rw e
u y
29
,
ˆ, ,
, , 0.
r
r r
rr
y y xr r r r r r Q r r rc
c r yQ Q r Q
rc cr r r Q r c
Q
A B K L C B C L D Dxx A
x B C A
LB C K C D
B
⎡ ⎤+ + + +⎡ ⎤= = ⎢ ⎥⎢ ⎥
⎢ ⎥⎣ ⎦ ⎣ ⎦−⎡ ⎤
⎡ ⎤= = =⎢ ⎥ ⎣ ⎦−⎣ ⎦
(2.8)
Let ˆx x x= − denote the state estimation error. It follows that the resulting closed loop
system involving the plant (2.1) and the regulator given by (2.3) and (2.8) can be written as
follows:
0 ,0
:
[ 0 0] .
r
r r r
xr r r r r r Q r
y x y x yr r r r r r r r r r
y yQ Q r Q Q Q rcl
r
e er r
Q
A B K B K B Cx x Dx A L C D L D x D L D wx B C A x B D
xe C x D w
x
⎧ +⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤⎪ ⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥= + + + − −⎪ ⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥⎪ ⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥ −⎪⎣ ⎦ ⎣ ⎦ ⎣ ⎦⎣ ⎦∑ ⎨
⎡ ⎤⎪⎢ ⎥⎪ = +⎢ ⎥⎪⎢ ⎥⎪ ⎣ ⎦⎩
(2.9)
Let 1
TT T TQ N
x x xχ×
⎡ ⎤= ⎣ ⎦ denote the state vector of the resulting closed loop system with
2N n q= + , and let
ˆ 00
r
r r
r r r r r r Q
y x yr r r r r r r
yQ r Q
A B K B K B C
A A L C D L DB C A
+⎡ ⎤⎢ ⎥
= + +⎢ ⎥⎢ ⎥⎣ ⎦
,
( ) ( ) ( )r
TTx x y T y Tr r r r r Q rE D D L D B D⎡ ⎤= − + −⎢ ⎥⎣ ⎦
,
0 0ex eC C⎡ ⎤= ⎣ ⎦ .
The resulting parameterized switched closed loop system dynamics can then be expressed as:
ˆ , ,
:1 if ,
2 if .
r r rex e
r rclr ex e
r rex e
r r
A E we C D w
e C D wr
e C D w
χ χχ
χ δχ δ
⎧ = +⎪
= +⎪∑ ⎨⎧ = + ≤⎪ = ⎨⎪ = + >⎩⎩
(2.10)
30
In the following, both the internal stability as well as the input-output stability properties of
the system (2.10) are analyzed.
2.3.3. Stability of the parameterized switched closed loop system
In this section, it is desired to study the stability properties of the system (2.10). The stability
analysis involves two steps, namely internal stability analysis and input-output stability
analysis. In the first step, the internal stability of the closed loop system is studied by
considering the system (2.10) in the absence of the signal rw , and studying the stability
properties of the origin for the resulting unforced switched system. The second step builds on
the internal stability results and presents input-output stability results for the system (2.10).
Consider first the system (2.10) in the absence of the signal rw . The state equation for the
resulting system is given by:
ˆrAχ χ= . (2.11)
Note that there are no impulsive changes in the state variables of the above switched system
at the switching times. The internal stability of the system (2.11) is then given by the
following well known result.
Lemma 2.1 [48]
If there exists a matrix 0TP P= > and a constant α such that the following inequalities hold:
{ }ˆ ˆ 2 0, 0, 1,2 ,Tr rA P PA P rα α+ + < > ∈ (2.12)
then the origin is an exponentially stable equilibrium point for the switched system (3.8) with
arbitrary switching, and:
( ) (0)tt ce αχ χ−< , (2.13)
31
where max
min
( )( )PcP
λλ
= .
Remark 2.2
The matrix inequality (2.12) can be solved using numerical algorithms. Here, we search for a
common Lyapunov function for the switched systems (2.11) to guarantee the stability of the
origin for the system (2.11) under arbitrary switching, since the switching surface
{ } : n e er rS x e C x D w δ= ∈ = + =R is not a fixed function of the state vector x and changes
with er rD w .
In the following, we will consider the input-output stability properties of the system (2.10)
with the external input signal rw . Let { }m 1,20
max ( )r rrt
w E w t∈≥
= . Then we have the following
stability result.
Lemma 2.2
Assume the origin for the switched system (2.11) is an exponentially stable equilibrium point.
Then the state vector in (2.10) is bounded, and the states will ultimately evolve inside the
bounded set Bβ given by:
{ }: ,NBβ χ χ β= ∈ ≤R (2.14)
where mcwβ εα
= + , and where ε is an arbitrary small positive number, i.e. 0 1ε< << .
32
Proof: Let 0( , )t tφ be the state transition matrix of the switched system (2.11) [49]. Without
loss of generality, let 0 0t = . The state trajectory of the system (2.10) is given by:
0( ) ( ,0) (0) ( , ) ( )
t
r rt t t E w dχ φ χ φ τ τ τ= + ∫ .
Based on (2.13), we have
( ,0) (0) (0) tt c e αφ χ χ −≤ .
It follows that:
0
( )m0
m
( ) ( ,0) (0) ( , ) ( )
(0)
1 (0) .
t
r r
tt t
tt
t t t E w d
c e ce w d
ec e cw
α α τ
αα
χ φ χ φ τ τ τ
χ τ
χα
− − −
−−
≤ +
≤ +
−≤ +
∫
∫
Hence, ( )tχ is ultimately bounded and we have:
mlim ( )t
cwtχ β εα→∞
≤ = − . (2.15)
Therefore, after a long enough finite time, the state trajectory will enter into the bounded set
Bβ given by (2.14) and will continue to evolve in Bβ thereafter.
The input-output stability of the closed loop system follows immediately from the above
result.
2.4. Regulation conditions for the switched system
The purpose of introducing the parameterized controllers is to find, for each { }1, 2r∈ ,
appropriate rQA ,
rQB and rQC in (2.7) to solve the output feedback regulation problem for the
33
switched system (2.10). The solution to the regulation problem for the switched system (2.10)
is presented in two steps. In the first step, regulation conditions for each of the individual
closed loop systems clrΣ , { }1, 2r∈ , are presented. In the second step, two sufficient
conditions for regulation in the switched closed loop systems clrΣ are derived. To derive the
sufficient conditions for regulation, a coordinate transformation is defined first, allowing the
forced switched closed loop system to be transformed into an unforced impulsive switched
system. Hence, the regulation problem is transformed into a stability analysis problem for the
origin of the resulting impulsive switched system. Conditions for achieving asymptotic
stability in the new impulsive switched system are then presented, which is equivalent to
achieving regulation in the original switched closed loop system.
2.4.1. Regulation conditions for 1clΣ and 2
clΣ
Let c
r r rr c y c
r r r
A B CA
B C A⎡ ⎤
= ⎢ ⎥⎣ ⎦
and xr
r c yr r
DE
B D⎡ ⎤
= ⎢ ⎥⎣ ⎦
in (2.4). For each of the systems rΣ , { }1, 2r∈ ,
and using a controller of the form (2.3), regulation conditions are given by the following
lemma.
Lemma 2.3 [52]:
For each { }1, 2r∈ , consider the system rΣ in (2.1) and a controller r in (2.3), and assume
that rA is a stability matrix. Then, for each { }1, 2r∈ , clrΣ in (2.4) achieves regulation if and
only if there exists a matrix rΠ that solves the linear matrix equations:
,
0 .r r r r r
ex er r
H A E
C D
Π = Π +
= Π + (2.16)
34
Similarly, using the parameterized feedback regulator r given in (2.3) and (2.8), regulation
conditions for each of the closed loop systems clrΣ , { }1, 2r∈ , in (2.10) can be derived and are
presented in the following Theorem.
Theorem 2.1
Assume that for each { }1, 2r∈ , the closed loop system clrΣ given in (2.10) is exponentially
stable. Then each of the systems 1clΣ and 2
clΣ achieves regulation only if, for each { }1, 2r∈ ,
there exists a pair of matrices ( ),r rΩ Ψ which satisfy the following equations:
,
0 .
xr r r r r r r
e er r
H A B D
C D
Ω = Ω + Ψ +
= Ω + (2.17)
Furthermore, if the linear matrix equations in (2.17) admit a solution ( ),r rΩ Ψ , and the
matrices rQA and
rQC in (2.6) are taken to be of the form
,
,r r
r
yQ r Q r
Q r r r
A H B D
C K
= +
= Ψ − Ω (2.18)
then, for each { }1, 2r∈ , clrΣ in (2.10) achieves regulation.
Proof: Equation (2.16) can be rewritten as
,
0 0 .
c xr r r r
r r rc y c c yr r r r r
e er r
A B C DH
B C A B D
C D
⎡ ⎤ ⎡ ⎤Π = Π +⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦⎡ ⎤= Π +⎣ ⎦
Partition rΠ as rr
r
Ω⎡ ⎤Π = ⎢ ⎥Φ⎣ ⎦
. Then (2.16) is the same as
35
( ),
,
0 .
c xr r r r r r r r
c c y yr r r r r r r r
e er r
H A B C D
H A B C D
C D
Ω = Ω + Φ +
Φ = Φ + Ω +
= Ω +
(2.19)
Letting cr r rCΨ = Φ , equation (2.17) follows immediately.
To prove the second part of the theorem, note that by assumption, for each { }1, 2r∈ , ˆrA
in (2.10) is a stability matrix. Moreover, since rA is related to ˆrA by a similarity
transformation, then rA is also a stability matrix. Now suppose rΩ and rΨ satisfy (2.17)
and that rQA and
rQC satisfy (2.18). To show that the closed loop system achieves regulation,
let rr I
Ω⎡ ⎤Φ = ⎢ ⎥
⎣ ⎦. Using the expression for c
rA in (2.8) and substituting in the expression for
rQA in (2.18) yields:
,
.00
r
r r
rr
y y xr r r r r r Q r r rc
r y yQ r r Q r
xrr y yr r
r Q r rQr
A B K L C B C L D DA
B C H B D
LBA DK C C D
BH
⎡ ⎤+ + + += ⎢ ⎥+⎢ ⎥⎣ ⎦
⎡ ⎤⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤= + + ⎢ ⎥⎢ ⎥ ⎢ ⎥ ⎣ ⎦⎣ ⎦⎣ ⎦⎣ ⎦ ⎣ ⎦
It follows that:
( )
( ) ( )
,0
.
r
r
r
xrr r r r Qc y yr r r
r r r r rQr
xr r r r r Q r c y y
r r r r
r
LB K B CA DA C D
BH
A B K C DB C D
H
Ω + ⎡ ⎤⎡ ⎤Ω + ⎡ ⎤Φ = + + Ω +⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣ ⎦⎣ ⎦ ⎣ ⎦⎡ ⎤Ω + Ω + +
= − Ω +⎢ ⎥⎢ ⎥⎣ ⎦
Given that rr Q r rC KΨ = + Ω from (2.18), then based on (2.17) we have:
( )
( )
,
.
xc c y yr r r r rr r r r r r
r
c y yr r r r r r
A B DA B C D
H
H B C D
⎡ ⎤Ω + Ψ +Φ = − Ω +⎢ ⎥
⎣ ⎦
= Φ − Ω +
36
Therefore, equation (2.19) is satisfied, which implies (2.16) is satisfied and regulation can be
achieved.
The results presented above represent regulation conditions for each of the closed loop
systems in (2.10), individually. In the following sections, regulation conditions for the system
(2.10) subject to switching are discussed.
2.4.2. Equivalent impulsive switched closed loop system model
Based on Lemma 2.3, and using a properly defined coordinate transformation, the forced
switched closed loop system (2.10) can be transformed into an unforced impulsive switched
system. To introduce the appropriate coordinate transformation, note that the original
regulation condition (2.16) can be rewritten as:
ˆˆ ˆ ,ˆ 0 ,
r r r r r
ex er r
H A E
C D
Π = Π +
= Π + (2.20)
where
0 0ˆ 0
0 0r r
II I
I
⎡ ⎤⎢ ⎥Π = − Π⎢ ⎥⎢ ⎥⎣ ⎦
. (2.21)
Consider now the following coordinate transformation:
ˆ .r rwχ χ= −Π (2.22)
Using (2.20), the forced closed loop switched system (2.10) can be transformed into an
unforced switched system given by:
37
ˆ ,,
:1 if ,2 if .
rex
clr ex
ex
Ae C
e Cr
e C
χ χχ
χ δχ δ
⎧ =⎪
=⎪∑ ⎨⎧ = ≤⎪ = ⎨⎪ = >⎩⎩
(2.23)
It should be noted from (2.22) that the coordinate transformation varies depending on the
value of { }1,2r∈ . Consequently, the states ( )tχ in the new system (2.23) undergo impulsive
changes at the switching times. Let sT be the set of switching times kτ , 1, 2,k = … .
Therefore, for a given switching time k sτ ∈T , we have that for kt τ= :
( ) ( ) ( )t t tχ χ χ+ −= + Δ ,
where ( ) ( ) ( ) ( )
ˆ ˆ( ) ( ) ( )r t r t r t r t
t w t w tχ − − + +Δ = Π −Π . Hence, the impulsive switched closed loop
system is given by:
ˆ ,
( ) ( ) ( ), , : ,
1 if ,
2 if .
r
scl exr
ex
ex
A
t t t te C
e Cr
e C
χ χ
χ χ χχ
χ δχ δ
+ −
⎧ =⎪
= + Δ ∈⎪⎪∑ =⎨⎪ ⎧ = ≤⎪ = ⎨⎪ = >⎩⎩
T (2.24)
In the new coordinate system, the switching surface S given in (2.2) is fixed and is
expressed as follows:
{ }: .N exS e Cχ χ δ= ∈ = =R (2.25)
Using the coordinate transformation in (2.22), the original output feedback regulation
problem for the switched system (2.10) is transformed into an asymptotic stability analysis
problem for the origin of the unforced impulsive switched system (2.24).
38
2.4.3. Regulation conditions for the switched system
First, let 0δ denote the distance between the switching surface S and the origin. Then we
have 0 exCδ δ= . Define:
{ }m 1,20
ˆmax ( )r rrt
w tβ∈≥
Δ = Π . (2.26)
Let mγ β β ε= + Δ + , where 1ε , and define the set Bγ as follows:
{ }:NBγ χ χ γ= ∈ ≤R . (2.27)
Using the definition for χ in (2.22), and the fact that the original state vector χ will
ultimately evolve inside a bounded set Bβ given by (2.14), it follows that there exists a finite
time Tγ at which χ will enter the set Bγ and continue to evolve in Bγ thereafter. A
sufficient condition for regulation is then given as follows:
Theorem 2.2
Assume that the switched closed loop system clrΣ given in (2.10) is internally stable under
arbitrary switching, and that for each { }1, 2r∈ , clrΣ given in (2.10) achieves output
regulation. If
0 ,δ γ> (2.28)
then the origin is an asymptotically stable equilibrium point for the impulsive switched
system (2.24), implying that the switched system (2.10) achieves regulation.
Proof: If 0δ γ> , then the switching surface S given by (2.25) does not intersect the bounded
set Bγ . The trajectory of the state vector χ will enter the set Bγ at t Tγ= and continue to
39
evolve inside Bγ thereafter. Consequently, there will be no more switching. Since the system
clrΣ that is active in the half-space containing Bγ is such that ˆ
rA is a stability matrix, and
since 0 Bγ∈ , the trajectory of the state vector χ asymptotically converges to the origin,
implying that regulation in the original switched closed loop system (2.10) is achieved.
If condition (2.28) is not satisfied, i.e. 0δ γ≤ , then the switching surface S must intersect
with the bounded set Bγ . To present regulation conditions in this case, define the matrices
11
ˆ
ex
s ex
CA
C A
⎡ ⎤= ⎢ ⎥⎢ ⎥⎣ ⎦
, 22
ˆ
ex
s ex
CA
C A
⎡ ⎤= ⎢ ⎥⎢ ⎥⎣ ⎦
and 0
bδ⎡ ⎤
= ⎢ ⎥⎣ ⎦
. Sufficient conditions for regulation are presented
in the following theorem.
Theorem 2.3
Assume the switched closed loop system clrΣ given in (2.10) is internally stable under
arbitrary switching and that, for each { }1, 2r∈ , clrΣ given in (2.10) achieves output
regulation. Moreover, assume that 0δ γ≤ , 1 2( ) ( ) 2s srank A rank A= = , and that the direction
of vector exC is not parallel to that of 1ˆexC A and 2
ˆexC A . If the following conditions are
satisfied,
( )( ) 1 2
ˆ 0,
,
1 if 0,2 if 0,
Tex exr
T Tsr sr
C C A
b A A b
rr
γ
δδ
−
<
>
= >⎧⎨ = <⎩
(2.29)
or
40
( )( ) 1 2
ˆ 0,
,
2 if 0,1 if 0,
Tex ex
r
T Tsr sr
C C A
b A A b
rr
γ
δδ
−
>
>
= >⎧⎨ = <⎩
(2.30)
then the origin is an asymptotically stable equilibrium point for the impulsive switched
system (2.24), implying that the switched system (2.10) achieves regulation.
Proof: The following proof is presented for the case of 0δ > . The case of 0δ < is treated
using the same ideas. Let region 1 and region 2 denote the half-spaces where the systems 1clΣ
and 2clΣ are active, respectively (See Figures 2.2 and 2.3). Hence, the origin is in region 1
where the system 1clΣ is active. Define the hypersurfaces { }1 1
ˆ: 0N exvS C Aχ χ= ∈ =R and
{ }2 2ˆ: 0N ex
vS C Aχ χ= ∈ =R . Then for all points in 1vS , 1ˆ 0exe C A χ= = and for all points in
2vS , 2ˆ 0exe C A χ= = . Since, by assumption, 0δ γ≤ , the switching surface S intersects with
the bounded set Bγ . Let { }: and N exS Cγ χ χ δ χ γ= ∈ = ≤R be the intersection of the
switching surface S with Bγ . By assumption, the direction of vector exC is not parallel to
that of 1ˆexC A and 2
ˆexC A . Therefore, the switching surface S must intersect with either of the
hypersurfaces 1vS or 2vS . Define the intersection of S with the hypersurfaces 1vS and 2vS as
{ }1 1ˆ: 0 and N ex exL C A Cχ χ χ δ= ∈ = =R
and
{ }2 2ˆ: 0 and N ex exL C A Cχ χ χ δ= ∈ = =R .
41
For each { }1, 2r∈ , let minr
r Ll
χχ
∈= denote the distance between the origin and the set rL .
Then, rl can be obtained by solving the problem of minimizing χ subject to the constraint
srA bχ = . Since, by assumption, 1 2( ) ( ) 2s srank A rank A= = , the solution to the above
problem is ( ) 1T Tr sr srl b A A b
−= , { }1, 2r∈ . If ( ) 12 2T T
r sr srl b A A b γ−
= > , then the sets rL ,
{ }1, 2r∈ , are outside the bounded set Bγ . In this situation, the set Sγ does not intersect with
the sets 1vS or 2vS . If ( )1ˆ 0
Tex exC C A < and ( ) 12 2
1 1 1T T
s sl b A A b γ−
= > , then the set Sγ will be
located in one of the two half-spaces defined by the hypersurface 1vS and where all the points
satisfy 1ˆ 0exe C A χ= < . Similarly, if ( )2
ˆ 0T
ex exC C A > and ( ) 12 22 2 2
T Ts sl b A A b γ
−= > , then the
set Sγ will be located in one of the two half-spaces defined by the hypersurface 2vS and
where all the points satisfy 2ˆ 0exe C A χ= > . By deriving conditions on e , and using the
stability properties of 1clΣ and 2
clΣ , it is possible to conclude about the asymptotic stability of
the origin of the impulsive switched closed loop system (2.24). In the following, two cases
are considered in detail.
• Case 1: ( )1ˆ 0
Tex exC C A < and ( ) 1 2
1 1T T
s sb A A b γ−
> (See Figure 2.2). In this case, all the
state trajectories leaving the set Sγ and entering region 1 have 0e < . Therefore, if the
state trajectory leaves the set Sγ and enters region 1, then it will not hit it again. Two
possible cases can be considered here depending on whether the state trajectory enters the
bounded set Bγ from region 1 or region 2 at time Tγ . In the case the state trajectory enters
the set Bγ from region 1, then the state trajectory cannot hit the switching surface S since
42
0e < for all states on the switching surface within the set Bγ . Therefore, the state
trajectory is confined to evolve in region 1 inside the set Bγ . Since 1clΣ is a stable system,
the state trajectory will converge to the origin asymptotically, which implies that
regulation for the switched system (2.10) is achieved. Consider now the case where the
state trajectory enters the set Bγ from region 2. Since the state trajectory is confined to
evolve inside the set Bγ and since the system 2clΣ is asymptotically stable, the state
trajectory must hit the switching surface and switching must take place. Following
switching, and based on the dynamics of the impulsive switched closed system as
presented in (2.24), the state vector immediately following switching will be in region 1.
Since 0e < for all states on the switching surface within the set Bγ for the system 1clΣ , the
state trajectory cannot cross the switching surface again and will continue to evolve in
region 1 thereafter. The state trajectory will approach the origin asymptotically, which
implies that regulation will be achieved.
• Case 2: ( )2ˆ 0
Tex exC C A > and ( ) 1 2
2 2T T
s sb A A b γ−
> (See Figure 2.3). In this case, all the
state trajectories leaving the set Sγ and entering region 2 have 0e > . Assume the state
trajectory enters the set Bγ from region 2 at the time Tγ . Given that the system 2clΣ is
asymptotically stable, and that the trajectories are confined to evolve inside Bγ , the state
trajectory must cross the switching surface to approach the origin. However, this is not
possible since, with respect to the system 2clΣ , 0e > for all Sγχ ∈ . Therefore, it is not
possible to have the state trajectory enter the set Bγ from region 2 at time Tγ . Consider
now the case where the state trajectory enters the set Bγ from region 1. If the state
43
trajectory crosses the switching surface to enter into region 2, then that will result in a
contradiction similar to that discussed for the case where the state trajectory enters the set
Bγ from region 2. Therefore, once the state trajectory enters the set Bγ from region 1, it
will continue to evolve in region 1 forever and will never cross the switching surface to
enter into region 2 for t Tγ> . Since 1clΣ is a stable system, then the state trajectory will
converge to the origin asymptotically, which implies that regulation for the switched
system (2.10) is achieved.
Therefore, based on the above analysis, if the conditions for Case 1 given by (2.29) or the
conditions for Case 2 given by (2.30) are satisfied, then regulation in the switched system
(2.10) can be achieved.
2.5. Regulator synthesis for the switched system
Based on the regulation conditions for the switched system proposed in the previous section,
a regulator synthesis approach is presented in this section. The proposed synthesis approach
is based on solving a set of properly formulated bilinear matrix inequalities. The main idea
behind the regulator synthesis approach is as follows. Consider an output feedback controller
r as given in (2.3) and (2.8), where r r
yQ r Q rA H B D= + and
rQ r r rC K= Ψ − Ω . Since
( ),r rΩ Ψ is the solution to the Sylvester equation (2.17), r is only parameterized in the
unknown matrices rK , rL and rQB . Assume r is such that the resulting closed system
satisfies (2.12). Then, based on Lemma 2.1 and Theorem 2.1, the switched closed loop
system clrΣ given in (2.10) is internally stable under arbitrary switching and, for each
{ }1, 2r∈ , clrΣ given in (2.10) also achieves output regulation. In this case, if any of (2.28),
44
Figure 2.2. State trajectories of 1clΣ with initial conditions on the switching surface inside the
set Bγ for case 1.
Figure 2.3. State trajectories of 2clΣ with initial conditions on the switching surface inside the
set Bγ for case 2.
1vS
Bounded set Bγ
1ˆexC A
Region 2: 2clΣ
Region 1: 1clΣ
S
S
2vS
exC
Region 2: 2clΣ
Region 1: 1clΣ
Bounded set Bγ
2ˆexC A
exC
45
(2.29) and (2.30) is satisfied, then based on Theorems 2.2 and 2.3, regulation in the switched
system (2.10) is achieved. The set of conditions in Lemma 2.1 and Theorems 2.2 and 2.3
yield bilinear matrix inequalities in the unknown parameters rK , rL and rQB . In the
following, a solution procedure for the formulated bilinear matrix inequalities is proposed to
determine the unknown parameters rK , rL and rQB in r .
The sufficient regulation condition (2.28) given in Theorem 2.2 is equivalent to:
0 mβ δ β ε< −Δ − . (2.31)
Using mcwβα
= and letting 0
m
m
wδ β εκ −Δ −
= , condition (2.28) is equivalent to:
c κα< . (2.32)
If (2.32) cannot be satisfied, the switching surface S intersects the set Bγ and conditions
(2.29) and (2.30) in Theorem 2.3 need to be verified. For the switched closed loop system
(2.10), the parameters c and α can be found using the following matrix inequalities [94]:
{ }2
ˆ ˆ 2 0, 0, 1, 2 ,
, 0.
Tr rA P PA P r
I P c I c
α α+ + < > ∈
< < > (2.33)
If the matrices rK , rL , rQB and
rQC in ˆrA are unknown, then the above equations define a
bilinear matrix inequality. Combining (2.32) and (2.33), the synthesis procedure for the
controller r can be realized by solving the following BMIs:
0 00 0
r r r r
T
r r r r r r r r r r r r r r r r r r r ry x y y x y
r r r r r r r r r r r ry y y y
Q r r Q r Q r r Q r
A B K B K B B K A B K B K B B KA L C D L D P P A L C D L D
B C H B D B C H B D
⎡ ⎤ ⎡ ⎤+ Ψ − Ω + Ψ − Ω⎢ ⎥ ⎢ ⎥+ + + + +⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥+ +⎣ ⎦ ⎣ ⎦
{ } 2 0, 0, 1,2 ,P rα α+ < > ∈
(2.34)
46
2 , 0,I P c I c< < > (2.35)
,c κα< (2.36)
where the unknown parameters are rK , rL , rQB , P , α and c . Once rK , rL ,
rQB are
determined based on the above BMIs, the matrices rQA and
rQC can be calculated using (2.18)
as given in Theorem 2.1.
In this chapter, a regulator synthesis algorithm, referred to as the P - iteration algorithm,
will be used to find a parameterized regulator that satisfies either one of the regulation
conditions (2.28), (2.29), or (2.30). The P - algorithm iteratively solves for the Lyapunov
matrix P and the parameters in the controller r given in (2.3). The basic idea of the P -
iteration is that a BMI can be converted into an LMI when some of the parameters in the
BMI are fixed. The approach for solving BMI problems is to alternate between two
optimization problems subject to LMIs, which are related to the matrix P in (2.34) and (2.35)
and the parameters in the controller r in (2.3), respectively. In the algorithm, the input data
is represented by rA , rB , yrC , eC , x
rD , yrD , e
rD , rH , δ and mw , whereas the unknown
variables to be determined are rK , rL , rQB ,
rQA , rQC P , α and c . The algorithm is
summarized below, where irK , i
rL , r
iQB , iP , iα and ic denote the solutions rK , rL ,
rQB , P ,
α and c obtained at the thi iteration of the algorithm.
47
1) Calculating κ : Determine ( ),r rΩ Ψ by solving the Sylvester equation (2.17) and let
rr
r
Ω⎡ ⎤Π = ⎢ ⎥Φ⎣ ⎦
with rr I
Ω⎡ ⎤Φ = ⎢ ⎥
⎣ ⎦. Let ˆ
rΠ be as in (2.21). Determine mβΔ using (2.26),
then calculate 0 m
mwδ β εκ −Δ −
= .
2) Initializing r : Initialize the controller parameters 0rK , 0
rL and 0rQB to make the
switched closed loop system (2.10) internally stable, which can be realized by solving
the following two LMIs separately for the unknown matrices rK , rL , rQB , KP and LP
with preset constants 0Kα ≥ and 0Lα ≥ .
{ }[ ] [ ] 0, 1, 2 ,Tr r r K K r r r K KA B K P P A B K P rα+ + + + < ∈ (2.37)
{ }0, 1, 2 .r r r r
Ty x y y x yr r r r r r r r r r r r
L L L Ly y y yQ r r Q r Q r r Q r
A L C D L D A L C D L DP P P r
B C H B D B C H B Dα
⎡ ⎤ ⎡ ⎤+ + + ++ + < ∈⎢ ⎥ ⎢ ⎥+ +⎢ ⎥ ⎢ ⎥⎣ ⎦ ⎣ ⎦
(2.38)
3) Initializing 0α : Based on the initial controller parameters 0rK , 0
rL and 0rQB from step
2, initialize the maximum decay rate 0α by solving the following optimization problem:
max α
subject to (2.34)
4) P -Step: At the thi iteration, 1,2,i = , given the matrices 1irK − , 1i
rL − and 1r
iQB − and
the scalar 1iα − , solve the following optimization problem for iP and ic :
min c
subject to (2.34) and (2.35)
5) -Step: At the thi iteration, 1,2,i = , given the matrix iP and ic from step 4, solve
the following optimization problem for irK , i
rL , r
iQB and iα :
48
max α
subject to (2.34) and (2.35)
6) Verification: Verify the constraints (2.36), (2.29) and (2.30), and if any of the three
conditions is satisfied, compute rQA and
rQC using (2.18), then stop the algorithm. If
none of (2.36), (2.29) and (2.30) is satisfied, then go to step 4.
The iterative loop is repeatedly executed until a solution is found, or there is no major
reduction in /c α relative to the previous iteration. The algorithm converges to a local
solution of /c α , since at each thi iteration of the algorithm, we have 1i ic c −≤ and 1i iα α −≥ .
Therefore, if no solution can satisfy the regulation conditions (2.36), (2.29) or (2.30) by the
iterative procedure described above, then the initial parameters 0rK , 0
rL and 0rQB can be
adjusted by changing the decay rates Kα and Lα in (2.37) and (2.38), and restarting the
iterative procedure again. If the above algorithm yields a solution, then, according to
Theorems 2.2 or 2.3, the switched closed loop system will achieve regulation.
2.6. Numerical example
In this section, the regulator synthesis method proposed in this chapter will be used to design
a controller that cancels the contact vibrations in a mechanical system. Figure 2.4 shows the
diagram of such a system consisting of a mass m , a contact surface cS , and their respective
coordinates. The mass m is attached to a spring with stiffness k and a damper with damping
coefficient c . The mass m moves only in the vertical direction whereas the contact surface
underneath it moves to the left. The mass m may enter into intermittent contact with the
49
Figure 2.4. Diagram of a mechanical system with switched dynamics.
surface cS , resulting in contact vibrations. When the mass m enters into contact with the
surface cS , the contact characteristics are represented by a spring with stiffness ck and a
damper with damping coefficient cc . The force F represents the external force used to
control the mass m whereas aF represents a disturbance force. This model can be found in
many applications, such as the interface between the read/write head and the disk surface in
hard disk drive systems. In the following, and with respect to the system shown in Figure 2.4,
the control objective is for the mass m to follow the displacement of the contact surface cS
while maintaining a desired constant separation in the vertical direction. Let v be the
deviation of the mass from its equilibrium position. Then the equations of motion of the mass
can be written as:
( )( )
when noncontact mode
when contact modea s
a c s
mv cv kv F F v v
mv cv kv F F F v v
+ + = − + >
+ + = − + + ≤
where cF is the contact force expressed as:
ck
Contact surface cSmaFv
ev sv
ckcc
F
Direction of surface motion
50
( ) ( )c c s c sF c v v k v v= − + − ,
and where sv is the displacement of the contact surface. The wavy contact surface profile is
expressed as a linear combination of sinusoidal functions:
( )coss k k k ek
v c t vω φ= + +∑ , (2.39)
with amplitudes kc , frequencies kω , phases kφ , 1, 2,k = … , and constant offset ev . Therefore,
the distance between the mass m and contact surface cS is sv v− . The control signal is
defined as /u F m= − for the non-contact situation and ( ) /c eu F k v m= − − when contact
takes place. Let ev− denote the desired distance between the mass m and contact surface cS .
The output y , to be fed to the controller, is defined as
y v= .
The performance variable e is defined to be the difference between the actual distance sv v−
and the desired distance ev− :
s ee v v v= − + .
Therefore, the system will switch between the contact and non-contact modes according to
the value of the performance variable e . If ee vδ> = , the system will operate in the non-
contact mode, and if ee vδ≤ = , the system will operate in the contact mode. Let 1x v= and
2x v= , then the switched system model is given by:
51
[ ]
[ ] [ ]
1 1
2 2
1
2
1
2
0 0 01 1 0
1 0:
1 0 0 1
1 if ,2 if ,
r r
r
r
e
e
x xA u d
x x
xy
x
xe d
x
e vr
e v
⎧⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤= + +⎪⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦⎣ ⎦ ⎣ ⎦⎪⎪ ⎡ ⎤⎪ = ⎢ ⎥⎪ ⎣ ⎦∑ ⎨
⎡ ⎤⎪ = + −⎢ ⎥⎪ ⎣ ⎦⎪⎪ ≤⎧
= ⎨⎪ >⎩⎩
(2.40)
where:
1
0 1( ) ( )c c
A k k c cm m
⎡ ⎤⎢ ⎥= + +⎢ ⎥− −⎣ ⎦
, ( )1
1c s c s c e a
s e
c v k v k v Fd m
v v
⎡ ⎤+ − +⎢ ⎥=⎢ ⎥
−⎢ ⎥⎣ ⎦
2
0 1A k c
m m
⎡ ⎤⎢ ⎥=⎢ ⎥− −⎣ ⎦
, 2
/a
s e
F md
v v⎡ ⎤
= ⎢ ⎥−⎣ ⎦.
The surface profile is given by ( )( )6cos 20 10 ms ev t vπ −= × + , and the force
( ) 51.5sin(20 ) 1.5cos(20 ) 10 NaF t tπ π −= + × . Let m , k , c , ck , cc and ev be 200mg, 2 N/m,
21 10−× N/m/Sec, 20 N/m, and 31 10−× N/m/Sec, and 630 10 m−− × , respectively. In the
following, a regulator for the switched system is designed using the synthesis procedure
described in the previous section. Simulation results will illustrate the performance of the
proposed regulator in maintaining the desired system output despite the presence of
switching.
In order to find a feasible solution for the BMIs, we introduce a coordinate transformation
as 1 1
2 2
x xT
x x′⎡ ⎤ ⎡ ⎤=⎢ ⎥ ⎢ ⎥′⎣ ⎦ ⎣ ⎦
, where 5
1 00 1 10
T −
⎡ ⎤= ⎢ ⎥×⎣ ⎦
. The motivation of this coordinate transformation
52
is to make the weight of state 2x less than 1x in χ since the performance variable e is only
defined based on the state 1x . Rewriting (2.40) in the state space form (2.1) results in:
5
10 1 101.1 10
A⎡ ⎤×
= ⎢ ⎥− −⎣ ⎦,
5
20 1 100.1 5
A⎡ ⎤×
= ⎢ ⎥− −⎣ ⎦,
1 2
0 2020 0
H Hπ
π⎡ ⎤
= = ⎢ ⎥−⎣ ⎦, 10 20 6
01 10
w w −
⎡ ⎤= = ⎢ ⎥×⎣ ⎦
,
1 2 5
01 10
B B −
⎡ ⎤= = ⎢ ⎥×⎣ ⎦
,
1
0 00.7469 1.75
xD⎡ ⎤
= ⎢ ⎥⎣ ⎦
, 2
0 00.75 0.75
xD⎡ ⎤
= ⎢ ⎥⎣ ⎦
,
[ ]1 2 1 0y y eC C C= = = , 1 2 0y yD D= = , [ ]1 2 0 1e eD D= = − .
It is desired to design a regulator that can reject the disturbance in the switched system.
Based on Theorem 2.1, solving the Sylvester equation (2.17) yields:
[ ]1 75314.1592653590 68947.8417604358Ψ = − − ,
[ ]2 75314.1592653590 68947.8417604357Ψ = − − ,
1 2 4
0 16.2831 10 0−
⎡ ⎤Ω = Ω = ⎢ ⎥×⎣ ⎦
, 1
1 2 1
2 2I ×
Ω⎡ ⎤⎢ ⎥Π = Π = Ω⎢ ⎥⎢ ⎥⎣ ⎦
.
Based on (2.26), we obtain 62 10mβ−Δ = × . Since 630 10δ −= − × and 6
m 2.09 10w −= × , and
with 0.01ε = , we have 13.3950κ = in (2.32). Using the P - algorithm proposed above,
we obtain:
3 91 123.583428456762 10 8.54599379771142 10K ⎡ ⎤= − × − ×⎣ ⎦ ,
3 92 123.582172573746 10 8.54575846443053 10K ⎡ ⎤= − × − ×⎣ ⎦ ,
53
3
1 3
409.395433365613 10581.725609783952 10
L−⎡ ⎤− ×
= ⎢ ⎥− ×⎣ ⎦,
3
2 3
409.395415548061 10581.726610025785 10
L−⎡ ⎤− ×
= ⎢ ⎥− ×⎣ ⎦,
1
3128.47291192622610
194.074510473731QB−⎡ ⎤
= ×⎢ ⎥−⎣ ⎦,
2
3128.47291197947110
194.074510543232QB−⎡ ⎤
×⎢ ⎥−⎣ ⎦,
4.9728c = , 0.3808α = , 13.0601cα= .
Then, based on (2.18), we have
[ ]1
35444.92042576818 54.6355866963264 10QC = − × ,
[ ]2
35444.77256150691 54.6343308133107 10QC = − × .
Since 13.0601< 13.3950c κα= = , condition (2.36) is satisfied. Therefore, regulation can be
achieved in the switched closed loop system using the above designed controller.
The simulation results of the response of the closed loop system under switching are
illustrated in Figures 2.5 and 2.6. It can be seen that if the performance variable is smaller
than -30 micrometers, the mass enters into contact with the contact surface cS and the model
of the system switches. It can also be seen that the disturbance rd changes at the switching
times. But even in the presence of switching, the switched system performance variable e
still converges to zero, which means the mass m asymptotically follows the contact surface
cS at the desired separation 30 micrometers.
2.7. Conclusion
The problem of regulation in switched bimodal systems against known disturbance or
reference signals is discussed. A regulator design approach based on the parameterization of
54
a set of controllers that can achieve regulation for the switched system is presented. The
forced switched closed loop system involving the parameterized controller is first
transformed into an unforced impulsive switched system. Consequently, the original
regulation problem is transformed into a stability analysis problem for the origin of the
impulsive switched system. Then sufficient conditions for regulation in switched systems are
derived and a regulator synthesis method based on solving a set of bilinear matrix
inequalities is presented. A simulation example involving a mechanical system with switched
dynamics is used to illustrate the effectiveness of the proposed regulator designed method.
The parameterization of the regulators considered in this paper was performed without
imposing any stability constraints on the rQ parameters. Although this advantage offers more
flexibility in the design of the desired rQ parameters, the derived structure of the rQ
parameters needed for regulation is constrained in terms of the matrices used in the state
space representation of the rQ parameters. This structure is not desirable when considering
the adaptive regulation problem to be discussed later in the thesis, since it does not allow for
the rQ parameters to be suitably parameterized for the purpose of tuning them online. In the
following chapters, the rQ parameters are represented as linear combination of stable basis
functions, which offers significant advantages when considering the adaptive regulation
problem.
55
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4-50
0
50
e (m
icro
met
ers)
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4-0.5
0
0.5
1
d r(1,1
) (m
/sec
2 )
Time(sec)
Figure 2.5. Simulation results for the case of 30ev = − micrometers showing the performance variable e and the switching component (1,1)rd in the disturbance rd .
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
-40
-30
-20
-10
0
10
20
30
40
v an
d v s (m
icro
met
ers)
Time(sec)
mass height vcontact surface vs
Figure 2.6. Simulation results for the case of 30ev = − micrometers showing the regulated mass height v and the contact surface displacement sv .
56
CHAPTER 3
Regulation in Switched Bimodal Systems
Based on Linear Matrix Inequalities
3.1. Introduction
In this chapter, a Q -parameterized regulator design approach is proposed for switched
bimodal systems where it is desired to reject known disturbance signals and track known
reference inputs simultaneously. The proposed method consists of three steps. First, a set of
Q -parameterized observer-based stabilizing controllers for the switched system is
constructed and the stability of the resulting closed loop switched system is analyzed. The Q
parameters are selected to be a linear combination of stable basis functions. Second, for each
subsystem in the bimodal system, a set of Q -parameterized regulators is derived. In the third
step, regulation conditions for the switched closed loop system are developed using the two
approaches summarized in the following.
• In the first approach, sufficient regulation conditions are derived based on the closed loop
system input-output properties using a common Lyapunov function. Then, using a set of
regulation conditions and appropriate coordinate transformations, the forced switched
57
closed loop system is transformed into an unforced impulsive switched system. Hence,
the regulation problem for the original switched closed loop system is transformed into a
stability analysis problem for the origin of an impulsive switched system. Sufficient
conditions for the stability of the origin of the impulsive switched system are presented,
which are equivalent to regulation in the original switched closed loop system. A
regulator synthesis method is consequently outlined based on solving properly formulated
linear matrix inequalities (LMIs).
