Hydraulic Servo-systems
Mohieddine Mali and Andreas Kroll
Windup in Control
Peter Hippe
Nonlinear H2 /H Constrained Feedback
Control
Murad Abu-Khalaf, Jie Huang
and Frank L. Lewis
Practical Grey-box Process Identification
Torsten Bohlin
Control of Traffic Systems in Buildings
Sandor Markon, Hajime Kita, Hiroshi Kise
and Thomas Bartz-Beielstein
Wind Turbine Control Systems
Fernando D. Bianchi, Hernn De Battista
and Ricardo J. Mantz
Advanced Fuzzy Logic Technologies
in Industrial Applications
Ying Bai, Hanqi Zhuang and Dali Wang
(Eds.)
Practical PID Control
Antonio Visioli
(continued after Index)
Reset Control
Systems
Alfonso Baos
Fac. Informtica
Depto. Informtica y Sistemas
Grupo de Informtica Industrial
Universidad de Murcia
Murcia 30071
Spain
abanos@um.es
Antonio Barreiro
Depto. Ingeniera de Sistemas y Automtica
ETSII
Universidad de Vigo
Campus de Lagoas-Marcosende
Vigo 36310
Spain
abarreiro@uvigo.es
The series Advances in Industrial Control aims to report and encourage technology
transfer in control engineering. The rapid development of control technology has
an impact on all areas of the control discipline. New theory, new controllers, actuators, sensors, new industrial processes, computer methods, new applications, new
philosophies all lead to new challenges. Much of this development work resides in
industrial reports, feasibility study papers and the reports of advanced collaborative
projects. The series offers an opportunity for researchers to present an extended exposition of such new work in all aspects of industrial control for wider and rapid
dissemination.
The name of Isaac Horowitz is usually associated with quantitative feedback theory (QFT), but his first-order reset element (FORE) is probably not so well known.
This Horowitz FORE compensator was a generalisation of the seminal Clegg Integrator (1958) in which integral control was coupled with a reset mechanism to inculcate additional performance and design benefits. These two compensators are the
original inspiration and source for the results and applications reported in this Advances in Industrial Control monograph, Reset Control Systems, by Alfonso Baos
and Antonio Barreiro. In fact, the authors whole reset control research project
seems to have been much motivated by original meetings in Spain with Professor
Horowitz.
The monograph is a comprehensive presentation of reset control systems, covering the fundamental theoretical foundations, design rules and applications experience. The monograph has four chapters on theoretical fundamentals. Opening these
is an invaluable introductory chapter (Chap. 1) that guides the reader through the
historical development of reset control. This particular chapter also sets out the theoretical context of reset systems and identifies their relationship with other related
theoretical fields, for example impulsive and hybrid systems. Chapters. 2 to 4 contain the core theoretical results of the monograph and set out, possibly for the first
time, a complete systems analysis framework for reset control.
Design for reset control is treated in Chap. 5, where the focus is on identifying
the type of systems for which reset control offers good design and performance
benefits. Alongside this; sets of useful compensator tuning rules are presented for
vii
viii
different compensator and system types. The chapter also investigates the properties
of the authors own novel PI + CI reset compensator.
In control engineering, the assessment of usefulness, effectiveness, design ease
and performance gain is through real applications studies. In this respect, the case
studies of Chap. 6 provide a trio of demanding benchmark applications. The authors expose all the steps of trying to accomplish successful reset control system
designs. The applications are from process control (a pilot heat exchanger plant),
control by tele-operation (a remote controlled process application that uses the Internet for communication) and utility process operations (a solar collector field control
problem). This chapter allows both the practitioner and the theoretician to see how
reset control works in practice, and the insights provided will undoubtedly inspire
new developments for the reset control field.
This invaluable and comprehensive study of reset control systems by Professors
Baos and Barreiro joins several other recent Advances in Industrial Control monographs that give more fundamental presentations of new potential applicable topics.
One such monograph is Process Control by Jie Bao and Peter L. Lee (ISBN 978-184628-892-0, 2007) that was received enthusiastically; another is Fractional-order
Systems and Controls by Concepcin A. Monje, YangQuan Chen, Blas M. Vinagre,
Dingy Xue and Vicente Feliu (ISBN 978-1-84996-334-3, 2010), and a third is,
Internet-based Control Systems by Shuang-Hua Yang (ISBN 978-1-84996-358-9,
2011). The industrial control engineer, the control academic or control postgraduate
researcher may find these three fundamentals with applications texts along with,
Reset Control Systems essential volumes for inclusion in a complete control library.
Industrial Control Centre, Glasgow, Scotland, UK
M.J. Grimble
M.A. Johnson
Foreword
As part of the small community of control researchers, often referred to as the QFTers, I first met Alfonso Baos in 1997 at an international QFT conference in Glasgow, Scotland. We kept in touch throughout the years via emails, and in 2008 he
arranged for his PhD student, Joaqun Carrasco, to spend a few months in my research lab. Joaquns thesis focused on reset control, and he wanted to learn about
our activities in that area. Since then, I continued to follow their work and have
come to fully appreciate their contributions to the area of reset control.
My foray into reset control research was somewhat accidental. While working
on challenging industrial robust control problems, we stumbled upon a series of
papers from the 1970s written by the late professor I. Horowitz, the founder of QFT,
and his students. I recall that right away we noticed something unique about these
papers. Horowitz presented nontraditional solutions to overcome the limitations
of linear, time-invariant (LTI) control. Horowitz was already familiar with Cleggintegrator and similar concepts that were introduced a few decades earlier on purely
empirical grounds. Horowitzs results on Clegg-integrator and First-Order Reset Element (FORE) offered the practicing control engineer a systematic and transparent
procedure for using such nonlinear elements in feedback control design. In our own
work, we initially attempted to use these ideas in experiments, and motivated by
successful results, we decided to investigate if one can explain such results theoretically. Orhan Beker, a PhD student of CV Hollot and I, was able to prove for the first
time that reset control can overcome classic LTI performance limitations in a feedback system comprising a plant with an integrator. I believe that these encouraging
results may have spurned other groups, including Baoss and Barreiros, to probe
deeper into this area which falls under the more general subjects of hybrid control
systems and impulsive control.
This book presents a compilation of numerous results derived by the authors and
colleagues. The material is written in self-contained manner and offers the interested
reader an exposition of current state-of-the-art results. The authors present stability
theory in progression starting from reset-times independent results and reset-times
dependent results, to the more general passivity-based results. A whole chapter is
devoted to time-delay systems where LTI limitations are especially harsh and stabilix
Foreword
ity conditions take the form of linear matrix inequalities. Numerical examples are
sprinkled throughout the presentation so to bring life into the theory.
In addition to the books theoretical focus, the authors appear to have made deliberate effort to remain true to the original motivation behind Horowitzs original
work to provide the control engineer with a practical technique for design of reset
control systems in applications. They achieve this goal using describing function
analysis as a design tool, extend the notion of reset to a partial reset compensator
where only some of the compensator states are reset, and introduce a new reset compensator that consists of a PI compensator whose integrator is reset only at a fraction
of Clegg integrator reset times. An entire chapter presents experimental applications
of reset control systems in a temperature control of a heat exchanger, a bilateral teleoperation, and a temperature control of a solar collector field.
In closing, I would like to thank the authors for asking me to write this forward.
This well written book strikes a fine balance between sophisticated theory and engineering design and therefore should be accessible to most readers. In writing this
book, the authors have succeeded in continuing Horowitzs philosophy for feedback
control research.
Amherst, USA
Yossi Chait
Preface
Although the origin of reset control systems goes back to 1958 with the founding
work of Clegg, the subject is still in its infancy. The seminal works of Horowitz
and coworkers in the 1970s were the first attempts to build a synthesis theory for
basic reset compensators such as the Clegg integrator or the FORE (first-order reset
element), but the lack of a control theory, approaching basic questions such as wellposedness and stability, for several decades has put aside of the mainstream further
developments in reset control, both in theory and applications. It is not until the
late 1990s that reset control starts to develop with the works of Chait, Hollot and
coworkers, giving a significative impetus to the field. In the meantime, the area of
hybrid control systems and impulsive control also give numerous results, having
clear connections with reset control. For the last decade, a number of international
groups have been working actively on reset control, and fortunately, reset control
has started to be seen as an attractive control design technique with a significant
potential for practical control applications.
The first direct contact of the authors with reset control was in September 2000,
when Isaac Horowitz was visiting our group in Spain. We were actively working in
nonlinear QFT at that time, that uses linear compensation, and in several fruitful discussions he gave us clear arguments for the many benefits of using nonlinear/reset
compensation to attack the nonlinear tiger and also to overcome fundamental limitations of LTI systems. We started to seriously work on reset control in 2006, and
since then our goal has been to develop a formal theory to cover basic theoretical
aspects, and also to define simple compensator structures and tuning rules for them,
with the focus on some practical problems, including applications in process control
and teleoperation, and in general systems with time-delays.
The first chapter of the book is an introduction to reset control systems, pursuing
two objectives. The first objective is to give a quick and simple description of what
a reset control is, and to provide basic explanations on why and when it is convenient to use this strategy. This objective is covered by the first two sections and is
summarized in this key idea: a reset control is a simple nonlinear control technique
very effective for linear plants subject to fundamental design limitations. The second objective of the chapter is to give a brief survey on the literature on analysis
xi
xii
Preface
and design of reset control systems. The historical perspective begins with the early
ideas on reset control, including the popular Clegg integrator and the first-order reset element (FORE) introduced by Horowitz and coworkers. In addition, the first
series of rigorous results on analysis and design of reset controllers using a statespace description are given, including full reset and partial reset compensators. In
addition, relationships between reset control and the wider field of impulsive and
hybrid control systems will be analyzed from different points of view.
In Chap. 2, a definition of a reset control system, or a reset system in general, is
given. In general, as it is common in impulsive systems, reset systems may exhibit
different types of solutions, in particular having complex patterns such as beating,
deadlock, and Zenoness. In control practice, this type of behavior is considered
pathological and thus several conditions will be given for reset control systems to
be well-posed. On the other hand, important properties of reset systems may be
derived by analyzing the reset instants that correspond to a given initial condition.
These patterns will be also analyzed, and their relationship with the observability
and reachability of the base linear system will be shown.
Chapter 3 is devoted to the stability problem of reset control systems with finitedimensional base systems. The stability problem is addressed from different, complementary points of view: (i) internal or Lyapunov stability, (ii) external or input
output stability with passivity analysis, and (iii) stability by the describing function
method. Internal stability techniques are subdivided into techniques giving rise to
stability conditions that do not depend directly on the reset instants (reset-times independent), or alternatively, are reset-times dependent. The first case is obtained directly using continuous time Lyapunov functions (that gives rise to the so-called H
condition), while the second case (reset-times dependent) requires a discretization
at the after-reset values and a subsequent discrete-time Lyapunov analysis. Then,
the inputoutput L2 stability is studied, and a number of results are presented in
connection with passivity and dissipativity properties of reset feedback loops. Finally, the standard describing function tool is used for approximately predicting the
appearance or absence of oscillations.
Stability of time-delay systems under reset control is approached in Chap. 4.