• In the second approach, based on the properties of the time-varying switching surface, the
whole closed loop system state space is partitioned into three adjacent regions. Based on
the resulting state space partitions, a regulation condition for the switched closed loop
system is formulated using a multiple Lyapunov function approach, then with the
regulation interpolation conditions for each subsystem derived in the frequency domain,
the corresponding regulator synthesis method is proposed using an iterative LMI
algorithm.
The main reason for considering Q parameters that are linear combinations of stable basis
functions is to simplify the controller synthesis procedure and reduce the number of
unknowns to be determined in the desired Q parameters. Moreover, such representation is
convenient to work with in the adaptive regulation case, as outlined later in the thesis. With
respect to the two regulator synthesis approaches summarized above, the main reason for
considering the second approach involving multiple Lyapunov functions is to attempt to
reduce the conservatism associated with the first approach based on a common Lyapunov
function.
58
The rest of the chapter is organized as follows. Section 3.2 introduces the switched
bimodal system model and presents the general regulation problem for the switched system.
In section 3.3, the construction of a Q -parameterized set of stabilizing controllers for the
switched system is discussed and the stability properties of the resulting closed loop switched
system are analyzed. In section 3.4, regulation conditions and a regulator synthesis method
for the switched system are proposed using a common Lyapunov function approach. In
section 3.5, regulation conditions and a regulator synthesis method for the switched system
are proposed using a multiple Lyapunov function approach, followed by the conclusion in
section 3.6.
3.2. The regulation problem for switched bimodal systems
This section considers switched bimodal systems subject to external inputs representing
disturbance and/or reference signals. Both the system dynamics as well as the external input
signals are assumed to switch according to a switching law defined by a switching surface.
Consider the switched system given by the following state space representation:
0
0
, (0) , , (0) ,
, :
,
1 if ,
2 if ,
xr r r r
r r r r ry yr r r
r e er r
e er r
e er r
x A x B u D w x xw H w w w
y C x D we C x D w
e C x D wr
e C x D wδδ
= + + == =
= +∑
= +
⎧ = + ≤= ⎨
= + >⎩
⎧⎪⎪⎪⎪⎨⎪⎪⎪⎪⎩
(3.1)
where nx∈R is the state vector, u∈R is the control input, y∈R is the measurement signal
to be fed to the controller, e∈R is the performance variable to be regulated and is assumed
to be measurable, hrw ∈R is the state vector of the exogenous systems generating the signal
59
rw , h hrH ×∈R have simple eigenvalues on the imaginary axis, { }1,2r∈ is the index of the
system rΣ under consideration at time t , and δ is a constant satisfying 0δ > . The
switching between the systems 1Σ and 2Σ is performed according to the value of the
performance variable e , and is determined based on the location of x with respect to a
switching surface S given by:
{ }: n e er rS x e C x D w δ= ∈ = + =R . (3.2)
In the following, it is assumed that for any given 0t ≥ , the system must operate in only one
of the two modes corresponding to { }1,2r∈ . For the switched system (3.1), it is desired to
construct an output feedback controller to regulate the performance variable e of the
switched system against the external input signal rw . Given the switching nature of the plant,
the output feedback controller is also chosen to be a switching feedback controller r ,
{ }1, 2r∈ , with a state space representation given by:
,:
,
c cc r c r
r c cr c r
x A x B y
u C x D y
⎧ = +⎪⎨
= +⎪⎩ (3.3)
where cc
nx ∈R , and where the switching among controllers is to obey the same rule given in
(3.1) and decided by the performance variable e as for switching between the two plant
models. In the following section, a framework of designing an output feedback regulator for
the switched system (3.1) is presented, within which regulation conditions and regulator
synthesis procedure are proposed.
60
3.3 Parameterization of a set of stabilizing controllers
The controller design approach presented in this chapter relies on the construction of a Q -
parameterized set of output feedback stabilizing controllers for the switched system. In this
section, the construction of sets of Q -parameterized output feedback stabilizing controllers
for the switched system (3.1) is first discussed. The stability of the resulting switched closed
loop system is then analyzed. As in the case of linear time invariant systems, the construction
of a Q -parameterized set of stabilizing controllers involves two steps. The first step involves
the construction of a central controller in the form of a switched observer-based state
feedback controller. The second step involves augmenting the central controller with stable
dynamics to construct a Q -parameterized set of stabilizing controllers.
3.3.1. Q - parameterized controller
Consider the following observer-based state feedback controller for the switched system (3.1):
( ) 0ˆ ˆ ˆ ˆ ˆ, (0) ,:
ˆ,o r r rr
r
x A x B u L y y x xu K x
⎧ = + + − =⎪⎨
=⎪⎩ (3.4)
where x is an estimate of the plant state vector x and ˆ ˆry C x= , { }1, 2r∈ is an estimate of
the plant output y . The state feedback gains rK , { }1, 2r∈ , and the observer gains rL ,
{ }1, 2r∈ , are assumed to switch according to the rule given in (3.1). Moreover, it is assumed
that there are no impulsive changes in the controller states at the switching times. The
construction of a Q -parameterized set of stabilizing controllers for the switched system
(3.1) proceeds along the same lines as in the case of linear time-invariant systems [91-93].
Each controller is expressed as a linear fractional transformation involving a fixed system rJ
61
and a proper asymptotically stable system rQ that could be chosen as desired (See Figure
2.1). The state space representation of the system rJ is given by:
0ˆ ˆ ˆ ˆ( ) , (0) ,ˆ: ,
ˆ ˆ.
r r r r r r r Q
r r Q
r
x A L C B K x L y B y x x
J u K x y
y y y C x
⎧ = + + − + =⎪⎪ = +⎨⎪ − = −⎪⎩
(3.5)
The rQ parameter is given by:
( ) 0ˆ , (0) ,:
,r
Q Q Q Q Q Qr
Q Q Q
x A x B y y x xQ
y C x
⎧ = + − =⎪⎨
=⎪⎩ (3.6)
where qnQx ∈ . In particular, throughout the rest of the paper, the rQ parameter is such that
QA is a fixed stability matrix, QB is a fixed matrix, and the matrix rQC changes with
{ }1, 2r∈ .
Remark 3.1
For a given { }1,2r∈ , the controller formulated in (3.5) and (3.6) is a parameterized
stabilizing controller for the system rΣ , where the system rQ represents a free design
parameter that can be chosen as desired as long as it is an asymptotically stable system. A
suitable choice for rQ can be a parameterization of the form 1
( ) ( )qn
ir r i
i
Q s sθ ψ=
= ∑ , where
( )i sψ are stable basis functions. Such representation can be used to approximate any proper
stable real rational transfer function. With such representation for rQ , a corresponding
realization is as given in (3.6), where QA and QB are fixed matrices, and where the
62
parameters , 1ir qi nθ = … , appear in
rQC . Here the rQC matrices are selected to achieve
regulation according to the regulation conditions to be discussed in sections 3.4 and 3.5.
Let ˆx x x= − denote the state estimation error and 1
TT T TQ N
x x xχ×
⎡ ⎤= ⎣ ⎦ denote the state
vector for the resulting closed loop system with 2 qN n n= + . The resulting closed loop
system is given by the following state space representation:
, ,
:1 if ,2 if ,
r r rex e
r rclr ex e
r rex e
r r
A E we C D w
e C D wr
e C D w
χ χχ
χ δχ δ
⎧ = +⎪ = +⎪∑ ⎨
⎧ = + ≤⎪ = ⎨⎪ = + >⎩⎩
(3.7)
where 00 0
rr r r r Q r r
r Q Q r
r r r
A B K B C B K
A A B CA L C
+⎡ ⎤⎢ ⎥
= −⎢ ⎥⎢ ⎥+⎣ ⎦
,
( ) ( ) ( )TTTx y T x y
r r Q r r r rE D B D D L D⎡ ⎤= − +⎢ ⎥⎣ ⎦, 0 0ex eC C⎡ ⎤= ⎣ ⎦ .
3.3.2. Internal stability of the Q -parameterized switched closed loop system
In the following, it is desired to study the stability properties of the system (3.7). The stability
analysis involves two steps, namely internal stability analysis and input-output stability
analysis. In the first step, the internal stability of the closed loop system is studied by
considering the system (3.7) in the absence of the signal rw , and studying the stability
properties of the origin for the resulting unforced switched system. The second step builds on
the internal stability results and presents input-output stability results for the system (3.7).
63
Consider first the system (3.7) in the absence of the signal rw . The state equation for the
resulting system is given by:
0 .0 0
rr r r r Q r r
Q Q Q r Q
r r r
A B K B C B Kx xx A B C xx A L C x
+⎡ ⎤⎡ ⎤ ⎡ ⎤⎢ ⎥⎢ ⎥ ⎢ ⎥= −⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥⎢ ⎥ ⎢ ⎥+⎣ ⎦ ⎣ ⎦⎣ ⎦
(3.8)
Consider now the three switched subsystems given by the following state equations:
( ) ,r r rx A B K x= + (3.9)
( ) ,r r rx A L C x= + (3.10)
.Q Q Qx A x= (3.11)
Then, the internal stability of the switched closed loop system (3.8) is given by the following
result.
Lemma 3.1
If the nominal switched systems (3.9) and (3.10) are each asymptotically stable under
arbitrary switching and QA is a stability matrix, then the switched system (3.8) is also
asymptotically stable under arbitrary switching.
Proof: In the following, assume the initial time 0 0t = . If the switched system (3.10) is
asymptotically stable, then according to the definition and lemma 3.1, we have:
0( ) ( , ) (0) (0) tx t t t x c x e λφ −= ≤ , (3.12)
64
for some , 0c λ > , where 0( , )t tφ is the state transition matrix of the switched system (3.10).
Then, in (3.12), we have 0x → as t →∞ for any (0)x . Therefore, for the state variable x in
(3.8), we obtain
lim ( ) 0t
x t→∞
= , (0) nx∀ ∈R . (3.13)
Similarly, if the switched system (3.11) is asymptotically stable, then we have that:
0( ) ( , ) (0) (0) QtQ Q Q Q Qx t t t x c x e λφ −= ≤ ,
for some , 0Q Qc λ > , and where 0( , )Q t tφ is the state transition matrix for the switched system
(3.11). The response of the state variable of Qx in (3.8) is then given by:
00( ) ( , ) (0) ( , ) ( )
t yQ Q Q Q Q rt
x t t t x t B C x dφ φ ξ ξ ξ= − ∫ .
It follows that
00( ) ( , ) (0) ( , ) ( )
t yQ Q Q Q Q rt
x t t t x t B C x dφ φ ξ ξ ξ≤ + ∫
( )
max0( ) (0) (0)Q Q
tt t yQ Q Q Q Q rx t c x e c e B C c x e dλ λ ξ λξ ξ− − − −⇒ ≤ + ∫
( )
max 0( ) (0) (0)Q Q
tt tyQ Q Q Q Q rx t c x e c c B C x e e dλ λ ξ λξ ξ− − − −⇒ ≤ + ∫
max
max
2 (0)( ) (0) if
( ) (0) (0) if
m
m
yQ Q r t
Q Q Q QQ
tyQ Q Q Q Q r Q
c c B C xx t c x e
x t c x c ct B C x e
λ
λ
λ λλ λ
λ λ
−
−
⎧ ⎡ ⎤⎪ ⎢ ⎥≤ + ≠⎪ ⎢ ⎥−⇒ ⎨ ⎣ ⎦⎪
⎡ ⎤≤ + =⎪ ⎣ ⎦⎩
where min( , )m Qλ λ λ= . Therefore, for the state variable Qx in (3.8), we have
lim 0Qtx
→∞= , (0) Qn
Qx∀ ∈R . (3.14)
Similarly, if the switched system (3.9) is asymptotically stable, then
65
0( ) ( , ) (0) (0) tx t t t x c x e λφ −= ≤ ,
for some , 0c λ > , where 0( , )t tφ is the state transition matrix for the switched system (3.9).
The response of the state variable of x in (3.8) is given by:
0 00( ) ( , ) (0) ( , ) ( ) ( , ) ( )
t tQr r Q r rt t
x t t t x t B C x d t B K x dφ φ ξ ξ ξ φ ξ ξ ξ= + +∫ ∫
Let max
max
2 (0)(0) if
(0) (0) if
yQ Q r
Q Q QQm
yQ Q Q Q r Q
c c B C xc x
c
c x c ct B C x
λ λλ λ
λ λ
⎧⎪ + ≠⎪ −= ⎨⎪
+ =⎪⎩
. Then we obtain:
0 00( ) ( , ) (0) ( , ) ( ) ( , ) ( )
t tQr r Q r rt t
x t t t x t B C x d t B K x dφ φ ξ ξ ξ φ ξ ξ ξ≤ + +∫ ∫
( ) ( )
max0 0( ) (0) (0)m
t tt t Q tr r m r rx t c x e ce B C c e d ce B K ce x dλ ξλ λ ξ λ ξ λξξ ξ−− − − − − −⇒ ≤ + +∫ ∫
( ) ( )
max 0 0( ) (0) (0)m
t tt Q t tm r r r rx t c x e cc B C e e d cc B K x e e dλ ξλ λ ξ λ ξ λξξ ξ−− − − − − −⇒ ≤ + +∫ ∫
maxm
mmax
mmax
2 2 (0)(0) if
2 (0)( ) (0) if
(0) (0) if
n
n
n
Qm r r r r t
m
r r tQm r r
tQm r r r r
cc B C cc B K xc x e
cc B K xx t c x cc t B C e
c x cc t B C cct B K x e
λ
λ
λ
λ λ λλ λ λ λ
λ λ λλ λ
λ λ λ
−
−
−
⎧⎡ ⎤⎪⎢ ⎥+ + ≠ ≠
−⎪⎢ ⎥−⎣ ⎦⎪⎪ ⎡ ⎤⎪ ⎢ ⎥⇒ ≤ + + = ≠⎨⎢ ⎥−⎪ ⎣ ⎦⎡ ⎤+ + = =⎣ ⎦
⎩
⎪⎪⎪⎪
(3.15)
where nλ is the minimum value of λ , Qλ and λ . It is easy to see that 0x → as t →∞ for
any (0)x in (3.15). Therefore, for the state variable x in (3.8), we have
lim ( ) 0t
x t→∞
= , (0) nx∀ ∈R . (3.16)
Then based on equalities (3.13), (3.14) and (3.16), it shows that the switched system (3.8) is
asymptotically stable.
66
Based on Lemma 3.1, a parameterized set of stabilizing controllers can be designed by
separately designing the individual switched systems (3.9), (3.10) and the system (3.11). First,
using the following Lemma, the gains rK and rL , { }1, 2r∈ , are chosen to make the switched
systems (3.9) and (3.10) asymptotically stable.
Lemma 3.2 [94]
The origin is an asymptotically stable equilibrium point for the switched systems (3.9) and
(3.10) under arbitrary switching if there exist matrices 0TK KP P= > , 0T
L LP P= > and
matrices rK , rL , { }1, 2r∈ , such that:
( ) ( ) { }( ) ( ) { }
0, 1, 2 ,
0, 1,2 .
Tr r r K K r r r
Tr r r L L r r r
A B K P P A B K r
A L C P P A L C r
+ + + < ∈
+ + + < ∈ (3.17)
By a simple change of variables, the matrix inequalities (3.17) can be transformed to LMIs
and solved using numerical algorithms. Here, we search for a common Lyapunov function
for each of the switched systems (3.9) and (3.10) to guarantee the stability of the systems (3.9)
and (3.10) under arbitrary switching. Moreover, based on the above choice for the matrix QA
in (3.6), the origin for the system (3.11) is an asymptotically stable equilibrium point.
Therefore, if the conditions of Lemma 3.2 are satisfied, then based on the results of Lemma
3.1, the origin is an asymptotically stable equilibrium point for the closed loop switched
system (3.8) under arbitrary switching.
67
3.4 Regulation of bimodal systems: a common Lyapunov function
approach
In this section, a regulator design method for bimodal systems is proposed using a common
Lyapunov function approach. First, the input-output stability property of the switched system
is presented based on a common quadratic Lyapunov function and the state space regulation
conditions for each closed loop subsystem 1clΣ and 2
clΣ are derived, based on which two
sufficient regulation conditions for the switched system are presented. Then a regulator
synthesis procedure is proposed using the properly formulated LMIs and a numerical
example is used to illustrate the effectiveness of the proposed regulator design method.
3.4.1 Input-output stability
In the following, we will consider the input-output stability properties of the switched system
(3.7) in the presence of an external bounded input signal rw based on a common Lyapunov
function. Since system (3.11) is asymptotically stable, there exists a positive definite matrix
QP such that:
Q
TQ Q Q Q nA P P A I+ = − (3.18)
Then based on KP , LP , and QP , we have the following result.
Proposition 3.1
Consider the switched system (3.8), if there exist matrices 0KP > , 0LP > , and 0QP > such
that (3.17) and (3.18) are satisfied, then there must exist a common quadratic Lyapunov
68
function for the switched system (3.8), for example, a common quadratic Lyapunov function
matrix can be blockdiag( , , )K L QP P P Pμ η= , where 0μ > and 0η > .
Proof: Consider the quadratic function ( ) ( ) ( )TV t P tχ χ χ= , where
blockdiag( , , )K L QP P P Pμ η= , and μ and η being positive parameters, we define the
following matrix
( )x Tr r rW A P PA= − +
Based on (3.17), we can have
{ }{ }
[ ] [ ] , 1,2
[ ] [ ] , 1, 2
T Kr r r K K r r r r
T Lr r L L r r r
A B K P P A B K W r
A L C P P A L C W r
+ + + < − ∈
+ + + < − ∈
where KrW and L
rW , { }1, 2r∈ , are positive definite matrices. In view of the block diagonal
form of P , then we have xrW as:
( )( )
1 2
13
23
Q
Kr r r
Txr r n
T T Lr r
W R R
W R I R
R R W
μ μ
μ η
⎡ ⎤⎢ ⎥⎢ ⎥= ⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
where 1rR , 2
rR and 3R are the following matrices
1 [ ] [ ]Q T Qr r r K K r rR B C P P B C= +
2 [ ] [ ]Tr r r K K r rR B K P P B K= +
3 [ ] [ ]y T yQ r Q Q Q rR B C P P B C= − + −
Then, let { }
11 1,2
max rrRψ
∈= ,
{ }2
2 1,2max rr
Rψ∈
= , { }1 min1,2
min ( )Krr
Wλ∈
⎡ ⎤ϒ = ⎣ ⎦ , { }2 min1,2
min ( )Lrr
Wλ∈
⎡ ⎤ϒ = ⎣ ⎦ , and
3 3Rψ = . We have
69
( )( )
1 2
13
23
( )
Q
T xr
kr r r
TT T TQ r n Q
T T Lr r
V W
W R R xx x x R I R x
xR R W
χ χ χ
μ μ
μ η
= −
⎡ ⎤⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎡ ⎤= − ⎣ ⎦ ⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥ ⎣ ⎦
⎢ ⎥⎣ ⎦
( ) ( )1 2 1 23 3r Q
T TT T T T T T T T T T Lk Q r r r Q n r Q r Q
xx Q x R x R x R x I x R x R x R x W x
xμ μ μ η
⎡ ⎤⎢ ⎥⎡ ⎤= − + + + + + + ⎢ ⎥⎢ ⎥⎣ ⎦⎢ ⎥⎣ ⎦
( ) ( )1 2 1 23 3r Q
T TT T T T T T T T T T Lk Q r r r Q Q n Q Q r Q rx Q x x R x x R x x R x x I x x R x x R x x R x x W xμ μ μ η⎡ ⎤= − + + + + + + + +⎢ ⎥⎣ ⎦
22 21 2 1 2 32 2 2Q Q Qx x x x x x x x xμ η ψ ψ μ ψ⎡ ⎤≤ − ϒ + + ϒ − − −⎢ ⎥⎣ ⎦
1 1 2
1 3
2 3 2
Q Q
x
x x x x
x
ψ ψψ μ μψψ μψ η
⎡ ⎤ϒ − −⎡ ⎤⎢ ⎥⎢ ⎥⎡ ⎤= − − − ⎢ ⎥⎣ ⎦ ⎢ ⎥⎢ ⎥⎢ ⎥− − ϒ⎣ ⎦ ⎣ ⎦
ˆ T
Q Qx x x Q x x x⎡ ⎤ ⎡ ⎤= − ⎣ ⎦ ⎣ ⎦ .
where
1 1 2
1 3
2 3 2
Qψ ψ
ψ μ μψψ μψ η
ϒ − −⎡ ⎤⎢ ⎥= − −⎢ ⎥⎢ ⎥− − ϒ⎣ ⎦
.
Clearly, we need to find proper μ and η to make Q positive definite. The three leading
principal minors of Q are:
1ϒ , 21 1μ ψϒ − ,
2 2 21 2 3 1 1 2 2 2 1 3 2( ) ( ) ( )μη μ ψ ψ η ψ ψ ψ ψ ψ μϒ ϒ − + ϒ ϒ − − + .
Therefore, by selecting
21
1
ψμ >ϒ
, 2 2 2
1 3 1 2 3 2 1 2
1 2 1 2 1
( )ψ μ ψ ψ ψ ψ μ ψ ψημ ψ
ϒ + + +>
ϒ ϒ + ϒ ϒ,
the matrix Q is positive definite.
70
If there exists a common Lyapunov function for the switched system (3.8), then we have the
following input-output stability property for the switched closed loop system (3.7).
Theorem 3.1
Consider the closed loop system (3.7) subject to a bounded input rw . Let { }1,20
max ( ) 0rrt
w tκ∈≥
= ≠ ,
0α > a preset constant and 0 1ε< . If there exist 0TP P= > and positive scalars μ and
β such that the following matrix inequalities are satisfied for all { }1, 2r∈ :
( )2
01
Tr r r
Tr
A P PA P PE
E P I
αε μακ
⎡ ⎤+ +⎢ ⎥ ≤−⎢ ⎥−⎢ ⎥⎣ ⎦
, (3.19)
( )
( )
2 0
0 0
Tex
Ter
ex er
P C
I D
C D I
κ
β μκ
βκ
⎡ ⎤⎢ ⎥⎢ ⎥
⎛ ⎞⎢ ⎥− ≥⎜ ⎟⎢ ⎥⎝ ⎠⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
, (3.20)
then there exits a finite time BT at which the states enter into and continue to evolve inside
the bounded set xB given by:
{ }: ,N TxB Pχ χ χ μ= ∈ ≤R (3.21)
and the output ( )e t in (3.7) satisfies:
( ) , Be t t Tβ≤ ∀ ≥ . (3.22)
71
Proof: Consider the state space representation (3.7) and the quadratic function ( ) TV Pχ χ χ= .
Inequality (3.19) implies that for any nonzero vector N h
rwχ +⎡ ⎤
∈⎢ ⎥⎣ ⎦
R , we have:
( )2
01
TT r r r
Tr rr
A P PA P PE
w wE P I
αχ χε μακ
⎡ ⎤+ +⎡ ⎤ ⎡ ⎤⎢ ⎥ ≤−⎢ ⎥ ⎢ ⎥⎢ ⎥−⎣ ⎦ ⎣ ⎦⎢ ⎥⎣ ⎦
.
It follows from the above inequality that:
( )2
1( ( )) ( ( )) ( ) ( ) 0T
r rd V t V t w t w tdt
ε μαχ α χ
κ−
+ − ≤ . (3.23)
Let ( )2
1( ) ( ) ( )T
r rp t w t w tε μακ−
= , then ( )( ) 1p t ε μα≤ − . Assume the initial time is 0 and let
0( (0))V Vχ = . Using the comparison principle, ( ( ))V tχ then satisfies:
( ) ( )
( ) ( )
( ) ( )
0 0
0 0
0
( ( ))
1
1 1 .
t tt
t tt
t t
V t V e e p d
V e e d
V e e
α τα
α τα
α α
χ τ τ
ε μα τ
ε μ
− −−
− −−
− −
≤ +
≤ + −
≤ + − −
∫∫
Consequently, we have:
( )lim ( ( )) 1t
V tχ ε μ→∞
≤ − .
Therefore, there exists a finite time BT such that
( ( )) , BV t t Tχ μ≤ ∀ ≥ . (3.24)
Based on inequality (3.20), we have:
( )( )
2 00
0
Tex
ex erTe
r
P CC D
I D
κκ
β βμκ
⎡ ⎤ ⎡ ⎤⎢ ⎥ ⎢ ⎥ ⎡ ⎤− ≥⎛ ⎞ ⎣ ⎦⎢ ⎥ ⎢ ⎥−⎜ ⎟⎢ ⎥ ⎢ ⎥⎣ ⎦⎝ ⎠⎣ ⎦
. (3.25)
72
Inequality (3.20) also implies that 0β μκ− ≥ . Multiplying (3.25) from the left by
T
rwχ⎡ ⎤
⎢ ⎥⎣ ⎦
and
from the right by rwχ⎡ ⎤
⎢ ⎥⎣ ⎦
yields:
( )
( )
2 2
2 2
( ) ( ) ( ) ( )
( ) .
Tr re t V t w t w t
V t
β βκ χ μκ κ
β βκ χ μ κκ κ
⎛ ⎞⎛ ⎞≤ + −⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠
⎛ ⎞⎛ ⎞≤ + −⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠
Based on (3.24), it follows that:
2 2 2
2
( )
, .B
e t
t T
β βμκ μ κκ κ
β
⎛ ⎞⎛ ⎞≤ + −⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠
= ∀ ≥
(3.26)
Since 0β > , ( )e t is ultimately bounded and we have ( ) , Be t t Tβ≤ ∀ ≥ .
Remark 3.3
The advantage of considering a common P matrix in (3.19) and (3.20) is that the set xB is
defined independently of the closed loop system that would be in operation, Bt T∀ ≥ . This
property will be very helpful in deriving the regulation conditions presented in section 3.5.
In the following sections the matrices rQC , { }1, 2r∈ , will be designed to achieve regulation
for the switched closed loop system (3.7).
73
3.4.2 State space regulation conditions for 1clΣ and 2
clΣ
The purpose of introducing the Q -parameterized controllers is to find appropriate matrices
QA , QB and rQC , { }1, 2r∈ , to solve the output feedback regulation problem for the switched
system (3.7). In this section, for a given closed loop system clrΣ , { }1, 2r∈ , a set of rQ
parameters is defined such that the corresponding closed loop system clrΣ achieves regulation.
These sets of rQ parameters will be used in the synthesis of regulators for the original
switched system. The construction of the sets of rQ parameters is achieved by solving a set
of linear matrix equations corresponding to standard regulation conditions for linear systems.
In the following, regulation conditions for each of the closed loop systems are presented and
the construction of the sets of rQ parameters is discussed in detail.
The plant rΣ and the block rJ shown in Figure 2.1 can be combined into a single system
rT . The dynamics of the system rT , { }1,2r∈ , can be represented as follows:
[ ]
,0 0
: 0 0 ,
ˆ 0 0 .
xrr r r r r r r
x yQr r r r r r
re er r
Q
ryr r
Q
wA B K B Kx x D ByA L Cx x D L D
wxT e C D
yx
wxy y C D
yx
⎧ + ⎡ ⎤⎡ ⎤⎡ ⎤⎡ ⎤ ⎡ ⎤= +⎪ ⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥ ⎢ ⎥+ − −⎣ ⎦ ⎣ ⎦⎣ ⎦ ⎣ ⎦ ⎣ ⎦⎪
⎪ ⎡ ⎤⎡ ⎤⎪ ⎡ ⎤ ⎡ ⎤= +⎨ ⎢ ⎥⎢ ⎥⎣ ⎦ ⎣ ⎦⎣ ⎦ ⎣ ⎦⎪⎪ ⎡ ⎤⎡ ⎤⎪ ⎡ ⎤− = − + ⎢ ⎥⎢ ⎥ ⎣ ⎦⎪ ⎣ ⎦ ⎣ ⎦⎩
(3.27)
Therefore, we have that:
ˆr
rQ
weT
yy y⎡ ⎤⎡ ⎤
= ⎢ ⎥⎢ ⎥−⎣ ⎦ ⎣ ⎦, (3.28)
where
74
11 12
21 22r r
rr r
T TT
T T⎡ ⎤
= ⎢ ⎥⎣ ⎦
,
and where, based on (3.27), the state space representations for 11rT , 12
rT , 21rT and 22
rT are as
follows:
11 1111
11 11: 00
xr r r r r r
r r x yr r r r r r r
r r e er
A B K B K DA E
T A L C D L DC D
C D
⎡ ⎤+⎡ ⎤ ⎢ ⎥
= + − −⎢ ⎥ ⎢ ⎥⎢ ⎥⎣ ⎦ ⎢ ⎥
⎣ ⎦
,
12 1212
12 12: 0r r r rr r
err r
A B K BA ET CC D
+⎡ ⎤ ⎡ ⎤=⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎣ ⎦⎣ ⎦,
21 2121
21 21:x y
r r r r r r r rr y
r r r r
A E A L C D L DT
C D C D⎡ ⎤ ⎡ ⎤+ − −
=⎢ ⎥ ⎢ ⎥−⎢ ⎥ ⎢ ⎥⎣ ⎦ ⎣ ⎦, 22 0.rT =
Let ( )rW s and ( )E s denote the Laplace transforms of the inputs rw and the performance
variable e , respectively. Let , ( )r rT QF s denote the closed loop system transfer function
relating ( )rW s to ( )E s . Based on (3.28), the closed-loop system is given by:
,( ) ( ) ( ),r rT Q rE s F s W s= (3.29)
where, for each { }1, 2r∈ ,
11 12 21,
11 12 21
( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ).r rT Q r r r r
r r r r
F s T s T s Q s T s
T s Q s T s T s
= +
= + (3.30)
Based on the definition of 11rT , 12
rT and 21rT in (3.28), a realization of the closed loop system
is given as follows:
75
11 11
21 21
12 21 12 12 21
12
11 11
0 0 00 0 0ˆ ˆ0 0
ˆ ˆ0 0 0
0 0r
r r
r rr r
r r r r r
r rQ r Q
r Q r
A EA E
A EE C A E D
C D B C AC C D
⎡ ⎤⎢ ⎥⎢ ⎥⎡ ⎤⎢ ⎥⎢ ⎥ =⎢ ⎥⎢ ⎥⎣ ⎦ ⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
. (3.31)
Standard regulation conditions for the system (3.31) are then presented as follows.
Lemma 3.3 [52]
For each { }1,2r∈ , assume the closed loop system clr∑ in (3.7) is asymptotically stable. Then,
for each { }1,2r∈ , i.e. in the absence of switching, clr∑ in (3.7) achieves regulation if and
only if there exists a matrix ˆrΠ which solves the linear matrix equations:
ˆ ˆˆ ˆ ,ˆ ˆˆ 0 .
r r r r r
r r r
H A E
C D
Π = Π +
= Π + (3.32)
Remark 3.4
The main idea behind considering the state space realization of (3.31) instead of (3.7) in
Lemma 3.3 is to express the regulation condition (3.32) as a linear matrix constraint. First
the parameters rK and rL in ˆrA can be determined by solving (3.17), then ˆ
rΠ can be
determined by solving the first equation in (3.32). The second equation in (3.32) becomes a
linear matrix equation in rQC , which appears in ˆ
rC .
76
3.4.3 Regulation conditions for the switched system
For a given { }1, 2r∈ , if the closed loop system clrΣ in (3.7) is asymptotically stable and (3.32)
is satisfied, then regulation in the closed loop system clrΣ can be achieved. However, that
does not guarantee achieving regulation in the switched closed loop system. In this section,
regulation conditions for the switched closed loop system are developed using two
approaches. In the first approach, sufficient regulation conditions are derived based on the
closed loop system input-output properties. In the second approach, the forced switched
closed loop system is transformed into an unforced impulsive switched system using an
appropriate coordinate transformation. Hence, the regulation problem for the switched closed
loop system is transformed into a stability analysis problem for the origin of an impulsive
switched system. A regulator synthesis method based on solving some linear matrix
inequalities is then proposed.
Firstly, the input-output stability properties of the switched system given in Theorem 3.1
are used to provide the following sufficient condition for regulation.
Theorem 3.2
Assume that the switched closed loop system (3.7) is internally stable under arbitrary
switching and that the conditions of Theorem 3.1 are satisfied. If for { }1, 2r∈ , (3.32) is
satisfied and
,β δ< (3.33)
then the switched closed loop system (3.7) achieves regulation.
Proof : Based on Theorem 1, we have that:
77
( ) , Be t t Tβ≤ ∀ ≥ .
If β δ< , then for all Bt T≥ , the output ( )e t will always be such that ( )e t δ< and there
will be no more switching. Since the systems (3.7) is asymptotically stable and since (3.32) is
satisfied, then regulation can be achieved.
The physical meaning of the above theorem is obvious. However, condition (3.33) cannot be
easily satisfied if δ is arbitrarily small. Another alternative to determine regulation
conditions for the system (3.7) is by examining the behavior of the state vector χ . To study
the regulation properties of the closed loop system, an appropriate coordinate transformation
will be used to transform the forced switched closed loop system (3.7) into an unforced
impulsive switched system. As a result, the regulation problem for the switched closed loop
system is transformed into a stability analysis problem for the origin of an impulsive
switched system.
For each { }1,2r∈ , if the closed loop system clr∑ in (3.7) is asymptotically stable and
(3.32) is satisfied, then there must exist a matrix rΠ which solves the following linear matrix
equations:
,
0 .r r r r r
ex er r
H A E
C D
Π = Π +
= Π + (3.34)
Now consider the following coordinate transformation:
r rwχ χ= −Π , (3.35)
Substituting the expression for χ from (3.35) into (3.7), and using (3.34), the forced closed-
loop switched system (3.7) can be transformed into an unforced switched system given by:
78
, ,
:1 if ,2 if .
rex
clr ex
ex
Ae C
e Cr
e C
χ χχ
χ δχ δ
⎧ =⎪
=⎪∑ ⎨⎧ = ≤⎪ = ⎨⎪ = >⎩⎩
Let τS be the set of switching times kτ , 1, 2,k = … . Since for each subsystem in (3.7), a
different coordinate transformation is used, then at a switching time kt ττ= ∈S , we have
( ) ( ) ( ),t t tχ χ χ+ −= + Δ
where ( ) ( ) ( ) ( )
( ) ( ) ( )r t r t r t r t
t w t w tχ − − + +Δ = Π −Π . Hence, the resulting impulsive switched closed
loop system is given by:
,
( ) ( ) ( ), : ,
1 if ,2 if .
r
kcl exr
ex
ex
A
t t t te C
e Cr
e C
τ
χ χ
χ χ χ τχ
χ δχ δ
+ −
⎧ =⎪
= +Δ = ∈⎪⎪∑ =⎨⎪ ⎧ = ≤⎪ = ⎨⎪ = >⎩⎩
S (3.36)
In the new coordinate system, the switching surface S given in (3.2) is fixed and can be
expressed as follows:
{ }: .N exS e Cχ χ δ= ∈ = =R (3.37)
Using the coordinate transformation (3.35), the original output feedback regulation problem
for the switched system (3.7) is transformed into an asymptotic stability analysis problem for
the origin of system (3.36). Note that under the conditions of Theorem 3.1, the original state
vector χ will ultimately evolve inside a bounded set xB given by (3.21). Define
{ }1,20
max ( )r rrt
w tη∈≥
= Π and max ex
BC
ηχγ χ
∈= with { }: NBη χ χ η= ∈ ≤R . A sufficient condition
for regulation is then given as follows.
79
Theorem 3.3
Assume that the switched closed loop system (3.7) is internally stable under arbitrary
switching and that the conditions of Theorem 3.1 are satisfied. If for { }1, 2r∈ , (3.32) is
satisfied, δ γ> and
( )( )2 01
Tex
ex
P C
C δ γμ
⎡ ⎤⎢ ⎥
>⎢ ⎥−⎢ ⎥
⎣ ⎦
, (3.38)
then the switched closed loop system (3.7) achieves regulation.
Proof : Adding ( )exr rC w tΠ to both sides of (3.37) yields
( )( ) ( ) ( )ex exr r r rC t w t C w tχ δ+Π = + Π .
Therefore, the equation for the switching surface S can be rewritten as:
( ) ( )ex exr rC t C w tχ δ= + Π . (3.39)
The term ( )exr rC w tΠ is time-varying and satisfies ( )ex
r rC w t γΠ ≤ . If δ γ> , then the
switching surface S would be closest to the origin when
( ) ( )exC t signχ δ δ γ= − . (3.40)
The switching surface (3.40) does not intersect with xB if an only if
( ){ }min = T exP C signχ χ χ δ δ γ μ− > . (3.41)
The solution to the left hand side of the above inequality is
( )( ) ( )( )1 21 Tex exC P C signδ δ γ−
− − .
80
It follows that (3.41) is equivalent to
( )( ) ( )( )1 21 Tex exC P C signδ δ γ μ−
− − > .
Since ( ) ( )22( )signδ δ γ δ γ− = − , the above inequality becomes
( ) ( )21 1Tex exC P C δ γμ
− < − ,
which is equivalent to (3.38). Then if (3.38) is satisfied, the switching surface S in (3.39)
will never intersect the bounded set xB for all possible ( )exr rC w tΠ , 0t ≥ . Since the
conditions in Theorem 3.1 are satisfied, then for Bt T≥ , the trajectory of the state vector χ
will enter the set xB and continue to evolve inside xB thereafter. Consequently, if (3.38) is
satisfied, there will be no more switching in the set xB . Since the system (3.7) is
asymptotically stable and since (3.32) is satisfied, then regulation will be achieved.
For xBχ ∈ , χ satisfies:
( ) ( )( ) ( )T
r r r rw t P w tχ χ μ+Π +Π ≤ .
It follows that if the conditions of Theorem 3.1 are satisfied, then for Bt T≥ , χ will evolve
inside the bounded set xB given by:
( ) ( ){ }: , TNxB c P c c Bηχ χ χ μ= ∈ + + ≤ ∈R .
Remark 3.5
The bounded set xB includes the union of all ellipsoids with center c− , c Bη∈ . The set xB is
a convex set. This can be verified by considering 1χ and 2χ in xB and 1c and 2c in Bη such
81
that ( ) ( )1 1 1 1Tc P cχ χ μ+ + ≤ and ( ) ( )2 2 2 2
Tc P cχ χ μ+ + ≤ . It can then be shown that
( )1 21χ θχ θ χ= + − , 0 1θ≤ ≤ , is in the set xB since ( ) ( )Tc P cχ χ μ+ + ≤ where
( )1 21c c cθ θ= + − is an element of Bη given that Bη is a convex set.
If condition (3.38) is not satisfied, then the switching surface S in (3.37) intersects with the
bounded set xB . The proof of this statement can be presented by considering the following
arguments. The switching surface S in (3.37) does not intersect with the bounded set xB if
and only if ( ) ( )21 1Tex ex exC P C C cδμ
− < + , for all c Bη∈ . This result can be derived by
considering ( )cχ + as a new coordinate and rewriting the expression for the switching
surface (3.37) in the new coordinate. Given that min ex
c BC c
η
δ δ γ∈
+ ≤ − , if (3.38) is not
satisfied, which means ( ) ( )21 1Tex exC P C δ γμ
− ≥ − , then there must be a point c Bη∈ such
that ( ) ( )21 1Tex ex exC P C C cδμ
− ≥ + . The latter inequality implies that the switching surface S
in (3.37) intersects with the bounded set xB . To present regulation conditions in this case,
define the matrices 11
ex
s ex
CA
C A⎡ ⎤
= ⎢ ⎥⎣ ⎦
, 22
ex
s ex
CA
C A⎡ ⎤
= ⎢ ⎥⎣ ⎦
and 0
bδ⎡ ⎤
= ⎢ ⎥⎣ ⎦
. Sufficient conditions for
regulation are then presented in the following theorem.
Theorem 3.4
Assume that the switched closed loop system clrΣ given in (3.7) is internally stable under
arbitrary switching, that the conditions of Theorem 3.1 are satisfied, and that, for each
82
{ }1, 2r∈ , (3.32) is satisfied. Moreover, assume that (3.38) is not satisfied and that
1 2( ) ( ) 2s srank A rank A= = . If the conditions:
( ) ( )( ) ( ) ( ){ }
11
111 1 1 1
0,
min ,
Tex ex
T Ts s s sc
C A P C
b A c A P A b A cη
δ
μ
−
−−
≤
<
+ + > (3.42)
or
( ) ( )( ) ( ) ( ){ }
12
112 2 2 2
0,
min ,
Tex ex
T Ts s s sc
C A P C
b A c A P A b A cη
δ
μ
−
−−
≤
>
+ + > (3.43)
are satisfied, then the origin is an asymptotically stable equilibrium point for the impulsive
switched system (3.36), implying that the switched system (3.7) achieves regulation.
Proof: In the following, we assume 0δ > . The proof for the case of 0δ < is similar. The
proof consists of two parts. The first part presents condition (3.42) so that 0e < on the
switching surface inside xB for system 1Σ and condition (3.43) so that 0e > on the
switching surface inside xB for system 2Σ . In the second part, and based on the sign of e on
the switching surface as given by conditions (3.42) and (3.43), the asymptotic stability of the
origin for the system (3.36), which is equivalent to achieving regulation in the system (3.7),
is studied.
Part 1: Conditions for a uniform sign of e on the switching surface inside xB .