Since reset control is able to overcome fundamental limitations, and time-delay
is one source of such limitations, then it is of great interest to study the problem
of delayed reset systems. The stability is addressed by choosing an appropriate
LyapunovKrasovskii functional, and by imposing that the functional should decrease in the continuous and reset modes. The resulting conditions take the form of
linear matrix inequalities, and, depending on the chosen functional, these LMIs can
be delay-dependent or delay-independent. In both cases, those LMIs, derived from
time-domain stability conditions, are translated into equivalent frequency-domain
conditions by means of adequate tools, like the KalmanYakubovichPopov lemma,
or passivity techniques. From the latter frequency-domain conditions, useful interpretations are exhibited regarding the achieved robustness, in terms of scaled smallgain or positive realness of certain subsystems. Finally, several examples illustrate
the application of the stability conditions, showing the potentials of reset control
when applied to time-delay systems.
Preface
xiii
Alfonso Baos
Antonio Barreiro
Contents
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 What Is Reset Control? . . . . . . . . . . . . . . . . . . . . . . .
1.2 Why Reset Control? . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.1 Overcoming Fundamental Limitations in the Time Domain
1.2.2 Overcoming Limitations in the Frequency Domain . . . . .
1.3 Early Ideas on Reset Control . . . . . . . . . . . . . . . . . . . .
1.3.1 Reset Control with the Clegg Integrator . . . . . . . . . . .
1.3.2 Reset Control with the First-Order-Reset-Element (FORE)
1.4 First General Approaches to Analysis of Reset Control Systems . .
1.4.1 General Setup and Asymptotic Stability . . . . . . . . . . .
1.5 Reset Systems as Impulsive Systems . . . . . . . . . . . . . . . .
1.5.1 Theory of Impulsive Differential Equations . . . . . . . . .
1.5.2 Impulsive Control Theory . . . . . . . . . . . . . . . . . .
1.5.3 Impulsive Dynamical Systems . . . . . . . . . . . . . . . .
1.6 Reset Systems as Hybrid Systems . . . . . . . . . . . . . . . . . .
1.7 Other Recent Results on Reset Control . . . . . . . . . . . . . . .
1.7.1 Design Based on L2 and H2 Performance . . . . . . . . .
1.7.2 Design Based on L2 Gain and Nonzero References . . . .
1.7.3 Design Based on Fixed Reset Instants tk . . . . . . . . . .
1.8 Preview of the Chapters . . . . . . . . . . . . . . . . . . . . . . .
1
1
8
9
11
20
20
23
25
25
27
29
31
34
37
43
44
47
48
50
57
57
59
60
63
64
67
73
74
76
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
xv
xvi
Contents
2.3.3 dim(MR ) 3 . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Reset Control Systems with Exogenous Inputs . . . . . . . . . . .
2.4.1 A Well-posed Reset Control System with Exogenous Input
2.4.2 A Reset Control System with Zeno Solutions . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
80
82
85
86
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
93
93
94
94
99
105
113
117
120
123
125
127
128
129
136
139
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
147
147
149
153
155
156
158
159
162
164
168
175
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
181
181
184
188
195
198
202
202
203
205
206
208
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Contents
Application Cases . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1 Temperature Control in a Heat Exchanger . . . . . . . . . . .
6.1.1 Process Model . . . . . . . . . . . . . . . . . . . . .
6.1.2 PI + CI Design . . . . . . . . . . . . . . . . . . . . .
6.2 Teleoperation of a Gantry Crane . . . . . . . . . . . . . . . .
6.2.1 Passive Teleoperation and System Overview . . . . .
6.2.2 Reset Procedure . . . . . . . . . . . . . . . . . . . .
6.2.3 Examples . . . . . . . . . . . . . . . . . . . . . . . .
6.3 Control of Solar Collector Fields . . . . . . . . . . . . . . .
6.3.1 The Solar Collector Field . . . . . . . . . . . . . . .
6.3.2 PI + CI Design . . . . . . . . . . . . . . . . . . . . .
6.3.3 PI + CI Compensator with Variable Reset Percentage
xvii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
211
211
212
216
224
225
228
231
237
238
239
242
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Chapter 1
Introduction
The first equation describes the continuous dynamics or the flow mode. The second equation defines the discrete or impulsive dynamics, also called jump mode
because at the time instants tk (k = 1, 2, . . .) when the error crosses zero (e(tk ) = 0)
the controller state jumps from u(tk ) to u(tk+ ) = 0.
Now let us see what happens in this reset control system, and compare its behavior with the purely linear base system. Let us take the same values as in [18],
K = a = 1 for the controller and P (s) = (s + 1)/(s(s + 0.2)) for the plant. The
A. Baos, A. Barreiro, Reset Control Systems, Advances in Industrial Control,
DOI 10.1007/978-1-4471-2250-0_1, Springer-Verlag London Limited 2012
Introduction
responses y(t) to a unit step reference r(t) applied at t = 1 are plotted in Fig. 1.2.
This figure shows that the base linear system is very underdamped, with a large overshoot, greater than 70%. The reset system response (dotted line in Fig. 1.2) is much
more damped, the main overshoot decreases to 40%, and the remaining oscillations
disappear very quickly.
Figure 1.3 plots the corresponding control signals, showing that the convergence
to zero is much faster for the reset system than for the linear system. This figure
illustrates more clearly the discontinuities or jumps of the reset system from u(tk )
to u(tk+ ) = 0, where tk are the reset times.
The reset times are the instants tk when e(tk ) = 0 or, looking at Fig. 1.2, when
y(tk ) = r(tk ) = 1. This figure shows the improvement achieved with reset control:
an acceptable, bounded overshoot, but with the same speed of response (and small
rise time) than the linear system, much more underdamped. In this way, reset control
Fig. 1.2 Step response y(t) of the reset system (solid) and base linear system (dotted)
Fig. 1.3 Control action u(t) of the reset system (solid) and base linear system (dotted)
appears as a solution for overcoming the typical linear design trade-offs between
bandwidth and robustness, or other competing objectives. We will discuss in the
next section the subject of fundamental limitations and reset control.
An intuitive explanation of the benefits of reset control can be understood looking
at Fig. 1.2. The initial part of the step response of both the base linear system and
the reset system is very fast, which is a good and desirable property. But after the
first time the output reaches the reference, that is y(t1 ) = r(t1 ) = 1, and the behavior
of the linear and reset systems is completely different. The linear system has a lot
of inertia, the linear states are far from the steady state values y = 1, y = 0, u = 0
and consequently a large overshoot appears.
On the other hand, the reset controller performs a strong resetting action from
u(t1 ) = 0 to u(t1+ ) = 0. This has the effect of discharging or emptying the controller state so that the step response produces a much smaller overshoot. A similar
favorable effect appears at the second reset time t2 , and so on. The reset system
reaches steady state after the fourth reset time t4 (with a very small error) but the
linear system is still very oscillatory at t = t4 and its settling time is much worse.
Although the previous example shows more advantages in favor of reset control,
several skeptical doubts could arise: (i) It might be that there exists another linear
controller, different from C(s) = 1/(s + 1), that performs equally well or better than
Introduction
the reset controller, and (ii) It might be that, starting from a well-tuned linear control
system, the inclusion of a reset provokes instability.
Actually, these two grounds for concern are justified. Regarding (ii), one can
find examples of stable linear control systems that become unstable if some reset
is applied. This tells us that reset should be designed with great care, if we want to
improve system performance and robustness. Regarding (i), it is true that there exist
powerful design methods for linear controllers that easily find solutions C(s) that
satisfy the given design objectives. In control practice, reset control may perform
better, but it also may perform worse than a well designed linear compensator, and
thus in general reset should be used carefully.
The key point is that there exist some classes of linear plants P (s) that are very
difficult to control with linear controllers and are subject to linearly unsolvable
trade-offs between competing design objectives. These problematic classes of systems include plants containing integrators, right half-plane poles or zeros, or time
delays. The corresponding unsolvable trade-offs are called linear fundamental limitations. These fundamental limitations are well known in the literature (see, e.g.,
[2], [27], [37]) and impose hard restrictions on the system performance. For example, if a system has an open-loop positive zero s = c > 0, then for any linear
controller the closed-loop step response exhibits an undershoot Mu , and (see [27])
Mu (1 )/(exp(cts ) 1) where ts is the settling time into the band [1 , 1 + ].
So we cannot make ts and Mu (good speed and robustness) simultaneously small,
and things get worse as c 0. These limitations are derived assuming linear plants
and controllers, thus in principle it might be possible that a nonlinear controller (like
a reset controller) overcomes the limitation and outperforms any linear solution.
This is the main point that justifies the use of reset control. In the following section,
the relations between fundamental limitations and reset control will be presented in
more detail.
Before that, let us look at some other possible reset controllers in order to gain
insight into certain dynamical aspects of reset systems. This presentation will be
intuitive and informal; the formal treatment of well-posedness, analysis and design
will be addressed later in this book. The basic block for simulation or block diagram
realization of reset systems is presented in Fig. 1.4. In the same way as standard
integrators are basic dynamic blocks for realization of finite-dimensional systems,
the reset systems require two extra inputs apart from the usual input e(t), the signal
that is being integrated. As Fig. 1.4 shows, there is a second input c(t), the reset
condition, that has the effect that the zero crossings of a signal c(t), that is, c(tk ) = 0,
define the reset instants t = tk (k = 1, 2, . . .) and trigger the reset action. The third
input a(t) defines the after-reset value to be applied at the output: u(tk+ ) = a(tk ).
Fig. 1.5 Simulation of a reset integrator: e(t) is the standard input, c(t) gives the reset condition,
and a(t) the after-reset value for the output u(t)
= e(t)
when c(t) = 0,
u(t)
+
(1.3)
u(t ) = a(t) when c(t) = 0,
u(0) = u0 .
If the input signals are sufficiently regular (forget for the moment possible difficulties arising from discontinuities or pathologies in them) then the simulation of
this reset integrator is very easy. To illustrate the role of each signal, Fig. 1.5 plots
a simulation example: the output u(t) integrates the main input (a slow sinusoidal),
but resets the output value at the zero crossings of c(t), to the after-reset values given
at those moments by a(t). The initial condition is u(0) = 0.
Normally, the after-reset value a(t) is 0. Only in special cases is this third input
required, for example, in partial reset systems, that is, when u(tk+ ) = p u(tk ) for
some number p, called the percentage of reset. The realization of this strategy requires the use of the third input, and some easy arrangement for circumventing the
algebraic loop a = pu. When the after-reset input is omitted, the default value a = 0
is assumed and the reset becomes zero reset. Furthermore, if we identify the two in-
Introduction
puts c(t) = e(t), that is, the reset instants appear at the zero crossings of the same
signal that is integrated, then we reach the well-known reset integrator or Clegg
integrator [22]. Figure 1.6 shows the block representing the Clegg integrator.
The Clegg integrator is simply given by
u(t)
= e(t) when e(t) = 0,
(1.4)
u(t + ) = 0 when e(t) = 0
with u(0) = u0 . Figure 1.7 shows the response of the integrator to a sinusoidal input,
and with u0 = 0.
Using Clegg integrators, every finite-dimensional reset system can be built in the
form of a block diagram realization based on this primitive block. For example, the
FORE system in (1.2) can be represented by the block diagram in Fig. 1.8.
Clegg integrators and FOREs are very simple, low-order, particular cases of reset
controllers. A general reset controller with linear base dynamics can be described
by the set of equations:
x(t + ) = A x(t)
when e(t) = 0,
(1.5)
x(0) = x0
where e(t), u(t) are the input and output signals, and x(t) Rn is the state vector.
The base linear system of (1.5) is given by the transfer function C(s) = C(sI
A)1 B + D, defined by the first and third equation, describing the flow mode and
holding almost all the time, when e(t) = 0. The impulsive dynamics, or the jump
mode, is triggered by e(t) = 0. If some of the states are affected by reset and some
are not, we can, without loss of generality, reorder states, for example, non-reset
states first and reset states last. In this way, the reset matrix A takes the form
A = diag(1, . . . , 1, 0, . . . , 0)
n1
(1.6)
n2
Introduction
controllers contains as particular case the subset of linear ones (obtained putting, for
example, c(t) = 1 t, so that reset condition is never met). In this way, reset controllers form a superset, or extension, of linear ones, so the set of achievable design
objectives is obviously larger. A recent rigorous study on performance improvement
by means of reset can be found in [1].
1
,
1 + C(s)P (s)
T (s) =
C(s)P (s)
,
1 + C(s)P (s)
(1.7)
and satisfying S(s) + T (s) = 1. It follows that S(s) is the closed-loop transfer
function from noise n(t) to tracking error e(t), and T (s) is the closed-loop transfer
function from reference r(t) to plant output y(t).
Leaving aside the restrictions arising from actual limitations in the real physical
components (for example, precision and range of sensors or actuators), there are
structural limitations affecting the closed-loop in Fig. 1.9, appearing even in ideal
or nominal conditions. A detailed exposition of linear limitations can be found in
[27], [37]. We present here a selection of results taken from [27, Sect. 8.6].
These restrictions arise from interpolation constraints, that is, from transmission
of poles or zeros from the open-loop transfer functions P (s) or C(s) to the closedloop ones. These roots will be called uncanceled poles or uncanceled zeros. It can
be seen that certain types of uncanceled poles or zeros transmitted to closed-loop
impose severe restriction on certain closed-loop time responses.
Notice that the attempt to cancel a plant pole or zero (with a controller zero or
pole, respectively), when lying in the right complex half-plane, is impossible with
perfect accuracy, and hence totally unacceptable as it leads to instability in practice.
On the other hand, although such approximate cancellation in the left half-plane
10
Introduction
does not produce instability, in [35] it is shown that there is a trade-off between input disturbance rejection (better when not canceling but shifting the slow undesired
poles) and robustness (which is better when canceling and not shifting the poles).
According to [27] (Sect. 8.6), we have the following interpolation constraints in
the sensitivity functions:
(i) The complementary sensitivity T (s) has a zero at all uncanceled zeros of P (s).
(ii) The sensitivity S(s) has a zero at all uncanceled poles of P (s).
These two basic properties can be particularized to open-loop integrators, openloop general poles or zeros, or imaginary-axis poles and zeros. For concreteness, let
us see what happens with open-loop general poles and zeros, reproducing part of
Lemma 8.3 in [27]:
Proposition 1.1 Consider the feedback system in Fig. 1.9, with all the closed-loop
poles to the left of < 0. Assume that the controller has at least one pole at the
origin. For an uncanceled plant zero z0 or plant pole 0 satisfying Re(z0 ) > or
Re(0 ) > , and for a unit reference step, we have the following:
(i) 0 e(t)ez0 t dt = z10 ;
(ii) 0 e(t)e0 t dt = 0;
(iii) 0 y(t)ez0 t dt = 0.
The following observations are consequences of the previous constraints [27].
First, if < z0 < 0, that is, z0 is an uncanceled minimum phase zero, then the
error must change sign because the initial value of the integral (i) is positive (since
e(0) = 1) and the total integral (from 0 to ) is negative. In other words, we have
the following linear limitation:
L1 Any uncanceled minimum phase zero z0 < 0, to the right of the closed-loop
poles, implies overshoot in the step response.
A similar discussion in the case z0 > 0 leads to the following limitation:
L2 Any uncanceled non-minimum phase zero gives rise to undershoot in the step
response and, if z0 > 0 is small, to a positive large value of the error integral (i).
Finally, from (ii):
L3 Any uncanceled pole 0 , to the right of the closed-loop poles, must produce
overshoot in the step response.
Using the previous limitations,
introducing
t a suitably defined settling time ts , partitioning the integral in (i) as 0 = 0s + ts , and manipulating bounds, more quantitative results are reached (see Lemmas 4.2 and 4.3 in [27] or Proposition 3 in
[35]) showing that there are restrictions affecting the overshoot Mp (or undershoot
Mu ) and the settling time ts . In other words, if we make good ts (small ts and fast
response) then the overshoot Mp (or undershoot Mu ) becomes bad, and vice versa.
It can be shown that reset control systems provide a very simple mechanism
for overcoming the previous linear limitations. The results that follow are based
11
on [17], where a simple example is presented, showing that reset control does not
suffer from the time-domain limitations of linear controllers. So, consider as in [17]
a linear feedback system of Fig. 1.9 with a plant:
1
P (s) =
s
and a controller C(s) that stabilizes the closed loop. It is known that the tracking
error e(t) to a unit step reference r(t) = 1 t 0 satisfies:
1
e(t) dt =
,
K
v
0
where the velocity gain is given by Kv = lims0 sC(s)P (s). Now, consider the
definition of rise time tr given by
t
tr = sup : y(t) t [0, ] .
12
Introduction
Fig. 1.10 Example in [17]. Unit step responses: (top) reset control and (bottom) linear control
Likewise, this section introduces their frequency domain counterparts, and shows
how reset controllers can overcome these limitations. The presentation is based on
[2] and [3]. The reader can also consult [27] and, for a more detailed treatment, [37].
The frequency domain limitations, for the typical control system in Fig. 1.9, arise
from the structure of the sensitivities S(s) and T (s) (1.7) that impose certain restrictions on the modulus or phase of these complex-valued functions. Some restrictions
are consequences of the Bodes integral formulas, and some others are related to the
Bodes gainphase relation.
Let us see first the Bodes integral formulas and their effect on performance limitations in the frequency domain [3].
Proposition 1.2 Consider the feedback system in Fig. 1.9. Assume that the loop
transfer function L(s) = C(s)P (s) satisfies
lim sL(s) = 0.
s0
1
satisfies the integral identity
Then the sensitivity function S(s) = 1+L(s)
log |S(i)| d =
pk ,
0
(1.10)
13
Equation (1.10) implies, following [3], that the sensitivity function S(i) has to
distribute is attenuation over different frequencies. Recall that |S(i)| is the closedloop frequency attenuation from noise n(t) to error e(t) (or from disturbance d(t) to
plant input) in Fig. 1.9. So if S(i) is made smaller for some frequencies, then necessarily is has to become larger at other frequencies, in order that the total integral
in (1.10) remains constant.
This means [3] that if disturbance attenuation is improved in one frequency range,
then it will be worse at another. This property is known as the waterbed effect. The
total integral in (1.10) is greater (worse attenuation ability) as the number of right
half-plane poles pk increases, or as their values pk increase, so that fast unstable
poles are worse than slow ones, as intuitively predicted. In the best case, when L(s)
has no poles in the right half-plane, we have:
log |S(i)| d = 0.
(1.11)
The Bodes integral limitation can be given a nice geometrical interpretation [3]
as shown in Fig. 1.11 for the case L(s) = 10/((s + 2)(s 1)). The figure plots
log |S(i)| versus in linear scale. The horizontal line log |S(i)| = 0 is also plotted as a reference. Since L(s) has a right half-plane pole at p = 1, then the total
integral in (1.10) is equal to pk = . Thus, the area of the curve above the
horizontal line log |S(i)| = 0 (positive contribution to the integral) minus the area
below the horizontal line (negative contribution) must be equal to pk = .
This means that if we want to make |S(i)| < 1 at lower frequencies (which
is typically desired for disturbance attenuation) then the appearance of disturbance
amplification |S(i)| > 1 is unavoidable at higher frequencies in order to balance
the integral constraint (1.10) (waterbed effect). Even in the best case (1.11) with
no poles of L(s) in the right half-plane, the negative value of the integral at low
frequencies has to be compensated by a positive value at high frequencies.
If we consider the complementary sensitivity function T (s) = L(s)/(1 + L(s)),
there exists as well a similar Bodes integral formula [3]:
0
1
log |T (i)|
d
=
,
zi
2
(1.12)
where the summation is taken over all the right half-plane zeros of L(s).
Now, an alternative to the Bodes integral constraints, in order to make evident
the fundamental limitations of linear design in the frequency domain, is to consider
a well-known restriction affecting minimum phase systems: the Bodes gainphase
relation. In what follows, an exposition on fundamental limitations derived from
the Bodes gainphase relation is presented based on the results in [2] (see also [3]).
Subsequently, some examples will be given, showing that reset control systems are
able to overcome these limitations and outperform linear controllers.
14
Introduction
Fig. 1.11 Waterbed effect: plot of log |S(i)| versus for L(s) = 10/((s + 2)(s 1))
Proposition 1.3 (Bodes gainphase relation) Assume that the transfer function
G(s) is minimum phase, that is, it does not have poles and zeros in the right halfplane. Then the phase is uniquely given by the gain curve by means of the relation
d log |G(i)|
d log ,
(1.13)
f (, 0 )
arg G(i0 ) =
2 0
d log
where the weighting kernel is
+ 0
2
.
f (, 0 ) = 2 log
0
Notice that the kernel weights the frequencies close to , for if 0 then
f (, 0 ) , and if then f (, 0 ) 0. From this interpretation of
the weight, it can be seen that the relation (1.13) gives rise to the approximate Bodes
gainphase relation, for minimum phase systems:
arg G(i0 )
d log |G(i)|
.
2
d log
(1.14)
d log |G(i)|
d log
(1.15)
If we define
n() :=
15
as the slope of the gain curve, in logarithmic scale, then the approximate Bodes
gainphase relation, for minimum phase systems, says:
and let nc = n(c ) be the crossover slope. Two measures of robust stability are
given by the guaranteed phase margin m and by the crossover slope nc . The negative feedback loop formed around L(s) is more robustly stable as m is larger (60
better than 45) and nc is more negative (20 dB/dec better than 10 dB/dec). Two
relations hold:
16
Introduction
arg L(ic ) + m ,
(1.17)
(1.20)
arg La (ic ) m + nc =: c(m , nc ),
2
is called the design inequality in [2]. If we particularize La (ic ), we typically obtain
restrictions on the crossover frequency c that limit in some way the achievable
bandwidth of the linear control system.
For example, if the plant has a non-minimum phase zero and the all-pass factor
is La (s) = (z s)/(z + s) then, from arg La (ic ) = 2 arctan(c /z), we reach
c(m , nc )
m + nc /2
= z tan
.