Let region 1 and region 2 denote the half-spaces where the systems 1clΣ and 2
clΣ are active,
respectively. In Figure 3.1 and Figure 3.2, we assume 0δ > , so that the origin is in region 1
where the system 1clΣ is active. Define two hypersurfaces as
83
{ }1 1: 0N exvS C Aχ χ= ∈ =R ,
{ }2 2: 0N exvS C Aχ χ= ∈ =R .
Then for all points in 1vS , 1 0exe C A χ= = and for all points in 2vS , 2 0exe C A χ= = . By
assumption, the switching surface intersects the bounded set xB . Let
( ) ( ){ }: and , TN exxS C c P c c Bηχ χ δ χ χ μ= ∈ = + + ≤ ∈R
denote the convex set representing the intersection of the switching surface S with the
bounded set xB . If the direction of the vector exC is not parallel to that of 1exC A or 2
exC A ,
then the switching surface S must intersect with the hypersurface 1vS or 2vS . Let
{ }1 1: 0 and N ex exL C A Cχ χ χ δ= ∈ = =R
and
{ }2 2: 0 and N ex exL C A Cχ χ χ δ= ∈ = =R
denote the intersection of S with the two hypersurfaces 1vS and 2vS , respectively. For each
{ }1, 2r∈ and for a given c Bη∈ , let ( ) ( )( ) minr
Tr L
l c c P cχ
χ χ∈
= + + . Then ( )rl c can be
obtained by solving the problem of minimizing ( ) ( )Tc P cχ χ+ + subject to srA bχ = . Since,
by assumption, 1 2( ) ( ) 2s srank A rank A= = , the above problem can be solved analytically and
the point r rLχ∗ ∈ providing the minimum value of ( ) ( )Tc P cχ χ+ + is given by:
( ) ( )11 1T Tr sr sr sr srP A A P A b A c cχ
−∗ − −= + − .
It follows that
( ) ( ) ( )11( ) T Tr sr sr sr srl c b A c A P A b A c
−−= + + , { }1, 2r∈ .
84
Therefore, for all c Bη∈ , a lower bound on the value of rl can be obtained by solving the
following quadratic optimization problem:
( ) ( ) ( ){ }11min ( ) min T Tr sr sr sr src B c
l c b A c A P A b A cη η
−−
∈ ≤= + + .
The above problem can be solved numerically. If
( ) ( ) ( ){ }11min T Tsr sr sr src
b A c A P A b A cη
μ−−
≤+ + > , { }1, 2r∈ ,
then the sets rL , { }1, 2r∈ , are outside the bounded set xB . In this situation, the convex set
xS does not intersect with the sets 1vS or 2vS . Let xSχ′∈ be the intersection point of xS and
an ellipse T Pχ χ ρ= , 0ρ > , tangent to S . Such point is given by
( ) ( )( ) 11 1T Tex ex
exP C C P Cχ δ−
− −′ = . Therefore, if ( )1 0exC A χ′ < , which is equivalent to
( ) ( )11 0
Tex exC A P C δ− < , then we have 0e < in xS for system 1Σ . If ( )2 0exC A χ′ > , which
is equivalent to ( ) ( )12 0
Tex exC A P C δ− > , then we have 0e > in xS for the system 2Σ . In the
following, the above two cases are discussed and it is shown that if any of the conditions
(3.42) or (3.43) is satisfied, then regulation for the switched system can be achieved.
Part 2: Regulation results for the case where either (3.42) or (3.43) is satisfied.
In the rest of the paper, the time BT ′ represents the time at which the χ state trajectory
enters the set xB and continues to evolve in xB thereafter. It is easy to see that B BT T′ ≤ since
x xB B⊂ .
• Case 1: In this case, condition (3.42) is assumed to be satisfied (See Figure 3.1). The set
xS is such that all the state trajectories leaving the set xS and entering region 1 have
85
0e < . Therefore, if the state trajectory leaves the set xS and enters region 1, it will not hit
the set xS again. Two possible cases can be considered here depending on whether the
state trajectory enters the bounded set xB from region 1 or region 2 at time Bt T ′= . In the
case the state trajectory enters the set xB from region 1, then the state trajectory cannot
hit the switching surface S since 0e < for all states on the switching surface within the
set xB . Therefore, the state trajectory is confined to evolve in region 1 inside the set xB .
Since 1clΣ is an asymptotically stable system, the state trajectory will converge to the
origin asymptotically, which implies that regulation in the switched system (3.7) is
achieved. Consider now the case where the state trajectory enters the set xB from region
2. Since the state trajectory is confined to evolve inside the set xB and since the system
2clΣ is asymptotically stable, the state trajectory must hit the switching surface and
switching must take place. Following switching, and based on the dynamics of the
impulsive switched closed system as presented in (3.36), the state vector immediately
following switching will be in region 1. Since 0e < for all states on the switching
surface within the set xB for the system 1Σ , the state trajectory cannot cross the
switching surface again and will continue to evolve in region 1 thereafter. The state
trajectory will approach the origin asymptotically, which implies that regulation will be
achieved.
• Case 2: In this case, condition (3.43) is assumed to be satisfied (See Figure 3.2). The set
xS is such that all the state trajectories leaving the set xS and entering region 2 have
0e > . Assume the state trajectory enters the set xB from region 2 at time Bt T ′= . Given
86
that the system 2clΣ is asymptotically stable, and that the trajectories are confined to
evolve inside xB , the state trajectory must cross the switching surface to approach the
origin. However, this is not possible since, with respect to the system 2clΣ , 0e > for all
xSχ ∈ . Therefore, it is not possible to have the state trajectory enter the set xB from
region 2 at time Bt T ′= . Consider now the case where the state trajectory enters the set xB
from region 1. If the state trajectory crosses the switching surface to enter into region 2,
then that will result in a contradiction similar to that discussed for the case where the state
trajectory enters the set xB from region 2. Therefore, once the state trajectory enters the
set xB from region 1 at time Bt T ′= , it will continue to evolve in region 1 forever and will
never cross the switching surface to enter into region 2. Since 1clΣ is an asymptotically
stable system, then the state trajectory will converge to the origin asymptotically, which
implies that regulation for the switched system (3.7) is achieved.
Therefore, based on the above analysis, if either the conditions for case 1 expressed in (3.42)
or the conditions for case 2 expressed in (3.43) are satisfied, then regulation for the switched
system (3.7) can be achieved.
Remark 3.6
The solution to the regulation problem presented above depends on the maximum magnitude
of the signal rw . This is expected since although switching in the closed loop system takes
place between two linear systems, the resulting switched system is nonlinear. Therefore both
the shape and maximum magnitude of the external inputs are expected in general to affect the
regulator design procedure.
87
Figure 3.1. State trajectories of 1clΣ with initial conditions on the switching surface inside the
set xB for case 1.
Figure 3.2. State trajectories of 2clΣ with initial conditions on the switching surface inside the
set xB for case 2.
S
2vSBounded set xB
ex2C A
exCRegion 2: 2
clΣ
Region 1: 1clΣ
χ′
2L
2χ∗
S
1vS
Bounded set xB
ex1C A
exCRegion 2: 2
clΣ
Region 1: 1clΣ
χ′
1L
1χ∗
88
3.4.4. Regulator synthesis for the switched system
Based on the regulation conditions presented above for the switched system (3.7), a regulator
synthesis method is presented in this section. The proposed method aims at finding a
regulator such that the sufficient regulation conditions of either Theorem 3.2, Theorem 3.3 or
Theorem 3.4 are satisfied. The first step in the synthesis algorithm is to design the gains rK
and rL using (3.17) such that the switched systems (3.9) and (3.10) are asymptotically stable.
Since QA is a fixed stability matrix, it follows from Lemma 3.1 that the switched closed loop
system (3.7) is internally stable under arbitrary switching. With the gains rK and rL already
designed using (3.17), the only remaining unknown parameter in (3.31) is rQC , which
appears in ˆrC . Based on the first equation of (3.32), ˆ
rΠ can be determined first. Therefore, if
rQC is such that
ˆ0 0 0r
e eQ r rC C D⎡ ⎤Π + =⎣ ⎦… , (3.44)
then (3.32) will be satisfied. To avoid having an empty set of solutions rQC in (3.32), we
should have qn h≥ . To determine the unknown variable rQC in the controller so that
regulation is achieved in the switched closed loop system (3.7), conditions (3.19) and (3.20)
of Theorem 3.1 and either one of the sufficient conditions (3.33), (3.38), (3.42) or (3.43) for
regulation given in Theorems 3.2, 3.3 and 3.4 need to be satisfied. Since condition (3.19) is a
bilinear matrix inequality that is generally very hard to solve, a congruence transformation
will be used in the following to transform (3.19) into a linear matrix inequality first.
89
We partition P in (3.19) as [ ]1 2
2 3
n nT
P P
P P×
⎛ ⎞⎜ ⎟⎜ ⎟⎝ ⎠
and define 11 1P−Ω = , 1
2 1 2P P−Ω = ,
13 3 2 1 2
TP P P P−Ω = − and 1
2
0T I
Ω⎡ ⎤Ω = ⎢ ⎥−Ω⎣ ⎦
. Then we have:
121
32
121 2
133 2 1 2
111
133 2 1 2
0,
0
,00
00,
00
T
T
TT
IP
I
II P PP
P P P P
PP
P P P P
−
−
−
−
−
ΩΩ ⎡ ⎤⎡ ⎤= ⎢ ⎥⎢ ⎥ Ω−Ω⎣ ⎦ ⎣ ⎦
Ω⎡ ⎤ ⎡ ⎤Ω = =⎢ ⎥ ⎢ ⎥Ω− ⎣ ⎦⎣ ⎦
Ω⎡ ⎤ ⎡ ⎤Ω Ω = =⎢ ⎥ ⎢ ⎥Ω− ⎣ ⎦⎣ ⎦
(3.45)
and
[ ] [ ]
( )
1 2 2
3
1 2 3
0,
00
, , , .
r
r
Q Q rr r r r r r r Q r r
r r rTr
Q Q r
r r r
r Q
A B CA B K A B K B C B K
A L CPA
A B CA L C
M C
⎡ ⎤−⎡ ⎤ ⎡ ⎤+ Ω Ω − + Ω +⎢ ⎥⎢ ⎥ ⎣ ⎦+⎣ ⎦⎢ ⎥Ω Ω = ⎢ ⎥−⎡ ⎤⎢ ⎥Ω ⎢ ⎥+⎢ ⎥⎣ ⎦⎣ ⎦
Ω Ω Ω
It is obvious that if 1 0Ω > and 3 0Ω > , then 0P > . Multiplying (3.19) from the left side by
( ),diag IΩ and from the right side by ( ),Tdiag IΩ yields:
( )
( ) ( )( )
2
1 21 2 3 1 2 3
3 3
22 3
1
0, , , , , ,
0 00.
0 1
r r
T T T Tr r r
T Tr
Tr Q r Q r
Tr T
A P PA P PE
E P I
IM C M C E
IE I
αε μακ
α
ε μακ
⎡ ⎤Ω Ω +Ω Ω + Ω Ω Ω⎢ ⎥
−⎢ ⎥Ω −⎢ ⎥⎣ ⎦⎡ ⎤Ω Ω⎡ ⎤ ⎡ ⎤
Ω Ω Ω + Ω Ω Ω +⎢ ⎥⎢ ⎥ ⎢ ⎥Ω Ω⎣ ⎦ ⎣ ⎦⎢ ⎥= ≤⎢ ⎥−⎡ ⎤⎢ ⎥−⎢ ⎥Ω Ω⎢ ⎥⎣ ⎦⎣ ⎦
(3.46)
90
Similarly, multiplying (3.20) from the left side by ( ), ,diag I IΩ and from the right side
by ( ), ,Tdiag I IΩ yields:
( )
( )
1 12
3 2
1 2
0 00
0
0 0
0
TexT
Ter
ex er
CI
I D
C D II
κ
β μκ
βκ
⎡ ⎤Ω Ω⎡ ⎤ ⎡ ⎤⎢ ⎥⎢ ⎥ ⎢ ⎥Ω −Ω⎣ ⎦⎣ ⎦⎢ ⎥⎢ ⎥⎛ ⎞⎢ ⎥− ≥⎜ ⎟
⎝ ⎠⎢ ⎥⎢ ⎥Ω −Ω⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦⎣ ⎦
(3.47)
The main steps in the proposed regulator synthesis procedure can be summarized below.
1) Design of an internally stabilizing controller: Design the controller parameters
rK and rL to make the switched closed loop system (3.7) asymptotically stable, by
solving the LMIs (3.17) for the unknown matrices 0TK KP P= > , 0T
L LP P= > and
matrices rK , rL , { }1, 2r∈ .
2) Design of rQ : Determine ˆrΠ by solving the first equation in (3.32), then design the
controller parameter rQC by solving the following optimization problem in the
unknown scalars 0β > , 0μ > , and matrices 1 0Ω > , 3 0Ω > , 2Ω and rQC , { }1, 2r∈ ,
based on a preset positive scalar α and 0 1ε< :
Minimize β
Subject to (3.46) and (3.47),
ˆ0 0 0.r
ee Q r rC C D
⎧⎪⎨⎡ ⎤Π + =⎪⎣ ⎦⎩ …
(3.48)
3) Verification: If β δ< , then based on condition (3.33) in Theorem 3.2, the
resulting controller given by (3.5) and (3.6) achieves regulation. If β δ≥ , the
condition (3.33) cannot be satisfied and conditions (3.38), (3.42) and (3.43) need to
91
be further checked. Based on the resulting closed loop system, solve for P using
(3.45) and solve for Π using (3.34), then determine the values of η and γ . If any of
the regulation conditions (3.38), (3.42) and (3.43) are satisfied, the designed
controller yields regulation.
Note that (3.46) is linear in the unknown parameters β , μ , 1Ω , 2Ω , 3Ω and rQC only if α
is prefixed. Let mα denote an upper bound on the value of α such that
( ) ( ) 11 2 3 1 2 3
3
0, , , , , , 0
0r r
Tr Q r QM C M C α
Ω⎡ ⎤Ω Ω Ω + Ω Ω Ω + ≤⎢ ⎥Ω⎣ ⎦
,
with unknown 1 0Ω > , 3 0Ω > , 2Ω and rQC . The problem to solve mα is not linear, but it is
easily seen to be quasi-convex. Hence mα can be estimated using the bisection algorithm. A
search for a solution to the LMIs in step 2 and 3 can be performed iteratively by considering
values of α in the interval ( )0, mα α∈ .
In the following section, a numerical example is used to illustrate the regulator synthesis
procedure proposed in this paper.
3.4.5. Numerical example
In this section, the regulator synthesis method proposed above will be used to design a
controller that cancels the contact vibrations in the mechanical system as shown in section
2.6. First, rewriting (2.40) in the state space form given in (3.1) results in:
1
0 1110000 10
A⎡ ⎤
= ⎢ ⎥− −⎣ ⎦, 2
0 110000 5
A⎡ ⎤
= ⎢ ⎥− −⎣ ⎦,
1 2
0 2020 0
H Hπ
π⎡ ⎤
= = ⎢ ⎥−⎣ ⎦, 10 20
0100
w w⎡ ⎤
= = ⎢ ⎥⎣ ⎦
, 1 2
01
B B⎡ ⎤
= = ⎢ ⎥⎣ ⎦
,
92
[ ]1 2 1 0eC C C= = = , 31
0 010
0.8 1.7xD −⎡ ⎤= ×⎢ ⎥⎣ ⎦
, 32
0 010
0.75 0.75xD −⎡ ⎤= ×⎢ ⎥⎣ ⎦
,
[ ] 81 2 1 2 0 1 10y y e eD D D D −= = = = − × .
To construct a parameterized set of stabilizing controllers, the gains rK , rL , { }1, 2r∈ , are
designed based on (3.17) with
[ ] 71 2 1.5081 0.0001 10K K= = − − × , [ ] 5
1 2 0.0018 1186.81 10L L= = − × .
Since 3 3rH ×∈R , the base stabilizing controller is augmented with a parameter rQ with
4 3qn h= > = and here given by ,1 ,2 ,3 ,42 3 4( 1) ( 1) ( 1) ( 1)
r r r rrQ
s s s sθ θ θ θ
= + + ++ + + +
. The corresponding
state space representation for rQ is:
1 0 0 01 1 0 00 1 1 00 0 1 1
QA
−⎡ ⎤⎢ ⎥−⎢ ⎥=⎢ ⎥−⎢ ⎥−⎣ ⎦
,
1000
QB
⎡ ⎤⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎣ ⎦
,
,1 ,2 ,3 ,4rQ r r r rC θ θ θ θ⎡ ⎤= ⎣ ⎦ . (3.49)
By solving the first Sylvester equation in (3.32), we have:
10 8
7 9
11 8
7 8
11 8
7 81
18 18
16
1.1374 10 1.0102 106.3478 10 7.1466 106.3469 10 1.0058 10
6.2055 10 1.4305 106.3469 10 1.0058 10
ˆ 6.2055 10 1.4305 10 4.1911 10 3.7735 10
2.3710 10 2
− −
− −
− −
− −
− −
− −
− −
−
× ×− × ×− × − ×
× − ×− × − ×
Π = × − ×× ×
− × 16
20 20
21 22
23 23
.6334 10 6.1103 10 6.5731 10
1.0304 10 9.8888 101.5995 10 1.6145 10
−
− −
− −
− −
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥×⎢ ⎥× − ×⎢ ⎥⎢ ⎥− × − ×⎢ ⎥− × ×⎣ ⎦
,
10 8
7 9
11 8
7 8
11 8
7 82
18 18
16
1.1408 10 1.0104 106.3485 10 7.1684 106.3479 10 1.0058 10
6.2058 10 1.4393 106.3479 10 1.0058 10
ˆ 6.2058 10 1.4393 10 4.2196 10 3.8310 10
2.4071 10 2
− −
− −
− −
− −
− −
− −
− −
−
× ×− × ×− × − ×
× − ×− × − ×
Π = × − ×× ×
− × 16
20 20
21 21
23 23
.6513 10 6.2025 10 6.6170 10
1.0372 10 1.0037 101.6233 10 1.6249 10
−
− −
− −
− −
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥×⎢ ⎥× − ×⎢ ⎥⎢ ⎥− × − ×⎢ ⎥− × ×⎣ ⎦
93
First calculate { }1,20
max ( ) 100rrt
w tκ∈≥
= = . Then, solving the LMIs in (3.46)-(3.48) for the
minimum value of β yields:
[ ]1
3740.5811 796.4368 134.3644 10QC = − × ,
[ ]2
31656.8 647.0492 210.3280 10QC = − − × ,
75.5253 10μ −= × , 632.1251 10β −= × .
Since 6 632.1251 10 30 10β δ− −= × > = × , regulation condition (3.33) cannot be satisfied.
Solving the Sylvester equation (3.34) and evaluating η results in 68.8398 10η −= × .
Condition (3.38) is then verified but cannot be satisfied. Verifying condition (3.43), we have:
( ) ( )( ) ( ) ( ){ }
1 72
11 7 72 2 2 2
4.8559 10 0,
min 6.3438 10 5.5253 10 ,
Tex ex
T Ts s s sc
C A P C
b A c A P A b A cη
δ
μ
− −
−− − −
≤
= × >
+ + = × > = ×
which means that (3.43) is satisfied. Then, based on Theorem 3.4, the designed regulator can
achieve regulation for the switched system (2.40).
The simulation results of the response for the closed loop system under switching are
illustrated in Figures 3.3-3.4. It can be seen that if the performance variable is smaller than
630 10 m−− × , the mass enters into contact with the contact surface cS and the model of the
system switches. It can also be seen that the disturbance rd changes at the switching times.
But even in the presence of switching, the switched system performance variable e still
converges to zero, which means the mass m asymptotically follows the contact surface cS at
the desired separation 30 micrometers.
94
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1-100
-50
0
50
100
e (m
icro
met
ers)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1-0.2
-0.1
0
0.1
0.2
d r(1,1
) (m
/sec
2 )
Time(sec)
Figure 3.3. Simulation results for the case of 30ev = − micrometers showing the performance variable e and the switching component (1,1)rd in the disturbance rd .
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-40
-30
-20
-10
0
10
20
v an
d v s (m
icro
met
ers)
Time(sec)
mass height vcontact surface vs
Figure 3.4. Simulation results for the case of 30ev = − micrometers showing the regulated mass height v and the contact surface displacement sv .
95
3.5 Regulation of bimodal systems: a multiple Lyapunov function
approach
In the previous section, a common Lyapunov function approach was used in the analysis of
the input-output stability of the switched system and a regulator synthesis method was also
presented. The approach based on a common Lyapunov function is a conservative approach
and may not yield the desired regulator. In this section, a regulator synthesis approach based
on multiple Lyapunov functions is proposed. Based on the properties of the time-varying
switching surface S in (3.2), the whole closed loop system state space is partitioned into
three adjacent regions. Using the resulting state space partitions, a regulation condition for
the switched closed loop system is formulated using a multiple Lyapunov function approach
and a corresponding regulator synthesis method is then proposed using an iterative LMI
algorithm. Furthermore, the regulation interpolation conditions for 1clΣ and 2
clΣ will be
derived in the frequency domain and are expressed in affine linear equalities with smaller
dimension as compared with the one in (3.44) and, therefore, are easier for the LMI software
tools to solve.
The switched bimodal system is given by:
0, (0) , , , :
1 if ,2 if ,
dr r r ry ydr r re ed
r rre ed
r re ed
r r
x A x B u E d x xy C x D de C x D d
e C x D dr
e C x D dδδ
⎧ = + + =⎪ = +⎪⎪ = +Σ ⎨⎪ ⎧ = + ≤⎪ = ⎨⎪ = + >⎩⎩
(3.50)
where nx∈R is the state vector, u∈R is the control input, y∈R is the measurement signal
to be fed to the controller, e∈R is the performance variable to be regulated and is assumed
to be measurable, { }1,2r∈ is the index of the system rΣ under consideration at time t , and
96
δ is a constant satisfying 0δ > . Switching between 1Σ and 2Σ is performed according to
the performance variable e which defines a time-varying switching surface S given by:
{ }: n e edr rS x e C x D d δ= ∈ = + =R . (3.51)
The external signal hrd ∈R , ,1 ,[ ]T
r r r hd d d= , representing disturbance and/or reference
signals, is also assumed to switch according to the rule given in (3.50), and where each ,r jd
is of the form:
( ) { }, , , ,, 1
1( ) cos , 1, 2 , 1, ,
j
j
kr j r j r j r j
r j i i i ki
d t c t c r j hω φ +=
= + + ∈ =∑ , (3.52)
with known amplitudes ,r jic , frequencies ,r j
iω , and phases ,r jiφ , 1, , ji k= … ; 1, ,j h= … . The
model in (3.50) can be expressed as in (3.1) easily. The signals ,r jd , { }1, 2r∈ ; 1, ,j h= ;
can be considered to be the outputs of exosystems given by:
0, , , , ,
,, ,
, (0) , :
,r j r j r j r j r jes
r jr j j r j
w H w w w
d C w
⎧ = =⎪∑ ⎨=⎪⎩
where 2,
jkr jw ∈ , and
( )
( )
2,1
,
2,
0 1[0] [0]
0
[0] [0] [0]0 1
[0]0
[0] [0] 0j
r j
r j
r jk
H
ω
ω
⎡ ⎤⎡ ⎤⎢ ⎥⎢ ⎥−⎢ ⎥⎢ ⎥⎣ ⎦⎢ ⎥
⎢ ⎥= ⎢ ⎥⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥−⎢ ⎥⎣ ⎦⎢ ⎥⎣ ⎦
…
…
97
( )( )
( )( )
, ,1 1
, , ,1 1 1
, ,,
, , ,
,1
sin
cos
(0) sin
cos
j j
j j j
j
r j r j
r j r j r j
r j r jr jk k
r j r j r jk k k
r jk
c
c
w c
c
c
φ
ω φ
φ
ω φ
+
⎡ ⎤⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦⎢ ⎥⎢ ⎥
= ⎢ ⎥⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦⎢ ⎥⎢ ⎥⎣ ⎦
[ ] [ ]1 (2 1)
1 0 1 0 1j
j kC
× +⎡ ⎤= ⎣ ⎦
Define the state vector ,1 , 1[ ]T T Tr r r h mw w w ×= , where
1
(2 1)h
jj
m k=
= +∑ . The exosystem esr∑
generating the signal rd is then given by:
0, (0): r r r r res
rr w r
w H w w wd C w
= =⎧∑ ⎨ =⎩
(3.53)
where
,1
,
0 00 00 0
r
r
r h
HH
H
⎡ ⎤⎢ ⎥= ⎢ ⎥⎢ ⎥⎣ ⎦
, 1 0 0
0 00 0
w
h
CC
C
⎡ ⎤⎢ ⎥= ⎢ ⎥⎢ ⎥⎣ ⎦
, and
0,1
00,
r
r
r h
ww
w
⎡ ⎤⎢ ⎥= ⎢ ⎥⎢ ⎥⎣ ⎦
.
It worth noting that there are many ways to obtain matrices rH and wC in (3.53). The model
in (3.50) is considered in this section since it is more suitable to work with to derive
regulation conditions in the frequency domain.
In the following, an output feedback controller will be constructed to regulate the
performance variable e against the external input signal rd such that lim ( ) 0t
e t→∞
= . In this
section, Q parameterized stabilizing controllers, as presented in the case of system (3.1), will
be used with the switched system (3.50) to achieve regulation. The resulting closed loop
system is given by the following state space representation:
98
, ,
:1 if ,2 if ,
r r rex ed
r rclr ex ed
r rex ed
r r
A E de C D d
e C D dr
e C D d
χ χχ
χ δχ δ
⎧ = +⎪
= +⎪∑ ⎨⎧ = + ≤⎪ = ⎨⎪ = + >⎩⎩
(3.54)
where 00 0
rr r r r Q r r
yr Q Q r
yr r r
A B K B C B K
A A B CA L C
+⎡ ⎤⎢ ⎥
= −⎢ ⎥⎢ ⎥+⎣ ⎦
, ( ) ( ) ( )TTTd yd T d yd
r r Q r r r rE E B D E L D⎡ ⎤= − +⎢ ⎥⎣ ⎦,
0 0ex eC C⎡ ⎤= ⎣ ⎦ .
The purpose of introducing the Q -parameterized controllers is to find appropriate
matrices QA , QB and rQC , { }1, 2r∈ , to solve the output feedback regulation problem for the
switched system (3.54). In the following, frequency domain regulation conditions for each of
the closed loop systems are presented first, then a regulation condition for the switched
closed loop system and a corresponding regulator synthesis algorithm are discussed.
3.5.1 Regulation conditions for 1clΣ and 2
clΣ in the frequency domain
Combining the plant rΣ in (3.1) and the block rJ in (3.6) into a single system rT , we have:
11 12
21 22ˆr
rr r
Qr r
T
de T Tyy y T T
⎡ ⎤ ⎡ ⎤⎡ ⎤= ⎢ ⎥ ⎢ ⎥⎢ ⎥−⎣ ⎦ ⎣ ⎦ ⎣ ⎦
, (3.55)
where 11rT , 12
rT , 21rT , 22
rT , { }1, 2r∈ , are as in presented in section 3.4.2. Let ( )rD s , , ( )r jD s
and ( )E s denote the Laplace transforms of the inputs rd , ,r jd and the performance variable
e , respectively. Let ,d jrE , ,ed j
rD , and ,yd jrD denote the thj columns of the matrices d
rE , edrD ,
and ydrD , respectively. Let 1
, , ,( ) ( ) ( )r r r r r r
hT Q T Q T QF s F s F s⎡ ⎤= ⎣ ⎦ denote the closed loop
99
system transfer function relating ( )rD s to ( )E s . Then based on (3.28) and (3.6), for each
{ }1, 2r∈ , the transfer function , ( )r rT QF s can be written as:
11 12 21, ( ) ( ) ( ) ( ) ( )
r rT Q r r r rF s T s T s Q s T s= + . (3.56)
Therefore, we have
, ,1
( ) ( ) ( ),r r
hj
T Q r jj
E s F s D s=
=∑
where
11, 12 21,, ( ) ( ) ( ) ( ) ( )
r r
j j jT Q r r r rF s T s T s Q s T s= + ,
and
,
11, , ,
,
: 00
d jr r r r r r
j y d j dy jr r r r r r r
e ed jr
A B K B K ET A L C E L D
C D
⎡ ⎤+⎢ ⎥
+ − −⎢ ⎥⎢ ⎥⎣ ⎦
, 12 12
1212 12: 0
r r r rr rer
r r
A B K BA ET CC D
+⎡ ⎤ ⎡ ⎤=⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎣ ⎦⎣ ⎦
, ,21,
, :y d j yd j
r r r r r rjr y yd j
r r
A L C E L DT
C D⎡ ⎤+ − −⎢ ⎥−⎢ ⎥⎣ ⎦
, 1, ,j h= .
According to the properties of ,r jd in (3.52), all the poles of , ( )r jD s , 1, ,j h= , are simple
poles located on the imaginary axis of the complex plane. For a given { }1, 2r∈ , some of the
disturbances , ( )r jD s , 1, ,j h= , may have common poles. Let prn denote the total number
of the different poles in the expression for ( )rD s . Let rp , 1, , ,prn= denote such poles, rS
denote the set of disturbance indices j such that rp is a pole of , ( )r jD s , and , ( )r j rR p
denote the residue of , ( )r jD s at the pole rp . Then we can state the following interpolation
conditions for regulation.
100
Theorem 3.5
Assume that, for each { }1, 2r∈ , the closed loop system clrΣ in (3.54) is internally stable.
Consider the closed loop system transfer function , ( )r rT QF s given in (3.56). For any given
{ }1,2r∈ , regulation in the resulting closed loop system (3.54) is achieved if and only if the
following interpolation conditions are satisfied
( )11, 12 21,, ( ) ( ) ( ) ( ) ( ) 0, 1, , .
r
j j pr j r r r r r r r r r r
j
R p T p T p Q p T p n∈
+ = =∑S
(3.57)
Proof: The Laplace transform of the closed loop system is given by
, ,1
( ) ( ) ( )r r
hj
T Q r jj
E s F s D s=
= ∑ .
According to (3.52), the poles of , ( )r jD s are all simple poles located on the imaginary axis.
Using partial fraction expansion, the response ( )e t can be expressed as follows:
1, , 0
1
1( ) ( ) ( ) ( )pr
r r
r
nj
r j r T Q rjr
e t L R p F p e ts p
−
= ∈
⎡ ⎤⎛ ⎞= +⎢ ⎥⎜ ⎟⎜ ⎟−⎢ ⎥⎝ ⎠⎣ ⎦
∑ ∑S
, (3.58)
where 1L− is the inverse Laplace transform operator, 0 ( )e t denotes the sum of responses
corresponding to partial fractions with , ( )r rT QF s poles and the response to non-zero initial
conditions. Since for each { }1, 2r∈ the closed loop system is asymptotically stable, the
response to nonzero initial conditions converges asymptotically to zero. Also, since
, ( )r rT QF s RH∞∈ , the space of real rational stable transfer matrices, then the responses
corresponding to partial fractions with , ( )r rT QF s poles asymptotically converge to zero.
101
Therefore, the asymptotic properties of the disturbance response are determined by the first
term on the RHS of (3.58):
1, ,
1
1( ) ( ) ( )pr
r r
r
nj
r j r T Q rjr
e t L R p F ps p
−
= ∈
⎡ ⎤⎛ ⎞′ = ⎢ ⎥⎜ ⎟⎜ ⎟−⎢ ⎥⎝ ⎠⎣ ⎦
∑ ∑S
.
Necessity: Since all the poles rp are on the imaginary axis and are distinct, if lim ( ) 0t
e t→∞
′ = ,
then:
, ,( ) ( ) 0, 1,r r
r
j pr j r T Q r r
j
R p F p n∈
= =∑S
.
It follows that:
( )11, 12 21,, ( ) ( ) ( ) ( ) ( ) 0, 1, ,
r
j j pr j r r r r r r r r r r
j
R p T p T p Q p T p n∈
+ = =∑S
,
which means that the interpolation conditions (3.57) must be satisfied.
Sufficiency: Based on the above discussion, satisfying the interpolation conditions (3.57)
implies that ( )e t in (3.58) asymptotically decays to zero.
Let ,1
( ) ( )qn
r r i ii
Q s sθ ψ=
=∑ , where ( ) ( ) ii s s pψ −= + , 1, , qi n= , and 0p > . Then, the state
space representation of rQ is as given in (3.6) with:
0 01
00 1
q q
Q
n n
pp
A
p×
−⎡ ⎤⎢ ⎥−⎢ ⎥=⎢ ⎥⎢ ⎥−⎣ ⎦
,
1
10
0q
Q
n
B
×
⎡ ⎤⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎣ ⎦
,
,1 ,2 , 1.
r qq
Q r r r n nC θ θ θ
×⎡ ⎤= ⎣ ⎦ (3.59)
102
Assume ( )rD s has 0k distinct pairs of complex conjugate poles on the imaginary axis and
one pole at zero (the pole related to the disturbance offset in (3.52)), so that the total number
of the different poles is 02 1prn k= + . Let rp + , 01, ,k= , denote the 0k distinct complex
poles in the top half of the complex plane. Then we have the following result.
Theorem 3.6
Assume, for each { }1, 2r∈ , the closed loop system clrΣ in (3.7) is internally stable. For any
given { }1, 2r∈ , the set rQS of
rQC matrices in (3.49) which yield regulation in the closed
loop system clrΣ in (3.7) can be defined by the following affine constraint on
rQC :
{ }: 0q
r r r r r
nT TQ Q QC A C Bθ θ= ∈ + =S , (3.60)
where the matrix r
Aθ and the vector r
Bθ are given by:
0 0
0 0
1 1 1, ,
1 1 1, ,
1, ,
1, ,
1
( ) ( )
( ) ( )
,( ) ( )
( ) ( )
(0) (0)
q
q
r q
q
q
nr re r r re r
nr im r r im r
nk kr re r r re r
nk kr im r r im r
nr r
V p V p
V p V p
AV p V p
V p V p
V V
θ
+ +
+ +
+ +
+ +
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥= ⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
0
0
0 1,0 1,
0,0,
0
( )( )
,( )( )(0)
r
r re r
r im r
kr re r
kr im r
r
V pV p
BV pV p
V
θ
+
+
+
+
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥
= ⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
(3.61)
and where , ( )ir re rV p + and , ( )i
r im rV p + denote the real and imaginary parts of ( )ir rV p + given by:
12 21,, 0( ) ( ) ( ) ( ) ( ) , 1, , , 1, , ,
r
i jr r i r r j r r r r r q
j
V p p R p T p T p k i nψ+ + + + +
∈
⎛ ⎞= = =⎜ ⎟⎜ ⎟
⎝ ⎠∑S
0 11,, 0( ) ( ) ( ), 1, , .
r
jr r r j r r r
j
V p R p T p k+ + +
∈
= =∑S
103
Proof: (Necessity) Using the expression for rQC given in (3.49), the interpolation conditions
can be rewritten as:
11, 12 21,, ,
1( ) ( ) ( ) ( ) ( ) 0, 1, ,
q
r
nj j p
r j r r r r i r r i r r r rij
R p T p T p p T p nθ ψ=∈
⎛ ⎞+ = =⎜ ⎟⎜ ⎟
⎝ ⎠∑ ∑S
. (3.62)
Define:
0 11,,( ) ( ) ( ), 1, ,
r
j pr r r j r r r r
j
V p R p T p n∈
= =∑S
,
12 21,,( ) ( ) ( ) ( ) ( ), 1, , 1, ,
r
i j pr r r j r r r i r r r r q
j
V p R p T p p T p n i nψ∈
= = =∑S
.
Then (3.62) can be written as
0,
1( ) ( ) 0, 1, , 1, ,
qni p
r r r i r r r qi
V p V p n i nθ=
+ = = =∑ .
Since we have ( ) ( )i ir rV s V s= , 1, , qi n= , where s is the complex conjugate of s , the
interpolation conditions corresponding to the 02k complex conjugate poles are equivalent to
the 0k interpolation conditions corresponding to the 0k poles rp + , 01, ,k= . For any
complex number s , let , ( )ir reV s and , ( )i
r imV s denote, respectively, the real and imaginary parts
of the functions ( )irV s , 1, , qi n= . The interpolation condition evaluated at a pole rp + ,
01, , k= , on the imaginary axis can be written as:
0 0, , , , , ,
1 1
( ) ( ) ( ) ( ) 0q qn n
i ir re r r i r re r r im r r i r im r
i i
V p V p j V p V pθ θ+ + + +
= =
⎡ ⎤ ⎡ ⎤+ + + =⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦∑ ∑ .
Hence, for each such pole, the above equation yields two linear equations of the form:
104
0, , ,
1
0, , ,
1
( ) ( ) 0,
( ) ( ) 0.
q
q
ni
r re r i r rei
ni
r im r i r imi
V s V s
V s V s
θ
θ
=
=
⎡ ⎤+ =⎢ ⎥
⎣ ⎦⎡ ⎤
+ =⎢ ⎥⎣ ⎦
∑
∑ (3.63)
Therefore, considering the 0k poles rp + selected as mentioned above yields 02k linear
equations in the unknowns ,r iθ , 1, , qi n= . On the other hand, the interpolation condition
evaluated at the pole at 0 can be written in the form:
0,
1(0) (0) 0
qni
r r i ri
V Vθ=
⎡ ⎤+ =⎢ ⎥
⎣ ⎦∑ .
Therefore, the interpolation conditions (3.57) are equivalent to the following affine constraint
on the matrix rQC :
0r r r
TQA C Bθ θ+ = ,
where rQC is given by (3.49), and where the p
r qn n× matrix r
Aθ and the 1prn × vector
rBθ are
given in (3.61).
(Sufficiency) For a given ,r rT QF and according to the expressions given above for r
Aθ and r
Bθ ,
satisfying (3.60) implies that (3.62) is satisfied and that regulation can be achieved.
3.5.2 Regulation condition and regulator synthesis for the switched system
The previous section presented the construction of a set rQS of
rQC matrices that yield
regulation for each of the closed loop systems clrΣ , { }1, 2r∈ . If qn is such that p
q rn n≥ , then
the set rQS will not be empty. However, simply selecting
rQC , { }1, 2r∈ , in the sets rQS does
not guarantee achieving regulation in the switched closed loop system. In this section,
regulation conditions for the switched closed loop system and a corresponding regulator
105
synthesis procedure based on piecewise quadratic Lyapunov functions will be presented. Let
{ }1,20
max ( )edr rr
t
D d tσ∈≥
= , δ δ σ− = − and δ δ σ+ = + . Based on the switching condition in (3.54),
the whole closed loop system state space can be partitioned into three adjacent regions 1ℜ ,
2ℜ and 3ℜ (as shown in Figure 3.5), where
{ }1N
exCχ χ δ−ℜ = ∈ ≤R ,
{ }2N
exCχ χ δ+ℜ = ∈ >R ,
3exN
ex
CC
δχ χ
δ−
+
⎧ − ⎫−⎡ ⎤ ⎡ ⎤⎪ ⎪ℜ = ∈ ≤⎨ ⎬⎢ ⎥ ⎢ ⎥⎪ ⎪⎣ ⎦⎣ ⎦⎩ ⎭
R .
It is easy to see that only the system 1clΣ can be active in the region 1ℜ and that only the
system 2clΣ can be active in the region 2ℜ , while any one of the two systems 1
clΣ and 2clΣ can
be active in the region 3ℜ . Furthermore, the boundaries between the three regions are the
two hypersurfaces
{ }13N
exS Cχ χ δ−= ∈ =R
and
{ }23N
exS Cχ χ δ+= ∈ =R .
Let 1 1, , , , 0ex j N NC c c c ×⎡ ⎤= ≠⎣ ⎦ and 1, , , ,T
j Nx x xχ ⎡ ⎤= ⎣ ⎦ , then there must exist an index
{ }1, ,j N∈ … such that 0jc ≠ . Consequently, for 13Sχ ∈ or 23Sχ ∈ , jx can be expressed as
106
Figure 3.5. State space partitions where the vector Tex exC C= .
j exj
x Ccδχ −= + or j ex
j
x Ccδχ += + , respectively, where 1 1 1, , , , ,
T
j j Nx x x xχ − +⎡ ⎤= ⎣ ⎦ and
1 1 11 , , , , ,ex j j N
j
C c c c cc − +⎡ ⎤= − ⎣ ⎦ . Therefore, the hypersurfaces 13S and 23S can be expressed
as
{ }113 1, NS H fχ χ −= + ∈R ,
{ }123 2 , NS H fχ χ −= + ∈R ,
where ( ) ( )
( ) ( )
1 1j j
ex
N j N j
I
H CI
− × −
− × −
⎡ ⎤⎢ ⎥
= ⎢ ⎥⎢ ⎥⎣ ⎦
,
( )
( )
1 1
1
1
0
0
j
j
N j
fcδ− ×
−
− ×
⎡ ⎤⎢ ⎥⎢ ⎥
= ⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
and
( )
( )
1 1
2
1
0
0
j
j
N j
fcδ− ×
+
− ×
⎡ ⎤⎢ ⎥⎢ ⎥
= ⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
.
We define a piecewise quadratic Lyapunov function ( ) 2 0T Ti i iV P q gχ χ χ χ= + + > for
each iχ ∈ℜ , { }1, 2,3i∈ , where Ti iP P= . For ( )V χ to be continuous on the hypersurfaces
13S and 23S , we should have
2ℜ
1ℜ3ℜ
exC
exC χ δ+=
exC χ δ−=
exC χ δ=
107
( ) ( ) ( )( ) ( ) ( )1 1 1 1 1 1
1 3 1 3 1 3
2
2 ,
T T
T T
H f P H f q H f g
H f P H f q H f g
χ χ χ
χ χ χ
+ + + + +
= + + + + +
( ) ( ) ( )( ) ( ) ( )
2 2 2 2 2 2
2 3 2 3 2 3
2
2 ,
T T
T T
H f P H f q H f g
H f P H f q H f g
χ χ χ
χ χ χ
+ + + + +
= + + + + +
which result in the following constraints, for { }1, 2r∈ ,
( )( ) ( )( ) ( ) ( )
3
3 3
3 3 3
0,
0,
2 0.
Tr
T Tr r r
TTr r r r r r
H P P H
H P P f H q q
f P P f q q f g g
− =
− + − =
− + − + − =
(3.64)
Let 1 exC C= , 2 exC C= − , 3ex
ex
CC
C−⎡ ⎤
= ⎢ ⎥⎣ ⎦
, 1δ δ−= , 2δ δ+= − and 3
δδ
δ−
+
−⎡ ⎤= ⎢ ⎥⎣ ⎦
. We then have the
following result.
Theorem 3.7
Consider the closed loop system (3.54) subject to a bounded input rd . Let
{ }1,20
max ( ) 0rrt
d tγ∈≥
= ≠ . If there exist matrices 1 1TP P= , 2 2
TP P= , 3 3TP P= , 1 2 1 2, , , 0η η ζ ζ ≥ ∈R
and 23 3, 0η ζ ≥ ∈R , and positive scalars α , κ and β such that the following matrix
inequalities are satisfied:
2 0,
T T Tr r r r r r r r r r r r
T T Tr r r r r r r r r r
T Tr r r r
A P P A P A q q C P Eq A q C g q E
E P E q I
α α ζα ζ α ζ δ
κ
⎡ ⎤+ + + −⎢ ⎥
+ − + ≤⎢ ⎥⎢ ⎥−⎣ ⎦
{ }1, 2r∈ (3.65)
3 3 3 3 3 3 3 3
3 3 3 3 3 3 3 3
3 3
2 0,
T T Tr r r r
T T T T Tr r
T Tr r
A P P A P A q q C P Eq A q C g q E
E P E q I
α α ζα ζ α ζ δ
κ
⎡ ⎤+ + + −⎢ ⎥
+ − + ≤⎢ ⎥⎢ ⎥−⎣ ⎦
{ }1, 2r∈ (3.66)
108
02 0 0
00 0
0
T Ti i i i ex
T T Ti i i i i i
Ted
ex ed
P q C Cq C g
I D
C D I
α α ηα η α η δ
β κγ
βγ
⎡ ⎤+⎢ ⎥+ −⎢ ⎥⎢ ⎥⎛ ⎞ >⎢ ⎥−⎜ ⎟⎢ ⎥⎝ ⎠⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
, { }1, 2,3i∈ (3.67)
1 if 0,0,
2 if 0.r r r
TQA C B rθ θ
δδ>⎧
+ = = ⎨ <⎩ (3.68)
(3.64) and ,β δ≤ (3.69)
then the switched closed loop system (3.54) achieves regulation.
Proof: First, (3.67) implies
02
Ti i i i
T T Ti i i i i i
P q Cq C gα α η
α η α η δ⎡ ⎤+
>⎢ ⎥+ −⎣ ⎦, { }1, 2,3i∈ .
Multiplying from the left by 1Tχ⎡ ⎤⎣ ⎦ and from the right by 1TTχ⎡ ⎤⎣ ⎦ yields,
( ) ( )( )( ) 2 ( ) 0Ti i iV t C tα χ η χ δ+ − > , iχ ∈ℜ .
Based on the definition of iη , iC and iδ , we have
( )( ) 0Ti i iC tη χ δ− ≤ , iχ ∈ℜ .
It follows that ( )( ) 0V tχ > for each iχ ∈ℜ .
Now consider the state space representation (3.54). Inequalities (3.65) and (3.66) imply
that for any vector 1TT T
rdχ⎡ ⎤⎣ ⎦ , we have
1 2 1 0
T T T Tr r r r r r r r r r r r
T T Tr r r r r r r r r r
T Tr r r r r r
A P P A P A q q C P Eq A q C g q E
d E P E q I d
χ α α ζ χα ζ α ζ δ
κ
⎡ ⎤+ + + −⎡ ⎤ ⎡ ⎤⎢ ⎥⎢ ⎥ ⎢ ⎥+ − + ≤⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥⎢ ⎥ ⎢ ⎥−⎣ ⎦ ⎣ ⎦⎣ ⎦
109
and
3 3 3 3 3 3 3 3
3 3 3 3 3 3 3 3
3 3
1 2 1 0
T T T Tr r r r
T T T T Tr r
T Tr r r r
A P P A P A q q C P Eq A q C g q E
d E P E q I d
χ α α ζ χα ζ α ζ δ
κ
⎡ ⎤+ + + −⎡ ⎤ ⎡ ⎤⎢ ⎥⎢ ⎥ ⎢ ⎥+ − + ≤⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥⎢ ⎥ ⎢ ⎥−⎣ ⎦ ⎣ ⎦⎣ ⎦
, { }1, 2r∈ .
It follows that, for each iχ ∈ℜ , { }1, 2,3i∈ , we have
( )( )( ( )) ( ( )) 2 ( ) ( ) ( ) 0T Ti i i r r
d V t V t C t d t d tdt
χ α χ ζ χ δ κ+ − − − ≤
and
( )( ) 0Ti i iC tζ χ δ− ≤ .
Condition (3.64) implies that the Lyapunov function ( ( ))V tχ is continuous on the boundary
hypersurfaces 13S and 23S . Hence, ( ( )) 0d V tdt
χ ≤ holds whenever
2( ( )) ( ) ( )Tr rV t d t d tα χ κγ κ≥ ≥ in the whole state space. Consequently, ( ( ))V tχ cannot
ultimately exceed the value 2κ γα
, namely,
2limsup ( ( ))t
V t κχ γα→∞
≤ .
Inequality (3.67) is equivalent to
( )[ ]
02 0 0 0 0
0 0
T Ti i i i ex
T T Ti i i i i i ex ed
Ted
P q C Cq C g C D
I D
α α ηγα η α η δβ
β γ κ
⎡ ⎤ ⎡ ⎤+⎢ ⎥ ⎢ ⎥+ − − >⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥−⎣ ⎦ ⎣ ⎦
.
Multiplying from the left by 1T Trdχ⎡ ⎤⎣ ⎦ and from the right by 1
TT Trdχ⎡ ⎤⎣ ⎦ yields, for
each iχ ∈ℜ ,
110
( ) ( )( )
( )
2
2
( ) ( ) 2 ( ) ( ) ( )
( ) .
T Ti i i r re t V t C t d t d t
V t
β βα χ η χ δ κγ γ
β βα χ κ γγ γ
⎛ ⎞⎛ ⎞< + − + −⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠⎛ ⎞⎛ ⎞
≤ + −⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠
Since 2limsup ( ( ))t
V t κχ γα→∞
≤ as shown before, we have
2 2 2 2limsup ( )t
e t β βκγ κ γ βγ γ→∞
⎛ ⎞⎛ ⎞< + − =⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠, iχ ∈ℜ .
Hence, we have
limsup ( )t
e t β→∞
< .
If β δ≤ as in (3.69), then after a long enough time, the output ( )e t will always be such that
( )e t β δ< ≤ and there will be no more switching. For the case of 0δ > , the system 1cl∑ is
the active closed loop system once switching stops. Since the systems 1cl∑ is asymptotically
stable and since 1 1Q QC ∈S based on (3.68), then regulation in the switched closed loop system
is achieved. The analysis is similar for the case of 0δ < .
In the following, a regulator synthesis method based on Theorem 3.7 and that makes use
of an iterative LMI algorithm will be presented. Since QA and QB in (3.6) are fixed, then the
synthesis method aims at finding controller gains rK , rL and matrices r rQ QC ∈S , { }1, 2r∈ ,
so that conditions (3.65)-(3.69) are satisfied. Since (3.65) includes the unknown parameters
rK , rL , rP , rq , rg , rη , rζ , α , κ and β , (3.65) is nonlinear in the unknown variables.
Therefore, the gains rK and rL will be designed first to yield an internally stable switched
111
closed loop system (3.54) under arbitrary switching. Using the common Lyapunov function
method, the gains rK and rL can be designed as follows:
[ ] [ ] { }
{ }
0, 0, 0, 1, 2 ,
0, 0, 0, 1, 2 .
Tr r r K K r r r k k k K
Ty yr r r L L r r r L L L L
A B K P P A B K P P r
A L C P P A L C P P r
α α
α α
+ + + + < > > ∈
⎡ ⎤ ⎡ ⎤+ + + + < > > ∈⎣ ⎦ ⎣ ⎦ (3.70)
Less conservative methods based on piecewise quadratic Lyapunov functions or polyhedral
Lyapunov functions can also be used [32,33,34]. Once the gains rK and rL are designed,
then the regulator synthesis is completed by finding matrices r rQ QC ∈S , { }1,2r∈ , that satisfy
the regulation conditions in Theorem 3.7. Let ,i nP , ,i nq , ,i ng , ,i nη , ,i nζ , nα , nκ , nβ , 1
nQC and
2
nQC denote the values of iP , iq , ig , iη , iζ , α , κ , β ,
1QC and 2QC in the thn step in the
proposed iterative synthesis algorithm, where { }1, 2,3i∈ . The main steps in the proposed
regulator synthesis procedure are summarized below.
1) Design an internally stabilizing controller: Design the controller gains rK and rL
to make the switched closed loop system (3.54) internally stable by solving (3.70) for
the unknown matrices 0TK KP P= > , 0T
L LP P= > and the gains rK , rL , { }1, 2r∈ .
2) Initialize rQ : For the case of 0δ > , first find the set 1QS by computing
1Aθ and
1Bθ
in (3.61), then initialize 1QC as
1 1
0Q QC ∈S and initialize
2QC as 2
10 qnQC ×∈R . For
example, we can set 1
0QC by solving
1min QC subject to
1 1 10T
QA C Bθ θ+ = and set
2
010
qQ nC ×= . For the case of 0δ < , follow a similar procedure.
3) Initialize α : One way to select an initial value 0 0α > for α can be as follows. First,
find an upper bound mα on the value of α by solving maxα subject to
112
0T Tr r r
T Tr
A P PA P A q qq A q g
α αα α
⎡ ⎤+ + +<⎢ ⎥
+⎣ ⎦, 0T
P qq g⎡ ⎤
>⎢ ⎥⎣ ⎦
, { }1, 2r∈ , with the unknowns TP P= ,
q and g ; and with known matrices rK , rL , and 0rQC , which is a generalized
eigenvalue problem. Then select an initial value 0α in the interval ( )0 0, mα α∈ .
4) (start the thn step). Minimize nβ subject to (3.64)-(3.67) with the unknowns ,i nP ,
,i nq , ,i ng , ,i nη , ,i nζ , nκ , and nβ ; and with known 1nα − , 1
1nQC − and
2
1nQC − .
5) Minimize nβ subject to (3.65)-(3.68) with the unknowns nα , 1
nQC ,
2
nQC , ,i nη , ,i nζ , nκ
and nβ ; and with known ,i nP , ,i nq , ,i ng .
6) If nβ δ≤ or 1n n ββ β σ− − < , where βσ is a prescribed tolerance, stop the algorithm.
Else set 1n n= + , then go to step 4.
For a given set of initial values for rK , rL , 0rQC and 0α , the above algorithm converges to a
local solution since β is guaranteed to decrease or stay the same in every iteration. Therefore,
if there is no feasible solution that satisfies nβ δ≤ , the initial values can be modified and
the algorithm can be restarted again. In the following, a numerical example is presented to
show the effectiveness of the proposed regulator synthesis method.
3.5.3 Numerical example
In this section, the regulator synthesis method will be used to design a controller that cancels
the contact vibrations in the same mechanical system as in section 2.6.
To construct a parameterized set of stabilizing controllers, the gains rK , rL , { }1, 2r∈ , are
first designed based on (3.70), which results in
113
7 21 2 1.5088 10 7.5351 10K K ⎡ ⎤= = − × − ×⎣ ⎦ ,
2 71 2 1.7969 10 1.1874 10
TL L ⎡ ⎤= = − × − ×⎣ ⎦ .
Based on the analysis of the poles of rd , we have 0 1k = and 02 2prn k= = . A rQ parameter
as in (3.49) with 4qn = is selected and is given by
,1 ,2 ,3 ,42 3 4( )
10 ( 10) ( 10) ( 10)r r r r
rQ ss s s sθ θ θ θ
= + + ++ + + +
.
Therefore, we have
10 0 0 01 10 0 00 1 10 00 0 1 10
QA
−⎡ ⎤⎢ ⎥−⎢ ⎥=⎢ ⎥−⎢ ⎥−⎣ ⎦
and
1000
QB
⎡ ⎤⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎣ ⎦
,
Since 0δ < , condition (3.60) needs to be satisfied for 2r = and we consider only the set 2QS
given by (3.60) where
2
19 21 23 25
20 21 23 24
2.6002 10 1.4806 10 5.6920 10 6.4702 105.4004 10 3.9026 10 3.2624 10 8.0294 10
Aθ
− − − −
− − − −
⎡ ⎤− × − × × ×= ⎢ ⎥− × × × − ×⎣ ⎦ ,
2
11
10
2.2489 104.5740 10
Bθ
−
−
⎡ ⎤− ×= ⎢ ⎥×⎣ ⎦
.
The initial values are set to
0 0.3α = , [ ]1
0 0 0 0 0QC = ,
2
0 8 11 12 118.0689 10 1.1509 10 1.0859 10 1.1404 10QC ⎡ ⎤= × − × × ×⎣ ⎦ .
Using the iterative algorithm in the regulator synthesis procedure, after 16 iterations, one
solution is found as
0.2637α = ,
114
1
6 11 11 113.3746 10 7.9973 10 8.3783 10 7.3212 10QC ⎡ ⎤= × − × − × ×⎣ ⎦ ,
2
8 11 12 117.2437 10 2.3512 10 1.0859 10 1.3756 10QC ⎡ ⎤= × − × × ×⎣ ⎦ ,
and 6 628.6357 10 30 10β δ− −= × < = × . Since the regulation conditions (3.19)-(3.33) are
satisfied, then the designed regulator can achieve regulation. The simulation results of the
response of the closed loop system under switching are illustrated in Figure 3.6 and 3.7,
which shows the performance variable e converging to zero even in the presence of
switching. Hence, the position v of the mass m asymptotically follows the contact surface
sv with the desired separation d ev v− .
3.6 Conclusion
The problem of regulation in bimodal switched systems against known disturbance or
reference signals is discussed. Regulator design approaches using the parameterization of a
set of stabilizing controllers for the switched system are presented using properly formulated
linear matrix inequalities. Two regulator synthesis approaches are considered based on
different regulation conditions in this chapter. The first approach is based on working with a
common Lyapunov function for the closed loop switched system. In this case, the forced
switched closed loop system involving a parameterized controller is transformed into an
unforced impulsive switched system. Consequently, the original regulation problem is
transformed into a stability analysis problem for the origin of the impulsive switched system.
Then, three sufficient conditions for regulation are presented based on a common Lyapunov
function and a regulator synthesis method is proposed by solving properly formulated LMIs.
The second approach is based on considering multiple Lyapunov functions for the closed
115
loop switched system. In this case, the regulation interpolation conditions for each subsystem
are first derived in the frequency domain. Then, based on the property of the switching
surface in the switched closed loop system, a sufficient regulation condition for the switched
closed loop system is proposed using multiple Lyapunov functions. A regulator synthesis
algorithm is finally developed by iteratively solving LMIs. A simulation example of a
mechanical system subject to contact vibrations is used to illustrate the effectiveness of the
both proposed regulator synthesis methods.
116
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1-50
0
50
e (m
icro
met
ers)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1-0.2
-0.1
0
0.1
0.2
d r(1,1
) (m
/sec
2 )
Time(sec)
Figure 3.6. Simulation results for the case of 30ev = − micrometers showing the performance variable e and the switching component (1,1)rd in the disturbance rd .
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-40
-30
-20
-10
0
10
20
v an
d v s (m
icro
met
ers)
Time(sec)
mass height vcontact surface vs
Figure 3.7. Simulation results for the case of 30ev = − micrometers showing the regulated mass height v and the contact surface displacement sv .
117
CHAPTER 4
Regulation in Switched Bimodal Systems with
an 2H Performance Constraint
4.1. Introduction
This chapter considers a regulation problem for discrete-time switched bimodal linear
systems where it is desired to achieve output regulation against known deterministic
disturbance or reference signals in the presence of unknown random disturbances. The
development of the proposed regulation method involves two main steps. In the first step, the
regulation problem for the switched system against the known deterministic disturbance or
reference signals is approached by constructing a set of observer-based Q -parameterized
stabilizing controllers that satisfy a sufficient regulation condition for the switched system. In
the second step, an 2H performance constraint is added to identify, from among the already
constructed regulators, those that provide the best 2H performance against the unknown
random disturbances. The proposed regulator is successfully evaluated on a bimodal
switched mechanical system experimental setup, hence demonstrating the effectiveness of
118
the proposed regulation approach in dealing simultaneously with known deterministic and
unknown random exogenous inputs.
The rest of the chapter is organized as follows. In section 4.2, the general regulation
problem for the switched bimodal system subject to known deterministic exogenous inputs
and unknown random exogenous inputs is presented. In section 4.3, the construction of a Q -
parameterized stabilizing set of switched controllers for the switched system is discussed. In
section 4.4, regulation conditions for each subsystem in the plant model and regulation
conditions for the resulting switched closed loop system are presented. Then, a set of
parameterized 2H controllers is constructed to deal with the unknown random exogenous
inputs. In section 4.5, a regulator synthesis method is proposed by solving some properly
formulated linear matrix inequalities. In section 4.6, a numerical example is presented to
show the effectiveness of the proposed designed method. In section 4.7, the proposed
regulator is evaluated on an experimental setup motivated by the flying height regulation
problem in hard disk drives. In the experimental setup, the tip of a flexible beam is supposed
to maintain a constant separation with respect to a surface with a known profile, while also
being subject to disturbance forces having a known deterministic component and an
unknown random component. The system exhibits a switching behavior depending on
whether contact takes place between the surface to be tracked and the tip of the beam. The
experimental results successfully demonstrate the effectiveness of the proposed approach in
achieving exact output regulation against known sinusoidal exogenous inputs and in
minimizing the effects of the random disturbance input in the presence of switching in the
system dynamics. Concluding remarks are presented in section 4.8.
119
4.2. The regulation problem for the switched bimodal system
Consider the discrete-time switched system given by the following state space representation:
0( 1) ( ) ( ) ( ) ( ), (0) , ( ) ( ) ( ) ( ), ( ) ( ) ( ) ( ), :1 if
x xr r r r r wy y yr r r r we e er r r r wr
x k A x k B u k D d k F d k x xy k C x k D d k F d ke k C x k D d k F d k
er
+ = + + + == + += + +Σ
=( ) ( ) ( ) ( ) ,
2 if ( ) ( ) ( ) ( ) ,
e e er r r r we e er r r r w
k C x k D d k F d ke k C x k D d k F d k
δδ
⎧⎪⎪⎪⎨⎪ ⎧ = + + ≤⎪ ⎨⎪ = + + >⎩⎩
(4.1)
where nx∈R is the state vector, u∈R is the control input, y∈R is the measurement signal
to be fed to the controller, e∈R is the performance variable to be regulated and is assumed
to be measurable, { }1,2r∈ is the index of the system rΣ under consideration at time k , and
δ is a constant satisfying 0δ > . The external signal lwd ∈R represents unknown signals
such as measurement noise and/or random disturbances. The external signal hrd ∈R ,
representing the known disturbance and/or reference signals, is also assumed to switch
according to the rule given in (4.1). The signal rd is given by:
1[ ]h Tr r rd d d= , (4.2)
where each jrd is of the form:
( ) { }, , , ,1
1( ) ( ) cos ( ) ( ) , 1, 2 , 1, ,
j
j
kj r j r j r j r j
r i i i ki
d k c k k k k c r j hω φ +=
= + + ∈ =∑ , (4.3)
with known amplitudes ,r jic , frequencies ,r j
iω , and phases ,r jiφ , 1, , ji k= … ; 1, ,j h= … . In
the following, it is assumed that for any given 0k ≥ , the system must operate in only one of
the two modes corresponding to { }1,2r∈ . For the switched system (4.1), it is desired to
120
construct an output feedback controller to regulate the performance variable e of the
switched system against the external input signals rd and wd .
4.3. Parameterization of a set of stabilizing controllers
The controller design approach presented in this chapter relies on the construction of a Q -
parameterized set of output feedback stabilizing controllers for the switched system. In this
section, the construction of sets of Q -parameterized output feedback stabilizing controllers
for the switched system (4.1) is first discussed. The stability of the resulting switched closed
loop system is then analyzed. As in the case of linear time invariant systems, the construction
of a Q -parameterized set of stabilizing controllers involves two steps. The first step involves
the construction of a central controller in the form of a switched observer-based state
feedback controller. The second step involves augmenting the central controller with stable
dynamics to construct a Q -parameterized set of stabilizing controllers.
4.3.1. Q -parameterized controller
Consider following observer-based state feedback controller for the switched system (4.1):
( ) 0ˆ ˆ ˆ ˆ ˆ( 1) ( ) ( ) ( ) ( ) , (0) ,:
ˆ( ) ( ),r r ro
rr
x k A x k B u k L y k y k x xu k K x k⎧ + = + + − =⎪⎨
=⎪⎩ (4.4)
where ˆ( )x k is an estimate of the plant state vector ( )x k and ˆ ˆ( ) ( )yry k C x k= is an estimate of
the plant output ( )y k . The state feedback gains rK , { }1, 2r∈ , and the observer gains rL ,
{ }1, 2r∈ , are assumed to switch according to the rule given in (4.1). The construction of a
Q -parameterized set of stabilizing controllers for the switched system (4.1) proceeds along
121
the same lines as in the case of linear time-invariant systems as shown in chapter 3. Each
controller is expressed as a linear fractional transformation involving a fixed system rJ , and
a proper stable parameter rQ that could be chosen as desired (See Figure 4.1). The state
space representation of the system rJ is given by:
0ˆ ˆ ˆ ˆ( 1) ( ) ( ) ( ) ( ), (0) ,ˆ: ( ) ( ) ( ),
ˆ ˆ( ) ( ) ( ) ( ).
yr r r r r r r Q
r r Q
yr
x k A L C B K x k L y k B y k x x
J u k K x k y k
y k y k y k C x k
⎧ + = + + − + =⎪
= +⎨⎪
− = −⎩
(4.5)
The parameter rQ is given by:
( ) 0ˆ( 1) ( ) ( ) ( ) , (0) ,:
( ) ( ),r
Q Q Q Q Q Qr
Q Q Q
x k A x k B y k y k x xQ
y k C x k
⎧ + = + − =⎪⎨
=⎪⎩ (4.6)
where QnQx ∈ . In particular, throughout the rest of the paper, the rQ parameter is such that
QA is a fixed stability matrix, QB is a fixed matrix, and the matrix rQC changes with
{ }1, 2r∈ . For a given { }1,2r∈ , the controller formulated in (4.5) and (4.6) is a
parameterized stabilizing controller for the system rΣ , where the system rQ represents a free
design parameter that can be chosen as desired as long as it is an asymptotically stable
system. A suitable choice for rQ can be a parameterization of the form ( )1
( )qn
ir r i
i
Q z zθ ψ=
=∑ ,
where ( )i zψ are stable basis functions. Such representation can be used to approximate any
proper stable real rational transfer function. With such representation for rQ , a corresponding
realization is as given in (4.6), where QA and QB are fixed matrices, and where the
parameters , 1ir qi nθ = … , appear in
rQC . Here the rQC matrices are selected to achieve
122
Figure 4.1. Closed loop system with a Q -parameterized controller.
regulation according to regulation conditions and an 2H performance criterion to be
discussed later.
Let ˆ( ) ( ) ( )x k x k x k= − denote the state estimation error and 1
TT T TQ N
x x xχ×
⎡ ⎤= ⎣ ⎦
denote the state vector for the resulting closed loop system with 2 QN n n= + . The resulting
closed loop system is given by the following state space representation:
( 1) ( ) ( ) ( ), ( ) ( ) ( ) ( ),
:1 if ( ) ( ) ( ) ( ) ,2 if ( ) ( ) ( ) ( ) ,
d wr r r r w
ex e er r r r wcl
r ex e er r r r wex e er r r r w
k A k E d k E d ke k C k D d k F d k
e k C k D d k F d kr
e k C k D d k F d k
χ χχ
χ δχ δ
⎧ + = + +⎪
= + +⎪∑ ⎨⎧ = + + ≤⎪ = ⎨⎪ = + + >⎩⎩
(4.7)
where 00 0
rr r r r Q r r
yr Q Q r
yr r r
A B K B C B K
A A B CA L C
+⎡ ⎤⎢ ⎥
= −⎢ ⎥⎢ ⎥+⎣ ⎦
,
( ) ( ) ( )TTTd x y T x y
r r Q r r r rE D B D D L D⎡ ⎤= − +⎢ ⎥⎣ ⎦, ( ) ( ) ( )
TTTw x y T x yr r Q r r r rE F B F F L F⎡ ⎤= − +⎢ ⎥⎣ ⎦
,
0 0ex er rC C⎡ ⎤= ⎣ ⎦ .
Plant r∑
rQ
ˆy y−Qy
rJ
rd e
u y
wd
123
4.3.2 Stability of the Q - parameterized switched closed loop system
In this section, the internal stability of the closed loop system is studied by considering the
system (4.1) in the absence of the signals rd and wd , and studying the stability properties of
the origin for the resulting unforced switched system. Consider first the system (4.1) in the
absence of the signal rd and wd . The state equation for the resulting system is given by:
( 1) ( )( 1) 0 ( ) .( 1) 0 0 ( )
rr r r r Q r r
yQ Q Q r Q
yr r r
A B K B C B Kx k x kx k A B C x kx k A L C x k
+⎡ ⎤+⎡ ⎤ ⎡ ⎤⎢ ⎥⎢ ⎥ ⎢ ⎥+ = −⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥⎢ ⎥ ⎢ ⎥+ +⎣ ⎦ ⎣ ⎦⎣ ⎦
(4.8)
Based on (4.8), define the following three subsystems:
( 1) [ ] ( ),r r rx k A B K x k+ = + (4.9)
( 1) [ ] ( ),yr r rx k A L C x k+ = + (4.10)
( 1) ( ).Q Q Qx k A x k+ = (4.11)
Then, the internal stability of the switched closed loop system (4.8) is given by the following
result.
Lemma 4.1
If the switched systems (4.9) and (4.10) are each asymptotically stable under arbitrary
switching and QA is a stability matrix, then the switched system (4.8) is also asymptotically
stable under arbitrary switching.
Proof: If the switched system (4.10) is asymptotically stable under arbitrary switching signal,
then we have:
( ) ( , 0) (0) (0) kx k k x c x aφ= ≤ ,
124
for some 0c > and (0,1)a∈ where ( ,0)kφ is the state transition matrix of the switched
system (4.10). Therefore, 0x → as k →∞ for any (0)x . Similarly, if the system (4.11) is
asymptotically stable, then we have:
( ) ( ,0) (0) (0) kQ Q Q Q Q Qx k k x c x aφ= ≤
for some 0Qc > and (0,1)Qa ∈ , and where ( ,0)Q kφ is the state transition matrix of the
system (4.11). The expression of the state variable Qx in (4.8) is given by:
1
0( ) ( ,0) (0) ( , ) ( )
ky
Q Q Q Q Q rj
x k k x k j B C x jφ φ−
=
= −∑ .
It follows
1
0
1
max0
( ) ( ,0) (0) ( , ) ( )
(0) (0)
ky
Q Q Q Q Q rj
kk y k j j
Q Q Q Q Q r Qj
x k k x k j B C x j
c x a c c x B C a a
φ φ−
=
−−
=
≤ +
≤ +
∑
∑.
Hence,
max
max
(0)( ) (0) ( ), if
1( ) (0) (0) , if
yQ Q Q rk k k
Q Q Q Q Q Q
k y kQ Q Q Q Q Q r Q Q
c ca x B Cx k c x a a a a a
ax k c x a c c x B C ka a a
⎧⎪ ≤ + − ≠⎪
−⎨⎪ ≤ + =⎪⎩
,
which yields
max2 (0)
( ) (0)1
yQ Q r k
Q Q Q m
c c x B Cx k c x a
a
⎡ ⎤⎢ ⎥≤ +
−⎢ ⎥⎣ ⎦, Qa a≠ ,
where max
yQ rB C is the maximum value of y
Q rB C and max( , )m Qa a a= . Therefore, we
have 0Qx → as k →∞ for any (0)Qx . Similarly, we have that 0x → as k →∞ for any
(0)x in (4.8).
125
Consider now the closed loop system (4.7), then we have the following result.
Lemma 4.2
Assume the system (4.8) is asymptotically stable under arbitrary switching. If the exogenous
inputs rd and wd are bounded, then the state vector χ of the system (4.7) is also bounded.
Proof: If the system (4.8) is asymptotically stable under arbitrary switching, then we have:
( ) ( ,0) (0) (0) kx x xk k c aχ φ χ χ= ≤ ,
for some 0xc > , (0,1)xa ∈ , and where ( ,0)x kφ is the state transition matrix of the switched
system (4.8). Then the state vector ( )kχ of system (4.7) is given by
1
0
( )( ) ( ,0) (0) ( , )
( )
krd w
x x r rj w
d kk k k j E E
d kχ φ χ φ
−
=
⎡ ⎤⎡ ⎤= + ⎢ ⎥⎣ ⎦
⎣ ⎦∑ .
Therefore,
1
max 0
( ) (0)k
rk d w k jx x x r r Q
j w
dk c a c E E a
dχ χ
−−
=
⎡ ⎤⎡ ⎤≤ + ⎢ ⎥⎣ ⎦
⎣ ⎦∑ ,
where max
d wr rE E⎡ ⎤⎣ ⎦ are the maximum value of d w
r rE E⎡ ⎤⎣ ⎦ , { }1, 2r∈ . From this we can
prove that ( )kχ is bounded if rd and wd are bounded.
The switched system (4.8) can be designed to be asymptotically stable using the following
result.
126
Theorem 4.1
The origin is an asymptotically stable equilibrium point for the switched system (4.8) under
arbitrary switching if QA is a stability matrix and there exist matrices 0TK KP P= > ,
0TL LP P= > , and matrices rK , rL , { }1, 2r∈ , such that:
{ }{ }
[ ] [ ] 0, , 1, 2 ,
[ ] [ ] 0, , 1, 2 .
Tr r r K r r r K
y T yr r r L r r r L
A B K P A B K P i j
A L C P A L C P i j
+ + − < ∈
+ + − < ∈ (4.12)
Proof: Condition (4.12) implies that each of the switched systems (4.9) and (4.10) admits a
common Lyapunov function. Hence, each of those two systems is asymptotically stable
under arbitrary switching. Based on Lemma 4.1, it follows that the switched system (4.8) is
asymptotically stable under arbitrary switching.
Remark 4.1
The main reason for considering common Lyapunov functions as opposed to multiple
Lyapunov functions in (4.12) is a follows. The existence of a common Lyapunov function for
each of the switched systems (4.9) and (4.10) implies that there exists a common Lyapunov
function for the system (4.8). As outlined in Section 4.4, the regulation conditions are
developed based on a common Lyapunov function for the switched closed loop system. The
use of a common Lyapunov function allows the regulation conditions to be transformed into
linear matrix inequalities (LMIs) in the regulator synthesis algorithm using an appropriate
congruence transformation, while considering multiple Lyapunov functions would have
made such transformation very difficult or impossible.
127
Since the matrix QA in (4.6) is chosen to be a stability matrix, then a parameterized set of
stabilizing controllers can be designed by solving (4.12) numerically. In the following, the
parameters in rQC in (4.6) will be chosen to satisfy the regulation conditions and an 2H
performance criterion simultaneously.
4.4. Regulation Conditions for switched system
This section presents regulation conditions for the switched closed loop system (4.7) against
the external input signals rd and wd . Due to the presence of the unknown random signal wd ,
it is not possible to achieve exact output regulation. Consequently, the goal behind the
regulator design is to achieve exact output regulation against the known deterministic
exogenous input rd and minimize the 2H norm of the closed loop system relating the input
wd to the performance variable e . It is proposed in this section to achieve these goals by
exploiting the flexibility in the selection of the Q parameter in the parameterized set of
stabilizing controllers for the switched system. More specifically, by properly selecting the
matrices rQC , { }1, 2r∈ , in (4.6), it will be possible to develop regulation conditions
corresponding to the above stated goals. The development of the regulation conditions is
performed in three steps. First, regulation conditions for each of the closed loop systems clrΣ ,
{ }1, 2r∈ , against the known deterministic input rd and in the absence of the signal wd are
reviewed. The second step determines conditions on the switched closed loop system such
that, with the controllers designed in step 1, switching in the closed loop system would stop
in a finite time in the presence of the both rd and wd . The conditions obtained in step two
128
would guarantee exact regulation in the closed loop system against rd . In the third step, the
conditions from step two are augmented with additional constraints so that the closed loop
system obtained once switching ends satisfies an 2H norm constraint. Each of the three steps
is discussed in detail in the following.
4.4.1 Regulation Conditions for 1Σ and 2Σ
This section provides the regulation conditions for each of the closed loop systems 1clΣ and
2clΣ in the presence of the deterministic signal rd and with 0wd = . The derivations presented
in the following parallel those for the continuous-time case and are included here for
completeness.
The plant rΣ and the block rJ shown in Figure 4.1 can be combined into a single system
rT . In the absence of wd , the dynamics of the system rT , { }1, 2r∈ , can be represented as
follows:
( )( 1) ( ),
( )0( 1) ( ) 0
( )( ): ( ) 0 0 ,
( )( )
( )ˆ( ) ( ) 0 0
( )
xrr r r r r r r
y x yQr r r r r r
re er r r
Q
y yr r
d kA B K B Kx k x k D By kA L Cx k x k D L D
d kx kT e k C D
y kx k
x ky k y k C D
x k
++ ⎡ ⎤⎡ ⎤⎡ ⎤⎡ ⎤ ⎡ ⎤= + ⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥ ⎢ ⎥++ − −⎣ ⎦ ⎣ ⎦⎣ ⎦ ⎣ ⎦ ⎣ ⎦
⎡ ⎤⎡ ⎤⎡ ⎤ ⎡ ⎤= + ⎢ ⎥⎢ ⎥⎣ ⎦ ⎣ ⎦⎣ ⎦ ⎣ ⎦
⎡ ⎤⎡ ⎤ ⎡− = − +⎢ ⎥⎣ ⎦ ⎣⎣ ⎦
( ).
( )r
Q
d ky k
⎧⎪⎪⎪⎪⎨⎪⎪ ⎡ ⎤⎪ ⎤ ⎢ ⎥⎦⎪ ⎣ ⎦⎩
(4.13)
Therefore, we have that:
( )( )( )ˆ( ) ( )
rr
Q
d ke kT
y ky k y k⎡ ⎤⎡ ⎤
= ⎢ ⎥⎢ ⎥−⎣ ⎦ ⎣ ⎦, (4.14)
where
129
11 12
21 22r r
rr r
T TT
T T⎡ ⎤
= ⎢ ⎥⎣ ⎦
, (4.15)
and where, based on (4.13), the state space representations for 11rT , 12
rT , 21rT and 22
rT are as
11 1111
11 11: 00
xr r r r r r
r r y x yr r r r r r r
r r e er r
A B K B K DA E
T A L C D L DC D
C D
⎡ ⎤+⎡ ⎤ ⎢ ⎥
= + − −⎢ ⎥ ⎢ ⎥⎢ ⎥⎣ ⎦ ⎢ ⎥
⎣ ⎦
,
12 1212
12 12: 0r r r rr r
errr r
A B K BA ET CC D
+⎡ ⎤ ⎡ ⎤=⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎣ ⎦⎣ ⎦,
21 2121
21 21:y x y
r r r r r r r rr y y
r r r r
A E A L C D L DT
C D C D⎡ ⎤ ⎡ ⎤+ − −
=⎢ ⎥ ⎢ ⎥−⎢ ⎥ ⎢ ⎥⎣ ⎦ ⎣ ⎦,
22 2222
22 22
0: .
0
yr r r r r
r yr r r
A E A L CT
C D C⎡ ⎤ ⎡ ⎤+
=⎢ ⎥ ⎢ ⎥−⎢ ⎥ ⎢ ⎥⎣ ⎦ ⎣ ⎦
Let ( )rD z , ( )jrD z and ( )E z denote the Z transforms of the inputs rd , j
rd and the
performance variable e , respectively. Let ,x jrD , ,e j
rD , and ,y jrD denote the thj columns of
the matrices xrD , e
rD , and yrD respectively. Let 1
, , ,( ) ( ), , ( )r r r r r r
hT Q T Q T QF z F z F z⎡ ⎤= ⎣ ⎦ denote
the closed loop system transfer function relating ( )rD z to ( )E z . It follows that:
,
,1
( ) ( ) ( ),
( ) ( ),
r r
r r
T Q r
hj j
T Q rj
E z F z D z
F z D z=
=
=∑ (4.16)
where 11, 12 21,, ( ) ( ) ( ) ( ) ( )
r r
j j jT Q r r r rF z T z T z Q z T z= + , and where the corresponding state space
representations for 11, jrT , 12 ( )rT z and 21, j
rT are as given similarly in section 3.4.2. Consider
the following representation of ( )rQ z :
1( ) ( )
qni
r r ii
Q z zθ ψ=
=∑ ,
130
where ( ) ( )1 ii z z F zψ −= , 1, qi n= , are stable basis functions, and where
11
11
( )m
mm m
m
b z bF zz a z a
−
−
+=
+ + is a stable function used to adjust the dynamic properties of ( )rQ z .
Let 1, , qTn
r r rθ θ θ⎡ ⎤= ⎣ ⎦… and 1Q qn m n= + − . A realization of ( )rQ z can then be given as
follows:
1
2
1
0 0 0 0 0 01 0 0 0 0 0
0 1 0 0 0 00 0 1 0 0 00 0 0 1 0 0 00 0 0 0 0 00 0 0 0 0 1 0
Q Q
m
m
Q
n n
aa
aA
a
−
×
−⎡ ⎤⎢ ⎥−⎢ ⎥⎢ ⎥⎢ ⎥−⎢ ⎥= ⎢ ⎥−⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
,
1
2
1
1
0
0Q
m
m
Q
n
bb
bB
b
−
×
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥= ⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
,
( )1 10Q Tr rmC θ× −
⎡ ⎤= ⎣ ⎦ . (4.17)
According to the properties of jrd in (4.3), all the poles of ( )j
rD z , 1, ,j h= , are located on
the unit circle. For a given { }1, 2r∈ , some of the disturbances ( )jrD z , 1, ,j h= , may have
common poles. Let prn denote the total number of the different poles in the expression for
( )rD z . Let rp , 1, , ,prn= denote such poles. Let rS denote the set of disturbance indices
j such that rp is a pole of ( )jrD z . Assume the function ( )rD z has 0k distinct pairs of
complex conjugate poles on the unit circle and one pole at 1 (the pole related to the
disturbance offset in (4.3)), so that the total number of the different poles is 02 1prn k= + . Let
rp + , 01, ,k= , denote the 0k distinct complex poles in the top half of the complex plane.
131
Denote ( )jr rR p the residue of ( )j
rD z at the pole rp . Then we have the following
interpolation conditions for regulation.
Theorem 4.2
Assume, for each { }1, 2r∈ , the closed loop system clrΣ in (4.7) is internally stable. For any
given { }1, 2r∈ , the closed loop system clrΣ in (4.7) achieves regulation if and only if the
parameter vector rθ in rQC satisfies the following affine constraint:
0r rrA Bθ θθ + = , (4.18)
where the matrix r
Aθ and the vector r
Bθ are given by:
00 0
0 0
1 1 10 1, ,,
1 1 1 0 1, , ,
01,, ,
1,, ,
1
( ) ( ) ( )( ) ( ) ( )
, ( )( ) ( )
( ) ( )
(1) (1)
q
q
r rq
q
q
nr re r r re r
r re rn
r im r r im r r im r
n kk kr re rr re r r re r
nk kr ir im r r im r
nr r
V p V p V pV p V p V p
A BV pV p V pVV p V p
V V
θ θ
+ ++
+ + +
++ +
+ +
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥= =⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
00
0
,
( )(1)
km r
r
pV
+
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
(4.19)
and where , ( )ir re rV p + and , ( )i
r im rV p + denote, respectively, the real and imaginary parts of
( )ir rV p + given by:
12 21,0( ) ( ) ( ) ( ) ( ), 1, , , 1, , ,
r
i j jr r r r r r i r r r q
j
V p R p T p p T p k i nψ+ + + + +
∈
= = =∑S
0 11,0( ) ( ) ( ), 1, , .
r
j jr r r r r r
j
V p R p T p k+ + +
∈
= =∑S
Proof: The Z transform of the closed loop system is given by
132
,1
( ) ( ) ( )r r
hj j
T Q rj
E z F z D z=
= ∑ .