(1.21)
c z tan
2
2
Notice that the limitation imposed by the coefficient c(,) becomes more restrictive
as the robustness requirements (m , nc ) are stronger. For example, if (m , nc ) =
(45, 10 dB/dec), then c = /2 1.57, but if (m , nc ) = (60, 20 dB/dec) then
c = /6 0.52. The restriction (1.21) also becomes stronger as the zero approaches
the origin (slow non-minimum phase zeros are bad) [2].
Similar limitations could be derived from unstable open-loop poles or from time
delays, see [2]. Therefore, the fact that a plant contains right half-plane poles or
zeros, or time delays, makes it very difficult to control these types of systems, at
least if linear controllers are used.
17
arg Lr () nr () ,
2
(1.22)
1.62 i38
e
,
(1.23)
that is, arg LCI (i) = 38 and nCI () = 20 dB/dec, which means a valuable
improvement of about 52 in phase lag, compared to the linear integrator, with phase
lag 90.
Consider a very simple example: Let us compare two negative feedback loops,
the first one containing a standard integrator Lm (s) = 1/s and a pure delay La (s) =
esh ; and the second one with a Clegg integrator Lr = LCI and a delay La (s) =
esh as well. In both cases, the modulus slope has to be nc = 20 dB/dec.
Let us particularize (1.19), with m = /3 = 60, having in mind that
arg La (i) = arg eih = h. Thus we reach
c h
(1.24)
(1.25)
Thus, for the same phase margin and delay, the reset system outperforms the
linear one. The bandwidth limitation c h 0.52 is improved to c h 1.43 by a
factor 1.43/0.52 = 2.75 (almost three times larger) provided by reset control.
Going further with this idea, one can think if it is possible to improve the Clegg
integrator, providing also 20 dB/dec but with a phase lag still less negative than
38. One simple idea for achieving this is replacing the reset condition e(t) = 0 by
a new condition u(t) = 0 where u(t) is obtained passing e(t) through some block
Fc (s) providing phase advance. Thus, consider now the Clegg Integrator with advanced reset [24] (or with variable reset band [42]):
x (t) = e,
+
x = 0,
u(t) = 0,
u(t) = 0,
(1.26)
18
Introduction
The particular case when = 0 stays constant recovers the standard Clegg integrator (1.23). If = 0 is held approximately constant over a frequency range then
|LACI | = k/, and we obtain 20 dB/dec as the standard integrator. The key feature is the behavior of the phase, the plot of arg LACI versus reveals that the less
negative phase is obtained for +30 giving arg LACI 25.
Thus, by designing the phase lead block Fc (s) so that it provides +30 along
the frequency range of interest (frequency where gain margin is defined), we will
obtain LACI 1 ei25 , which is the same attenuation as that of an integrator but
with 90 25 = 65 of phase improvement, a very important advantage, regarding
fundamental limitations.
Example 1.1 To conclude this section, let us translate the ideas above into a simple
but quantitative design problem. Let us consider as shown in Fig. 1.12 a one-second
time-delay plant P (s) = es and three possible controllers with adjustable gain: an
integrator, a Clegg integrator, and an advanced Clegg integrator. These are not examples of real reset control systems that work well in practice because among other
reasons the reset action destroys the steady state property for tracking external references. Other real examples of useful practical application of reset will be presented
later in the book. So, consider that we are only interested in the stability of the zero
solution, and how restrictive the trade-off between speed (c ) and robustness (m )
is in each of the three cases.
If the design objective is chosen to be a gain crossover frequency c of, say,
4/9 rad/s = 80/s, then it is easy to see that the corresponding gains for the integrator and Clegg are respectively k = kI = 4/9, k = kCI = (4/9)/1.62, and the
resulting phase margins m are respectively 10 and 62.
19
Fig. 1.13 Bode plots of the delay and three integrators example
On the other hand, to take full advantage of the advanced Clegg, one has to
choose an adequate lead filter Fc (s), for example, a PD: Fc (s) = Kp + Kd s. As
Fc (s) defines the signal so that its zero crossings induce resetting, Fc (s) can be normalized as Fc (s) = 1 + sT . The phase advance is () = arg(1 + j T ). As the
maximum advance is at = 30 = /6, and we would like to exploit this advance
at c (for maximizing the phase margin), the best choice is (c ) = 30 which is
solved by c T = 1/ 3 implying T = 9/(4 3). Having fixed the lead compensator Fc (s), it is easy to obtain the gain kACI = 1.32 achieving c = 4/9 crossover
frequency.
Figure 1.13 plots the resulting open-loop responses of the solutions with a standard integrator, a Clegg, and an advanced Clegg. All the three achieve the same
crossover frequency c = 4/9 = 1.4 rad/s, but the corresponding margins are respectively 10, 62 and 75, showing the advantages provided by (advanced) resetting for overcoming fundamental trade-offs. The corresponding time plots, with a
zero reference and x(0) = 1, are shown in Fig. 1.14.
20
Introduction
Fig. 1.14 Time plots of the delay and three integrators example
21
the considered control system is shown in Fig. 1.15 (top), where we can see a twodegree-of-freedom structure with prefilter F (s) acting on the reference r(t), output
disturbance d(t), and measurement noise n(t). The plant (possibly including linear
factors from the controller) is given by P (s), and G is the controller that could be
linear or nonlinear.
The justification of the use of reset control is presented in [32] as a simple mechanism to alleviate the limitations of linear design. More precisely, the effect of sensor
noise n(t) at the plant input u(t) in Fig. 1.15 (top) is given, in the linear case, by
B(s) :=
G(s)
U (s)
=
,
N (s)
1 + L(s)
where L(s) = G(s)P (s). Typically, L(s) is designed such that |L(j )| 1 over
a sufficiently large frequency range. This may require |G(j )| 1 over a larger
frequency range. As a consequence, large G in the range where L is small (and
1 + L 1) causes, from the above equation, large B and amplification of noise
tending to saturate the plant.
Hence, it is desirable to attenuate L as fast as possible outside the frequency
range where specifications require large L. However, from Bode gainphase relation, a fast attenuation of |L| is accompanied by a very negative phase, which
compromises stability.
These linear limitations justify the use of a nonlinear element whose describing
function has a smaller phase than that of a linear element with the same magnitude
characteristic, and such a nonlinear element is the Clegg integrator. More precisely,
Krishnan and Horowitz propose as nonlinear controller G in Fig. 1.15 (top), the
configuration in Fig. 1.15 (bottom), formed by a linear integrator b/s in parallel
with the Clegg integrator (CI), and a linear compensator G(s). It was well known
that the describing function of the CI is (see Sect. 3.4):
CI (j ) :=
1.62ej 38 1.62 j 52
=
e
,
that is, the same magnitude slope as an integrator, but with phase 38, which
means 52 less than that of the linear integrator. One simple example shows the
benefit of a Clegg integrator [32]: Consider an unforced typical negative feedback
loop with two linear integrators L(s) = (1/s)(1/s). Then, the autonomous response
22
Introduction
is oscillating and does not converge to zero. On the other hand, if either of the
integrators is replaced by a Clegg integrator, that is, L = (1/s) (CI ) or L = (CI )
(1/s), then the system is stable and all initial conditions tend to zero in finite time.
Notice that in order that the sinusoidal describing function CI (j ) is representative of the approximate response of the CI, the input should have a mean value of
zero. For this reason, [32] proposes the parallel configuration in Fig. 1.15 (bottom),
containing a linear integrator b/s.
Another interesting interpretation of the dynamic response of the Clegg integrator
is given by Krishnan and Horowitz in [32], and consists of noticing that it can be
understood as a linear integrator which is affected by an adequate train of corrective
Dirac impulses, as an additional input. More precisely, let 0 t1 < t2 < < tn be
the zero crossings of e(t), the input to the CI. Then x(t), its output, is
t
t
n tk+1
x(t) =
e( ) d =
e( ) d
e( ) d.
0
tn
x(t) =
0
tk+1
e( )
tk
n
k=1 tk
e( ) d , we reach
xk (t tk ) d.
(1.28)
k=1
If x() is a stationary random process, the interval v(t) is itself a stationary random
process. The computations of the statistics of v() as a function of the statistics of
x() have been reported only for a few processes. If x() is a zero mean Gaussian
Markov process with autocorrelation
Rxx ( ) = ae| | ,
23
2ev
,
1 e2v
implying the mean value v = log 2/ and variance v2 = /(12 2 ). Thus for highly
uncorrelated processes ( ) the variance tends to zero, and it seems reasonable
to replace v(t) by v in the integral defining the CI output which leads, among other
results, to
Syx () 1 ej v
=
,
Sxx ()
j
(1.29)
the right-hand side being equal to the random describing function of the CI. These
results on noise analysis are extended to the parallel of CI with an integrator in
Fig. 1.15 (bottom), and finally to the noise response of the whole loop. In this way,
Krishnan and Horowitz present a complete design procedure which clearly reveals
that the optimum nonlinear solution with a Clegg integrator (the feasible controller
that minimizes the rms value of noise at the plant input) outperforms the optimum
solution among only linear controllers.
24
Introduction
three inertias J3 , J2 , J1 connected with flexible shafts. The servo motor drives the
inertia J3 and a tachometer measures the speed of the inertia J1 , resulting in the
fifth-order linear plant:
46083950
.
(s + 1.524)(s 2 + 3.1s + 2820)(s 2 + 3.62s + 9846)
The following specifications were proposed to highlight the trade-offs in linear
design and the improvement achieved with reset control:
P (s) =
25
To conclude this section, let us quote some statements in [21] on the pioneering
work of Horowitz on reset controllers in the 1970s: History shows that Prof. Isaac
Horowitz was often ahead of the curve in his feedback control research . . . Horowitz
motivated their use [reset controllers] by showing that with qualitative design, they
can exhibit better performance trade-offs than those in linear, time-invariant systems.
26
Introduction
arbitrary order and partial reset) were addressed. To deal with these general cases,
the authors in [18] and [21] adopt a modeling setup based on state space instead of
transfer functions.
They showed that a reset control system like the one in Fig. 1.16 can be represented by a closed-loop model based on state space variables in the form:
x (t) = Acl x(t) + Bcl w(t) if x(t)
/ M (t),
+
if x(t) M (t),
x t = AR x(t)
(1.30)
(1.31)
with n 0s. Another interesting proposal in [18] is the definition of the so-called
reset surface M (t):
M (t) = Rn : e(t) = 0, (I AR ) = 0
(1.32)
where the feedback tracking error is
e(t) = r(t) n(t) d(t) yp (t) = w(t) Ccl x(t).
The first condition in the definition of M (t) is e(t) = 0, which shows the dependence on exogenous signals via e(t) or w(t) and also shows the time-varying
nature of M (t) when the exogenous signals are not constant. The second condition,
(I AR ) = 0 is imposed for well-posedness, to avoid the so-called re-resetting,
27
that is, the fact that immediately after a reset, the state vector could satisfy again
the reset condition, implying formally an infinite sequence of resettings to be performed instantaneously, also referred to as beating. Thus, to avoid beating, the term
(I AR ) = 0 is added in the definition of M (t). As a consequence of this definition,
/ M (t).
x(t) M (t) x t +
In addition, the stability of reset control systems is approached, firstly from an
internal or asymptotic point of view, and secondly from an external or inputoutput
point of view. Regarding internal stability, the unforced or autonomous version of
the reset system (1.30) is considered:
x (t) = Acl x(t) if x(t)
/ M,
+
x t = AR x(t) if x(t) M ,
where
(1.33)
M = Rn : Ccl = 0, (I AR ) = 0 .