According to (4.3), the poles of ( )jrD z are all located on the unit circle. Using partial fraction
expansion, the response ( )e k can be expressed as follows:
1, 0
1
1( ) ( ) ( ) ( )pr
r r
r
nj j
r r T Q rjr
e k Z R p F p e kz p
−
= ∈
⎡ ⎤⎛ ⎞= +⎢ ⎥⎜ ⎟⎜ ⎟−⎢ ⎥⎝ ⎠⎣ ⎦
∑ ∑S
, (4.20)
where 1Z − is the inverse Z transform operator, 0 ( )e k denotes the sum of responses
corresponding to partial fractions with , ( )r rT QF z poles and the response to non-zero initial
conditions. Since for each { }1, 2r∈ the closed loop system is asymptotically stable, the
response to nonzero initial conditions converges asymptotically to zero. Also, since
, ( )r rT QF z RH∞∈ , the space of real rational stable transfer matrices, then the responses
corresponding to partial fractions with , ( )r rT QF z poles asymptotically converge to zero.
Therefore, the asymptotic properties of the disturbance response are determined by the first
term on the right hand side of (4.20):
1,
1
1( ) ( ) ( )pr
r r
r
nj j
r r T Q rjr
e k Z R p F pz p
−
= ∈
⎡ ⎤⎛ ⎞′ = ⎢ ⎥⎜ ⎟⎜ ⎟−⎢ ⎥⎝ ⎠⎣ ⎦
∑ ∑S
.
Necessity: Since all the poles rp are on the unit circle and are distinct, if lim ( ) 0k
e k→∞
′ = , then
it follows that:
,( ) ( ) 0, 1,r r
r
j j pr r T Q r r
j
R p F p n∈
= =∑S
,
or equivalently
( )11, 12 21,( ) ( ) ( ) ( ) ( ) 0, 1, ,r
j j j pr r r r r r r r r r r
j
R p T p T p Q p T p n∈
+ = =∑S
,
133
which means that the following interpolation conditions must be satisfied.
( ) { }11, 12 21,( ) ( ) ( ) ( ) ( ) 0, 1, , , 1, 2r
j j j pr r r r r r r r r r r
j
R p T p T p Q p T p n r∈
+ = = ∈∑S
, (4.21)
Using the expression for rQ given in (4.21), the interpolation conditions can be rewritten as:
11, 12 21,
1( ) ( ) ( ) ( ) ( ) 0, 1, ,
q
r
nj j i j p
r r r r r r r i r r r rij
R p T p T p p T p nθ ψ=∈
⎛ ⎞+ = =⎜ ⎟⎜ ⎟
⎝ ⎠∑ ∑S
. (4.22)
Define:
0 11,( ) ( ) ( ), 1, ,r
j j pr r r r r r r
j
V p R p T p n∈
= =∑S
,
12 21,( ) ( ) ( ) ( ) ( ), 1, , , 1, ,r
i j j pr r r r r r i r r r r q
j
V p R p T p p T p n i nψ∈
= = =∑S
.
Then (4.22) can be written as
0
1( ) ( ) 0, 1, , 1, ,
qni i p
r r r r r r qi
V p V p n i nθ=
+ = = =∑ .
Since we have ( ) ( )i ir rV z V z= , 1, , qi n= , where z is the complex conjugate of z , the
interpolation conditions corresponding to the 02k complex conjugate poles are equivalent to
the 0k interpolation conditions corresponding to the 0k poles rp + , 01, ,k= . For any
complex number z , let , ( )ir reV z and , ( )i
r imV z denote, respectively, the real and imaginary parts
of the functions ( )irV z , 1, , qi n= . The interpolation condition evaluated at a pole rp + ,
01, , k= , can be written as:
0 0, , , ,
1 1( ) ( ) ( ) ( ) 0
q qn ni i i i
r re r r r re r r im r r r im ri i
V p V p j V p V pθ θ+ + + +
= =
⎡ ⎤ ⎡ ⎤+ + + =⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦∑ ∑ .
Hence, for each such pole, the above equation yields two linear equations of the form:
134
0, ,
1
0, ,
1
( ) ( ) 0,
( ) ( ) 0.
q
q
ni i
r re r r r re ri
ni i
r im r r r im ri
V p V p
V p V p
θ
θ
+ +
=
+ +
=
⎡ ⎤+ =⎢ ⎥
⎣ ⎦⎡ ⎤
+ =⎢ ⎥⎣ ⎦
∑
∑ (4.23)
Therefore, considering the 0k poles rp + selected as mentioned above yields 02k linear
equations in the unknowns irθ , 1, , qi n= . On the other hand, the interpolation condition
evaluated at the pole at 1 can be written in the form:
0
1(1) (1) 0
qni i
r r ri
V Vθ=
⎡ ⎤+ =⎢ ⎥
⎣ ⎦∑ .
Therefore, the interpolation conditions (4.21) are equivalent to the following affine constraint
on the parameter vector rθ :
0r rrA Bθ θθ + = ,
where the pr qn n× matrix
rAθ and the 1p
rn × vector r
Bθ are given in (4.19).
Sufficiency: For a given ,r rT QF and according to the expressions given above for r
Aθ and r
Bθ ,
satisfying (4.18) implies that (4.20) is satisfied and that regulation can be achieved.
4.4.2 Regulation condition for the switched system
The previous section presented the construction of rQC matrices that yield regulation in each
of the closed loop systems clrΣ , { }1, 2r∈ , against the deterministic exogenous input rd .
However, such rQC matrices do not guarantee achieving regulation in the switched closed
loop system, even when 0wd = . In this section, based on the regulation condition (4.18) for
135
each of the two subsystems in (4.7), regulation conditions for the switched closed loop
system (4.7) are presented in the following result.
Theorem 4.3
Consider the closed loop system (4.7) subject to bounded inputs rd and wd . Let
{ } ( )2 2
1,20
max ( ) ( ) 0r wrk
d k d kκ∈≥
= + ≠ , and 0α > a preset constant and 0 1ε< . If for
{ }1, 2r∈ , there exist a matrix 0P > and positive scalars μ and β such that the following
matrix inequalities are satisfied,
( ) ( )( ) 0,1
T T d wr r r r r
T Td w d w d wr r r r r r r
A PA P P A P E E
E E PA E E P E E I
α
ε μκ
⎡ ⎤⎡ ⎤− + ⎣ ⎦⎢ ⎥<⎢ ⎥−
⎢ ⎥⎡ ⎤ ⎡ ⎤ ⎡ ⎤ −⎣ ⎦ ⎣ ⎦ ⎣ ⎦⎣ ⎦
(4.24)
( )( ) ( )
( )
0
0 0
Texr
Ter
Ter
ex e er r r
P C
DI
F
C D F I
α
β μκ
β
⎡ ⎤⎢ ⎥⎢ ⎥⎡ ⎤−⎢ ⎥⎢ ⎥ >⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦⎢ ⎥
⎡ ⎤⎢ ⎥⎣ ⎦⎣ ⎦
(4.25)
1 if 0,0,
2 if 0,r rrA B rθ θ
δθ
δ>⎧
+ = = ⎨ <⎩ (4.26)
β δ≤ . (4.27)
then the switched closed loop system (4.7) achieves regulation against rd .
136
Proof: Consider the state space representation (4.7) and the quadratic function ( ) TV Pχ χ χ= .
Then for any nonzero vector 2 1
( )( )( )
N hr
w
kd kd k
χ+ +
⎡ ⎤⎢ ⎥∈⎢ ⎥⎢ ⎥⎣ ⎦
R , we have:
( ) ( )( )
( ) ( )( ) ( ) 01( ) ( )
T T T d wr r r r r
r rT Td w d w d wr r r r r r rw w
k kA PA P P A P E Ed k d k
E E PA E E P E E Id k d k
χ χα
ε μκ
⎡ ⎤⎡ ⎤ ⎡ ⎤⎡ ⎤− + ⎣ ⎦⎢ ⎥⎢ ⎥ ⎢ ⎥ <⎢ ⎥⎢ ⎥ ⎢ ⎥−⎢ ⎥⎡ ⎤ ⎡ ⎤ ⎡ ⎤ −⎢ ⎥ ⎢ ⎥⎣ ⎦ ⎣ ⎦ ⎣ ⎦⎣ ⎦ ⎣ ⎦⎣ ⎦
.
It follows from the above inequality that for switched system (4.7):
( ) ( ) ( )1( ( 1)) ( ( )) ( ( ))
( ) ( )
Tr r
w w
d k d kV k V k V k
d k d kε μ
χ χ α χκ− ⎡ ⎤ ⎡ ⎤
+ − < −⎢ ⎥ ⎢ ⎥⎣ ⎦ ⎣ ⎦
. (4.28)
Hence, ( ( 1)) ( ( )) 0V k V kχ χ+ − < holds whenever
( ) ( ) ( )2 21( ( )) 1 ( ) ( )r wV k d k d k
ε μα χ ε μ
κ−
≥ − ≥ + .
Consequently, ( ( ))V kχ cannot ultimately exceed the value ( )1 ε μα−
, and we have:
( )1lim ( ( ))t
V kε μ
χα→∞
−≤ .
Therefore, there exists a finite time bk such that
( ( )) , bV k k kμχα
≤ ∀ ≥ (4.29)
Using Schur complement formula, inequality (4.25) is equivalent to:
( )
( )( )( )
01 0
0
Texr
T ex e eer r rr
Ter
CPC D FD
IF
αβ μ βκ
⎡ ⎤⎢ ⎥⎡ ⎤⎢ ⎥⎢ ⎥ ⎡ ⎤− >− ⎣ ⎦⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦⎢ ⎥⎣ ⎦
.
137
Multiplying from the left by ( )( )( )
T
r
w
kd kd k
χ⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
and from the right by ( )( )( )
r
w
kd kd k
χ⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
yields:
( ) ( )
( )( )
2 ( ) ( )( ) ( )
( ) ( )
< ( ) .
Tr r
w w
d k d ke k V k
d k d k
V k
β μβ α χ
κ
β α χ β μ
⎛ ⎞− ⎡ ⎤ ⎡ ⎤⎜ ⎟< + ⎢ ⎥ ⎢ ⎥⎜ ⎟⎣ ⎦ ⎣ ⎦⎝ ⎠
+ −
Then, based on (4.29), we have:
2 2( ) , be k k kβ< ∀ ≥
Hence, ( )e k is ultimately bounded and we have:
( )e k β< , bk k∀ ≥ (4.30)
If (4.27) is satisfied, then after a long enough time bk , the output ( )e k will always be such
that ( )e k β δ< ≤ and there will be no more switching. For the case of 0δ > , the origin is
in the region where the system 1cl∑ is active. Since the systems 1
cl∑ is asymptotically stable
and since the parameters 1θ satisfy the regulation interpolation condition in (4.26), then
regulation for the switched system is achieved. The analysis is similar for the case of 0δ < .
4.4.3 Regulation condition for the switched closed loop system clrΣ with an 2H
performance constraint
The previous section presented conditions on the switched closed loop system so that the
states and the performance variable remain bounded and exact output regulation is achieved
against rd . However, the closed loop system may not achieve optimal performance against
wd . If the conditions of Theorem 4.3 are satisfied, then after a long enough time bk , there
138
will be no more switching in the closed loop system, and only one of the two systems 1cl∑ or
2cl∑ will be active for bk k≥ . Consequently, in order to minimize the effects of the unknown
random disturbance wd on the performance variable e , it is proposed to add an 2H
performance constraint in the design of the regulator for the closed loop system that is active
for bk k≥ .
In the following, some standard results on the 2H norm of systems are reviewed [94,95].
Consider in the closed loop system (4.7). Let wT denote the transfer function from wd to e
for 1cl∑ for 0δ > or 2
cl∑ for 0δ < . The 2H norm of wT is defined by
( )22
2 0
1 Tr ( ) ( )2
j jw w wT T e T e d
π ω ω ωπ
∗≡ ∫ . (4.31)
It is well known that 2wT γ< if and only if ( ) ( )( ) 2Tr
T Te e w wr r r rF F E XE γ+ < , where 0X >
satisfies
( ) 0TT ex ex
r r r rA XA X C C− + < , 1 if 0,2 if 0.
rδδ>⎧
= ⎨ <⎩
Using the Schur complement formula, we have the following result to calculate the 2H norm.
Lemma 4.3
The 2H norm of wT is lower than γ if and only if there exist real symmetric matrices 0X >
and 0S > such that
( )0
0
0
rTT ex
r r
exr
X XA
A X X C
C I
⎡ ⎤−⎢ ⎥⎢ ⎥− <⎢ ⎥
−⎢ ⎥⎣ ⎦
, 1 if 0,2 if 0,
rδδ>⎧
= ⎨ <⎩ (4.32)
139
( ) ( )
00 0
wr
er
T Tw er r
X XEI F
E X F S
⎡ ⎤⎢ ⎥⎢ ⎥ >⎢ ⎥⎢ ⎥⎣ ⎦
, (4.33)
( ) 2Tr S γ< . (4.34)
The 2H performance conditions (4.32)-(4.34) are then used in the following result to
summarize the performance of the closed loop switched system.
Theorem 4.4
Consider the closed loop system (4.7) subject to bounded inputs rd and wd . Assume the
conditions of Theorem 4.3 are satisfied. Moreover, assume there exist real symmetric
matrices 0X > and 0S > such that the 2H performance conditions (4.32)-(4.34) are
satisfied. Then the switched closed loop system achieves exact output regulation against the
known deterministic input rd and satisfies 2wT γ< with respect to the unknown random
input wd .
4.5. Regulator synthesis
Based on the above regulation conditions and 2H performance criterion, a regulator
synthesis method for the switched system is presented in the following. The synthesis method
aims at finding the proper rK , rL and rQC such that that conditions (4.24)-(4.27) and (4.32)-
(4.34) are satisfied. Since rA includes all the unknown parameters rK , rL and rQC , (4.24) is
nonlinear in the unknown variables. This poses a major difficulty in designing the regulator
using rA directly. But it is worth pointing out that, based on Theorem 4.1, a parameterized
set of stabilizing controllers can be designed by separately designing the individual systems
140
(4.9), (4.10) and (4.11). Therefore, the gains rK and rL can be designed first based on (4.12).
Since QA and QB in (4.24) are fixed, then it is only necessary to find the matrices rQC such
that the regulation conditions (4.24)-(4.27) and (4.33) are satisfied. Since condition (4.24)
and (4.33) are still bilinear matrix inequalities in the unknown variable rQC , a congruence
transformation will be used first to transform (4.24) and (4.33) into a linear matrix inequality.
First, using Schur complement formula, (4.24) is equivalent to
( )( )
0
10 0
T
r
Td wr r
d wr r r
P P A P
I E E P
PA P E E P
α
ε μκ
⎡ ⎤− +⎢ ⎥⎢ ⎥−⎢ ⎥⎡ ⎤− <⎣ ⎦⎢ ⎥⎢ ⎥⎡ ⎤ −⎣ ⎦⎢ ⎥⎣ ⎦
(4.35)
We partition P in (4.35) as [ ]1 2
2 3
n nT
P P
P P×
⎛ ⎞⎜ ⎟⎜ ⎟⎝ ⎠
and define 11 1P−Ω = , 1
2 1 2P P−Ω = ,
13 3 2 1 2
TP P P P−Ω = − and 1
2
0T I
Ω⎡ ⎤Ω = ⎢ ⎥−Ω⎣ ⎦
. Then we have:
121
32
121 2
133 2 1 2
111
133 2 1 2
0,
0
,00
00,
00
T
T
TT
IP
I
II P PP
P P P P
PP
P P P P
−
−
−
−
−
ΩΩ ⎡ ⎤⎡ ⎤= ⎢ ⎥⎢ ⎥ Ω−Ω⎣ ⎦ ⎣ ⎦
Ω⎡ ⎤ ⎡ ⎤Ω = =⎢ ⎥ ⎢ ⎥Ω− ⎣ ⎦⎣ ⎦
Ω⎡ ⎤ ⎡ ⎤Ω Ω = =⎢ ⎥ ⎢ ⎥Ω− ⎣ ⎦⎣ ⎦
(4.36)
and
141
[ ] [ ]
( )
1 2 2
3
1 2 3
0,
00
, , , .
r
r
Q Q rr r r r r r r Q r r
r r rTr
Q Q r
r r r
r Q
A B CA B K A B K B C B K
A L CPA
A B CA L C
M C
⎡ ⎤−⎡ ⎤ ⎡ ⎤+ Ω Ω − + Ω +⎢ ⎥⎢ ⎥ ⎣ ⎦+⎣ ⎦⎢ ⎥Ω Ω = ⎢ ⎥−⎡ ⎤⎢ ⎥Ω ⎢ ⎥+⎢ ⎥⎣ ⎦⎣ ⎦
Ω Ω Ω
It is obvious that if 1 0Ω > and 3 0Ω > , then 0P > . Multiplying (4.35) from the left side by
( ), ,diag IΩ Ω and from the right side by ( ), , Tdiag IΩ Ω yields:
( ) ( )( )
( )
11 2 3
3
2 3
2 11 2 3
3 3
01 0 , , ,
0
010 0
0, , ,
0 0
r
r
Tr Q
Td wr r T
d wr Q r r
M C
II E E
IM C E E
α
ε μκ
⎡ ⎤Ω⎡ ⎤− Ω Ω Ω⎢ ⎥⎢ ⎥Ω⎣ ⎦⎢ ⎥
⎢ ⎥− ⎡ ⎤⎢ ⎥⎡ ⎤− <⎢ ⎥⎣ ⎦ Ω Ω⎢ ⎥⎣ ⎦⎢ ⎥
Ω Ω⎡ ⎤ ⎡ ⎤⎢ ⎥⎡ ⎤Ω Ω Ω −⎢ ⎥ ⎢ ⎥⎣ ⎦⎢ ⎥Ω Ω⎣ ⎦ ⎣ ⎦⎣ ⎦
(4.37)
Multiplying (4.25) from the left side by ( ), ,diag I IΩ and from the right side by
( ), ,Tdiag I IΩ yields:
( )
( ) ( )( )
1 1
3 2
1 2
0 00
0
0 0
0
TexrT
Ter
Ter
ex e er r r
CI
DI
F
C D F II
α
β μκ
β
⎡ ⎤Ω Ω⎡ ⎤ ⎡ ⎤⎢ ⎥⎢ ⎥ ⎢ ⎥Ω −Ω⎣ ⎦⎣ ⎦⎢ ⎥⎢ ⎥⎡ ⎤⎢ ⎥− ⎢ ⎥ >⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦⎢ ⎥⎢ Ω −Ω ⎥⎡ ⎤
⎡ ⎤⎢ ⎥⎢ ⎥ ⎣ ⎦⎣ ⎦⎣ ⎦
(4.38)
142
Similarly, we partition X in (4.33) as [ ]1 2
2 3
n nT
X X
X X×
⎛ ⎞⎜ ⎟⎜ ⎟⎝ ⎠
and define 11 1X −Ψ = , 1
2 1 2X X−Ψ = ,
13 3 2 1 2
TX X X X−Ψ = − and 1
2
0T I
Ψ⎡ ⎤Ψ = ⎢ ⎥−Ψ⎣ ⎦
. Multiplying (4.32) from the left side by
( ), ,diag IΨ Ψ and from the right side by ( ), ,T Tdiag IΨ Ψ yields:
( )
( ) ( )
11 2 3
3
1 11 2 3
3 2
1 2
, , ,
, , , 0
r
r
r Q
TT exr Q rT
exr
M C
M C CI
C II
⎡ ⎤Ψ⎡ ⎤− Ψ Ψ Ψ⎢ ⎥⎢ ⎥Ψ⎣ ⎦⎢ ⎥
⎢ ⎥Ψ Ψ⎡ ⎤ ⎡ ⎤⎢ ⎥Ψ Ψ Ψ − <⎢ ⎥ ⎢ ⎥Ψ −Ψ⎢ ⎥⎣ ⎦⎣ ⎦⎢ ⎥
Ψ −Ψ⎡ ⎤⎢ ⎥−⎢ ⎥⎢ ⎥⎣ ⎦⎣ ⎦
00
0
0 00
00
,1 if 0,2 if 0.
rδδ>⎧
= ⎨ <⎩
(4.39)
Multiplying (4.33) from the left side by ( ), ,diag I IΨ and from the right side by
( ), ,Tdiag I IΨ yields:
( ) ( )
1 2
3 3
2 3
00
0
wr
we
T Tw wr eT
IE
I DI
E D S
⎡ Ψ Ψ ⎤⎡ ⎤ ⎡ ⎤⎢ ⎥⎢ ⎥ ⎢ ⎥Ψ Ψ⎣ ⎦ ⎣ ⎦⎢ ⎥⎢ ⎥ >⎢ ⎥
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥Ψ Ψ⎣ ⎦⎣ ⎦
00
00 , (4.40)
( ) 2Tr S γ< . (4.41)
Note that (4.37) and (4.38) are linear in the unknown parameters β , μ , 1Ω , 2Ω , 3Ω and
rQC only if α is prefixed. In the following, a regulator synthesis algorithm is presented to
obtain rQC matrices that simultaneously minimize the 2H performance γ and satisfy the
regulation conditions in Theorem 4.3. The iterative algorithm converges to a local solution
143
related to an initial value of α . Let nμ , 1,nΩ , 2,nΩ , 3,nΩ , 1,nΨ , 2,nΨ , 2,nΨ , nS , ,Qr nC and nα
denote the values of μ , 1Ω , 2Ω , 3Ω , 1Ψ , 2Ψ , 2Ψ , S , rQC and α in the thn iteration of the
algorithm. The main steps in the proposed regulator synthesis procedure are summarized
below.
1) Design of an internally stabilizing controller: Design the controller parameters
rK and rL to make the switched closed loop system (4.7) internally stable by
solving the LMIs (4.12) for the unknown matrices 0TK KP P= > , 0T
L LP P= > and
matrices rK , rL , { }1, 2r∈ .
2) Initialize α : Find an upper bound mα on the value of α in (4.37) with unknown
1 0Ω > , 2Ω , 3 0Ω > ,and rQC , which is a generalized eigenvalue problem and can be
solved efficiently in Matlab LMI toolbox. Select an initial value 0α for α in the
interval ( )0, mα α∈ .
3) Substitute β by δ in (4.38).
4) (start the thn iteration for nS ). Minimize ( )nTr S subject to (4.26), (4.27),
(4.37), (4.38), (4.39) and (4.40) in the unknown nμ , 1,nΩ , 2,nΩ , 3,nΩ , 1,nΨ , 2,nΨ ,
2,nΨ , nS , and ,Qr nC with known 1nα − .
5) Minimize ( )nTr S subject to (4.26), (4.27), (4.37), (4.38), (4.39) and (4.40) in the
unknown nμ , 1,nΩ , 2,nΩ , 3,nΩ , 1,nΨ , 2,nΨ , 2,nΨ , nS , ,Qr nC and nα with known 1,nΩ
and 3,nΩ .
6) If ( ) ( )1 2n nTr S Tr S σ−− < , a prescribed tolerance, stop the algorithm.
144
For a given initial value 0α of α , the above algorithm converges to a local solution since
( )nTr S is guaranteed to decrease or stay the same in every iteration. Therefore, if there is no
proper feasible solution that can be found, the initial value 0α can be reset and the algorithm
can be restarted again. In the following, a numerical example is presented to show the
effectiveness of the proposed regulator synthesis method.
4.6. Numerical example
In this section, the regulator synthesis method proposed in this chapter will be used to design
a controller that cancels the contact vibrations in the mechanical system discussed in section
2.6. It is assumed that the system is subject to both deterministic and random disturbances.
First, this system is discretized to obtain a zero-order-hold equivalent model with sampling
period 0.001sT s= as in (4.1) with
3 6
1 3
945.6832 10 976.8748 10107.4562 935.9144 10
A− −
−
⎡ ⎤× ×= ⎢ ⎥− ×⎣ ⎦
, 3 6
2 3
995.0125 10 995.8425 109.9584 990.0333 10
A− −
−
⎡ ⎤× ×= ⎢ ⎥− ×⎣ ⎦
,
9
1 6
493.7891 10976.8748 10
B−
−
⎡ ⎤×= ⎢ ⎥×⎣ ⎦
, 9
1 6
498.7520 10995.8425 10
B−
−
⎡ ⎤×= ⎢ ⎥×⎣ ⎦
,
9
1 6
493.7891 10 0976.8748 10 0
xD−
−
⎡ ⎤×= ⎢ ⎥×⎣ ⎦
, 9
2 6
498.7520 10 0995.8425 10 0
xD−
−
⎡ ⎤×= ⎢ ⎥×⎣ ⎦
,
10
1 7
493.7891 10976.8748 10
xF−
−
⎡ ⎤×= ⎢ ⎥×⎣ ⎦
, 10
2 7
498.7520 10995.8425 10
xF−
−
⎡ ⎤×= ⎢ ⎥×⎣ ⎦
,
[ ]1 2 1 2 1 0e e y yC C C C= = = = , 51 2 1 2 0 1 10e e y yD D D D −⎡ ⎤= = = = − ×⎣ ⎦ , [ ]1 2 1 2 0e e y yF F F F= = = = ,
1
0.0747sin(20 ) 0.175cos(20 )0.1cos(20 )
k kd
kπ π
π+⎡ ⎤
= ⎢ ⎥⎣ ⎦
, 2
0.075sin(20 ) 0.075cos(20 )0.1 cos(20 )
k kd
kπ π
π+⎡ ⎤
= ⎢ ⎥×⎣ ⎦.
145
The switching rule is defined as 6
6
1 if 50 102 if 50 10
er
e
−
−
⎧ ≤ − ×= ⎨
> − ×⎩ and wd is a unknown random
disturbance signal with 2max 0.2m / secwd ≤ . To construct a parameterized set of
stabilizing controllers, the gains rK , rL , { }1, 2r∈ , are designed based on (4.12) resulted in:
[ ] 31 42.7610 1.1425 10K = − × , [ ] 3
2 56.2912 1.1810 10K = − − × ,
[ ]1 1.6118 22.8644 TL = − , [ ]2 1.6405 76.0713 TL = − − .
Based on the analysis of the poles of rd , we have 0 1k = and 2prn = . Therefore, the base
stabilizing controller is augmented with a parameter rQ with dimension 3qn = and given by
1 2 3
2 3r r r
rQz z zθ θ θ
= + + . The parameters in the matrices rQC , { }1, 2r∈ , needed to achieve
regulation for each subsystem are computed based on the regulation condition (4.18). Since
0δ < , condition (4.18) needs to consider only the system 2Σ given by:
2
122.6190 2.7779 2.925710
2.6117 2.4421 2.2628Aθ
−− − −⎡ ⎤= ×⎢ ⎥− − −⎣ ⎦
, 2
65.055110
5.6455Bθ
−−⎡ ⎤= ×⎢ ⎥−⎣ ⎦
.
Based on the disturbance signals, we have 0.64κ = . First, without considering the 2H
performance criterion in the regulator synthesis procedure, the regulator is designed to only
satisfy the regulation conditions (4.26), (4.37) and (4.38). We have
1
6 3 62.7121 10 450.4848 10 1.2831 10QC ⎡ ⎤= − × − × ×⎣ ⎦ ,
2
6 3 62.6866 10 547.9511 10 1.1974 10QC ⎡ ⎤= − × − × ×⎣ ⎦ .
Based on the above designed controllers, the simulation results are shown in Figures 4.2-4.4.
Now consider the 2H performance criterion in the regulator synthesis procedure, then using
the iterative algorithm with an initial value 0 0.03α = , and after 8 iterations, we obtain
146
1
6 3 33.3810 10 709.5338 10 543.1366 10QC ⎡ ⎤= − × × ×⎣ ⎦ ,
2
6 3 33.3022 10 679.4347 10 583.2225 10QC ⎡ ⎤= − × × ×⎣ ⎦ ,
0.031α = , 61.5413 10γ −= × .
Based on the above designed optimal parameterized controllers, the simulation results are
shown in Figures 4.5 and 4.6. Compared with the results in Figure 4.2 and 4.4, it can be seen
that the noise attenuation in the performance variable e has been improved by using the
proposed optimal parameterized regulator within a class of 2H controllers.
147
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4-60
-50
-40
-30
-20
-10
0
10
20
e (m
icro
met
ers)
Time(sec)
Figure 4.2. Simulation results showing the performance variable e obtained using the controller designed without accounting for the 2H performance constraint.
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4-0.2
-0.1
0
0.1
0.2
d r(1,1
) (m
/sec
2 )
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4-0.2
-0.1
0
0.1
0.2
dw (m
/sec
2 )
Time(sec)
Figure 4.3. Simulation results showing the switching component (1,1)rd in the known input
rd and the random disturbance wd .
148
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4-60
-50
-40
-30
-20
-10
0
10
20
v an
d v s (m
icro
met
ers)
Time(sec)
mass height vcontact surface vs
Figure 4.4. Simulation results showing the displacement of the mass obtained using the controller designed without accounting for the 2H performance constraint.
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4-60
-50
-40
-30
-20
-10
0
10
20
e (m
icro
met
ers)
Time(sec)
Figure 4.5. Simulation results showing the performance variable e obtained using the controller designed based on the 2H performance constraint.
149
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4-60
-50
-40
-30
-20
-10
0
10
20
v an
d v s (m
icro
met
ers)
Time(sec)
mass height vcontact surface vs
Figure 4.6. Simulation results showing the displacement of the mass obtained using the controller designed based on the 2H performance constraint.
4.7. Experimental evaluation
In this section, a bimodal switched system experimental setup is used to evaluate the
performance of the proposed adaptive regulator. The experimental setup simulates a
mechanical system subject to intermittent contact with a contact surface, and is motivated by
the flying height regulation problem in hard disk drives. The experimental setup is first
described in the following section. The identification of the switched system model is then
discussed followed by a presentation of the regulation experimental results.
150
4.7.1. Description of the experimental setup
A schematic diagram of the experimental setup is shown in Figure 4.7. The experimental
setup consists of a flexible beam corresponding to a suspension beam taken from a hard disk
drive. The tip of the suspension beam is supposed to track, at a fixed distance, an unknown
surface profile, while being subject to an unknown external disturbance force and to an
intermittent force representing a contact force with the surface being tracked. A multilayer
Lead Zirconate Titanate (PZT) piezoelectric actuator (Model PL 122.11, Physic Instrumente)
is attached to the suspension beam and serves to adjust the position of the beam tip. The PZT
actuator is driven by a power amplifier (E-650 LVPZT amplifier, Physic Instrumente). The
real-time measurement of the absolute tip position as well as the tip velocity is performed
using a Laser Doppler Vibrometer (LDV) (Polytec OFV-072, OFV-552, OFV-5000). A small
size permanent magnet is attached to the tip of the suspension beam. A magnetic coil placed
underneath the beam tip is used to apply external forces on the beam tip. These external
forces simulate the intermittent contact force between the beam tip and a hypothetical contact
surface, as well as other external disturbance forces. The details on how the contact force is
calculated are discussed in the following section. A voltage-to-current converter is used to
generate the current flowing in the coil. A PCI 6040e input-output card from National
Instruments and a personal computer are used to implement the controller and to interface it
with the rest of the system. The adaptive regulation algorithm is implemented using Labview
real-time module and Labview simulation interface toolkit with Matlab Simulink. The
suspension beam, electromagnet, and Laser Doppler Vibrometer are all mounted on a
vibration isolation optical table to reduce the effects of external vibrations. A picture of the
151
actual experimental setup is shown in Figure 4.8(a), and a close-up view of the suspension
beam with the PZT actuator and the electromagnetic actuator is shown in Figure 4.8(b).
4.7.2. Simulation of the external disturbances and of the contact force
This section outlines how the external disturbances and the intermittent contact force applied
to the beam tip are generated using the magnetic coil and a tiny permanent magnet. When a
current i flows in the coil, the axial magnetic flux density B generated by the coil is given
by [96]:
( ) ( ) ( ) ( )( )
2 2 2 22 2 2 20
2 2 222 1 1 11 1
ln ln2
z L r r z r rNiz z L zr r z r rz L r r
μ ⎡ ⎤⎛ ⎞ ⎛ ⎞+ + + + +⎢ ⎥⎜ ⎟ ⎜ ⎟= + −⎢ ⎥⎜ ⎟ ⎜ ⎟− + ++ + + ⎝ ⎠⎢ ⎥⎝ ⎠⎣ ⎦
B , (4.42)
where 0μ is the permeability of free space, 2r and 1r are the inner and outer radii of the
magnetic coil, respectively, z is the distance measured from the top of the coil along its
longitudinal axis, N is the number of turns of wire per unit length in the coil and L is the
length of the coil. The direction of the magnetic field B is parallel to the longitudinal axis of
the coil. Assume that at a distance z h= above the coil there is a cylindrical magnet of radius
R and thickness l and having a fixed and uniform magnetization
sM= −M ,
where the direction of M is also along the longitudinal axis of the coil. The equivalent
volume current density mρ and the equivalent surface current density mσ of the magnet are
given by
0mρ = −∇⋅ =M ,
sm
s
M z hM z h l
σ=⎧
= ⎨− = +⎩ (4.43)
152
Figure 4.7. Schematic diagram of the experimental setup.
The electromagnetic force on the permanent magnet with magnetization M due to an
external magnetic field B is given by [96]:
m mV SF dV dSρ σ= +∫ ∫B B , (4.44)
where V and S are the volume and surface of the permanent magnet, respectively. From
equations (4.42), (4.43) and (4.44), we obtain the vertical force on the permanent magnet
with magnetization sM due to the external magnetic field B generated by the coil as
( ) ( )( )
( ) ( ) ( )( )
( ) ( )( )
( ) ( )( )
2
2 22 22 2 2 22 0
2 22 22 1 1 1 1 1
2 22 22 22 2
2 2 2 21 1 1 1
ln ln2
ln ln .
mS
s
s
F dS
M R h h l
h L r r h l r rNiM R h L h lr r h L r r h l r r
h l L r rh r rh h l L
h r r h l L r r
σ
π
μπ
=
= − +
⎡ ⎛ ⎞ ⎛ ⎞+ + + + + +⎢ ⎜ ⎟ ⎜ ⎟= + + +⎢ ⎜ ⎟ ⎜ ⎟− + + + + + +⎢ ⎝ ⎠ ⎝ ⎠⎣
⎤⎛ ⎞⎛ ⎞ + + + ++ + ⎥⎜ ⎟⎜ ⎟− − + +⎥⎜ ⎟⎜ ⎟+ + + + + +⎝ ⎠ ⎥⎝ ⎠⎦
∫ B
B B
(4.45)
Suspension beam
PZT actuator
Magnetic coil
Permanent magnet
PC Based Real
Time Controller
Laser Doppler
Vibrometer
PZT Amplifier
Laser beam
ty
Voltage-to-current converter
, v vy y
i
u v
153
In the experimental setup, we have 30.5 10R m−= × , 31.5 10l m−= × , 31 3 10r m−= × ,
32 8 10r m−= × , 310 10L m−= × and 0h h hδ= + , where 3
0 1 10h m−= − × and hδ is the
deviation of the position of the magnet from 0h when current flows in the coil. Therefore,
based on (4.45), if hδ is in the range of few tens of micrometers, the force F can be
approximated as a linear function of the coil current i , namely
iF k i≅ .
The coil current is produced by a voltage-to-current converter circuit such that vi k v= where
v is the input voltage to the circuit and 210 /vk A volt−= . In the following, let ty denote the
displacement of the suspension beam tip, and assume the contact surface profile is given by
( )( ) 6( ) 5sin 240 50 10cs t t mπ −= − × .
The electromagnet will be used to apply a force that simulates the contact force cF at the
beam tip as follows:
0 if ( ) ( ) if ,
c t c
c c c t c c t t c
F y sF k s y c s y y s
= >⎧⎨ = − + − ≤⎩
(4.46)
where ck and cc are the stiffness and damping coefficient associated with the contact surface.
Here we set 310c cc k −= × [1]. It follows that the input voltage cv needed to simulate the
contact force is given by:
3
0 if
( ) ( ) 10 if .
c t c
cc c t c t t c
i v
v y skv s y s y y s
k k−
= >⎧⎪⎨ ⎡ ⎤= − + − × ≤⎣ ⎦⎪⎩
154
(a)
(b)
Figure 4.8. The experimental setup: (a) General view, (b) Close-up view of the suspension beam with the actuators.
Suspension Beam
Coil
Permanent Magnet
PZT ActuatorPI PL112.11
Microscope AdapterPolytec OFV-072
NI Labview Real Time Module
NI Labview with Simulation Interface Toolkit
PZT AmplifierPI E-650.00
Vibrometer ControllerPolytec OFV-5000
Fiber Vibrometer Polytec OFV-552
PZT Actuated Suspension Beam
155
In the experiment, we set 48 10 /c
i v
k A mk k
= × . The disturbance force at the tip of the flexible
beam is also simulated by applying a voltage to the voltage-to-current convert circuit as
follows:
( )2cos(240 3 / 4) volts if 2cos(240 3 / 4) 4 volts if .
a t c
ac t c
v t y sv t y s
π ππ π
= + >⎧⎨ = + + ≤⎩
(4.47)
Therefore, in the contact mode, the total voltage needed to simulate both the external
disturbances and the contact force is given by 3( ) ( ) 10cc ac c t c t ac
i v
kv v s y s y vk k
−⎡ ⎤+ = − + − × +⎣ ⎦ .
Hence, the total voltage input v to the voltage-to-current converter corresponding to the two
modes of operation is then
( ) ( )3 3
if
10 10 if .
a c
c cc c ac t t c
i v i v
v v y sk kv s s v y y y s
k k k k− −
= >⎧⎪⎨ = + × + − + × ≤⎪⎩
4.7.3. Identification of the switched system model
In the following, the two subsystem models corresponding to the non-contact mode and to
the contact mode will be obtained and verified. In the experimental setup, the measurement is
a voltage signal vy provided by the Laser Doppler Vibrometer and related to the beam tip
displacement ty by v y ty k y= where 0.1v/yk um= . A second order low pass Butterworth
filter is used to process the measurement vy and minimize the effects of the spillover
associated with the use of a reduced-order model of the PZT actuated suspension beam [97-
99]. The filter is also used to attenuate any measurement noise from the laser interferometer
or from external high frequency disturbances from the environment. A low pass Butterworth
156
filter with input vy and output fy , and having a cut-off frequency of 240Hz, is used and is
given by the following transfer function
( )1 2
11 2
1 23.64 0.268 0.6282f
q qH qq q
− −−
− −
+ +=
− +.
The models developed using system identification relate the measured signal vy to two
inputs, namely the PZT control voltage u and an exogenous input signal d defined as
follows:
( )3
if
10 if .
a c
cc c ac c
i v
d v y skd s s v y s
k k−
= >⎧⎪⎨ = + × + ≤⎪⎩
Hence, the signal d v= in the non-contact mode, and is associated with the external
disturbance defined by av . However, in the contact mode, we have
( ) ( )3 310 10c ct t c c ac
i v i v
k kd v y y s s vk k k k
− −= + + × = + × + ,
which is therefore different from v . The reason why the term ( )310ct t
i v
k y yk k
−− + × is not
part of the exogenous input signal d in this case is that it is incorporated as part of the plant
dynamics corresponding to the contact mode, and is responsible for the change to the model
dynamics, as illustrated in the simulation example in section 2.6.
First, the transfer function for the non-contact mode is identified. The transfer function
relates the piezoelectric actuator control voltage u and the signal d to the measurement vy
of the tip vertical displacement. The excitation signals applied to the PZT actuator and to the
voltage-to-current circuit for system identification are selected as ( ) 10.25sin(2 )u t f t voltsπ=
and ( ) 22sin(2 )d t f t voltsπ= , where 1f and 2f are swept from 1 to 350Hz, respectively. The
157
sampling period is set to 1KHz. Using the Matlab System Identification Toolbox, the system
is identified as a second order system
( ) ( )( ) ( ) ( )
( ) ( )1 1
1 2
1 1
b bv
a a
H q H qy k u k d k
H q H q
− −
− −= + , (4.48)
where
( )1 1 21 0.2459 0.9313aH q q q− − −= + + ,
( )1 1 21 1.004 2.1601bH q q q− − −= + , ( )1 1 2
2 0.2187 0.2267bH q q q− − −= + .
Figure 4.9 illustrates the accuracy of the identified model, where the dashed line shows the
frequency response of the real system under the exciting input signals and the solid line
shows the bode plot of the analytical model. It can be seen that the identified model matches
the real system well in the frequency range from 1 to 300Hz.
To identify the contact mode model, the excitation signals are selected as
( ) 10.25sin(2 )u t f t voltsπ= and ( )23sin 2d f t voltsπ= , where 1f and 2f are swept from 1 to
350Hz. The model is identified as
( ) ( )( ) ( ) ( )
( ) ( )1 1
1 2
1 1
c cb b
v c ca a
H q H qy k u k d k
H q H q
− −
− −= + , (4.49)
where
( )1 1 21 0.5645 0.8735caH q q q− − −= + + ,
( )1 1 21 0.9125 1.805c
bH q q q− − −= + , ( )1 1 22 0.201 0.2021c
bH q q q− − −= + .
Figure 4.10 illustrates the accuracy of the identified model.
158
4.7.4. Experimental results
In the following, based on the models in non-contact and contact mode, respectively, a
regulator is designed so that the tip of the suspension beam tracks a contact surface
( )( ) 65sin 20 60 10cs t mπ −= − × at a desired separation height of 660 10 m−× . The output y , to
be fed to the controller, is defined to be the measurement voltage corresponding to the
difference between the actual separation t cy s− and the desired separation of 660 10 m−× :
( )660 10
= 0.5sin(240 ) volts.y t c
v
y k y s
y kπ
−= − − ×
−
Define the performance variable e to be the same as the output y :
0.5sin(240 ) voltsve y kπ= − .
First, based on the transfer functions (4.48) and (4.49), the state space representation as in
(4.1) is given as
1
0 0.87351 0.5645
A−⎡ ⎤
= ⎢ ⎥−⎣ ⎦, 2
0 0.93131 0.2459
A−⎡ ⎤
= ⎢ ⎥−⎣ ⎦, 1
1.80450.9125
B ⎡ ⎤= ⎢ ⎥⎣ ⎦
, 2
2.16011.0036
B ⎡ ⎤= ⎢ ⎥⎣ ⎦
,
1
0.2021 00.2010 0
xD ⎡ ⎤= ⎢ ⎥⎣ ⎦
, 1
0.20210.2010
xF ⎡ ⎤= ⎢ ⎥⎣ ⎦
, 2
0.2267 00.2187 0
xD ⎡ ⎤= ⎢ ⎥⎣ ⎦
, 2
0.22670.2187
xF ⎡ ⎤= ⎢ ⎥⎣ ⎦
,
[ ]1 1 0 1e yC C= = , [ ]2 2 0 1e yC C= = ,
[ ]1 2 1 2 0 1y y e eD D D D= = = = , [ ]1 2 1 2 0y y e eF F F F= = = = .
1
30.5cos(240 ) 0.4sin(240 ) 0.3cos(240 )4
0.5sin(240 )
k k kd volts
k
ππ π π
π
⎡ ⎤+ + +⎢ ⎥=⎢ ⎥⎣ ⎦
, 2
30.5cos(240 )4
0.5sin(240 )
kd volts
k
ππ
π
⎡ ⎤+⎢ ⎥=⎢ ⎥⎣ ⎦
.
159
101
102
103
104
10-1
100
101
102
Am
plitu
de
101
102
103
104
-500
-400
-300
-200
-100
0
100
Pha
se (d
egre
es)
Frequency (rad/sec) (a)
101
102
103
104
10-2
100
Am
plitu
de
101
102
103
104
-400
-300
-200
-100
0
100
Pha
se (d
egre
es)
Frequency (rad/sec) (b)
Figure 4.9. The frequency response of the identified analytical model (solid line) and the real system based on experimental data (dashed line) in the non-contact mode: (a) transfer function from u to vy , (b) transfer function from d to vy .
160
101
102
103
104
100
102
Am
plitu
de
101
102
103
104
-800
-600
-400
-200
0
Pha
se (d
egre
es)
Frequency (rad/sec) (a)
101
102
103
104
10-2
100
102
Am
plitu
de
101
102
103
104
-600
-400
-200
0
Pha
se (d
egre
es)
Frequency (rad/sec) (b)
Figure 4.10. The frequency response of the identified analytical model (solid line) and the real system based on experimental data (dashed line) in the contact mode: (a) Transfer function from u to vy , (b) transfer function from d to vy .
161
Let wd denote the unknown random disturbance with max 0.5wd volts≤ . The switching
rule is defined as 1 if 62 if 6
e voltsr
e volts≤ −⎧
= ⎨ > −⎩. To construct a parameterized set of stabilizing
controllers, the gains rK , rL , { }1, 2r∈ , designed based on (4.12) are given by :
[ ] 31 -159.1746 504.2805 10K −= × , [ ] 3
2 -126.6187 407.4727 10K −= × ,
[ ] 31 873.4887 564.5477 10TL −= × , [ ] 3
2 931.2618 245.8963 10TL −= × .
Based on the analysis of the poles of rd , we have 0 1k = and 2prn = . Therefore, the base
stabilizing controller is augmented with a parameter rQ with dimension 4qn = and given by
1 2 3 43
2 2 3 4
0.46 0.45 101.4 0.9
r r r rr
zQz z z z z z
θ θ θ θ⎛ ⎞+= + + + ×⎜ ⎟− + ⎝ ⎠
. Since 0δ < , condition (4.18) needs to
consider only the system 2Σ given by:
2
30.8619 8.8250 12.0044 8.676610
11.9739 8.1385 0.1084 8.2966Aθ
⎡ ⎤= ×⎢ ⎥− −⎣ ⎦
, 2
3 305.197110
261.1491Bθ
−⎡ ⎤= ×⎢ ⎥−⎣ ⎦
.
First, without considering the 2H performance constraint in the regulator synthesis procedure,
the regulator is designed based only on the regulation conditions (4.26), (4.27), (4.37) and
(4.38). We have
[ ]1
30 1.0448 0.4263 0.0905 0.0975 10QC −= − − × ,
[ ]2
30 0.5360 0.5121 1.3472 1.3250 10QC −= − − − × .
The experimental results corresponding to the above designed controller re shown in Figures
4.11-4.14, which include the two cases of with and without the unknown random disturbance
dw . It can be seen that the resulting system with the designed controller can exactly reject
the known disturbance in the presence of switching, however it cannot attenuate the
162
disturbance dw properly. By considering the 2H constraint in the regulator synthesis
procedure, we obtain
[ ]1
30 0.1376 0.2452 0.1665 0.0208 10QC −= − × ,
[ ]2
30 0.1129 0.2334 0.2105 0.1002 10QC −= − − × ,
2 0.118γ = .
The experimental results are shown in Figures 4.15 and 4.16, which indicate that the
disturbance dw has been attenuated effectively and the tip of the suspension beam tracks the
contact surface while maintaining the desired separation between the tip and the contact
surface.
4.8. Conclusion
The problem of regulation in bimodal switched systems against known deterministic
exogenous inputs and unknown random inputs is treated. A regulator design approach based
on the parameterization of a set of stabilizing controllers for the switched system is presented.
First, regulation conditions for the switched system are derived and are used to construct a set
of Q parameters needed to achieve regulation against the known deterministic exogenous
inputs. A set of 2H controllers from within the already constructed set of controllers is then
considered to achieve regulation also against the unknown random exogenous input. A
regulator synthesis algorithm is developed based on solving a set of linear matrix inequalities.
Finally, a switched bimodal mechanical system experimental setup involving a flexible beam
subject to contact vibrations is used to successfully demonstrate the performance of the
proposed regulator.
163
The results presented so far in the thesis assume that the deterministic exogenous inputs
are known. However, in some cases, the properties of the deterministic signals maybe
unknown or time-varying. To deal with these types of inputs, an adaptive regulation method
will be developed for discrete-time switched bimodal linear systems in the next chapter.
164
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-100
-50
0
50
100
Per
form
ance
var
iabl
e (m
icro
met
ers)
Time(sec)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-100
-50
0
50
100
Pos
ition
(mic
rom
eter
s)
Time(sec)
Tip positionContact surface
Figure 4.11. Experimental results showing the performance variable e and the position y for the case of 0dw = and obtained using the controller designed without accounting for the 2H performance constraint.
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-0.5
0
0.5
Ran
dom
dis
turb
ance
dw
(V)
Time(sec)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
0
50
100
Cur
rent
in th
e co
il (m
A)
Time(sec)
Figure 4.12. Experimental results showing the random disturbance dw and the total current in the coil for the case of 0dw = and obtained using the controller designed without accounting for the 2H performance constraint.
165
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-100
-50
0
50
100
Per
form
ance
var
iabl
e (m
icro
met
ers)
Time(sec)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-100
-50
0
50
100
Pos
ition
(mic
rom
eter
s)
Time(sec)
Tip positionContact surface
Figure 4.13. Experimental results showing the performance variable e and the position y in the presence of dw and obtained using the controller designed without accounting for the
2H performance constraint.
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-0.5
0
0.5
Ran
dom
dis
turb
ance
dw
(V)
Time(sec)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
0
50
100
Cur
rent
in th
e co
il (m
A)
Time(sec)
Figure 4.14. Experimental results showing the random disturbance dw and the total current in the coil obtained using the controller designed without accounting for the 2H performance constraint.
166
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-100
-50
0
50
100
Per
form
ance
var
iabl
e (m
icro
met
ers)
Time(sec)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-100
-50
0
50
100
Pos
ition
(mic
rom
eter
s)
Time(sec)
Tip positionContact surface
Figure 4.15. Experimental results showing the performance variable e and the position y in the presence of dw and obtained using the controller designed by accounting for the 2H performance constraint.
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-0.5
0
0.5
Ran
dom
dis
turb
ance
dw
(V)
Time(sec)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
0
50
100
Cur
rent
in th
e co
il (m
A)
Time(sec)
Figure 4.16. Experimental results showing the random disturbance dw and the total current in the coil obtained using the controller designed by accounting for the 2H performance constraint.
167
CHAPTER 5
Adaptive Regulation in Switched Bimodal
Linear Systems
5.1. Introduction
In this chapter, an adaptive regulation method is presented for discrete-time switched
bimodal systems where it is desired to reject unknown sinusoidal disturbance signals and/or
track unknown sinusoidal reference inputs. The primary practical motivation for this problem
came from considering the problem of flying height regulation in hard disk drives. The
switching nature of the system dynamics coupled with the unknown disk surface profile
could be addressed using an adaptive regulation approach. The regulator design approach
involves two main steps. In the first step, a set of switched Q -parameterized observer-based
state-feedback controllers for the switched linear system is constructed. Switching in the
controller is performed according to the same switching rule as in the plant. Assuming the
reference or disturbance signals are known, regulation conditions for the switched closed
loop system are presented in the form of properly formulated linear matrix inequalities
(LMIs). The regulation conditions are used to construct a set of Q parameters needed to
achieve regulation in the switched closed loop system. In the second step, an adaptation
168
algorithm is developed to tune, online, the Q parameter in the expression of the
parameterized controller for the case of unknown reference or disturbance signals. The
tuning is such that the Q parameter converges to the desired Q parameter that guarantees
regulation in the switched closed loop system.
The rest of the chapter is organized as follows. In section 5.2, the general regulation
problem for the switched bimodal system is presented. In section 5.3, the construction of a
Q -parameterized set of switched controllers for the switched system is discussed and
stability properties of the resulting switched closed loop system are analyzed. In section 5.4,
regulation conditions for the switched closed loop system are presented using properly
formulated LMIs. In section 5.5, an adaptation algorithm is developed to tune the Q
parameter in the controller and drive it to the desired Q parameter needed to achieve
regulation in the switched closed loop system. In section 5.6, numerical simulation results for
a simple switched mechanical system are used to illustrate the performance of the proposed
regulator. In section 5.7, the proposed adaptive regulator is evaluated on an experimental
setup, where the tip of a flexible beam is supposed to maintain a constant separation with
respect to a surface with an unknown profile, while also being subject to an unknown
disturbance force. The experimental results successfully demonstrate the effectiveness of the
proposed approach in achieving exact output regulation against unknown sinusoidal
exogenous inputs and in the presence of switching in the system dynamics. Concluding
remarks are presented in section 5.8.
169
5.2. The adaptive regulation problem for bimodal systems
This chapter considers discrete-time switched systems given by the following state space
representation:
0( 1) ( ) ( ) ( ), (0) , ( ) ( ) ( ), ( ) ( ) ( ), :1 if ( ) ( ) ( ) ,2 if
xi i i iy y
i i ie ei i ii
e ei i i
x k A x k B u k D d k x xy k C x k D d ke k C x k D d k
e k C x k D d ki
eδ
+ = + + == += +Σ
= + ≤=
( ) ( ) ( ) ,e ei i ik C x k D d k δ
⎧⎪⎪⎪⎨⎪ ⎧⎪ ⎨⎪ = + >⎩⎩
(5.1)
where nx∈R is the state vector, u∈R is the control input, y∈R is the measurement signal
to be fed to the controller, e∈R is the performance variable to be regulated and is assumed
to be measurable, δ is a constant satisfying 0δ > , and { }1,2i∈ is the index of the system
iΣ under consideration at time k .
The external signal hid ∈R , representing disturbance and/or reference signals, is also
assumed to switch according to the rule given in (5.1). The signal id is given by:
1( ) [ ( ), , ( )]h Ti i id k d k d k= … , (5.2)
where
( ) 1,, , ,
1
( ) cos , 1, ,s
s
kk ss s s s
i i i i id k c k c s hν ν ν
ν
ω φ +
=
= + + =∑ , (5.3)
with unknown amplitudes ,sicν , frequencies ,s
iνω , and phases ,s
iνφ , 1, , skν = … ; 1, ,s h= … .
For the switched system (5.1), it is desired to construct an output feedback controller to
regulate the performance variable e . More specifically, a Q -parameterized adaptive output
feedback controller will be developed in this chapter to asymptotically drive the performance
variable e to zero, i.e. lim ( ) 0k
e k→∞
= , in the presence of the unknown external input signal id .
170
Given the switching nature of the plant, the parameterized output feedback controller is also
chosen to be a switching feedback controller and where the switching between the two
controllers is to obey the same rule given in (5.1) as for switching between the two plant
models.
5.3. Parameterization of a set of stabilizing controllers
The development of the proposed adaptive regulator is based on considering a Q -
parameterized set of output feedback stabilizing controllers for the switched system. A
summary of the construction of such a set of controllers is presented followed by an internal
stability analysis for the resulting bimodal closed loop system.
5.3.1. Q -parameterized controller
In the following, a Q -parameterized output feedback controller for the switched system (5.1)
is used, where the controller is expressed as a linear fractional transformation involving a
fixed system iJ , and a proper stable parameter iQ that could be chosen as desired (See
Figure 2.1). The state space representation of the system iJ is given by:
0ˆ ˆ ˆ ˆ( 1) ( ) ( ) ( ) ( ), (0) ,ˆ: ( ) ( ) ( ),
ˆ ˆ( ) ( ) ( ) ( ).
yi i i i i i i Q
i i Q
yi
x k A LC B K x k L y k B y k x x
J u k K x k y k
y k y k y k C x k
⎧ + = + + − + =⎪
= +⎨⎪
− = −⎩
(5.4)
where ˆ( )x k is an estimate of the plant state vector ( )x k and ˆ ˆ( ) ( )yiy k C x k= is an estimate of
the plant output ( )y k . The state feedback gains iK , { }1, 2i∈ , and the observer gains iL ,
171
{ }1, 2i∈ , are assumed to switch according to the rule given in (5.1). The system iQ is given
by:
( ) 0ˆ( 1) ( ) ( ) ( ) , (0) ,:
( ) ( ),Q Q Q Q Q Q
i QQ i Q
x k A x k B y k y k x xQ
y k C x k
⎧ + = + − =⎪⎨
=⎪⎩ (5.5)
where QnQx ∈ . In particular, throughout the rest of the paper, the iQ parameter is chosen to
be of the form as
( )1
( )qn
qi i q
q
Q z zθ ψ=
=∑ , (5.6)
where ( ) ( )1 qq z z F zψ −= , 1, qq n= , are stable basis functions, and where
11
11
( )m
mm m
m
b z bF zz a z a
−
−
+=
+ + is a stable function used to adjust the dynamic properties of ( )iQ z .
Such representation can be used to approximate any proper stable real rational transfer
function. Let 1, , qTn
i i iθ θ θ⎡ ⎤= ⎣ ⎦… and 1Q qn m n= + − . A realization of ( )iQ z can then be given
as follows:
1
2
1
0 0 0 0 0 01 0 0 0 0 0
0 1 0 0 0 00 0 1 0 0 00 0 0 1 0 0 00 0 0 0 0 00 0 0 0 0 1 0
Q Q
m
m
Q
n n
aa
aA
a
−
×
−⎡ ⎤⎢ ⎥−⎢ ⎥⎢ ⎥⎢ ⎥−⎢ ⎥= ⎢ ⎥−⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
,
1
2
1
1
0
0Q
m
m
Q
n
bb
bB
b
−
×
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥= ⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
,
( )1 10Q Ti imC θ× −
⎡ ⎤= ⎣ ⎦ .
172
It follows that QA is a fixed stability matrix, QB is a fixed matrix, and the matrix QiC
changes with the switching signal { }1, 2i∈ . If the signal id is known, the parameter vectors
iθ , { }1, 2i∈ , can be designed according to the regulation conditions to be discussed in
sections 5.4 and that guarantee the switched closed loop system achieves regulation. In the
case where the properties of the input signal id are unknown, an adaptation algorithm will be
proposed to adjust and drive the parameter vectors iθ in the QiC matrices to the desired
parameter vectors that yield regulation in the adaptive switched closed loop system against
the unknown signal id .
Let ˆ( ) ( ) ( )x k x k x k= − denote the state estimation error and 1
TT T TQ N
x x xχ×
⎡ ⎤= ⎣ ⎦
denote the state vector for the resulting closed loop system with 2 QN n n= + . The resulting
closed loop system is given by the following state space representation:
( 1) ( ) ( ), ( ) ( ) ( ),
:1 if ( ) ( ) ( ) ,2 if ( ) ( ) ( ) ,
i i iex ei i icl
i ex ei i iex ei i i
k A k E d ke k C k D d k
e k C k D d ki
e k C k D d k
χ χχ
χ δχ δ
⎧ + = +⎪
= +⎪∑ ⎨⎧ = + ≤⎪ = ⎨⎪ = + >⎩⎩
(5.7)
where 00 0
Qi i i i i i i
yi Q Q i
yi i i
A B K B C B KA A B C
A L C
⎡ ⎤+⎢ ⎥= −⎢ ⎥⎢ ⎥+⎣ ⎦
, ( ) ( ) ( )TTTx y T x y
i i Q i i i iE D B D D L D⎡ ⎤= − +⎢ ⎥⎣ ⎦,
0 0ex ei iC C⎡ ⎤= ⎣ ⎦ .
173
5.3.2 Stability of the Q -parameterized switched closed loop system
In this section, the internal stability of the closed loop system is studied by considering the
system (5.7) in the absence of the signal id . With 0id = , the origin 0χ = of the unforced
switched closed loop system should be asymptotically stable. In the absence of the signal id ,
the state equation for the resulting closed loop system is given by:
( 1) ( )( 1) 0 ( ) .( 1) 0 0 ( )
Qi i i i i i i
yQ Q Q i Q
yi i i
x k A B K BC B K x kx k A B C x kx k A LC x k
⎡ ⎤+ +⎡ ⎤ ⎡ ⎤⎢ ⎥⎢ ⎥ ⎢ ⎥+ = −⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥⎢ ⎥ ⎢ ⎥+ +⎣ ⎦ ⎣ ⎦⎣ ⎦
(5.8)
Then, the internal stability of the switched closed loop system (5.8) is given by the following
result.
Theorem 5.1
The origin is an asymptotically stable equilibrium point for the switched system (5.8) under
arbitrary switching if QA is a stability matrix and there exist matrices 0kiP > , 0l
iP > and
matrices iK , iL , { }1, 2i∈ , such that:
[ ] [ ] { }0, , 1, 2T k ki i i j i i i iA B K P A B K P i j+ + − < ∈ , (5.9)
{ } 0, , 1, 2Ty l y l
i i i j i i i iA L C P A L C P i j⎡ ⎤ ⎡ ⎤+ + − < ∈⎣ ⎦ ⎣ ⎦ . (5.10)
Proof: Based on (5.8), define the following three subsystems for { }1, 2i∈ :
[ ]( 1) ( ),i i ix k A B K x k+ = + (5.11)
( 1) ( ),yi i ix k A LC x k⎡ ⎤+ = +⎣ ⎦ (5.12)
174
( 1) ( ).Q Q Qx k A x k+ = (5.13)
Systems (5.11) and (5.12) are switched systems. Define the piecewise quadratic Lyapunov
functions 2
1
( ( )) ( ) ( ) ( )T kx i i
i
V x k x k k P x kζ=
⎛ ⎞= ⎜ ⎟
⎝ ⎠∑ for system (5.11) and
2
1
( ( )) ( ) ( ) ( )T lx i i
i
V x k x k k P x kζ=
⎛ ⎞= ⎜ ⎟
⎝ ⎠∑ for system (5.12), respectively, where
1, when the subsystem is active at time ( )
0, otherwise
th
ii k
kζ⎧
= ⎨⎩
, { }1, 2i∈ . If (5.9) and (5.10) are satisfied,
the switched systems (5.11) and (5.12) admit continuously decreasing piecewise quadratic
Lyapunov functions xV and xV , which means that the switched systems (5.11) and (5.12) are
asymptotically stable under arbitrary switching. Since QA is a stability matrix, system (5.13)
is also asymptotically stable. Then, as in the proof for Lemma 4.1, it is easy to have that the
switched system (5.1) is asymptotically stable under arbitrary switching.
5.4. Regulation conditions for the switched system
In this section, regulation conditions for the switched closed loop system (5.7) are presented
under the assumption that the properties of the external input in (5.2) and (5.3), namely
frequencies, amplitudes and phases, are all known.
First, the performance variable ( )e k in the switched closed loop system cliΣ in (5.7) can
be written as:
( )11 12 210( ) ( ) ( )i i i i ie k T T QT d k e k= + + , (5.14)
175
where 12 220( ) ( )i i i Qe k T Q T y k= . Based on the state space realization of the system 22
iT given in
(4.15), the output of the system 22iT is not affected by the input ( )Qy k . Hence, 0( )e k
represents the response of the system 12 22i i iT Q T to its initial conditions, and will be shown
later in this section to be bounded by an exponentially decaying signal. First, the closed-loop
system transfer function can be written as:
( ),
11 12 21
( ) ( ) ( )
( ) ( ) ( ) ( ) ( ).i iT Q i
i i i i i
E z F z D z
T z T z Q z T z D z
=
= + (5.15)
Let 1( )e k denote the response of the switched system 11iT to the input ( )id k , 2 ( )e k denote
the response of the switched system 12iT to an arbitrary input 12 ( )id k and 3( )e k denote the
response of the switched system 21i iQT to the input ( )id k . Based on the definitions of 11
iT ,
12iT and 21
iT in (4.15), we have the following results.
Lemma 5.3
Consider the switched system 11iT subject to a bounded input id . Let
{ }1,20
max ( ) 0iik
d kγ∈≥
= ≠ and
10 1ε< . If there exist matrices 111 0H > , 11
2 0H > and positive scalars 1α , 1μ and 1β such
that the following matrix inequalities are satisfied:
( )( ) ( )
11 11 11 111
1 1 11 112
11 11 11 11 11
0
10 0,
T
i i i j
T
i j
j i j i j
H H A H
I E H
H A H E H
α
ε μγ
⎡ ⎤− +⎢ ⎥⎢ ⎥−⎢ ⎥− ≤⎢ ⎥⎢ ⎥−⎢ ⎥⎣ ⎦
( ) { } { }, 1, 2 1,2i j∀ ∈ × , (5.16)
176
( )( ) ( )
11 111
1 1 112
11 111
0
0 0
T
i i
T
i
i i
H C
I D
C D I
α
β μγ
β
⎡ ⎤⎢ ⎥⎢ ⎥−⎢ ⎥ ≥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
, { }1, 2i∀ ∈ , (5.17)
then there exists a finite time 1k such that 1( )e k satisfies:
1 1 1( ) , e k k kβ< ∀ ≥ . (5.18)
Proof: Consider the state space representation of 11iT and define the piecewise quadratic
Lyapunov function
211
11 11 11 111
( ( )) ( ) ( ) ( )TV k k k H kχ χ ζ χ=
⎛ ⎞= ⎜ ⎟
⎝ ⎠∑ ,
where 1, when the subsystem is active at time ( )
0, otherwise
th kkζ
⎧= ⎨⎩
, { }1, 2∈ , and 211( ) nkχ ∈R . First,
using the Schur complement formula, inequality (5.16) is equivalent to:
( ) ( )
( ) ( ) ( )
11 11 11 11 11 11 11 111
1 111 11 11 11 11 112
0,1
T T
i j i i i i j i
T T
i j i i j i
A H A H H A H E
E H A E H E I
α
ε μγ
⎡ ⎤− +⎢ ⎥
≤⎢ ⎥−⎢ ⎥−⎢ ⎥⎣ ⎦
(5.19)
Inequality (5.19) implies that for any nonzero vector 11 2( )( )
n h
i
kd kχ +⎡ ⎤
∈⎢ ⎥⎣ ⎦
R , we have:
( ) ( )
( ) ( ) ( )
11 11 11 11 11 11 11 111
11 11
1 111 11 11 11 11 112
( ) ( )01( ) ( )
T TT i j i i i i j i
T Ti ii j i i j i
A H A H H A H Ek k
d k d kE H A E H E I
αχ χ
ε μγ
⎡ ⎤− +⎢ ⎥⎡ ⎤ ⎡ ⎤
≤⎢ ⎥⎢ ⎥ ⎢ ⎥−⎣ ⎦ ⎣ ⎦⎢ ⎥−
⎢ ⎥⎣ ⎦
.
It follows that, for all ( ) { } { }, 1, 2 1, 2i j ∈ × ,
177
( ) ( )1 111 1111 11 1 11 11 2
1( 1) ( 1) 1 ( ) ( ) ( ) ( )T T T
j i i ik H k k H k d k d kε μ
χ χ α χ χγ−
+ + − − ≤ . (5.20)
Let { }1, 2i∈ denote the mode of the system that is active at time k and { }1, 2j∈ denote the
mode of the system that is active at time 1k + . Then, for any 0k ≥ , the above inequality
implies that
( ) ( )2 21 111 11
11 11 1 11 11 21 1
1( 1) ( 1) ( 1) 1 ( ) ( ) ( ) ( ) ( )T T T
i ik k H k k k H k d k d kε μ
χ ζ χ α χ ζ χγ= =
−⎛ ⎞ ⎛ ⎞+ + + − − ≤⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠∑ ∑ ,
which is equivalent to
( )1 111 11 11 11 1 11 112
1( ( 1)) ( ( )) ( ) ( ) ( ( ))T
i iV k V k d k d k V kε μ
χ χ α χγ−
+ − ≤ − .
Hence, 11 11 11 11( ( 1)) ( ( )) 0V k V kχ χ+ − ≤ holds whenever
( ) ( )1 1 1 111 11 2
1 1
1 1( ( )) ( ) ( )T
i iV k d k d kε μ ε μ
χα α γ− −
≥ ≥ .
Consequently, 11 11( ( ))V kχ cannot ultimately exceed the value ( )1 1
1
1 ε μα−
, and we have:
( )1 111 11
1
1lim ( ( ))k
V kε μ
χα→∞
−≤ . (5.21)
Therefore, there exists a finite time 1k such that
111 11 1
1
( ( )) , V k k kμχα
< ∀ ≥ . (5.22)
Inequality (5.17) is equivalent to:
( )( )( )
11111
11 111 1 1112
01 0
0
Ti
i
i iT
i
H CC D
I D
αβ μ
βγ
⎡ ⎤ ⎡ ⎤⎢ ⎥ ⎢ ⎥ ⎡ ⎤− ≥− ⎣ ⎦⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥⎣ ⎦⎣ ⎦
. (5.23)
178
Let { }1, 2i∈ denote the mode of the system that is active at time k . Then (5.23) can be
rewritten as:
( )( )( )
211
1111 11 11
1111 12
( ) 01 0
0
T
i
i iT
i
k H CC D
DI
α ζ
ββ μγ
=
⎡ ⎤⎛ ⎞⎡ ⎤⎢ ⎥⎜ ⎟
⎝ ⎠ ⎢ ⎥⎢ ⎥ ⎡ ⎤− ≥⎣ ⎦⎢ ⎥⎢ ⎥−⎢ ⎥⎢ ⎥ ⎣ ⎦
⎢ ⎥⎣ ⎦
∑.
Multiplying from the left by 11( )( )
T
i
kd kχ⎡ ⎤⎢ ⎥⎣ ⎦
and from the right by 11( )( )i
kd kχ⎡ ⎤⎢ ⎥⎣ ⎦
yields:
( ) ( )
( )( )
1 121 1 1 11 11 2
1 1 11 11 1 1
( ) ( ) ( ) ( )
( ) .
Ti ie k V k d k d k
V k
β μβ α χ
γ
β α χ β μ
⎛ ⎞−≤ +⎜ ⎟
⎝ ⎠≤ + −
(5.24)
Then based on (5.22), we have:
1 1 1( ) , e k k kβ< ∀ ≥ .
Lemma 5.4
Consider the switched system 12iT subject to a bounded input 12
id . Let { }
12
1,20
max ( ) 1iik
d k∈≥
= and
20 1ε< . If there exist matrices 121 0H > , 12
2 0H > and positive scalars 2α and 2β such that
the following matrix inequalities are satisfied:
( )( ) ( )
12 12 12 122
12 122
12 12 12 12 12
0
0 1 0,
T
i i i j
T
i j
j i j i j
H H A H
I E H
H A H E H
α
ε
⎡ ⎤− +⎢ ⎥⎢ ⎥− − ≤⎢ ⎥⎢ ⎥−⎢ ⎥⎣ ⎦
( ) { } { }, 1, 2 1,2i j∀ ∈ × , (5.25)
( )12 122
12 22
0T
i i
i
H C
C
α
β
⎡ ⎤⎢ ⎥ ≥⎢ ⎥⎣ ⎦
, { }1, 2i∀ ∈ , (5.26)
179
then there exists a finite time 2k such that 2 ( )e k satisfies:
2 2 2( ) , e k k kβ< ∀ ≥ . (5.27)
Proof: Consider the state space representation of 12iT and define the piecewise quadratic
Lyapunov function 2
1212 12 12 12
1
( ( )) ( ) ( ) ( )TV k k k H kχ χ ζ χ=
⎛ ⎞= ⎜ ⎟
⎝ ⎠∑ where 12 ( ) nkχ ∈R . Proceeding
as in the proof of Lemma 5.3 and based on inequality (5.25), there exists 2 0k > such that
12 12 22
1( ( )) , V k k kχα
< ∀ ≥ . (5.28)
Inequality (5.26) is equivalent to:
( )12 12 122 2
2
1 0T
i i iH C Cαβ
− ≥ , (5.29)
Let { }1, 2i∈ denote the mode of the system that is active at time k . Then (5.29) can be
rewritten as:
( )2
12 12 122 2
1 2
1( ) 0T
i ik H C Cα ζβ=
⎛ ⎞− ≥⎜ ⎟
⎝ ⎠∑ .
Multiplying from the left by 12 ( )T kχ and from the right by 12 ( )kχ yields:
( ) ( )212 212 2 2 12 12( ) ( )iC k V kχ α β χ≤ . (5.30)
Based on (5.28), it follows that:
2 2 2( ) , e k k kβ< ∀ ≥ .
180
Similarly, let 21
21
0iQi
Q i Q
AB C A⎡ ⎤⎢ ⎥⎢ ⎥⎣ ⎦
A and 21
21iQ
iQ i
EB D⎡ ⎤⎢ ⎥⎢ ⎥⎣ ⎦
E . Then we have the following result.
Lemma 5.5
Consider the switched system 21i iQ T subject to a bounded input id . Let
{ }1,20
max ( ) 0iik
d kγ∈≥
= ≠ ,
3 0α > be a preset constant and 30 1ε< . If there exist matrices 1 0QH > , 2 0QH > and a
positive scalar 3β such that the following matrix inequalities are satisfied:
( )( ) ( )
3
3 32
0
10 0,
TQ Q Q Qi i i j
TQ Qi j
Q Q Q Q Qj i j i j
H H H
I H
H H H
α
ε αγ
⎡ ⎤− +⎢ ⎥⎢ ⎥−⎢ ⎥− ≤⎢ ⎥⎢ ⎥−⎢ ⎥⎣ ⎦
A
E
A E
( ) { } { }, 1, 2 1,2i j∀ ∈ × , (5.31)
1
21 3
00
0
TQ Qi n i
Qn i
H C
C β
×
×
⎡ ⎤⎡ ⎤⎣ ⎦⎢ ⎥ ≥⎢ ⎥⎡ ⎤⎣ ⎦⎣ ⎦
, { }1, 2i∀ ∈ , (5.32)
then there exists a finite time 3k such that 3 ( )e k satisfies:
3 3 3( ) , e k k kβ< ∀ ≥ . (5.33)
Remark 5.2
The proof for the above result is similar to the proof of Lemma 5.4. It should also be noted
that Lemma 5.5 holds even if the matrix QiC is time-varying as long as (5.32) is satisfied
with the time-varying QiC matrix. This property will be used in the development and analysis
of the adaptation algorithm presented later.
181
Consider now the following assumption.
Assumption 5.1
There exists a unique parameter vector 0θ satisfying the interpolation condition (4.18) for
1i = in the case of 0δ > and 2i = in the case of 0δ < , which means the matrix iAθ in (4.19)
is square and nonsingular. Furthermore, the parameter vector 0θ satisfies 0l uθ θ θ≤ ≤ and
0θ η≤ , where lθ , uθ and η are known constants and are such that any qnθ ∈R satisfying
l uθ θ θ≤ ≤ also satisfies θ η≤ .
For assumption 5.1 to be satisfied, we should have 02 1qn k= + . It is also worth noting that by
properly selecting the functions ( )q zψ in (5.6), we can have η in any pre-specified desired
range. Let Q ql n n n= + − . Lemma 5.3 implies that the system 11iT is an exponentially stable
switched system. It follows that the system 22iT is also an exponentially stable system.
Moreover, given that QA is a stability matrix, then iQ is an exponentially stable switched
system. Hence, 0( )e k in (5.14) is bounded by an exponentially decaying signal. It follows
that for any arbitrary small positive number 00 1ε< , there exists a finite time 0k such that
0 ( )e k satisfies 0 0( )e k ε< , 0 k k∀ ≥ . Let Q ql n n n= + − . Based on Lemmas 5.3-5.5, a
sufficient regulation condition for the switched closed loop system (5.7) can be given as
follows.
182
Theorem 5.2
Assume the switched closed loop system (5.7) is internally stable under arbitrary switching
and that the conditions of Lemmas 5.3 and 5.4 are satisfied. Let 3 0α > be a preset constant
and 30 1ε< . If Assumption 5.1 is satisfied and 1β δ< , and if there exist matrices
1 0QH > , 2 0QH > such that:
2
22
1 0
(5.31)
0 0 1, if 0,
0 2, if 0q
q q q
l l l nQi
n l n n
iH
I iδβ
η δδ β ε× ×
× ×
⎧⎪⎪ ⎡ ⎤⎛ ⎞ = >⎧⎨ ≥ ⎢ ⎥⎜ ⎟ ⎨⎪ ⎜ ⎟ = <− − ⎢ ⎥ ⎩⎝ ⎠⎪ ⎣ ⎦⎩
(5.34)
then, with
( )
( )
1 0 2 11 1
1 1 2 01 1
0 0 , 0
0 0 , 0
Q
Q
Q T Qnm
Q Q Tn m
C and C if
C and C if
θ δ
θ δ
×× −
× × −
⎧ ⎡ ⎤= = >⎪ ⎣ ⎦⎨
⎡ ⎤= = <⎪ ⎣ ⎦⎩
(5.35)
the switched closed loop system (5.7) achieves regulation.
Proof: Since 0θ η≤ , we have 2
0 0q q
Tn nIη θ θ× ≥ . It follows from (5.34) that
2
2
0 01 0
0 0
0q
q
l l l nQi T
n l
H βθ θδ β ε
× ×
×
⎡ ⎤⎛ ⎞≥ ⎢ ⎥⎜ ⎟⎜ ⎟− − ⎢ ⎥⎝ ⎠ ⎣ ⎦
, 1, if 02, if 0
ii
δδ
= >⎧⎨ = <⎩
.
Hence, 2
121 0
01 0
00 0lQ T
i lH β θθδ β ε×
×
⎛ ⎞ ⎡ ⎤⎡ ⎤− ≥⎜ ⎟ ⎢ ⎥ ⎣ ⎦⎜ ⎟− − ⎣ ⎦⎝ ⎠
, which is equivalent to
1 0
21 0
1 02
0
00
TQ Ti l
Tl
H θ
δ β εθ
β
×
×
⎡ ⎤⎡ ⎤⎣ ⎦⎢ ⎥⎢ ⎥ ≥⎛ ⎞− −⎢ ⎥⎡ ⎤ ⎜ ⎟⎣ ⎦⎢ ⎥⎝ ⎠⎣ ⎦
, 1, if 02, if 0
ii
δδ
= >⎧⎨ = <⎩
. (5.36)
183
When QiC are designed as in (5.35), and if (5.36) is satisfied, then it is easy to see that (5.32)
is also satisfied with 1 03
2
δ β εβ
β− −
= . It follows that
1 03 3
2
( ) , e k k kδ β ε
β− −
< ∀ ≥ .
Let { }0 1 2 3max , ,k k k k k= + , then based on the definition of ( )e k in (5.15) and the results in
Lemmas 5.3-5.5 , we have 1 01 2 0
2
( )e kδ β ε
β β ε δβ
⎛ ⎞− −< + + <⎜ ⎟
⎝ ⎠, k k∀ ≥ . Hence, after
time k k= , there will be no more switching and the closed loop system will evolve in only
one of the two modes { }1, 2i∈ . For the case of 0δ > , the system 1clΣ is active for k k≥ .
Since regulation condition (4.18) is satisfied for the system 1clΣ with 1
QC chosen as in (5.35),
then regulation for the switched closed loop system (5.7) is achieved. The analysis is similar
for the case of 0δ < .
It can be seen in the conditions of Theorem 5.2 that the smaller the term 2
1 0
βδ β ε− −
in
(5.34) is, the easier it is for (5.34) to be satisfied. In Appendices A and B, an approach to the
design of the stabilizing controllers is presented and that will make it easier for (5.34) to be
satisfied. The approach is based on first designing the gains iK by solving the problem of
minimizing 2β subject to (5.25) and (5.26). Then, based on the known iK , the gains iL will
be designed by solving the problem of minimizing 1β subject to (5.16) and (5.17).
Furthermore, (5.16) is equivalent to (5.19) which implies that (5.9) and (5.10) can be
satisfied. Therefore the proposed method to design iK and iL yields an internally stable
184
switched closed loop system (5.7). Solving the two optimization problems mentioned above
is difficult given that (5.16) and (5.25) are nonlinear matrix inequalities. In Appendices A
and B, two algorithms are developed to find local solutions for the two optimization
problems based on iterative LMIs. The resulting gains iK and iL , { }1, 2i∈ , yield the
minimum value 1β∗ of 1β such that (5.16) and (5.17) are satisfied and the minimum value
2β∗ of 2β such that (5.25) and (5.26) are satisfied.
Consider now the following assumption which will be used in the development of the
adaptation algorithm in the next section for the case of exogenous inputs with unknown
properties.
Assumption 5.2
Consider the parameterized stabilizing controllers obtained using the gains iK and iL ,
{ }1, 2i∈ , provided by Algorithms 1 and 2 in Appendices A and B, respectively, and the
parameters iQ , { }1, 2i∈ , with 0θ η≤ . Condition (5.34) in Theorem 5.2 is satisfied with
1 1β β ∗= and 2 2β β ∗= .
Remark 5.3
If Assumption 5.2 is satisfied, then based on Theorem 5.2, as long as QiC is set as given in
(5.35), the switched closed loop system (5.7) can achieve regulation. Based on the matrices
QiH , { }1, 2i∈ , obtained in (5.34), define the following set:
185
1
2
1 01
2
01, if 0
= : 0, 2, if 00
q
TQ Ti l
n
Tl
Hiiθ
θδ
θ δ β ε δθβ
×
∗
× ∗
⎧ ⎫⎡ ⎤⎡ ⎤⎣ ⎦⎪ ⎪⎢ ⎥ = >⎧⎪ ⎪⎢ ⎥∈ ≥⎨ ⎨ ⎬⎛ ⎞− − = <⎢ ⎥ ⎩⎡ ⎤⎪ ⎪⎜ ⎟⎣ ⎦ ⎜ ⎟⎢ ⎥⎪ ⎪⎝ ⎠⎣ ⎦⎩ ⎭
S R . (5.37)
Equation (5.36) implies that 0θ is an element of the set θS . Furthermore, if the parameter
vector iθ , where 1i = in the case of 0δ > and 2i = in the case of 0δ < , is always selected
in θS , regardless of whether iθ changes as a function of time or not, and 10qj nθ ×= , where
2j = in the case of 0δ > and 1j = in the case of 0δ < , then we have ( )e k δ< , k k∀ ≥ ,
where { }0 1 2 3max , ,k k k k k= + . This result follows from the fact that if the first LMI in (5.34)
and the LMI in (5.37) are satisfied, then Lemma 5.5 implies that 1 03
2
( )e kδ β ε
β
∗
∗
− −≤ ,
k k∀ ≥ . Therefore, ( )e k δ< , k k∀ ≥ . Moreover, it should be noted that the set θS is a
convex set, a property that will be exploited in the development of the adaptation algorithm
for the switched closed loop system.