As a tool for addressing internal stability, a general theorem for arbitrary Lyapunov functions is given, followed by a particularization to quadratic stability.
This condition is called the H condition and is easily testable by using linear
matrix inequalities, or directly by frequency domain plots. When reset affects one
single state (n = 1), the problem is reduced to the search for the single variable
R.
A property that follows immediately from the H condition is that the class of
quadratically stable reset control systems requires their base-linear systems to be
stable, that is, Acl stable. On the other hand, the authors pointed out that reset can
stabilize an unstable base-linear system, and showed it with an example. Some other
examples show that there exist reset control systems that are stable but not quadratically stable. In spite of these examples, it is noticed that the class of quadratically
stable systems is rich and contains many common and practical cases of reset control applications. Furthermore, it can be seen that the H condition can be easily
incorporated into the Horowitz design scheme introduced in the previous subsection.
The H condition has been also proven to be useful in establishing other interesting properties of reset control systems, for example regarding some type of
input-state stability. There are an important number of results by Chait, Hollot, and
coworkers regarding other aspects of performance in the steady state and transient
responses, that cannot be included here, for brevity. The interested reader can find
this information in [18], [21], and in the references therein.
28
Introduction
feedback control systems with resetting applied at the zero-crossings of the tracking
error signal. At the same time, the field of reset control systems, as previously defined, could be classified as a subarea of the much wider field of impulsive systems.
This term of impulsive systems is used in different areas of applied mathematics
and dynamical systems theory, and there is an important body of results and a large
literature on the subject. It is out of the scope of this section to present a detailed
survey of results on impulsive systems.
Reset control systems can be considered as a particular class of impulsive systems characterized by the fact that the reset condition is triggered by the zero crossing of the tracking error signal, in a typical feedback loop. As the error signal is
a function of the state for unforced systems (and the forcing external signals can
be described by exosystems) we can say that in reset control systems the reset
events are time-independent and only state-dependent. On the other hand, the research in impulsive theory considers impulses applied at instants depending only
on time or both time-dependent and state-dependent. More generally, as reported in
Bainov and Simeonov (1989) [4], impulsive systems may be classified as (i) systems
with impulses at fixed instants, (ii) systems with impulse effect at variable instants,
and (iii) autonomous systems with impulse effects. Hence, the class (i) has timedependent impulses, the class (ii) has both time- and state-dependent impulses, and
class (iii) which contains reset control systems has only state-dependent impulses.
Traditionally, most of the research effort has been dedicated to cases (i) and (ii), and
thus case (iii), which includes reset control, is less developed.
Summarizing, reset control systems form a particular class of impulsive systems with a specific state-dependent mechanism that triggers reset. Furthermore,
the field of reset control considers implicitly as the main motivation the idea that
the reset action is an effective way for overcoming fundamental limitations in linear
feedback. This motivation regarding control limitations is not present in the general literature on impulsive control. In spite of the previous differences between the
points of view of reset control and impulsive systems, there are strong and deep
connections between the concrete problems addressed by both fields. Thus, the research on reset control can profit from employing results on impulsive systems as
addressed in the corresponding areas of applied mathematics and mathematical control theory.
Having this is mind, we include here three subsections based on three relevant
monographs in the field: Theory of impulsive differential equations by Lakshmikantham, Bainov, and Simeonov [33], Impulsive control theory by Yang [43], and Impulsive and hybrid dynamical systems by Haddad, Chellaboina, and Nersesov [30].
The following comments and discussions on these three books are focused only on
aspects regarding specifically the class of reset control systems, that is, we discuss
how reset control can be treated within the framework presented in the books, and
how general results could be particularized, if possible, to reset control.
29
lim k (x) =
for all x. Figure 1.18 shows a possible time response of one such a system. Notice that solutions may hit the same surface k (x) several times and also present
complex behaviors called pulse phenomenon and confluence. In general,
a simplifying assumption is to consider functions k to be continuous over Rn .
30
Introduction
The class (ii) can be extended to a more general class in the form:
x = f (t, x) if h(t, x) = 0,
x = Ik (x) if h(t, x) = 0,
(1.36)
(1.37)
where M is an arbitrary subset in the domain of definition of x Rn . Figure 1.20 plots a possible trajectory, in the one-dimensional case, n = 1 and
x R. Although the analysis of an autonomous system with impulse effects
may appear to be easier than the other cases, in general it has been proven
harder since in general the continuity assumption over functions k cannot be
achieved.
In summary, the classification proposed by [33] is based on the resetting mechanism, and considers three possibilities: only time-dependent reset (class (i), t = tk ),
only state-depending reset (class (iii), x M) and both, time- and state-reset (class
(ii), t = k (x)).
The reset control systems that we are interested in, as depicted in Fig. 1.16 or
defined by (1.30), fit well into the third class (iii). The unforced system (1.30) without exogenous input (w = 0) is directly in the form (1.37). The same is true in the
case of constant (step) exogenous inputs. Furthermore, for the typical control inputs
(step, ramp, sinusoid, etc.) it is always possible to express them as the output of
an appropriately defined system (the so-called exosystem). As a consequence, the
31
zero crossing of the error that defines the surface M (t) in (1.32) can be expressed
as a condition on the system states and on the states of the exosystem. In this way,
the reset condition can be made time-independent so that reset systems belong to the
class (iii) of autonomous systems with impulses.
Most of the results in [33] apply to the cases (i) and (ii). Regarding stability,
which is a basic property required for a control system, the results in [33] for class (i)
systems cannot be applied to reset control since the reset times in reset control are
not prescribed, but are state-dependent. In addition, most of the literature about
state-dependent impulsive systems has been developed for systems which can be
described by functions k with some specific properties, for example, it is usual to
assume that they are continuous over their domain. As it will be seen in detail in
Chap. 2, in general this is not the case for reset control systems.
To conclude this section, the book [33] presents a systematic treatment of impulsive systems, addressing well-posedness and continuity of solutions, stability criteria, comparison results, vector Lyapunov functions, and many other aspects. The
interested reader can find a detailed treatment of these topics in [33], [4], and in the
references therein.
32
Introduction
if t = k (x),
x = f (t, x)
x = U (k, y) if t = k (x),
(I)
(1.38)
y = g(t, x),
where the control input is implemented as sudden jumps in the state, given by the
impulsive control law U (k, y). A type-II impulsive control system is given by:
if t = k (x),
x = f (t, x, u)
(II)
(1.39)
x = U (k, y) if t = k (x),
y = g(t, x, u)
if u = (t, x),
so that there are actually two control laws, the impulsive control law U (k, y) and
the continuous control law u = (t, y). A type-III impulsive control system is given
by:
if t = k (x),
x = f (t, x, u)
(1.40)
(III)
x = jk (x)
if t = k (x),
y = g(t, x, u)
if u = (t, x),
where for this type of system the control u is continuous and what is impulsive is
the plant itself. Most of the results are presented for systems with impulses at fixed
t = tk or variable t = k (x) times with some regularity condition over functions k ,
but not for autonomous systems with impulses, hence they are not applicable to the
reset control systems as in Fig. 1.16 or defined by (1.30).
On the other hand, it is out the scope of this section to report a detailed account
of the results in [43] but, just to give an idea of one among the many approaches
presented, let us see a simple example. The example is taken from Chap. 6 and
illustrates the concept of practical stability and the comparison system.
Given a system of type-II (1.39) and given the pair (, ) with 0 < < , the
system is said to be practically stable when
x0 <
x(t, t0 , x0 ) < ,
where x(t, t0 , x0 ) is the system solution for t t0 starting at x(t0 ) = x0 . Hence, the
concept of practical stability is weaker than that of asymptotic stability and only imposes boundedness on the state evolution. Consider the following two-dimensional
chaotic system [43]:
x = y f (x),
(1.41)
y = x y + sin t,
where
1
f (x) = bx + (a b)(|x + 1| |x 1|),
2
with a = 1.02, b = 0.55, = 1, = 1.015, = 0.15, and = 0.75. This is a
chaotic system, with the chaotic attractor shown in Fig. 1.21.
33
34
Introduction
Fig. 1.22 Partial stability of the impulsively controlled chaotic system in [43]
d < e ,
|1 e |
< v.
1 de
A valid solution is = 1 and (d, ) = (0.06, 1.01). So we can conclude that no
matter how large the initial condition is, (x0 , y0 ) < < , the resetting controls
in (1.42) achieve the practical stability (boundedness) (x(t), y(t)) < = 1 around
the origin. The simulation results in Fig. 1.22 show how the uncontrolled attractor
is destroyed and the controlled solution remains within the specified bounds.
35
analysis and design of impulsive systems. The book also addresses impulsive systems from the point of view of nonnegative and compartmental systems, optimal
control, disturbance rejection, robustness against uncertainties, periodic solutions,
and so on.
In spite of this, again the developed results are not directly applicable to general
reset control systems (the approach is based on a number of regularity conditions
that give some restrictions over the functions k ). Some brief comments about dissipativity theory (modeling, definitions, and basic principles) will be given in the
following. It is out of scope of this subsection to give a detailed account of the
many aspects treated in [30]. However, to give an idea of the powerful technique of
energy-based control, we will include a simple but illustrative example.
An impulsive system is described, in its most general form, by the equations
x (t) = fc (x(t))
if (t, x(t))
/ S,
x(t) = fd (x(t))
if (t, x(t)) S ,
(1.44)
where the resetting set S can be both time- and state-dependent. In order to have
well-posedness of the resetting times, two assumptions are imposed. The first assumption A1 imposes that if a trajectory reaches the closure of S , at a point not
in S , then the trajectory must be directed away from S . The second assumption
A2 imposes that when a trajectory intersects S , it jumps out of S .
The general system (1.44) contains as a particular case the state-dependent reset
systems, obtained replacing the reset condition (t, x(t)) S by certain x(t) Z ,
where Z is the subset of the state-space where reset takes place. In any case, these
are autonomous or unforced systems, and in order to address dissipativity, it is necessary to include input and output signals, giving rise to models in the form:
x (t) = fc (x(t)) + Gc (x(t))uc (t)
if (x(t), uc (t))
/Z,
if (x(t), uc (t)) Z ,
if (x(t), uc (t))
/Z,
if (x(t), uc (t)) Z .
(1.45)
for all T t0 and for all input and output, continuous and discrete, signals
uc , ud , yc , yd , in their corresponding signal spaces. Basically, it is an extension of
the usual dissipation inequality in which the power into the system contains the
usual continuous term, typically sc (uc (t), yc (t)) = yc (t) uc (t), plus discrete energy
packets sd (ud (tk ), yd (tk )) applied at the reset instants tk .
A complete framework for dissipativity of impulsive systems is given, including generalizations of the available and required storage functions, Kalman
YakubovichPopov conditions, stability of feedback interconnections, port-
36
Introduction
controlled Hamiltonian systems, and so on. For brevity, it is not possible to include
more details but, before concluding this subsection, we will include an insightful example, just to give an idea of how powerful and intuitive this approach is. Consider
a second order Lienard plant:
q + f (q(t)) = u(t),
: Rn
(1.47)
and f
being infinitely differentiable, f (q) = 0 if and only if
with q
fj
fi
q = 0, and qj = qi for all i, j . The equations represent a mechanical system with
n particles of unit mass coupled by elastic forces given by f (). Defining p = q,
the
Hamiltonian is obtained as the sum of the kinetic and potential energies
q
1
H (p, q) = p p +
f ( ) d,
(1.48)
2
0
Rn
Rn
where the integral is independent of the path, from the assumptions on f . The proposed feedback controller is a copy or replica of the plant, but endowed with a reset
mechanism. More precisely,
/Z,
qc + g(qc (t) q(t)) = 0 if (q(t), p(t), qc (t), pc (t))
qc (t) = q(t) qc (t)
pc (t) = pc (t)
(1.49)
37
38
Introduction
Fig. 1.24 Example in Haddad, Chellaboina, and Nersesov [30]. Plant position and velocity
x = Ax + Bd d if x C,
x+ = Ar x
if x D,
y = Cx,
(1.56)
39
Fig. 1.25 Example in Haddad, Chellaboina, and Nersesov [30]. Controller position and velocity
used in order to the system solutions to be well defined for forward time. As a result,
the reset system is given by
= 1, x = Ax + Bd d if x C or ,
+ = 0, x+ = Ar x
if x D and ,
(1.58)
y = Cx.
For this formulation of reset control systems, several results about bounded-input
bounded-state and bounded-input bounded-output stability are developed. In particular, it is shown that if the following pair of LMIs
T
A P +P A+F M P B C T
T
< 0,
B P
I 0
(1.59)
C
0 I
T
AR P AR P R M 0
is feasible for some P = P T > 0, F , R 0, and > 0, then there exists some
> 0 small enough such as the reset control system (1.58) is finite gain L2 -stable
from d to y. It has been also shown that even for simple examples the above statement needs some refinement by using piecewise quadratic Lyapunov functions, and
that in general it may be necessary to solve a large number of LMIs to obtain a tight
bound of the gain.
40
Introduction
Fig. 1.26 Example in Haddad, Chellaboina, and Nersesov [30]. Control signal
= 1, x1 = x2 + d, x2 = x1 + x2 if x1 x2 0 or ,
+ = 0, x1 + = x1 , x2 + = 0
if x1 x2 0 and ,
(1.60)
y = x1 ,
0 1
where the flow and jump sets are obtained for M = 1
. Note that FORE has
0
1
a base linear system with transfer function s , and that the closed-loop base sys1
, thus the base closed-loop system is
tem is given by the transfer function s 2 s+1
unstable for 0. In addition, for 2 the dominant poles are real.
It has been shown that for any value of R the L2 -gain from d to y is welldefined and that it tends to zero as tends to infinity (for positive values of ,
the base closed-loop system is unstable, and thus it does not have a well-defined
L2 -gain). Figures 1.281.29 show the time evolution of the closed-loop state of the
reset control system for a rectangular pulse as disturbance, for = 2. In this case,
the obtained L2 -gain from d to y is 0.58. As a result, the reset control system can
beat the base linear control system, having less L2 -gain from d to y and thus having
a better disturbance rejection.
41
Fig. 1.27 Example in Haddad, Chellaboina, and Nersesov [30]. Plant, controller and total energy
It should be pointed out that for higher order systems, and especially partial reset
compensators, the reset condition based on the sign change in the product of the
input and output signals results in solutions of the reset control systems that are different from the reset condition based on the zero crossing of the input (which is the
classical condition used by Clegg and Horowitz), even though the initial condition
of the compensator state is zero. Consider, for example, (Fig. 1.30) a reset compensator given by a parallel connection of an integrator and a Clegg integrator (this is a
compensator structure that has been used, for example, in [32]).
This is a simple partial reset compensator, and may be described by (with reset
condition e(t) = 0)
1pr
x (t) = pr e(t)
x(t + ) = 10 00 x(t)
u = ( 1 1 ) x,
if e(t) = 0,
if e(t) = 0,
(1.61)
where the base linear compensator is an integrator, and pr is a parameter with value
0 pr 1. Note that for pr = 0 the reset compensator corresponds to a linear
42
Introduction
Fig. 1.28 Example 2 in Nesic, Zaccarian, and Teel [44]. Time evolution of states x1 and x2 for a
rectangular pulse disturbance
integrator, and that for pr = 1 it is a CI. On the other hand, using the alternative
reset condition, the reset compensator is given by
1pr
u = ( 1 1 ) x.
if e(t)u(t) 0 or ,
if e(t)u(t) 0 and ,
(1.62)
Figure 1.31 shows a time response of both implementations for a sinusoidal input
and zero initial conditions. Both responses are identical up to the first reset, then
the compensator (1.62) performs several resets every time units, and thus both
responses diverge up to next common reset. Note that due to the partial reset the
signs of the input and output are different for a finite time interval after the first
reset, and in fact there would be deadlock for = 0 with the reset condition based
on the sign change.
43
Fig. 1.29 Example 2 in Nesic, Zaccarian, and Teel [44]. Time evolution of disturbance d and
closed-loop output y = x1
Fig. 1.30 I + CI
compensator
44
Introduction
45
that is, the resets are applied when input and output have opposite signs. Figure 1.33
shows the reset set (in gray) or the set R in the (y, u) plane where reset is applied,
D :
y 0,
u 0,
or
y 0,
u 0,
and the flow set (white area) C or the set for the continuous mode of operation,
C :
y 0,
u 0,
or y 0,
u 0.
Both sets intersect on the lines u = 0 and y = 0 where reset takes precedence over
flow. Analogously to the developments given in [36, 44] and outlined in Sect. 1.6,
the flow set is divided into a number of angular regions C = C1 CN in order
to apply a piecewise quadratic (PWQ) Lyapunov analysis that will be commented
on later. Combining the plant and controller equations as in Fig. 1.32, the model of
the closed-loop system is obtained. This system is described by a base linear
closed loop x = Ax + Bw, z = Cx + Dw and a resetting law x + = AR x, where x
contains the plant and controller states and (A, B, C, D, AR ) are derived from the
plant and controller models.
The reset set D and flow set C for the closed-loop system can be expressed in
terms of the state x and the input w
D:
TR (x, w) 0,
or TR (x, w) 0,
C:
Tf (x, w) 0,
or Tf (x, w) 0,
w(0),
46
Introduction
developed using L2 and H2 analysis. For the L2 analysis, recall that the L2 gain
of the system is given by
=
z2
0<w2 < w2
sup
for any inputoutput pair (w, z) of the system , with initial state x(0) = 0. Thus, it
represents the inputoutput amplification in the sense of the L2 norms w2 , z2 .
Stabilization of the closed-loop reset system in Fig. 1.32 requires a finite gain
< , and optimal performance requires minimizing < with respect to the free parameters in the controller K. Unfortunately, the computation of
the gain of a nonlinear system is extremely difficult, in general. However, a technique for estimating upper bounds of is proposed based on dissipativity
with respect to quadratic Lyapunov storage functions. To this end, quadratic Lyapunov functions V (x) = x P x with P > 0 are postulated, resulting in two LMIs
(continuous and discrete). These two LMIs are easily computable in the two parameters , P . From a bisection search on , checking feasibility w.r.t. P > 0, a procedure for computing an upper bound of is obtained.
In order to improve accuracy in the bounding < and to avoid conservativeness, more general storage functions V (x) can be used. A detailed treatment for
piecewise quadratic (PWQ) Lyapunov functions is given. In this way, each angular
sector in the flow set in Fig. 1.33 (and the corresponding regions in the (w, x)-space)
are assigned to a Lyapunov function Vi (x) = x Pi x with i = 1, . . . , N . The treatment is more involved and the reduction of conservatism requires in some cases to
include strictly proper input weighting filters, but finally, a set of LMIs is obtained
that are applied to representative examples suggesting that are quite accurate for the
L2 gain characterization.
Regarding the H2 gain analysis, the first thing that had to be defined is the interpretation of the H2 norm for nonlinear reset systems. One of the possible interpretations of the H2 norm of linear SISO systems, given by stable strictly proper
transfer functions G(s) = C(sI A)1 B, is
1/2
2
|g(t)| dt
,
G2 =
0
that is, the total energy of g(t), the SISO impulse response of G(s). This impulsive
response can also be written as g(t) = CeAt B, so it can be identified with g(t) =
CeAt x0 , that is, the autonomous response with initial condition x0 = B Rn .
In the MIMO case, a similar interpretation holds based on initial conditions x0
equal to the columns Bi of B, or equal to sums of columns Bi , Bj , . . . of B. Inspired
by this interpretation, the following definition is proposed: the H2 norm of the reset
system in Fig. 1.32 corresponding to an initial value x0 Rn is the total output
energy defined as
1/2
2,x0 =
z (t)z(t) dt
,
0
47
with > 0 a small offset. A complete and detailed H2 analysis is developed based
on standard Lyapunov and PWQ Lyapunov functions and providing computable
LMIs so that, given the initial state x0 and the bound , the feasibility of the LMIs
implies the bound 2,x0 . Iterating with various initial conditions and/or controller parameters, and by a bisection search on , these LMIs provide accurate,
powerful tools for performance optimization on reset control systems.
48
Introduction
0
for some finite 1/. The target L2 tracking criterion is a bound in the form
22 ,
e22 + e
(1.63)
which is shown to imply asymptotic tracking, that is, limt e(t) = 0. Since r(t)
is assumed to be known a priori, it is directly included in the model that in this
way only considers the error response e(t) with respect to bounded disturbances
d(t). Let us briefly see the case of constant r = r0 (the case r(t) = r0 et is also
considered). The closed-loop flow and jump modes are respectively given by
x = Ax + Br0 ,
x+ = AJ x,
and thus the equilibrium state is xe = A1 Br0 . Introducing the coordinate change
x = x xe and considering disturbances d(t), the closed-loop equations take the
form (flow and jump):
e = C x .
x + = AJ x ,
x = Ax + Bd d,
This is a hybrid system from d to e, and it is analyzed by means of the Lyapunov
function V (x) = x P x and the quantity
1
e e + e e d d.
L = V (x) +
It is required that L < 0 along the continuous trajectories of the system (state in
the flow set) and non-increasing variation V (x) = V (x+ ) V (x) 0 along resets
(state in the jump set). By processing the inequalities for these two conditions, two
computable LMIs are obtained and thus their feasibility is a sufficient condition:
If d = 0, the reset control system is globally asymptotically stable.
If d = 0, then a relation like (1.63) holds in the form
0 + 2 d22 ,
22 x
e22 + e
0 Px
(1.64)
49
are that the reset instants tk are a priori fixed and known, and that the after reset
values of the controller state xr (tk+ ) are not zero, but calculated online for optimal
performance.