5.5. Adaptive regulation in the switched closed loop system
In the previous sections, it was assumed that the disturbance properties are known. Using the
proposed Q -parameterized controllers given by (5.4) and (5.5), conditions for regulation in
the resulting switched closed loop system are presented. In the case where the disturbance
input properties are unknown and possibly time-varying, it is desired to introduce adaptation
in the controller design process. After obtaining the controller parameters iK , iL and the
corresponding 1β∗ and 2β
∗ values in the previous sections, and if Assumption 5.2 is satisfied,
186
then adaptation can be introduced in the design of the proposed Q -parameterized controllers
given by (5.4) and (5.5). The aim of the adaptation is to tune the parameter vector iθ in QiC
so that it converges to the desired parameter vector 0θ as in (5.35) that yields regulation.
This section presents the development of an adaptation algorithm for the switched closed
loop system and presents the properties of the resulting adaptive switched closed loop system.
Based on the regulation condition in Theorem 4.2, 0θ is defined by only one set of
interpolation conditions in (4.18) corresponding to 1i = in the case of 0δ > and 2i = in the
case of 0δ < . Therefore, at any time k , when the system cliΣ is active, where 1i = in the
case of 0δ > and 2i = in the case of 0δ < , a parameter adaptation algorithm is turned on
to provide an estimate iθ of the parameter vector 0θ . Otherwise, the adaptation algorithm is
turned off and the estimated parameter vector iθ is frozen. In the rest of this section it is
assumed that 10qj nθ ×= where 2j = in the case of 0δ > and 1j = in the case of 0δ < .
Assume that the system cliΣ is active at 0k = . Let b
mk and emk , 1, 2,m = , denote the
beginning time and the end time of the time period during which cliΣ is active before the thm
switching from cliΣ to cl
jΣ (See Figure 5.1). Let e bm m mk kτ = − denote the corresponding dwell
time for cliΣ . Now let lz− denote the l time step delay operator and
1 1,
1
ˆ ( ) ( )qn
q qi k i
q
Q k z F zθ − −
=
= ∑ . Let 22 122( ) ( ) ( )i Qe k T z y k−= , which is an exponentially decaying
signal representing the response of the system 22iT to its initial conditions, regardless of what
( )Qy k is. The performance variable ( )e k in the adaptive closed loop system cliΣ is given by:
187
Figure 5.1. Diagram of the switching sequence for the closed loop system.
( )11 1 12 1 21 1, 22
11 1 12 1,
( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ),i i i i k i i
i i i i k
e k T z d k T z Q T z d k e k
T z d k T z Q d k
− − −
− −
= + +
= +
(5.38)
where,
21 122 ˆ( ) ( ) ( ) ( ) ( ) ( ).i id k T z d k e k y k y k−= + = −
Define a modified performance variable
12 1 12 1, ,( ) ( ) ( ) ( ) ( ) ( )i k i i i ke k e k Q T z d k T z Q d k− −= + − . (5.39)
Then we have
11 1 12 1,
11 1 1 1 12 1
1
( ) ( ) ( ) ( ) ( )
ˆ( ) ( ) ( ) ( ) ( ) ( ).q
i i i k i
nq q
i i i iq
e k T z d k Q T z d k
T z d k k z F z T z d kθ
− −
− − − −
=
= +
⎛ ⎞= + ⎜ ⎟
⎝ ⎠∑
(5.40)
Let
11 10
1 12 11
( ) ( ) ( ),
( ) ( ) ( ) ( ).i i
i
v k T z d k
v k F z T z d k
−
− −
=
= (5.41)
Then
10 1
1
ˆ( ) ( ) ( ) ( ),qn
q qi
q
e k v k k z v kθ −
=
= +∑
which can be rewritten as:
bmk e
mk 1bmk + 1
emk +
mτ 1mτ +
k
cliΣ is active
cljΣ is active
188
0ˆ( ) ( ) ( ) ( ),Tie k v k k kφ θ= − (5.42)
where
1 1
1
( ) ( ), , ( 1) ,
ˆ ˆ ˆ( ) ( ), , ( ) .q
T
q
Tni i i
k v k v k n
k k k
φ
θ θ θ
⎡ ⎤= − − − +⎣ ⎦
⎡ ⎤= ⎣ ⎦
(5.43)
It should be noted that the signal 1( )v k can be computed at each time step k since the signal
( )d k in (5.41) represents the second output of the iJ block in Figure 2.1, and is therefore
accessible. Similarly, the modified performance variable ( )e k in (5.39) can be computed
from ( )e k at each time step k .
Let 0θ be the nominal parameter vector satisfying the regulation interpolation conditions
(4.18) and 1 10 0
1
( )qn
q q
q
Q z F zθ − −
=
=∑ . Then define
( )11 1 12 1 1 21 10 22
11 1 1 1 12 10
1
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ).q
i i i i i
nq q
i i iq
e k T z d k T z Q z T z d k e k
T z d k z F z T z d kθ
− − − −∗
− − − −
=
= + +
= +∑ (5.44)
It follows that
0 0( ) ( ) ( ) .Te k v k kφ θ∗ = −
It should be noted that ( )e k∗ is bounded by an exponentially decaying signal. The modified
performance variable can be expressed as follows
( ) ( ) ( ) ( )
( ) ( 1) ( ),T
e k e k e k e k
k k e kφ θ∗ ∗
∗
= − +
= − + (5.45)
where
0ˆ( ) ( )ik kθ θ θ= − .
189
The modified error ( )e k is linearly parameterized in the parameter estimation error vector θ ,
and therefore, will be used instead of performance error ( )e k in the development and
analysis of the proposed adaptation algorithm.
The main idea behind the adaptation algorithm is as follows. If Assumption 5.2 is satisfied,
then we have 0 θθ ∈S , where θS defined in (5.37). If at every time step the estimate ( )i kθ
satisfies ˆ ( )i k θθ ∈S , and 10qj nθ ×= , then based on Remark 5.3, we have ( ) , e k k kδ< ∀ ≥ .
Hence for k k≥ , the system cliΣ is always active. Consequently, if the parameter estimate
ˆ ( )i kθ can be driven to 0θ asymptotically, then based on Theorem 5.2, the adaptive switched
closed loop system cliΣ achieves regulation. In the following, an adaptation algorithm based
on the recursive least squares algorithm with time-varying forgetting factor will be used to
asymptotically drive the parameter estimate ˆ ( )iθ ⋅ to 0θ . However, the recursive least squares
adaptation algorithm cannot guarantee ( )i θθ ⋅ ∈S for all 0k ≥ . Therefore, to deal with the
case where ( )i θθ ⋅ ∉S , it is proposed to use the recursive least squares algorithm with
projection onto the boundary of the set θS whenever ( )i θθ ⋅ ∉S . For [ , ]b em mk k k∈ , 1, 2,m = … ,
the algorithm is given as follows:
( ) ( 1)ˆ ˆ( 1) Proj ( ) ( 1) , 1 ( 1) ( ) ( 1)
1 ( 1) ( ) ( ) ( 1) ( ) ( 1) . ( ) 1 ( ) ( 1) ( )
i i T
T
T
P k kk k e kk P k k
P k k k P kP k P kk k P k k
φθ θφ φ
φ φρ φ φ
⎧ ⎫++ = + +⎨ ⎬+ + +⎩ ⎭
⎡ ⎤− −= − −⎢ ⎥+ −⎣ ⎦
(5.46)
190
where 1ˆ ˆ( ) ( )b ei m i mk kθ θ −= , 1( ) ( )b e
m mP k P k −= , 1 0( ) 0bP k P= > , ( )kρ is a time-varying forgetting
factor satisfying min max0 ( ) 1kρ ρ ρ< ≤ ≤ < , and where ( ){ }ˆProj iθ ⋅ denotes the orthogonal
projection of ( )iθ ⋅ onto the set θS and is computed according to the following steps:
1) If ( )i k θθ ∈S , then
{ }ˆ ˆProj ( ) ( )i ik kθ θ= . (5.47)
2) If ( )i k θθ ∉S , then
(a) Transform the coordinate basis for the parameter space by defining
1 2ˆ ˆ( ) ( ) ( )i ik P k kϑ θ−= ,
and denote by θS the image of θS under the linear transformation 1 2( )P k − .
(b) Orthogonally project the image ˆ ( )i kϑ of ˆ ( )i kθ onto the boundary of θS to yield
ˆ ( )i kϑ∗ by solving:
ˆ ˆmin ( ) ( )i ik kϑ ϑ∗ −
( )
( )
1 21
2
1 01 21
2
ˆ0 ( ) ( )
0ˆ0 ( ) ( )
TTQi l i
T
l i
H P k k
P k k
ϑ
δ β εϑ
β
∗×
∗∗
× ∗
⎡ ⎤⎡ ⎤⎢ ⎥⎢ ⎥⎣ ⎦⎢ ⎥ ≥⎢ ⎥⎛ ⎞− −⎡ ⎤⎢ ⎥⎜ ⎟⎜ ⎟⎢ ⎥⎣ ⎦⎢ ⎥⎝ ⎠⎣ ⎦
(c) Compute the estimated parameter vector as follows:
{ } 1 2ˆ ˆProj ( ) ( ) ( )i ik P k kθ ϑ∗= . (5.48)
As a first step in the adaptation process, an initial estimate ˆ (0)iθ in θS of the unknown
parameter vector is determined by solving the following LMI in the unknown ˆ (0)iθ ,
191
1
2
1 01
2
ˆ0 (0)
0ˆ0 (0)
TQi l i
l i
H θ
δ β εθ
β
×
∗
× ∗
⎡ ⎤⎡ ⎤⎣ ⎦⎢ ⎥
⎢ ⎥ ≥⎛ ⎞− −⎢ ⎥⎡ ⎤ ⎜ ⎟⎢ ⎥⎣ ⎦ ⎜ ⎟⎝ ⎠⎣ ⎦
, (5.49)
where QiH is obtained by solving (5.34). In the following, it will be shown that using the
adaptation algorithm in (5.46)-(5.48), and starting with an initial value ˆ (0)iθ that satisfies
(5.49), the estimated parameter vector ˆ ( )iθ ⋅ converges asymptotically to 0θ .
Based on Assumption 5.1, there exists a parameter vector 0θ satisfying the interpolation
conditions corresponding to the disturbance input properties. In order to analyze the
convergence of ( )kθ as k →∞ , we introduce a bound on the magnitude of the response e∗
in (5.44). Let 0 1β< < denote an upper bound on the rate of decay of the response e∗ and α
denote the maximum magnitude of the response e to the exogenous input id . Then, we have
( ) ( ) , [ , ], 1, 2,bmk k b e
m me k e k k k k mαβ −∗ ∗< = ∈ = … (5.50)
In the following, we will assume the constants α and β in (5.50) are known a priori. The
following assumption is used to derive some properties of the covariance matrix ( )P ⋅ in
(5.46).
Assumption 5.3
The signal ( )1v ⋅ in (5.41) is persistently exciting of order qn .
192
With 02 1qn k= + , and since 22 ( )e k is an exponentially decaying signal, then Assumption 5.3
is valid as long as no zeros of 12 21( ) ( ) ( )i iT z F z T z coincide with poles of ( )iD z . The following
Theorem presents the convergence properties of the adaptation algorithm.
Theorem 5.3
Assume that the switched closed loop system (5.7) is internally stable under arbitrary
switching. If Assumptions 5.2 and 5.3 are satisfied, then updating the parameters iθ in QiC
using the algorithm given by (5.46)-(5.48), where 1i = for 0δ > and 2i = for 0δ < , and
setting 10qj nθ ×= , j i≠ , we have:
( )0ˆlim ( ) 0ik
kθ θ→∞
− = , (5.51)
and the adaptive switched closed loop system achieves regulation.
Proof: The proof is given in Appendix C.
Hence, the adaptation algorithm is capable of driving the parameter estimates to the desired
parameter vector 0θ needed to achieve regulation for the switched system. In the following
section, a numerical example is used to illustrate the effectiveness of the proposed adaptation
algorithm.
193
5.6. Numerical example
In this section, the adaptive regulation method proposed in this chapter will be used to design
a controller that cancels the contact vibrations in the mechanical system discussed in section
2.6. First, the system is discretized to obtain a zero-order-hold equivalent model with
sampling period 0.001sT s= as in (5.1) with
3 6
1 3
945.6832 10 976.8748 10107.4562 935.9144 10
A− −
−
⎡ ⎤× ×= ⎢ ⎥− ×⎣ ⎦
, 3 6
2 3
995.0125 10 995.8425 109.9584 990.0333 10
A− −
−
⎡ ⎤× ×= ⎢ ⎥− ×⎣ ⎦
,
9
1 6
493.7891 10976.8748 10
B−
−
⎡ ⎤×= ⎢ ⎥×⎣ ⎦
, 9
1 6
498.7520 10995.8425 10
B−
−
⎡ ⎤×= ⎢ ⎥×⎣ ⎦
,
9
1 6
493.7891 10 0976.8748 10 0
xD−
−
⎡ ⎤×= ⎢ ⎥×⎣ ⎦
, 9
2 6
498.7520 10 0995.8425 10 0
xD−
−
⎡ ⎤×= ⎢ ⎥×⎣ ⎦
,
[ ]1 2 1 2 1 0e e y yC C C C= = = = , 51 2 1 2 0 1 10e e y yD D D D −⎡ ⎤= = = = − ×⎣ ⎦ ,
1
0.0747sin(20 ) 0.175cos(20 )0.1cos(20 )
k kd
kπ π
π+⎡ ⎤
= ⎢ ⎥⎣ ⎦
, 2
0.075sin(20 ) 0.075cos(20 )0.1 cos(20 )
k kd
kπ π
π+⎡ ⎤
= ⎢ ⎥×⎣ ⎦.
The switching rule is defined as 6
6
1 if 50 102 if 50 10
er
e
−
−
⎧ ≤ − ×= ⎨
> − ×⎩. To construct a parameterized set
of stabilizing controllers, the gains iK , { }1, 2i∈ , are designed with an initial value
2,0 0.02α = . After 10 iterations using Algorithm 1, we obtain
62 1.0838 10β ∗ −= × , 2 0.0256α = ,
6 31 1.9740 10 1.9487 10K ⎡ ⎤= − × − ×⎣ ⎦ , 6 3
2 2.0937 10 1.9816 10K ⎡ ⎤= − × − ×⎣ ⎦ .
194
Then, the gains iL , { }1, 2i∈ , are designed with initial values 1,0 0.02α = ,
[ ]1,0 0.9473 102.7668 TL = and [ ]2,0 0.9951 9.7338 TL = − . After 8 iterations using
Algorithm 2, we obtain
61 21.4066 10β ∗ −= × , 1 0.0249α = , [ ]1 0.9549 100.2073 TL = − , [ ]2 0.8834 1.3441 TL = − .
Based on the analysis of the poles of ( )iD z , we have 0 1k = , and 02 2pin k= = since the
offset in the external input signal is 0. The parameter iQ is selected to be of the form
( )1 2 1 32
0.5 0.5 102 1i i i
qQ qq q
θ θ − −+= + ×
− +. Since 650 10 0δ −= − × < , then according to Theorem 5.3
only the system 2Σ is considered to compute the nominal parameter vector 0θ needed to
achieve regulation. Using regulation condition (4.18), it follows that:
0
37.7747 38.5927
θ−⎡ ⎤
= ⎢ ⎥⎣ ⎦
, 72
5.0431 5.008010
0.4006 0.7165Aθ −− −⎡ ⎤
= ×⎢ ⎥⎣ ⎦
, 62
0.018710
1.2395Bθ −−⎡ ⎤
= ×⎢ ⎥−⎣ ⎦.
The corresponding state space realization of iQ as given in (5.6) is:
0 1 01 2 00 1 0
QA−⎡ ⎤
⎢ ⎥= ⎢ ⎥⎢ ⎥⎣ ⎦
, 3
0.50.5 100
QB −
⎡ ⎤⎢ ⎥= ×⎢ ⎥⎢ ⎥⎣ ⎦
,
1 20Qi i iC θ θ⎡ ⎤= ⎣ ⎦ .
Assuming 70η = , and solving the LMIs in (5.34) with 1β∗ and 2β
∗ yields
3 7 11 7
3 5 9 14 9
7 9 52
7 14 5
4.3293 4.9191 10 6.9764 10 8.1933 10 6.9755 104.9191 10 1.2463 10 3.1686 10 8.3198 10 3.1685 106.9764 10 3.1686 10 3.8281 2.6668 10 3.82818.1933 10 8.3198 10 2.6668 10 2.
QH
− − − −
− − − − −
− − −
− − −
− × − × − × − ×− × × × × ×
= − × × ×− × × × 7 5
7 9 5
8788 10 2.6362 106.9755 10 3.1685 10 3.8281 2.6362 10 3.8280
− −
− − −
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥
× ×⎢ ⎥⎢ ⎥− × × ×⎣ ⎦
.
Therefore, the convex set θS for the adaptive algorithm is formulated as
195
2 1 32
1 3
0= : 0
0 696.0531
TQ T
T
Hθ
θθ
θ
×
×
⎧ ⎫⎡ ⎤⎡ ⎤⎪ ⎪⎣ ⎦⎢ ⎥∈ ≥⎨ ⎬⎢ ⎥⎡ ⎤⎪ ⎪⎣ ⎦⎣ ⎦⎩ ⎭
S R .
Since the external input signal id is assumed to have unknown amplitude, frequency, and
phase, it is desired to search online for the desired parameter vector 0θ that yields the desired
regulator. The forgetting factor in the algorithm is a constant 0.9ρ = . The initial conditions
of the algorithm are ˆ(0) [0 0]Tθ = and 4(0) 10P I= where I is 2 2× identity matrix. In the
adaptation algorithm, the UD factorization algorithm is used in order to improve the
numerical properties of the algorithm in (5.46). The performance of the closed loop adaptive
system is illustrated in Figures 5.2-5.3. It can be seen that the parameter vector ˆ( )θ ⋅
converges to the nominal parameter 0θ , and that the performance variable e is driven to zero
asymptotically.
196
0 0.05 0.1 0.15 0.2 0.25 0.3-100
-50
0
50
e (m
icro
met
ers)
0 0.05 0.1 0.15 0.2 0.25 0.3-0.2
-0.1
0
0.1
0.2
d r(1,1
) (m
/sec
2 )
Time(sec)
Figure 5.2. Simulation results showing the performance variable e and the switching component (1,1)id in the input id .
0 0.05 0.1 0.15 0.2 0.25 0.3-100
-50
0
50
100
150
Time(sec)
Thet
a
estimated parameternominal parameter
Figure 5.3. Simulation results showing estimated parameter vector θ and nominal parameter vector 0θ .
197
5.7. Experimental evaluation
In this section, the experimental setup presented in section 4.7 will be used to evaluate the
proposed adaptive regulation method. In the following, based on the obtained models (4.48)
for the non-contact and (4.49) for the contact modes, an adaptive regulator will be designed
to maintain a constant separation of 650 10 m−× between the tip of the suspension beam and a
hypothetical contact surface ( )( ) 65sin 20 50 10cs t mπ −= − × . The output y , to be fed to the
controller, is defined to be the measurement voltage corresponding to the difference between
the actual separation t cy s− and the desired separation of 650 10 m−× :
( )650 10= 0.5sin(240 ) volts.
y t c
v
y k y sy kπ
−= − − ×
−
Define the performance variable e to be the same as the output y :
0.5sin(240 ) voltsve y kπ= − .
Then based on the transfer functions (4.48) and (4.49), the state space representation for the
switched bimodal system as in (5.1) is given by
1
0 0.87351 0.5645
A−⎡ ⎤
= ⎢ ⎥−⎣ ⎦, 2
0 0.93131 0.2459
A−⎡ ⎤
= ⎢ ⎥−⎣ ⎦,
1
1.80450.9125
B ⎡ ⎤= ⎢ ⎥⎣ ⎦
, 2
2.16011.0036
B ⎡ ⎤= ⎢ ⎥⎣ ⎦
,
1
0.2021 00.2010 0
xD ⎡ ⎤= ⎢ ⎥⎣ ⎦
, 2
0.2267 00.2187 0
xD ⎡ ⎤= ⎢ ⎥⎣ ⎦
,
[ ]1 1 0 1e yC C= = , [ ]2 2 0 1e yC C= = ,
198
[ ]1 2 1 2 0 1e e y yD D D D= = = = ,
111
( )( )
0.5sin(240 ) voltsd k
d kkπ
⎡ ⎤= ⎢ ⎥⎣ ⎦
, 212
( )( )
0.5sin(240 ) voltsd k
d kkπ
⎡ ⎤= ⎢ ⎥⎣ ⎦
,
where
( )311
310 2cos(240 ) 0.4sin(240 ) 0.3cos (240 ) volts4
cc c ac
i v
kd s s v k k kk k
ππ π π− ⎛ ⎞= + × + = + + +⎜ ⎟⎝ ⎠
,
2132cos (240 ) volts4ad v k ππ= = + .
Since it is desired to maintain the tip of the flexible beam at a distance of 650 10 m−× above
the contact surface, the switching rule is defined as 1 if 5 volts2 if 5 volts
er
e≤ −⎧
= ⎨ > −⎩. The parameter
iQ is selected to be of the form ( )1 2 1 32
4.8 4.8 101.5 1i i iqQ q
q qθ θ − −+
= + ×− +
. Since 5 volts 0δ = − < ,
then according to Theorem 5.3 only the system 2Σ is considered to compute the nominal
parameter vector 0θ needed to achieve regulation. The parameterized set of stabilizing
controllers is first designed so that conditions (5.16)-(5.26) are satisfied with
1 3.1753β = , 2 4.7455β = ,
[ ] 31 222.7372 509.4074 10K −= − × , [ ] 3
2 178.5739 397.7601 10K −= − × .
[ ] 31 843.0621 549.3770 10TL −= × , [ ] 3
2 931.2618 245.8963 10TL −= × .
Using regulation condition (4.18), it follows that:
30
4.260510
3.1377θ −⎡ ⎤= ×⎢ ⎥−⎣ ⎦
, 2
54.8736 50.209914.9131 26.6924
Aθ ⎡ ⎤= ⎢ ⎥−⎣ ⎦
, 32
76.245410
147.2916Bθ −−⎡ ⎤
= ×⎢ ⎥−⎣ ⎦.
199
Furthermore, the bounds on the nominal parameter vector 0θ outlined in Assumption 5.1 are
assumed to be given by 3 30
10 1010 10
10 10l uθ θ θ− −−⎡ ⎤ ⎡ ⎤× = ≤ ≤ = ×⎢ ⎥ ⎢ ⎥−⎣ ⎦ ⎣ ⎦
and 314.1421 10η −= × .
Therefore, conditions (5.31) and (5.34) are satisfied with
3 3 3 9 3
3 3 3 9 3
3 3 3 3 32
9
10.2086 10 3.8842 10 3.6034 10 7.4146 10 3.6031 103.8842 10 5.2234 10 202.3389 10 325.4124 10 202.3379 10
3.6034 10 202.3389 10 750.1231 10 1.3598 10 750.1203 107.4146 10 325.
QH
− − − − −
− − − − −
− − −
−
× − × × − × ×− × × × × ×
= × × × − × ×− × 9 3 3 3
3 3 3 3 3
4124 10 1.3598 10 10.2271 10 1.3673 103.6031 10 202.3379 10 750.1203 10 1.3673 10 750.1175 10
− − − −
− − −
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥
× − × × − ×⎢ ⎥⎢ ⎥× × × − × ×⎣ ⎦
.
In the experiment, the properties of contact surface cs and of the external disturbance are
assumed unknown. The proposed adaptation algorithm is used to tune the parameter iQ in
the controller to yield regulation. The projection algorithm ( ){ }ˆProj iθ ⋅ in the adaptive
algorithm is modified according to the following steps in order to simplify the calculations in
the real-time implementation:
1) If ( )ˆl i ukθ θ θ≤ ≤ , then
{ }ˆ ˆProj ( ) ( )i ik kθ θ= . (5.52)
2) Otherwise
(a) Transform the coordinate basis for the parameter space by defining
1 2ˆ ˆ( ) ( ) ( )i ik P k kϑ θ−= ,
(b) Solve the following optimization problem with the unknown variable ˆ ( )i kϑ∗ :
ˆ ˆmin ( ) ( )i ik kϑ ϑ∗ −
1 2 ˆ( ) ( )l i uP k kθ ϑ θ∗≤ ≤ .
(c) Compute the estimated parameter vector as follows:
200
{ } 1 2ˆ ˆProj ( ) ( ) ( )i ik P k kθ ϑ∗= . (5.53)
First, assume the external disturbance is known, the experimental results with the nominal
parameter 0θ are verified and the results are shown in Figure 5.4-5.7. Then the external
disturbances are assumed unknown and the adaptive regulation method is applied. A branch
and bound type optimization algorithm proposed [100,101] is used to find a solution for the
projection operation in the adaptation algorithm. The forgetting factor in the adaptation
algorithm is a constant 0.93ρ = . The initial conditions for the adaptation algorithm are
ˆ(0) [0 0]Tθ = and 6(0) 10P I= where I is 2 2× identity matrix. In the adaptation algorithm,
the UD factorization algorithm is used in order to improve the numerical properties of the
algorithm. The experimental results are presented in Figures 5.8-5.12. It can be seen that
despite the unknown properties of the contact surface cs and of the external disturbance, and
despite the presence of switching in the plant dynamics, the parameter vector ˆ( )θ ⋅ still
converges to the desired parameter vector 0θ and that the performance variable e is driven to
zero asymptotically using the proposed adaptation algorithm.
5.8. Conclusion
The problem of regulation in bimodal switched systems against unknown disturbance or
reference signals is discussed. An adaptive regulator design approach based on the
parameterization of a set of stabilizing controllers for the switched system is presented. First,
regulation conditions for the switched system are derived and are used to construct a set of
Q parameters needed to achieve regulation for the case of known disturbance or reference
201
signals. For the case of unknown disturbance or reference signals, an adaptation algorithm is
developed to tune online the Q parameter in the expression of the parameterized controller.
The tuning is such that the tuned Q parameter converges to the desired Q parameter that
guarantees regulation for the switched closed loop system. A numerical example of a
mechanical system subject to contact vibrations is used to illustrate the proposed adaptive
regulator design method. The performance of the proposed adaptive regulation system is also
successfully validated on a switched mechanical system experimental setup.
202
0 0.05 0.1 0.15 0.2 0.25 0.3
-100
-80
-60
-40
-20
0
20
40
60
80
100
Per
form
ance
erro
r (m
icro
met
ers)
Time(sec)
Figure 5.4. Experimental results showing the performance variable e for the case where the nominal parameter vector 0θ is used.
0 0.05 0.1 0.15 0.2 0.25 0.3
-100
-80
-60
-40
-20
0
20
40
60
80
100
Pos
ition
(mic
rom
eter
s)
Time(sec)
Tip positionContact surface
Figure 5.5. Experimental results showing the tip position ty of the suspension beam and the contact surface cs for the case where the nominal parameter vector 0θ is used.
203
0 0.05 0.1 0.15 0.2 0.25 0.3
-20
0
20
40
60
80
100
120
140
Cur
rent
in th
e co
il (m
A)
Time(sec)
Figure 5.6. Experimental results showing the current i in the coil used to simulate the disturbance force and the contact force for the case where the nominal parameter vector 0θ is used.
0 0.05 0.1 0.15 0.2 0.25 0.3-5
-4
-3
-2
-1
0
1
2
3
4
5
Con
trol i
nput
(vol
ts)
Time(sec)
Figure 5.7. Experimental results showing the control input signal u driving the piezoelectric actuator for the case where the nominal parameter vector 0θ is used.
204
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
-100
-80
-60
-40
-20
0
20
40
60
80
100
Per
form
ance
var
iabl
e (m
icro
met
ers)
Time(sec)
Figure 5.8. Experimental results showing the performance variable e for the case where the adaptive regulator is used.
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
-100
-80
-60
-40
-20
0
20
40
60
80
100
Pos
ition
(mic
rom
eter
s)
Time(sec)
Tip positionContact surface
Figure 5.9. Experimental results showing the tip position ty of the suspension beam and the contact surface cs for the case where the adaptive regulator is used.
205
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
-20
0
20
40
60
80
100
120
140
Cur
rent
in th
e co
il (m
A)
Time(sec)
Figure 5.10. Experimental results showing the current i in the coil used to simulate the disturbance force and the contact force for the case where the adaptive regulator is used.
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-5
-4
-3
-2
-1
0
1
2
3
4
5
Con
trol i
nput
(vol
ts)
Time(sec)
Figure 5.11. Experimental results showing the control input signal u driving the piezoelectric actuator for the case where the adaptive regulator is used.
206
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-0.015
-0.01
-0.005
0
0.005
0.01
0.015
Time(sec)
Thet
a
estimated parameternominal parameter
Figure 5.12. Experimental results showing the desired parameter vector 0θ (dashed line) and
the estimated parameter vector θ (solid line) for the case where the adaptive regulator is used.
207
CHAPTER 6
Conclusion
Motivated by the flying height regulation problem of the read/write head in hard disk drives,
this thesis addressed the regulation problem in switched bimodal linear systems. Three main
contributions are developed with respect to the different cases of the exogenous input signals
and are summarized in the following section. The last section in this chapter addresses future
work based on the research results presented in the thesis.
6.1. Contributions of the thesis
The main approach used in this thesis to solve the regulation problem in bimodal linear
systems is based on designing Q -parameterized output feedback controllers for the switched
system. The free Q parameter in the parameterized controller is selected properly to yield
regulation in the switched bimodal system. A summary of the three main contributions in the
thesis are summarized in the following.
208
6.1.1 Development of a regulator synthesis procedure in bimodal linear systems against
known deterministic exogenous inputs
Regulation methods for switched continuous-time bimodal linear systems against known
disturbance and reference signals are developed. Two main methods are proposed for this
case.
1) Regulator synthesis based on bilinear matrix inequalities: The regulator design
approach consists of two main steps. The first step is based on constructing a switched
observer-based state feedback central controller for the switched linear system, which is
augmented with additional dynamics to construct a Q -parameterized set of switched
controllers. The proposed Q -parameterized controller is different from the traditional Q -
parameterized controllers for linear systems in that the Q parameters are not necessarily
asymptotically stable systems. In the second step, two sufficient regulation conditions
are derived for the resulting parameterized switched closed loop system. The regulation
conditions present guidelines for the selection of parameters in the controller to yield the
desired regulator for the switched bimodal systems. A corresponding regulator synthesis
approach is proposed based on solving properly formulated bilinear matrix inequalities.
2) Regulator synthesis based on linear matrix inequalities: In this approach, the
construction of a set of parameterized controllers is performed using asymptotically
stable Q parameters that are expressed as linear combinations of stable basis functions.
The stability for the resulting closed loop switched system with the Q -parameterized
controllers is analyzed, including the internal stability and input-output stability in the
closed loop system. Regulation conditions for each of the two subsystems in the
resulting bimodal switched closed loop system are derived in both the state space domain
209
and the frequency domain. Regulation conditions for the switched closed loop system are
then proposed. Two approaches are used to develop regulation conditions for the
switched closed loop system, namely, the common Lyapunov function approach and the
multiple Lypunov function approach. In the common Lyapunov function approach, the
input-output stability of the switched system is developed using a common Lyapunov
function, then the corresponding regulator synthesis method is formulated using properly
formulated linear matrix inequalities. In the multiple Lyapunov function approach, the
input-output stability property of the switched closed loop system is analyzed using a
multiple Lyapunov function approach based on the properties of the switching surface. A
corresponding regulator synthesis method is proposed using an iterative linear matrix
inequalities algorithm.
6.1.2 Development of a regulator synthesis procedure in bimodal linear systems with an
2H performance constraint.
A regulator synthesis method for switched discrete-time bimodal linear systems subject to
known disturbance or reference signals and unknown random input signals is developed. The
design of the proposed regulation method involves two main steps. In the first step, a set of
parameterized discrete time controllers that yield exact output regulation against the known
deterministic exogenous inputs in the switched bimodal system is constructed. In the second
step, additional constraints are added to identify 2H controllers within the already
constructed set of parameterized controllers and that can yield optimal output regulation
210
against the unknown random exogenous input signal. A corresponding regulator synthesis
method is developed by solving properly formulated linear matrix inequalities.
6.1.3 Development of an adaptive regulation approach in bimodal linear systems against
unknown sinusoidal exogenous inputs
An adaptive regulation problem for switched discrete-time bimodal linear systems, where it
is desired to achieve exact output regulation against unknown deterministic disturbance or
reference signals, is addressed. Assuming the properties of the deterministic exogenous input
signals are known, a sufficient regulation condition for the resulting switched closed loop
system involving Q -parameterized controllers is presented. The switched Q -parameterized
stabilizing controller is designed by iteratively solving linear matrix inequalities. In the case
of exogenous inputs with unknown properties, an adaptation algorithm is developed to tune
the Q parameter in the expression of the parameterized controller and make it converge to
the desired Q parameter that guarantees regulation for the switched bimodal system.
A summary of the advantages and limitations of various algorithms is given in the table 6.1.
211
rw Regulator Advantages Limitations
With unconstrained Q parameter
-- More flexibility -- BMI synthesis algorithm -- Local solutions -- Q structure not suitable for
adaptation Known deterministic
With constrained Q parameter
-- LMI synthesis algorithm-- Suitable for adaptation
-- Less flexibility in Q selection
Known deterministic and unknown random
With constrained Q parameter and added
2H performance constraints
-- Address both deterministic and random inputs
-- LMI synthesis algorithm
-- Less flexibility in Q selection-- Performance with respect to
steady state system
Unknown deterministic
With constrained Q parameter
-- Estimated Q converge to the desired Q
-- Need to know the number of sinusoids
-- Need to know the convex set θS containing the desired
parameter vector
Table 6.1 Summary of advantages and limitations of the proposed algorithms
6.2. Future work
Based on the research results presented in this thesis, recommendations for future work
include:
• Regulation in more General Switched Bimodal Systems
In this thesis, the switching rule between the two plant models as well as between the
disturbance and reference signals is defined according to the performance variable,
which is expressed as a time-varying switching surface with 0δ ≠ . A possible
extension of this work may include the case of 0δ = . Another possible extension is
the case where the switching times are not known. Future research should also aim at
reducing the conservativeness associated with the proposed synthesis approaches,
212
since the current LMI solvers are not able to deal with high order systems and with
cases involving very small values of δ .
• Regulation in Switched Linear Systems Involving more than Two Subsystems
The regulation methods developed in this thesis are associated with switched bimodal
systems. A more general theoretical analysis is needed for switched systems with
arbitrary numbers of subsystems. Regulation conditions for switching systems should
also be developed for cases where regulation implies that switching will always take
place, i.e. regulation is associated with an infinite number of switching instants. The
cases of known and unknown switching times (i.e. whether, at any given instant in
time, the mode of the plant is known or unknown) should be treated.
• Regulation in Switched Nonlinear Systems of Practical Interest
The flying height regulation problem that motivated this research is actually
associated with a nonlinear plant model due to the nonlinear intermolecular forces
that exist between the read/write head and the disk surface at very small flying
heights. Future research can extend the proposed regulation approach to this and other
types of nonlinear switched systems of practical interest.
• Application to the Flying Height Regulation in Hard Disk Drives
In this thesis, an experimental setup is used to evaluate the performance of the
proposed adaptive regulation method. The experimental setup involves the control of
the tip position of a hard disk drive suspension beam that is actuated using a
piezoelectric actuator. For this particular application, future research may focus on: (1)
The development of in situ real-time flying height measurement method. (2) The
development of PZT thin-film based micro actuators. (3) The development of less
213
conservative regulation approaches involving very small values of δ (i.e. very
small separation between the read/write head and the disk surface).
214
APPENDIX
A. Algorithm to compute iK , { }1, 2i∈ , and 2β∗
First, we introduce the following result.
Lemma A.1
Consider the switched system 12iT and a preset constant 2 0α > . If there exist matrices
121 0U > , 12
2 0U > , 121V and 12
2V , and a positive scalar 2β , such that the following
optimization problem
2min β
subject to
( )( ) ( )
( )
12 12 12 122
2
12 12 12
0
0 1 0,
T
i i i i i i
Ti
i i i i i j
U U AU BV
I B
AU BV B U
α
ε
⎡ ⎤− + +⎢ ⎥⎢ ⎥− − ≤⎢ ⎥⎢ ⎥+ −⎢ ⎥⎣ ⎦
( ) { } { }, 1, 2 1,2i j∀ ∈ × , (A.1)
( )12 12 122
12 12 22
0,T
i i i
i i
U C U
C U
α
β
⎡ ⎤⎢ ⎥ ≥⎢ ⎥⎣ ⎦
{ }1, 2i∀ ∈ , (A.2)
admits an optimal solution 2β∗ , then 2β
∗ is also the optimal solution of (5.25) and (5.26) with
feedback gains iK given by
215
( ) 112 12i i iK V U
−= , { }1, 2i∈ . (A.3)
Proof: Letting ( ) 112 12i iU H
−= , from (A.3), we have ( ) 112 12
i i iV K H−
= . Substituting 12iU and
12iV into (A.1), we have
( ) ( ) ( ) ( )
( ) ( )( )( ) ( )
1 1 112 12 122
2 21 112 12
0
0 1 0.
Ti i i i i i
Ti
i i i i i j
H H H A B K
I B
A B K H B H
α
ε α
− − −
− −
⎡ ⎤− + +⎢ ⎥⎢ ⎥− − ≤⎢ ⎥⎢ ⎥+ −⎢ ⎥⎣ ⎦
(A.4)
Multiplying (A.4) from the left side and the right side by ( )12 12, ,i jdiag H I H , respectively,
and substituting the expressions for 12iA and 12
iE , we immediately have (5.25). Similarly
multiplying (A.2) from the left side and the right side by ( )( )112 ,idiag U I−
, we have (5.26).
Note that (A.1) and (A.2) are linear in the unknowns 12iU , 12
jU , 12iV , and 2β only if 2α is
fixed. In the following, an algorithm is presented to obtain a local solution to the
optimization problem in Lemma A.1. Let 12,i nU , 12
,j nU , 12,i nV , 2,nβ and 2,nα denote the values of
12iU , 12
jU , 12iV , 2β and 2α in the thn iteration of the algorithm.
Algorithm 1: Design the controller gains iK , { }1, 2i∈
Step 1. Find an upper bound 2mα on the value of 2α in (A.1) with unknown 12iU , 12
jU ,
and 12iV , which is a generalized eigenvalue problem and can be solved efficiently using
the Matlab LMI toolbox. Select an initial value 2,0α for 2α in the interval ( )2 20, mα α∈ .
216
Step 2. (start the thn iteration). Minimize 2,nβ subject to (A.1) and (A.2) in the
unknowns 12,i nU , 12
,j nU , 12,i nV and 2,nβ ; and with known 2, 1nα − .
Step 3. Minimize 2,nβ subject to (A.1) and (A.2) in the unknowns 2,nα , 12,i nV and 2,nβ ;
and with known 12,i nU and 12
,j nU .
Step 4. If 2, 2, 1 2n nβ β σ−− < , a prescribed tolerance, go to step 5. Else set 1n n= + , then
go to step 2.
Step 5. Set 2 2,nβ β∗ = and ( ) 112 12, ,i i n i nK V U
−= .
For a given initial value 2,0α of 2α , the above algorithm converges to a local solution since
2β is guaranteed to decrease or stay the same in every iteration.
217
B. Algorithm to compute iL , { }1, 2i∈ , and 1β∗
Based on the designed gains iK , { }1, 2i∈ , in the above algorithm, we can design the gains
iL , { }1,2i∈ , by solving the optimization problem 1min β subject to (5.16) and (5.17).
However, matrix inequalities (5.16) and (5.17) are nonlinear in the unknown parameters 111H ,
112H , 1L , 2L , 1μ , 1β , and 1α . In the following, an algorithm is presented to obtain a local
solution based on preset initial values of 1L , 2L , and 1α . First we introduce the following
result to compute an initial values of iL , { }1, 2i∈ , to make the switched system 11iT
asymptotically stable. Due to the block triangular structure of the 11iA , the switched system
11iT is asymptotically stable under arbitrary switching if and only if the switched systems
(5.11) and (5.12) are asymptotically stable under arbitrary switching. Since (A.1) is satisfied,
it follows that (5.25) and (5.9) are satisfied, and the feedback gains iK , { }1, 2i∈ , provided
by Algorithm 1 yield an asymptotically stable switched system (5.11). However, the observer
gains iL , { }1, 2i∈ , cannot be computed directly based on (5.10) since it is a nonlinear matrix
inequality. Therefore, the following result will be used to find observer gains iL , { }1, 2i∈ ,
such that the switched system (5.12) is asymptotically stable. The computed gains iL ,
{ }1, 2i∈ , will serve as initial values for the observer gains in the algorithm to solve the
optimization problem mentioned above.