For the sake of concreteness, let us follow the presentation in [45]. The system
considered is the typical reset control system like the one in Fig. 1.34. No disturbances are considered (d = 0), and in addition the reference input is assumed to be
constant r(t) = r. The closed-loop reset system is then described by:
x = Ax + Br
+
x tk = M k x + N k r
if t = tk ,
if t = tk ,
with pre-specified reset times t1 < t2 < . Note that in this case the reset control
system analysis is much simpler, indeed it is a particular type of an impulsive system
with impulses at fixed times (see Sect. 1.5.1), and thus a number of results including
well-posedness and stability can be directly applied. For example, internal stability, with zero reference r = 0, can be derived by using an auxiliary discrete-time
(referred to as induced difference system):
+
= Mk+1 eA(tk+1 tk ) x tk+ =: Lk+1 x tk+ .
x tk+1
If tk := tk tk1 < T for some fixed T < and for all k, then the reset
control system is (asymptotically) stable if and only if the induced difference system
is (asymptotically) stable. As a corollary, in the simplest case in which both tk =
and Mk = M are constant, the system is asymptotically stable if and only if all the
eigenvalues of MeA =: L have modulus strictly smaller than one, that is, L is Schur
stable.
The main idea in [45] and [28] is to let free the controller after-reset state xr (tk+ )
(not necessarily equal to zero), and to compute xr (tk+ ) in order to minimize a
quadratic performance index in the form:
tk+1
Jk = e (tk+1 )P0 e(tk+1 ) + e (tk+1 )Q0 e(t
k+1 ) +
e (s)P1 e(s) ds,
tk
where the tracking error is e(t) = r(t) y(t) = r(t) Cx(t). The index Jk can
be seen as a moving index for the kth time interval [tk , tk+1 ] containing two terminal contributions (weighted by P0 , Q0 ) and a distributed contribution (the one
weighted by P1 ). The choice of xr (tk+ ) at the beginning of the interval is the one
that minimizes Jk along the interval [tk , tk+1 ], in a moving horizon optimization
that resembles in part the idea in model-based predictive control (MPC).
For the solution, it is assumed that for any reference values r R there exists a
steady state value xss such that
Axss + Br = 0,
Cxss r = 0,
which, in the case of stable A, amounts to G(0) = 1, that is, unit dc gain of the
closed loop G(s) = C(sI A)1 B. The solution for xr (tk+ ) that minimizes Jk has
the form
1
xr tk+ = k22 k21 (xp xpss ) + xrss ,
50
Introduction
where the subscripts p, r, ss stand for the plant, reset controller, and steady-state,
respectively, and k21 , k21 are matrices that can be obtained from the problem data
(A, B, C, P0 , Q0 , P1 ).
To conclude this section, some simulation results that correspond to a real experimental application (a PZT microactuator positioning stage) are shown. The dynamics of the microactuator plant can be described by the state space model:
x1 (t) = x2 (t),
x2 (t) = a1 x1 (t) a2 x2 (t) + bu(t),
P :
y(t) = x1 (t),
where x1 , x2 are position and velocity of the moving stage and u is the control input,
and the parameters are a1 = 106 , a2 = 1810, and b = 3 106 . On the other hand,
the reset controller has a PI as base linear controller, and a periodic reset action:
if t = tk ,
xr (t) = e(t)
+
xr (tk ) = E1 x1 (t) + E2 x2 (t) + Gr if t = tk ,
R :
51
52
Introduction
Fig. 1.36 Control inputs for the step responses of the example in Zheng et al. [45]
phase lead obtained by reset compensation. This material is based on several published works [7, 9, 11, 13, 14, 19, 20].
Stability of time-delay systems under reset control is approached in Chap. 4.
Since reset control is able to overcome fundamental limitations, and time-delay is
one source of such limitations, it is of great interest to study the problem of delayed
reset systems. The stability is addressed by choosing an appropriate Lyapunov
Krasovskii functional, and by imposing that the functional should decrease in the
continuous and reset modes. The resulting conditions take the form of linear matrix
inequalities, and, depending on the chosen functional, these LMIs can be delaydependent or delay-independent. In both cases, those LMIs, derived from timedomain stability conditions, are translated into equivalent frequency-domain conditions by means of adequate tools, like the KalmanYakubovichPopov lemma, or
passivity techniques. From the latter frequency-domain conditions, useful interpretations are exhibited regarding the achieved robustness in terms of scaled small-gain
or positive realness of certain subsystems. Finally, several examples illustrate the
application of the stability conditions, showing the potentials of reset control when
applied to time-delay systems. The chapter is based on [5, 10, 12, 16].
In Chap. 5, reset compensation has been used to overcome limitations of LTI
compensation. In this chapter, a new reset compensator, referred to as PI + CI, is
introduced. It basically consists of adding a Clegg integrator to a PI compensator
with the goal of improving the closed-loop response by using the nonlinear charac-
References
53
teristic of this element. It turns out that by resetting a percentage of the integral term
in a PI compensator, a significant improvement can be obtained over a well-tuned
PI compensator in some relevant practical cases, such as systems with dominant lag
and integrating systems. The main goal is the development of PI + CI tuning rules
for basic dynamic systems in a wide range of applications, including first and higher
order plus dead time systems. In addition, a number of design improvements such
us the use of a fixed or variable reset band, the integration with QFT, and the use of a
variable reset percentage are discussed. This material is based on [6, 8, 24, 3841].
Finally, in Chap. 6, several practical applications of reset control systems will be
developed, all based on the PI + CI compensator: a temperature control system of
a heat exchanger, a bilateral teleoperation control system, and finally, a temperature
control of a solar collector field. The first two applications have been tested by
means of experiments in plants, while the third has been tested by using a (wellproven) simulator of the field. The chapter is based on [23, 24, 38, 39, 41, 42].
References
1. Aangenent, W.H.T.M., Witvoet, G., Heemels, W.P.M.H., van de Molengraft, M.J.G., Steinbuch, M.: Performance analysis of reset control systems. Int. J. Robust Nonlinear Control
20(11), 12131233 (2009)
2. strm, K.J.: Limitations on control system performance. Eur. J. Control 6, 220 (2000)
3. strm, K.J., Murray, R.M.: Feedback Systems. An Introduction for Scientists and Engineers.
Princeton University Press, Princeton (2008)
4. Bainov, D.D., Simeonov, P.S.: Systems with Impulse Effect: Stability, Theory and Applications. Ellis Horwood, Chichester (1989)
5. Baos, A., Barreiro, A.: Delay-independent stability of reset control systems. In: 32nd Annual
Conference of IEEE Industrial Electronics Society, Paris, France (2006)
6. Baos, A., Vidal, A.: Definition and tuning of PI + CI reset controller. In: European Control
Conference, Kos, Grecia (2007)
7. Baos, A., Carrasco, J., Barreiro, A.: Reset times-dependent stability of reset control system.
In: European Control Conference, Kos, Grecia (2007)
8. Baos, A., Vidal, A.: Design of PC + CI reset compensators for second order plants. In: IEEE
International Symposium on Industrial Electronics, Vigo, Espaa (2007)
9. Baos, A., Carrasco, J., Barreiro, A.: Reset times-dependent stability of reset control with
unstable base systems. In: IEEE International Symposium on Industrial Electronics, Vigo,
Espaa (2007)
10. Baos, A., Barreiro, A.: Delay dependent stability of reset control systems. In: American
Control Conference, New York, EE.UU (2007)
11. Baos, A., Dormido, S., Barreiro, A.: Stability analysis of reset control systems with reset
band. In: 3rd IFAC Conference on Analysis and Design of Hybrid Systems, Zaragoza, Espaa
(2009)
12. Baos, A., Barreiro, A.: Delay-independent stability of reset systems. IEEE Trans. Autom.
Control 54(2), 341346 (2009)
13. Baos, A., Barreiro, A.: Limit cycles analysis in reset systems with reset band. Nonlinear
Anal. Hybrid Syst. (2010). doi:10.1016/j.nahs.2010.07.004
14. Baos, A., Carrasco, J., Barreiro, A.: Reset times-dependent stability of reset control systems.
IEEE Trans. Autom. Control 56(1), pp. 217223 (2011)
15. Baos, A., Mulero, J.I.: On the well-posedness of reset control systems. Technical Report
TR-DIS-1-2011, University of Murcia (2011)
54
Introduction
16. Barreiro, A., Baos, A.: Delay-dependent stability of reset systems. Automatica 46(1), 216
221 (2010)
17. Beker, O., Hollot, C.V., Chait, Y., Han, H.: Plant with integrator: an example of reset control
overcoming limitations of linear systems. IEEE Trans. Autom. Control 46(11), 17971799
(2001)
18. Beker, O., Hollot, C.V., Chait, Y., Han, H.: Fundamental properties of reset control systems.
Automatica 40, 905915 (2004)
19. Carrasco, J., Baos, A., van der Schaft, A.: A passivity approach to reset control of nonlinear
systems. In: 34th Annual Conference of the IEEE Industrial Electronics Society, Orlando,
EE.UU (2008)
20. Carrasco, J., Baos, A., Barreiro, A.: Stability of reset control systems with inputs. In: 16th
IEEE Mediterranean Conference on Control and Automation, Ajaccio, Francia (2008)
21. Chait, Y., Hollot, C.V.: On Horowitzs contributions to reset control. Int. J. Robust Nonlinear
Control 12, 335355 (2002)
22. Clegg, J.C.: A nonlinear integrator for servomechanism. Trans. AIEE, Part II 77, 4142 (1958)
23. Fernndez, A.F., Barreiro, A., Baos, A., Carrasco, J.: Reset control for passive teleoperation
applications in process control. In: 34th Annual Conference of the IEEE Industrial Electronics
Society, Orlando, EE.UU (2008)
24. Fernndez, A.F., Barreiro, A., Baos, A., Carrasco, J.: Reset control for passive bilateral teleoperation. IEEE Trans. Ind. Electron. (2010). doi:10.1109/TIE.2010.2077610
25. Feuer, A., Goodwin, G.C., Salgado, M.: Potential benefits of hybrid control for linear time
invariant plants. In: Proc. Amer. Control Conf., Alburquerque, pp. 27902794 (1997)
26. Goebel, R., Sanfelice, R., Teel, A.R.: Hybrid dynamical systems. IEEE Control Syst. Mag.
(2009)
27. Goodwin, G.C., Graebe, S.F., Salgado, M.E.: Control System Design. Prentice-Hall, Upper
Saddle River (2001)
28. Guo, Y., Wang, Y., Zheng, J., Xie, L.: Stability analysis, design and application of reset control
systems. In: Proceedings of the IEEE Int. Conf. on Control and Automation, Guangzhou,
China, May 30June 1 (2007)
29. Haddad, W.M., Nersesov, S.G., Chellaboina, V.S.: Energy-based control for hybrid portcontrolled Hamiltonian systems. Automatica 39, 14251435 (2003)