218
Lemma B.1
Consider the switched system (5.12), and let 0α ≥ be a preset constant. If there exist
matrices 1 0U > , 2 0U > and iV , and scalars iS , { }1,2i∈ , such that the following matrix
inequalities
( ) 0Ty
i i i i i i
yi i i i j
U U AU V C
AU V C U
α⎡ ⎤− + +⎢ ⎥ <⎢ ⎥+ −⎣ ⎦
, ( ) { } { }, 1, 2 1, 2i j∀ ∈ × , (B.1)
y yi i i iS C C U= , { }1, 2i∀ ∈ , (B.2)
are satisfied, then the observer gains iL given by
1i i
i
L VS
= , { }1,2i∈ , (B.3)
yield an asymptotically stable switched system (5.12).
Proof: From (B.2) and (B.3), we have
y yi i i i iV C L C U= , { }1,2i∈ .
Substituting in (B.1) the above expression for yi iV C and multiplying (B.1) from the left side
and the right side by ( )1 1,i jdiag U U− − , we have
( )( )
1 1 1
1 10
Tyi i i i i j
yj i i i j
U U A L C U
U A L C U
α− − −
− −
⎡ ⎤− + +⎢ ⎥ <⎢ ⎥+ −⎣ ⎦
,
which is equivalent to
1 1 1 0Ty y
i i i j i i i i iA L C U A L C U Uα− − −⎡ ⎤ ⎡ ⎤+ + − + <⎣ ⎦ ⎣ ⎦ , { }, 1, 2i j∈ .
219
Let 1li iP U −= and 1l
j jP U −= , then from the above inequality, we obtain (5.10). Therefore, the
switched system (5.12) is asymptotically stable.
Starting with the initial values for iL , { }1, 2i∈ , computed above and with an initial value for
1α , the following algorithm is proposed to find a local solution to the optimization problem
1min β subject to (5.16) and (5.17). Let 11,i nH , 11
,j nH , 1,nμ , 1,nβ , 1,nα and ,i nL denote the values
of 11iH , 11
jH , 1μ , 1β , 1α and iL in the thn iteration of the algorithm.
Algorithm 2: Design of the controller gains iL , { }1, 2i∈ .
Step 1. Set the initial values ,0iL of iL as ,01
i ii
L VS
= by solving the matrix inequalities
(B.1) and (B.2) in the unknowns 0iU > , iV and iS with a preset constant 0α ≥ .
Step 2. Find an upper bound 1mα on the value of 1α in (5.16) with unknown 11iH , 11
jH ,
and 1μ , which is a generalized eigenvalue problem. Select an initial value 1,0α for
1α in the interval ( )1 10, mα α∈ .
Step 3. (Start the thn iteration). Minimize 1,nβ subject to (5.16) and (5.17) with
unknown 11,i nH , 11
,j nH , 1,nμ and 1,nβ , and with known 1, 1nα − and , 1i nL − .
Step 4. Minimize 1,nβ subject to (5.16) and (5.17) in the unknowns 1,nα , ,i nL , 1,nμ and
1,nβ with known 11,i nH and 11
,j nH .
Step 5. If 1, 1, 1 1n nβ β σ−− < , a prescribed tolerance, go to step 6. Else set 1n n= + , then
go to step 3.
220
Step 6. Set 1 1,nβ β∗ = and ,i i nL L= , { }1, 2i∈ .
For given initial values ,0iL of iL and 1,0α of 1α , the above algorithm converges to a local
solution since 1β is guaranteed to decrease or stay the same in every iteration.
221
C. Proof of Theorem 5.3
First, consider the following Lyapunov function for convergence analysis:
1( ) ( ) ( ) ( )TV k k P k kθ θ−= .
The coordinate transformation 1 2ˆ ˆ( ) ( ) ( )i ik P k kϑ θ−= in the projection algorithm has the effect
of yielding
10 0 0 0
ˆ ˆ ˆ ˆ( ) ( ) ( ) ( ) ( ) ( )T T
i i i iV k k P k k k kθ θ θ θ ϑ ϑ ϑ ϑ−⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤= − − = − −⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ,
where 1 20 0( )P k θϑ θ−= ∈S . Thus, since ˆ ( )i kϑ∗ is an orthogonal projection of ˆ ( )i kϑ onto θS
and since 0 θϑ ∈S , then we have 2 2
0 0ˆ ˆ( ) ( )i ik kϑ ϑ ϑ ϑ∗− ≤ − , which means
{ } { }1 10 0 0 0
ˆ ˆ ˆ ˆProj ( ) ( ) Proj ( ) ( ) ( ) ( )T T
i i i ik P k k k P k kθ θ θ θ θ θ θ θ− −⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤− − ≤ − −⎣ ⎦ ⎣ ⎦⎣ ⎦ ⎣ ⎦ . (C.1)
As indicated in [102], it follows from (C.1) that convergence analysis results based on the
function 1( ) ( ) ( ) ( )TV k k P k kθ θ−= for the recursive least squares algorithm without projection
apply also to the algorithm with projection.
The recursive least squares algorithm is given by
ˆ ˆ( 1) ( ) ( 1) ( 1) ( 1) ( 1),i ik k k P k k e kθ θ ρ φ+ = + + + + + (C.2)
1 1 ( 1) ( 1) ( ) ( 1) ( 1) . TP k k P k k kρ φ φ− −⎡ ⎤+ = + + + +⎣ ⎦ (C.3)
Using equation (C.2) and the expression for e in (5.45), we have:
( 1) ( 1) ( 1) ( 1) ( 1) ( ) ( 1) ( 1) ( 1) ( 1).Tk I k P k k k k k P k k e kθ ρ φ φ θ ρ φ ∗⎡ ⎤+ = − + + + + − + + + +⎣ ⎦
From equation (C.3), we have:
222
1( 1) ( 1) ( 1) ( 1) ( 1) ( 1) ( ).TI k P k k k k P k P kρ φ φ ρ −− + + + + = + +
Therefore, for [ , ], 1, 2,b em mk k k m∈ = … , the dynamics of the parameter estimation error
vector is given by:
1( 1) ( 1) ( 1) ( ) ( ) ( 1) ( 1) ( 1) ( 1),k k P k P k k k P k k e kθ ρ θ ρ φ−∗+ = + + − + + + + (C.4)
where 1( ) ( )b em mk kθ θ −= , 1( ) ( )b e
m mP k P k −= . It follows that:
1 1( 1) ( 1) ( 1) ( ) ( ) ( 1) ( 1) ( 1)P k k k P k k k k e kθ ρ θ ρ φ− −∗+ + = + − + + + . (C.5)
From (C.3) we also have
1 11( ) ( 1) ( 1) ( 1)( 1)
TP k P k k kk
φ φρ
− −= + − + ++
. (C.6)
Using (C.4) with (C.6) we obtain
( 1) ( ) ( 1) ( 1) ( 1) ( 1) ( ) ( 1) ( 1) ( 1) ( 1)
Tk k k P k k k kk P k k e k
θ θ ρ φ φ θρ φ ∗
+ = − + + + +− + + + +
. (C.7)
Based on the Lyapunov function 1( 1) ( 1) ( 1) ( 1)TV k k P k kθ θ−+ = + + + and the projection
property of (C.1), using (C.5) and (C.7), we have that
{
}
1
1
2 1
2
( 1) ( 1) ( 1) ( 1)
( 1) ( ) ( 1) ( ) ( 1)
( 1) ( ) ( 1) ( 1) ( 1) ( 1)
( 1) ( 1) ( 1) ( ) ( ) ( 1)
( 1) ( ) ( 1) ( 1) ( 1) ( ) ( )( 1
T
T
T T
T
T T
V k k P k k
k V k k k k
k k k k P k k
k k P k P k k e k
k k k k P k P k kk
θ θ
ρ ρ θ φ
ρ θ φ φ φ
ρ φ θ
ρ θ φ φ θ
ρ
−
−∗
−
+ = + + +
≤ + + + − +
+ + + + + +
− + + + +
− + + + +
+ + 2) ( 1) ( 1) ( 1) ( 1).T k P k k e kφ φ ∗+ + + +
Using the second equation in (5.46), the following identities can be easily derived
( ) ( 1) 1 ( 1) ( 1) ( 1) ( 1)
( ) ( 1) 1 ( 1) ( ) ( 1)
T T
T
T
k k k k P k k
k kk P k k
θ φ ρ φ φ
θ φφ φ
⎡ ⎤+ − + + + + +⎣ ⎦+
= −+ + +
,
223
1( 1) ( 1) ( 1) ( ) ( )
( ) ( 1) 1 ( 1) ( ) ( 1)
T
T
T
k k P k P k k
k kk P k k
ρ φ θ
θ φφ φ
−− + + +
+= −
+ + +
,
( )
1
2
( 1) ( ) ( 1) ( 1) ( 1) ( ) ( )
( ) ( 1)
1 ( 1) ( ) ( 1)
T T
T
T
k k k k P k P k k
k k
k P k k
ρ θ φ φ θ
θ φ
φ φ
−+ + + +
+=
+ + +
,
( 1) ( ) ( 1)( 1) ( 1) ( 1) ( 1)1 ( 1) ( ) ( 1)
TT
T
k P k kk k P k kk P k k
φ φρ φ φφ φ
+ ++ + + + =
+ + +.
Using the above identities in the expression for ( 1)V k + , we obtain
( )2
2
( 1) ( 1) ( )( 1) ( ) ( 1) 2 ( ) ( 1) ( 1)
1 ( 1) ( ) ( 1)( 1) ( ) ( 1) ( 1) ( 1).
1 ( 1) ( ) ( 1)
T TT
T
T
V k k V kk k k k k e k
k P k kk P k kk e k
k P k k
ρρ θ φ θ φ
φ φφ φρφ φ
∗
∗
+ ≤ ++ ⎡ ⎤− + + + +⎢ ⎥⎣ ⎦+ + +
+ ++ + +
+ + +
Completing the square in the above equation we get
2
2
2
( 1)( 1) ( 1) ( ) ( ) ( 1) ( 1)1 ( 1) ( ) ( 1)
( 1) ( 1)1 ( 1) ( ) ( 1)
( 1) ( ) ( 1) ( 1) ( 1)1 ( 1) ( ) ( 1)
(
TT
T
T
T
kV k k V k k k e kk P k k
k e kk P k k
k P k kk e kk P k k
k
ρρ θ φφ φ
ρφ φ
φ φρφ φ
ρ
∗
∗
∗
+ ⎡ ⎤+ ≤ + − + + +⎣ ⎦+ + ++
+ ++ + +
+ ++ + +
+ + +
≤ + 2
2
1) ( ) ( 1) ( 1)
( 1) ( ) ( 1) ( 1).
V k k e k
k V k k e k
ρ
ρ ρ∗
∗
+ + +
≤ + + + +
(C.8)
Using the above equation (C.8) iteratively, it follows that at time emk , we have
224
12
1
1 2 2max max
1
1 2 2max max
1
212
max maxmax
( ) ( 1) ( ) ( ) ( )
( ) ( )
( ) ( )
( )
e eem mm
bbmm
em e b
m m m
bm
mm m
m m
k kke bm m
j ii ki k
kk i i kb
mi k
jb jm
j
bm
V k i V k j e i
V k
V k
V k
τ
ττ τ
τ τ
ρ ρ
ρ ρ α β
ρ ρ α β
βρ α ρρ
−
∗== +=
+ − −
= +
+ −
=
+
⎧ ⎫⎡ ⎤ ⎡ ⎤⎪ ⎪≤ + +⎢ ⎥ ⎢ ⎥⎨ ⎬⎢ ⎥ ⎢ ⎥⎪ ⎪⎣ ⎦ ⎣ ⎦⎩ ⎭
≤ +
≤ +
⎡≤ +
∑∏ ∏
∑
∑
1
1 12 2
max max12max max 2
max
2( 1)22 max
max max 2max
22max2
max max 2max
( )1
( )
( ) .
m
m
m m
m mm
m m
m
j
j
bm
bm
bm
V k
V k
V k
τ
τ
τ τ
τ ττ
τ ττ
β βρ ρ
ρ α ρβρ
β ρ βρ α ρ
ρ β
β ρ βρ α ρ
ρ β
=
+
+
+
⎤⎢ ⎥⎣ ⎦
⎛ ⎞⎡ ⎤ ⎡ ⎤⎜ ⎟−⎢ ⎥ ⎢ ⎥⎜ ⎟⎣ ⎦ ⎣ ⎦⎝ ⎠≤ +
⎡ ⎤− ⎢ ⎥⎣ ⎦
−≤ +
−
⎡ ⎤−⎣ ⎦≤ +−
∑
According to the adaptation algorithm, we have
1ˆ ˆ( ) ( )b e
m mk kθ θ −= ,
1( ) ( )b em mP k P k −= ,
which yields
22max2
max 1 max 2max
( ) ( )m m
me em mV k V k
τ ττ
β ρ βρ α ρ
ρ β−
⎡ ⎤−⎣ ⎦≤ +−
. (C.9)
Let em denote the index for the last switching sequence. It follows that e
bmk k≤ and that
emτ →∞ . Hence,
22max2
max 1 max 2max
lim ( ) lim ( )m me e
me
e ee em me e
e em m
k kV k V k
τ τ
τβ ρ β
ρ α ρρ β−
→∞ →∞
⎛ ⎞⎡ ⎤−⎣ ⎦⎜ ⎟≤ +⎜ ⎟−⎝ ⎠. (C.10)
Using the inequality (C.9) iteratively, we have
225
1
1
1 1
2212max2
1 max 1 max max 21 1 max
22max2
max 2max
( ) ( )
mem meem
jm
e
m me e
mme bm
m j m
V k V kτ ττ
τ
τ τ
β ρ βρ ρ α ρ
ρ β
β ρ βα ρ
ρ β
−
=
− −
−−
−= = +
⎧ ⎫⎡ ⎤⎡ ⎤⎡ ⎤−∑ ⎪ ⎪⎣ ⎦⎢ ⎥⎢ ⎥≤ + ⎨ ⎬−⎢ ⎥⎢ ⎥⎪ ⎪⎣ ⎦⎣ ⎦⎩ ⎭
⎡ ⎤−⎣ ⎦+−
∑ ∏. (C.11)
Since k is finite, then for all em m< , the dwell time mτ for the system cliΣ is finite also. Let
minτ denote the minimum dwell time of the system cliΣ . It follows from inequality (C.11) that
1
1 min
1 1
1
1 min
222max( 1) 2
1 max 1 max max 21 max
22max2
max 2max
(max 1 max
( ) ( )
( )
mem mem
m e
e
m me e
me
mm e
mm me b
mm
mb
V k V k
V k
τ τττ
τ τ
ττ
β ρ βρ ρ α ρ
ρ β
β ρ βα ρ
ρ β
ρ ρ
−
=
− −
−
=
−− −
−=
⎧ ⎫⎡ ⎤⎡ ⎤−∑ ⎪ ⎪⎣ ⎦⎢ ⎥≤ + ⎨ ⎬−⎢ ⎥⎪ ⎪⎣ ⎦⎩ ⎭⎡ ⎤−⎣ ⎦+
−
∑≤ +
∑
min min
min
min min
1min min
1 min
222max1) 2
max max 21 max
22max2
max 2max
22max( 1)2
max 1 max maxm
( )
e
me
mm e
mm
m
mbV k
τ ττ
τ τ
τ τττ
β ρ βρ α ρ
ρ β
β ρ βα ρ
ρ β
β ρ βρ α ρ ρ
ρ
−
=
−− −
=
−
⎧ ⎫⎡ ⎤⎡ ⎤+⎪ ⎪⎣ ⎦⎢ ⎥⎨ ⎬−⎢ ⎥⎪ ⎪⎣ ⎦⎩ ⎭
⎡ ⎤+⎣ ⎦+−
⎡ ⎤+∑ ⎣ ⎦≤ +
∑
min
min min
2
21 maxax
22max2
max 2max
1
emm
mτ
τ τ
ρβ
β ρ βα ρ
ρ β
−
=
⎡ ⎤ ⎧ ⎫⎡ ⎤⎪ ⎪⎢ ⎥ ⎨ ⎬⎢ ⎥−⎢ ⎥ ⎣ ⎦⎪ ⎪⎣ ⎦ ⎩ ⎭⎡ ⎤+⎣ ⎦+
−
∑
1min min min min
1 min
min
min min
1
22max( 1) max max2
max 1 max max 2max
max
22max2
max 2max
1 1
( )11
e
me
mm e
m
mbV kτ τ τ ττ
τ
τ
τ τ
β ρ β ρ ρρ α ρ ρ
ρ βρ
β ρ βα ρ
ρ β
−
=
−
−
⎡ ⎤⎡ ⎤ ⎡ ⎤⎢ ⎥−⎢ ⎥ ⎢ ⎥⎡ ⎤⎡ ⎤+∑ ⎢ ⎥⎣ ⎦ ⎣ ⎦ ⎣ ⎦⎢ ⎥≤ + ⎢ ⎥⎡ ⎤−⎢ ⎥⎣ ⎦ ⎢ ⎥− ⎢ ⎥⎢ ⎥⎣ ⎦⎣ ⎦
⎡ ⎤+⎣ ⎦+−
226
1min min min min
1
min
min min
22 ( 1)max2 max max
max 1 max 2maxmax
22max2
max 2max
( )1
.
meem
m
mbV k
τ τ τ ττ
τ
τ τ
β ρ β ρ ρρ α ρρρ β
β ρ βα ρ
ρ β
−
=
−⎡ ⎤⎡ ⎤+∑ ⎡ ⎤−⎣ ⎦⎢ ⎥≤ + ⎢ ⎥−−⎢ ⎥ ⎣ ⎦⎣ ⎦⎡ ⎤+⎣ ⎦+
−
It follows from the above that 1( )e
emV k − is bounded. Hence, using inequality (C.10) and the
fact that emτ → ∞ , we get
( )21minlim [ ( )] ( ) lim ( ) 0
e ee e em me e
e e em m m
k kP k k V kλ θ−
→∞ →∞≤ ≤ .
It follows that
21minlim [ ( )] lim ( ) 0
e ee em me e
e em m
k kP k kλ θ−
→∞ →∞≤ .
Based on Assumption 5.3, there exists an integer 0N > and 0 m M< < < ∞ such that, for all
k k≥ , we have
1
( ) ( )k N
T
i k
mI i i MIφ φ+
= +
≤ ≤∑ . (C.12)
Consequently, the covariance matrix ( )P ⋅ satisfies [103]:
11 max
max maxmax min
1[ ( )] ( ) , 1
NN N
MP k I P k I k k Nm
ρρ λρ ρ
−
−⎡ ⎤+ ≤ ≤ ∀ ≥ +⎢ ⎥−⎣ ⎦
.
It follows from the above that
1 1 maxmin max max
max
( ) [ ( )] , 1
N NN
Mm I P k P k I k k Nρρ ρ λρ
− −⎡ ⎤≤ ≤ + ∀ ≥ +⎢ ⎥−⎣ ⎦
,
which implies that
1min min[ ( )] , NP k m k k Nλ ρ− ≥ ∀ > + (C.13)
227
1 1 maxmax max max
max
[ ( )] [ ( )] , .1
NN
MP k P k k k Nρλ ρ λρ
− −⎡ ⎤≤ + ∀ > +⎢ ⎥−⎣ ⎦
(C.14)
Therefore, 1minlim [ ( )]
e eme
em
kP kλ −
→∞ is bounded away from zero as indicated in (C.13) which yields
2lim ( ) 0e eme
em
kkθ
→∞= . Hence, the adaptation algorithm is capable of driving the parameter
estimates iθ to the desired parameter vector 0θ needed to achieve regulation.
228
BIBLIOGRAPHY
[1] Linda, K., Ono, K., and Yamane, M., “Dynamic Characteristics and Design
Consideration of a Tripad Slider in the Near-contact Regime,” ASME Journal of
Tribology, vol. 124, pp. 600-606, 2002.
[2] Xu, J., and Tsuchiyama, R., “Ultra-low-flying-height Design from the Viewpoint of
Contact Vibration,” Tribology International, vol. 36, pp. 459-466, 2003.
[3] Wu, Z., and Ben Amara, F., "Adaptive Regulation of the Flying Height for 2-DOF
Tripad Slider in Hard Disk Drives," 2005 IEEE International Conference on
Mechatronics and Automation, July 29 - August 1, Niagara Falls, Canada, pp. 160-165.
[4] Wu, Z., and Ben Amara, F., “Flying Height Control for a 2-DOF Tripad Slider in Hard
Disk Drives with Switched Regulators," 2005 ASME International Mechanical
Engineering Congress and Exposition, November 5 - 11, Orlando, Florida.
[5] Liberzon, D., and Morse, A.S., “Basic Problems in Stability and Design of Switched
Systems,” IEEE Control System Magazine, vol. 19, no. 5, pp. 59-70, 1999.
[6] Blondel, V.D., and Tsitsiklis, J.N., “Complexity of Stability and Controllability of
Elementary Hybrid System,” Automatica, vol. 35, pp. 479-489, 1999.
[7] Goncalves, J.M., Megretski, A., and Dahleh M.A., “Global Stability of Relay Feedback
Systems,” IEEE Transactions on Automatic Control, Vol. 46, No. 4, pp. 550-562, 2001.
[8] Murphey, T.D., “On Multiple Model Control for Multiple Contact Systems,” Automatica,
vol. 44, pp.451-458, 2008.
[9] Kim, D.K., Park, P.G., and Ko, J.W., “Output Feedback H∞ Control of Systems over
Communication Networks Using a Deterministic Switching System Approach,”
229
Automatica, vol. 40, pp. 1205-1212, 2004.
[10] Khalil, H.K., Nolinear Systems, 2nd edition, Upper Saddle River, NJ: Prentice-Hall, 1996.
[11] Narendra, K.S., Driollet, O.A., Feiler, M., and George, K., “Adaptive Control Using
Multiple Models, Switching and Tuning,” International Journal of Adaptive Control and
Signal Processing, vol. 17, no. 2, pp. 87–102, 2003.
[12] Savkin, A.V., Skafidas, E., and Evans, R.J., “Robust Output Feedback Stabilizability via
Controller Switching,” Automatica, vol. 35, no. 1, pp. 69–74, 1999.
[13] Skafidas, E., Evans, R. J., Savkin, A.V., and Petersen, I.R., “Stability Results for
Switched Controller Systems,” Automatica, vol. 35, no. 4, pp. 553–564, 1999.
[14] Ishii, H., and Francis, B.A., “Stabilizing a Linear System by Switching Control with
Dwell Time,” IEEE Transactions on Automatic Control, vol. 47, no. 12, pp. 1962–1973,
2002.
[15] Cronin, B., and Spong, M.W., “Switching Control for Multi-input Cascade Nonlinear
Systems,” Proceedings of the 42nd IEEE Conference on Decision and Control, pp.
4277–4282, 2003.
[16] Kosmatopoulos, E.B., and Ioannou, P.A., “Robust Switching Adaptive Control of Multi-
input Nonlinear Systems,” IEEE Transactions on Automatic Control, vol. 47, no.4, pp.
610–624, 2002.
[17] Zhao, J., and Spong, M.W., “Hybrid Control for Global Stabilization of the Cart-
pendulum System,” Automatica, vol. 37, no. 12, pp. 1941–1951, 2001.
[18] Ezzine, J., and Haddad, A.H., “Error Bounds in the Averaging of Hybrid Systems,”
IEEE Transactions on Automatic Control, vol. 42, no. 12, pp. 1188-1192, 1989.
[19] Pettersson, S., “Synthesis of Switched Linear Systems,” Proceedings of the 42nd IEEE
230
Conference on Decision and Control, pp. 5283-5288, 2003.
[20] Hespanha, J.P., “Logic-based Switching Algorithms in Control,” Ph.D. dissertation,
Dept. of Electrical Engineering, Yale University, 1998.
[21] Ji, Z., Wang, L., and Guo, X., “Design of Switching Sequences for Controllability
Realization of Switched Linear Systems,” Automatica, vol. 43, no.4, pp. 662–668, 2007.
[22] Ooba, T., and Funahashi, Y., “On a Common Quadratic Lyapunov Function for Widely
Distant Systems,” IEEE Transactions on Automatic Control, vol. 42, no. 12, pp. 1697-
1699, 1997.
[23] Shorten, R., Narendra, K.S., and Mason, O., “A Result on Common Quadratic Lyapunov
Functions,” IEEE Transactions on Automatic Control, vol. 48, no. 1, pp. 110-113, 2003.
[24] Dayawansa, W.P., and Martin, C.F., “A Converse Lyapunov Theorem for a Class of
Dynamical Systems which Undergo Switching,” IEEE Transactions on Automatic
Control, vol. 44, no. 4, pp. 751-760, 1999.
[25] Branicky, M.S., “Multiple Lyapunov Functions and Other Analysis Tools for Switched
and Hybrid Systems,” IEEE Transactions on Automatic Control, vol. 43, no. 4, pp. 475-
482, 1998.
[26] Branicky, M.S., “Studies in Hybrid Systems: Modeling, Analysis, and Control,” Ph.D.
dissertation, Mass. Inst. Technol., Cambridge, MA, 1995.
[27] Brayton, R.K., and Tong, C.H., “Constructive Stability and Asymptotic Stability of
Dynamic Systems,” IEEE Transactions on Circuits and Systems, vol. 27, no.11, pp.
1121-1130, 1980.
[28] Polanski, A., “On Absolute Stability Analysis by Polyhedral Lyapunov Functions,”
Automatica, vol. 36, no. 4, pp. 574-578, 2000.
231
[29] Daafouz, J., Riedinger, P., and Iung, C., “Stability Analysis and Control Synthesis for
Switched Systems: A Switched Lyapunov Function Approach,” IEEE Transactions on
Automatic Control, vol. 47, no. 11, pp. 1883-1887, 2002.
[30] Hespanha, J.P., and Morse, A.S., “Switching between Stabilizing Controllers,”
Automatica, vol. 38, pp. 1905-1917, 2002.
[31] Lee, S.H., Kim, T.H., and Lin J.T., “A New Stability Analysis of Switched Systems,”
Automaica, vol. 36, pp. 917-922, 2000.
[32] Blanchini, F., Miani, S., and Savorgnan, C., “Stability Results for Linear Parameter
Varying and Switching Systems,” Automatica, vol. 43, pp. 1817-1823, 2007.
[33] Johansson, M. and Rantzer, A., “Computation of Piecewise Quadratic Lyapunov
Functions for Hybrid Systems,” IEEE Transactions on Automatic Control, vol. 43, no. 4,
pp. 555-559, 1998.
[34] Hassibi, A., and Boyd, S., “Quadratic Stabilization and Control of Piecewise-linear
Systems,” Proceeding of America Control Conference, Philadelphia, Pennsylvania, June,
pp. 3659-3664, 1998.
[35] Goncalves, J.M., Megretski, A. and Dahleh M.A., “Global Analysis of Piecewise Linear
Systems Using Impact Maps and Surface Lyapunov Functions,” IEEE Transactions on
Automatic Control, vol. 48, no. 12, pp. 2089-2106, 2003.
[36] Rodrigues, L., and How, J.P., “Observer-based Control of Piecewise-affine Systems,”
International Journal of Control, vol. 76, no. 5, pp. 459-477, 2003.
[37] Boukas, E.K., and Benzaouia, A., “Stability of Discrete-time Linear Systems with
Markovian Jumping Parameters and Constrained Control,” IEEE Transactions on
Automatic Control, vol. 47, no. 3, pp. 516-521, 2002.
232
[38] Fang, Y., and Loparo, K.A., “Stabilization of Continuous-time Jump Linear Systems,”
IEEE Transactions on Automatic Control, vol. 47, no. 10, pp. 1590-1603, 2002.
[39] Costa, O.L.V., and Tuesta, E.F., “Finite Horizon Quadratic Optimal Control and a
Separation Principle for Markovian Jump Linear System,” IEEE Transactions on
Automatic Control, vol. 48, no. 10, pp. 1836-1842, 2003.
[40] Loparo, K.A., Aslanis, J.T., and Hajek, O., “Analysis of Switching Linear Systems in the
Plane, Part 2, Global Behavior of Trajectories, Controllability and Attainability,”
Journal of Optimization Theory and Application, vol. 52, no. 3, pp. 395-427, 1987.
[41] Ezzine, J., and Haddad, A.H., “Controllability and Observability of Hybrid Systems,”
International Journal of Control, vol. 49, no. 6, pp. 2045-2055, 1989.
[42] Szigeti, F., “ A Differential Algebraic Condition for Controllability and Observability of
Time Varying Linear System,” Proceedings of 31st IEEE Conference on Decision and
Control, Tueson, Arizona, pp. 3088-3090, 1992.
[43] Sun, Z. and Ge, S.S., “Analysis and Synthesis of Switched Linear Control System,”
Automatica, vol. 41, pp. 181-195, 2005.
[44] Xie, G.M.., Zheng, D.Z., and Wang, L., “Controllability of Switched Linear Systems,”
IEEE Transactions on Automatic Control, vol. 47, no. 8, pp. 1401-1405, 2002.
[45] Yang, Z., “An Algebraic Approach towards the Controllability of Controlled Switching
Linear Hybrid Systems,” Automatica, vol. 38, no. 7, pp. 1221-1228, 2002.
[46] Cheng, D., Guo, L., Lin, Y. and Wang, Y., “Stabilization of Switched Linear Systems,”
IEEE Transactions on Automatic Control, vol. 50, no. 5, pp. 661-666, 2005.
[47] Boukas, E.K., Stochastic Switching Systems: Analysis and Design, Birkhauser, 2006.
[48] Liberzon, D., Switching in Systems and Control, Birkhauser: Boston, 2003.
233
[49] Sun, Z., and Ge, S.S., Switched Linear Systems-Control and Design, Springer: London,
2005.
[50] Davison, E.J., “The Robust Control of a Servomechanism Problem for Linear Time-
invariant Multivariable Systems,” IEEE Transactions on Automatic Control, vol. 21, pp.
25-34, 1976.
[51] Francis, B.A., and Wonham, W.M., “The Internal Model Principle of Control Theory,”
Automatica, vol. 12, pp. 457-465, 1976.
[52] Francis, B.A., “The Linear Multivariable Regulator Problem,” SIAM Journal of Control
and Optimization, vol. 15, no. 3, pp. 486-505, 1977.
[53] Isidori, A., and Byrnes, C.I., “Output Regulation of Nonlinear Systems,” IEEE
Transactions on Automatic Control, vol. 48, no. 5, pp. 880-885, 1990.
[54] Byrnes, C.I., Priscoli, F.D., Isidori, A., and Kang, W., “Structurally Stable Output
Regulation of Nonlinear Systems,” Automatica, vol. 33, pp. 369-385, 1997.
[55] Huang, J., and Rugh, W.J., “Stabilization on Zero-error Manifolds and Nonlinear
Servomechanism Problem,” IEEE Transactions on Automatic Control, vol. 37, no. 7, pp.
1009–1013, 1992.
[56] Huang, J., and Rugh, W.J., “An Approximation Method for the Nonlinear
Servomechanism Problem,” IEEE Transactions on Automatic Control, vol. 37, no. 9, pp.
1395–1398, 1992.
[57] Huang, J., “Asymptotic Tracking of a Nonminimum Phase Nonlinear System with
Nonhyperbolic Zero Dynamics,” IEEE Transactions on Automatic Control, vol. 45, no.
3, pp. 542–546, 2000.
[58] Krener, A.J., “The Construction of Optimal Linear and Nonlinear Regulators,” Systems,
234
Models and Feedback, Theory and Applications, A. Isidori and T. J. Tarn, Eds.
Cambridge, MA: Birkhauser, 1992.
[59] Marconi, L., Isidori, A., and Serrani, A., “Autonomous Vertical Landing on an
Oscillating Platform: An Internal-model Based Approach,” Automatica, vol. 38, pp. 21–
32, 2001.
[60] Serrani, A. and Isidori, A., “Global Robust Output Regulation for a Class of Nonlinear
Systems,” System Control Letters, vol. 39, pp. 133–139, 2000.
[61] Devasia, S., Paden, B., and Rossi, C., “Exact-output Tracking Theory for Systems with
Parameter Jumps”, International Journal of Control, vol. 67, no. 1, pp. 117-131, 1997.
[62] Devasia, S., Paden, B., and Rossi, C., “Regulator Theory for Switched Systems,”
Proceeding of the 34th Conference on Decision and Control, New Orleans, LA,
December, 1995.
[63] Sakurama, K. and Sugie, T., “Trajectory Tracking Control of Bimodal Piecewise Affine
System”, International Journal of Control, vol. 78, no. 16, pp. 1314-1326, 2005.
[64] Van de Wouw, N. and Pavlov, A., “Tracking and Synchronization for a Class of PWA
Systems,” Automatica, vol. 44, pp. 2909-2915, 2008.
[65] Hespanha, J.P., “Root-mean-square Gains of Switched Linear Systems,” IEEE
Transactions on Automatic Control, vol. 48, pp. 2040–2045, 2003.
[66] Lee, J.W., and Dullerud, G.E., “Optimal Disturbance Attenuation for Discrete-time
Switched and Markovian Jump Linear Systems,” SIAM Journal of Control and
Optimization, vol. 45, no. 4, pp. 1329-1358, 2006.
[67] Lee, J.W., and Khargonekar, P.P., “Optimal Output Regulation for Discrete-time
Switched and Markovian Jump Linear Systems,” SIAM Journal of Control and
235
Optimization, vol. 47, no. 1, pp. 40-72, 2008.
[68] Rodrigues, L. and Boukas, E.K., “Piecewise-linear H∞ Controller Synthesis with
Applications to Inventory Control of Switched Production Systems,” Automatica, vol. 42,
pp. 1245-1254, 2006.
[69] Lin, H., and Antsaklis, P.J., “Hybrid State Feedback Stabilization with 2l Performance
for Discrete-time Switched Linear Systems,” International Journal of Control, vol. 81,
no.7, pp. 1114-1124, 2008.
[70] Geromel, J.C., Colaneri, P., and Bolzern, P., “Dynamic Output Feedback Control of
Switched Linear Systems,” IEEE Transactions on Automatic Control, vol. 53, no. 3, pp.
720-733, 2008.
[71] Sun, X., Zhao, J., and Hill, D.J., “Stability and L2-gain Analysis for Switched Delay
Systems: A Delay-dependent Method,” Automatica, vol. 42, no. 10, pp. 1769–1774,
2006.
[72] Xie, D., Wang, L., Hao, F., and Xie, G., “LMI Approach to L2-gain Analysis and
Control Synthesis of Uncertain Switched Systems,” IEE Proceedings on Control Theory
& Applications, vol. 151, no. 1, pp. 21–28, 2004.
[73] Zhang, L., Shi, P., Wang, C., and Gao, H., “Robust H∞ Filtering for Switched Linear
Discrete-time Systems with Polytopic Uncertainties,” International Journal of Adaptive
Control and Signal Processing, vol. 20, no. 6, pp. 291–304, 2006.
[74] Zhang, L., Shi, P., Boukas, E.K., and Wang, C., “ H∞ Control of Switched Linear
Discrete-time Systems with Polytopic Uncertainties,” Optimal Control, Applications and
Methods, vol. 27, no. 5, pp. 273–291, 2006.
[75] Cuzzola, F.A., and Morari, M., “An LMI Approach for H∞ Analysis and Control of
236
Discrete-time Piecewise Affine Systems,” International Journal of Control, vol. 75, no.
16, pp. 1293–1301, 2002.
[76] Marino, R., and Tomei, P., “Output Regulation for Linear Systems via Adaptive Internal
Model,” IEEE Transactions on Automatic Control, vol. 48, no. 12, pp. 2199–2202, 2003.
[77] Ben Amara, F., Kabamba, P.T., and Ulsoy, A.G., “Adaptive Band-limited Disturbance
Rejection in Linear Discrete-time Systems,” Mathematical Problems in Engineering, vol.
1, no. 2, pp. 139-177, 1995.
[78] Ben Amara, F., Kabamba, P.T., and Ulsoy, A.G., “Adaptive Sinusoidal Disturbance
Rejection in Linear Discrete-time Systems - Part I: Theory,” ASME Journal of Dynamic
Systems, Measurement, and Control, vol. 121, no. 4, pp. 648-654, 1999.
[79] Landau, I.D., Constantinescu, A., and Rey, D., “Adaptive Narrow Band Disturbance
Rejection Applied to an Active Suspension - An Internal Model Principle Approach,”
Automatica, vol. 41, pp. 563-574, 2005.
[80] Bodson, M., and Douglas, S.C., “Adaptive Algorithms for the Rejection of Sinusoidal
Disturbances with Unknown Frequency,” Automatica, vol. 33, no. 12, pp. 2213–2221,
1997.
[81] Tomizuka, M., Chew, K.K., and Yang, W.C., “Disturbance Rejection through an
External Model,” ASME Journal of Dynamic Systems, Measurement, and Control, vol.
112, no. 4, pp. 559-564, 1990.
[82] Serrani, A., “Rejection of Harmonic Disturbances at the Controller Input via Hybrid
Adaptive External Models,” Automatica, vol. 42, pp. 1977-1985, 2006.
[83] Serrani, A., and Isidori, A., and Marconi, L., “Semiglobal Nonlinear Output Regulation
with Adaptive Internal Model,” IEEE Transactions on Automatic Control, vol. 46, no. 8,
237
pp. 1178-1194, 2001.
[84] Ye, X.D. and Huang, J., “Decentralized Adaptive Output Regulation for a Class of
Large-scale Nonlinear Systems,” IEEE Transactions on Automatic Control, vol. 48, pp.
276–281, 2001.
[85] Priscoli, F.D., Marconi, L., and Isidori, A., “A New Approach to Adaptive Nonlinear
Regulation,” SIAM Journal of Control and Optimization, vol. 45, no. 3, pp. 829-855,
2006.
[86] Bastin, G., and Gevers, M.R., “Stable Adaptive Observers for Nonlinear Time-varying
Systems,” IEEE Transactions on Automatic Control, vol. 33, pp. 650–658, 1988.
[87] Marino, R., and Tomei, P., “Global Adaptive Observers for Nonlinear Systems via
Filtered Transformations,” IEEE Transactions on Automatic Control, vol. 37, pp. 1239–
1245, 1992.
[88] Byrnes, I., and Isidori, A., “Limit Sets, Zero Dynamics and Internal Models in the
problem of Nonlinear Output Regulation,” IEEE Transactions on Automatic Control, vol.
48, pp. 1712–1723, 2003.
[89] Xi, Z., and Ding, Z., “Global Adaptive Output Regulation of a Class of Nonlinear
Systems with Nonlinear Exosystems,” Automatica, vol. 43, pp. 143-149, 2007.
[90] Ding, Z., “Asymptotic Rejection of Unknown Sinusoidal Disturbances in Nonlinear
Systems,” Automatica, vol. 43, pp. 174-177, 2007.
[91] Youla, D.C., Jabr, H.A., and Bongiorno, J.J., "Modern Wiener-Hopf Design of Optimal
Controllers: Part I," IEEE Transactions on Automatic Control, vol. 21, pp. 3-13, 1976.
[92] Youla, D.C., Jabr, H.A., and Bongiorno, J.J. "Modern Wiener-Hopf Design of Optimal
Controllers: Part II," IEEE Transactions on Automatic Control, vol. 21, pp. 319-338,
238
1976.
[93] Zhou, K., Doyle, J.C., and Glover, K., Robust and Optimal Control. Prentice Hall, 1996.
[94] Boyd, S., Ghaoui, L., Feron, E., and Balakrishnan, V., Linear Matrix Inequalities in
System and Control Theory, SIAM Studies in Applied Mathematics, 1994.
[95] Scherer, C., Gahinet, P., and Chilali, M., “Multiobjective Output-feedback Control via
LMI Optimization,” IEEE Transactions on Automatic Control, vol. 42, no. 7, pp. 896-
911, 1997.
[96] Furlani, E.P., Permanent Magnet and Electromechanical Devices: Materials, Analysis
and Applications, Academic Press, 2001.
[97] Yang, S.M., and Liu, Y.C., “Frequency Domain Control of Flexible Beams with
Piezoelectric Actuator,” ASME Journal of Dynamic Systems, Measurement and Control,
vol. 117, pp.541-546, 1995.
[98] Hwang, C.L., and Chen, Y.M., “Discrete Sliding-mode Tracking Control of High-
Displacement Piezoelectric Actuator Systems,” ASME Journal of Dynamic Systems,
Measurement and Control, vol. 126, pp.721-731, 2004.
[99] Economou, D., Mavroidis, C., Antoniadis, I., and Lee, C., “Maximally Robust Input
Preconditioning for Residual Vibration Suppression Using Low-pass FIR Digital
Filters,” ASME Journal of Dynamic Systems, Measurement and Control, vol. 124, pp.85-
97, 2002.
[100] Huyer, W., and Neumaier, A., “Global Optimization by Multilevel Coordinate Search,”
Journal of Global Optimization, vol. 14, pp.331-355, 1999.
[101] Huyer, W., and Neumaier, A., “MINQ: a Matlab Program for Bound Constrained
Indefinite Quadratic Programming Based on Rank 1 Modifications,”
239
http://www.mat.univie.ac.at /~neum/software/minq.
[102] Goodwin, G.C., and Sin, K.S., Adaptive Filtering Prediction and Control, Prentice-
Hall, 1984.
[103] Canetti, R.M., and Espana, M.D., “Convergence Analysis of the Least-squares
Identification Algorithm with a Variable Forgetting Factor for Time-varying Linear
Systems,” Automatica, vol. 25, pp. 609-612, 1989.