30. Haddad, W.M., Chellaboina, V.S., Nersesov, S.G.: Impulsive and Hybrid Dynamical Systems.
Stability, Dissipativity and Control. Princeton University Press, Princeton (2006)
31. Horowitz, I.M., Rosenbaum, P.: Nonlinear design for cost of feedback reduction in systems
with large parameter uncertainty. Int. J. Control 24(6), 9771001 (1975)
32. Krishnan, K.R., Horowitz, I.M.: Synthesis of a nonlinear feedback system with significant
plant-ignorance for prescribed system tolerances. Int. J. Control 19(4), 689706 (1974)
33. Lakshmikantham, V., Bainov, D.D., Simeonov, P.S.: Theory of Impulsive Differential Equations. World Scientific, Singapore (1989)
34. Loquen, T., Tarbouriech, S., Prieur, Ch.: Stability of reset control systems with nonzero reference. In: Proceedings of the 47th IEEE Conf. on Decision and Control, Cancun, Mexico
(2008)
35. Middleton, R.H., Graebe, S.F.: Slow stable open loop poles: to cancel or not to cancel. Automatica 35, 877886 (1999)
36. Neic, D., Zaccarian, L., Teel, A.R.: Stability properties of reset systems. Automatica 44(8),
20192026 (2008)
37. Seron, M., Braslavsky, J.H., Goodwin, G.C.: Fundamental Limitations in Filtering and Control. Springer, London (1997)
38. Vidal, A., Baos, A., Moreno, J.C., Berenguel, M.: PI + CI compensation with variable reset: application on solar collector fields. In: 34th Annual Conference of the IEEE Industrial
Electronics Society, Orlando, EE.UU (2008)
39. Vidal, A., Baos, A.: QFT-based design of PI + CI reset compensator: applications in process control. In: 16th IEEE Mediterranean Conference on Control and Automation, Ajaccio,
Francia (2008)
References
55
40. Vidal, A., Baos, A.: Stability of reset control systems with variable reset: application to
PI + CI compensation. In: European Control Conference, Budapest, Hungary (2009)
41. Vidal, A., Baos, A.: Reset compensation applied on industrial heat exchangers. In: 14th IEEE
International Conference on Emerging Technologies and Factory Automation, Mallorca, Spain
(2009)
42. Vidal, A., Baos, A.: Reset compensation for temperature control: experimental applications
on heat exchangers. Chem. Eng. J. 159(13), 170181 (2010)
43. Yang, T.: Impulsive Control Theory. Lecture Notes in Control and Information Sciences,
vol. 272. Springer, Berlin (2001)
44. Zaccarian, L., Nesic, D., Teel, A.R.: First order reset elements and the Clegg integrator revisited. Proc. Am. Control Conf. 1, 563568 (2005)
45. Zheng, J., Guo, Y., Fu, M., Wang, Y., Xie, L.: Improved reset control design for a PZT positioning stage. In: Proceedings of the 16th IEEE Int. Conf. on Control Applications, Singapore
(2007)
Chapter 2
xp (0) = xp0 ,
y(t) = Cp xp (t)
(2.1)
and the reset compensator R, with state xr , is modeled in principle by the impulsive
differential equation
x r (t) = Ar xr (t) + Br e(t) if e(t) = 0,
(2.2)
(R)
if e(t) = 0,
xr (t + ) = A xr (t)
where xr (0) = xr0 and v(t) = Cr xr (t). Here np is the dimension of the state xp , and
nr is the dimension of the state xr . In addition, xr (t + ), or simply x+
r , is the value
xr (t + ) with 0+ . A is a diagonal matrix with diagonal elements (A )ii = 0 if
the compensator state component (xr )i is to be reset, and (A )ii = 1 otherwise, i =
1, . . . , nr . In general, it is assumed that the first n compensator state components
are not reset, and the last n compensator states are reset or set to zero at the reset
instants t in which the compensator input e(t) is zero. Thus, nr = n + n and A
is given by A = diag(In , On ). For the particular case corresponding to n = 0
A. Baos, A. Barreiro, Reset Control Systems, Advances in Industrial Control,
DOI 10.1007/978-1-4471-2250-0_2, Springer-Verlag London Limited 2012
57
58
and nr = n , that is, if A is the zero matrix, the reset compensator will be referred
to as full reset. Otherwise, it will be referred to as partial reset.
The closed-loop autonomous unforced system is given by e(t) = y(t), u(t) =
v(t), where by definition the closed-loop state of dimension n = np + nr is
xp
.
(2.3)
x=
xr
The result is the closed-loop system (without exogenous inputs)
x (t) = Ax(t)
if x(t)
/ M,
x(t + ) = AR x(t)
if x(t) M
(2.4)
A B C
x
with x(0) = x0 and y(t) = Cx(t), and where x0 = xp0
, A = BrpCp Ap r r , AR =
r0
diag(Inp , A ) = diag(Inp , (In , On )), and C = (Cp 0).
Therefore, to complete the closed-loop system equations, the set M , which will
be referred to later as the reset surface, needs to be defined. Another set, the afterreset surface MR , also plays an important role in the definition of closed-loop system solutions. Note that reset actions occur when the state x(t) contacts the reset surface M at some instant t, that is, x(t) M , and then the state jumps to
AR x(t) MR . In general, the set MR will be defined as
MR = R(AR ) N (C),
(2.5)
where R(X) and N (X) stand for the image and the null subspace of the linear
operator given by the matrix X, respectively. Thus, MR is the set of states x that
belong both to the null space of C (and then the output is y = Cx = 0) and to the
image space of AR (they are the after-reset states). In addition, the set M is defined
as
M = N (C) \ MR
(2.6)
where the after-reset states are removed from the reset surface. Otherwise, an infinite
number of resets would be produced after a reset action; in general, the reset system
(2.4) can exhibit rather complex behaviors that have been referred to as beating or
pulse phenomena in the literature on impulsive systems [1, 10]. They are related to
the fact that reset instants may not be well defined or may not be distinct; in addition, even if the reset instants are well defined and are distinct, they may converge
to a finite number and then the reset system exhibits Zenoness. This topic will be
analyzed in detail in Sect. 2.2.2; however, a common solution in practice to avoid
these behaviors is to use time regularization.
59
60
Proposition 2.1 For any initial condition x0 D , assume that 1 (x0 ) < 2 (x0 ) <
< k (x0 ) < , and k (x0 ) as k , then there exists a unique solution
x(, x0 ) (x() in short) to the reset control system (2.4) that can be written in the
form x(t, x0 ) = W (t, x0 )x0 for t > 0, where the transition matrix W (t, x0 ) is given
by
W (t, x0 ) = eA(tk (x0 )) AR eA(k (x0 )k1 (x0 )) AR AR eA1 (x0 )
(2.8)
61
this is a central problem in reset control where a base system has to be designed for
the reset system to perform crossings with a reset surface.
For the analysis of the crossings with the hyperplane defined by the row vector C,
that is, the instants t > 0 at which CeAt x0 = 0, it is convenient to use the equation
Ce(A+I )t x0 = 0 for some R, obtaining the same results [4]. This is a simple,
but key simplifying result because all the eigenvalues of A can be assumed to have
non-positive real parts without loss of generality, for some properly chosen. In
addition, it has also been shown [4] that if for some initial condition x0 the matrix A
does not have dominant real eigenvalues and (A, C, x0 ) is reduced, the continuous
SkolemPisot problem always has a solution for that initial condition.
By definition, an eigenvalue of A is dominant if it is the rightmost placed eigenvalue of A in the complex plane. In addition, the term CeAt x0 can always be split as
CeAt x0 = y1 (t) + y2 (t), where by definition the dominant term y1 (t) is not identically zero if (C, A, x0 ) is reduced, and in addition y2 (t) tends to zero exponentially
fast as t increases. This means that a reset control system with such a matrix A will
always produce crossings for the initial condition x0 . Note that the set of states x0
for which (A, C, x0 ) is not reduced has a linear subspace structure, and will be re Note that if the modes corresponding to dominant eigenvalues are
ferred to as R.
unobservable then (A, C, x0 ) is not reduced for any x0 Rn , and thus R = Rn ,
but in the case they are observable, (A, C, x0 ) may not be reduced for some initial
condition x0 Rn , and in general R is not the empty set.
In general, if the matrix A Rnn has eigenvalues 1 , 2 , . . . , s such that ki =
index(i ), then it can be expressed as
Q1
s
A = (P1 . . . Ps )J ... =
Pi J (i )Qi ,
(2.9)
i=0
Qs
where J is the Jordan form, and J (i ) the Jordan segment associated to the eigenvalue i . In addition, for a function f : R R such that f (i ), f (i ), . . . ,
f (ki 1) (i ) exist for each i = 1, . . . , s, the value of f (A) can be determined by
using a generalization of the spectral theorem for non-diagonalizable matrices
f (A) =
s k
i 1
f (j ) (i )
i=1 j =0
j!
(A i I )j Gi ,
(2.10)
s
f (i )Gi ,
(2.11)
i=1
where in addition the spectral projectors are simply given by Gi = vi wTi , with vi
and vj being the right-hand eigenvector and the left-hand eigenvector corresponding
to the eigenvalue i , respectively.
62
In the following, it will be shown how a reset control system produces an infinite
number of resets, and that the reset intervals are upper-bounded, if the base system
does not have real dominant eigenvalues.
Proposition 2.2 Consider the reset control system (2.4), where the matrix A does
not have real dominant eigenvalues, and a nonempty closed set of initial conditions
then reset intervals are uniformly upper-bounded, that is, k (x0 ) =
D Rn \ R,
k+1 (x0 ) k (x0 ) < M , k = 1, 2, . . . , for any x0 D and some finite constant
M > 0.
Proof It is assumed, without loss of generality, that A has dominant imaginary
eigenvalues. The case in which A has a pair of dominant eigenvalues will be
considered in the following (with multiplicity not necessarily equal to one), the
more general case of multiple dominant eigenvalues is a bit more involved but
follows a similar reasoning. The eigenvalues of A are ordered in such a way that
its spectrum is (A) = {1 , 2 , . . . , s } = {i1 , i1 , 3 + i3 , . . . , s + is }, and
l < 0, l = 3, . . . , s. In addition, let kl , l = 1, . . . , s, be the index of each eigenvalue.
Then, the output y(t) of the reset control system can be expressed (using (2.9) and
(2.10) for computing eAt ) as
y(t) =
s k
l 1 j l t
t e
l=1 j =0
j!
C(A l I )j Gl x0 ,
(2.12)
2 k
l 1 j il t
t e
l=1 j =0
j!
C(A il I )j Gl x0
(2.13)
and
y2 (t) =
s k
l 1 j l t il t
t e e
l=3 j =0
j!
C(A (l + il )I )j Gl x0 .
(2.14)
(2.15)
63
where P2 (t) = P1 (t), P1,R = 2 Real{P1 }, and P1,I = 2 Imag{P1 }. Now, for instants
tm = m 2
1 , with m N, the value of the dominant term is given simply by y1 (tm ) =
CP1,R (tm )x0 , and since (C, A, x0 ) is reduced for any x0 D , and D is closed,
there must exist some lower bound 1 > 0 such that |y1 (tm )| > 1 x0 , for any m
and x0 D . This can be shown by contradiction, if it were false then it would exist
a sequence of states {x01 , x02 , . . . } x0 such as for any t > 0, CP1,R (t)x0n 0
as n . Thus, since D is closed it must contain a state x0 such as (C, A, x0 ) is
not reduced, which is a contradiction. Also note that for an m1 large enough, the sign
of y1 (tm ) is constant for m > m1 (for example, choosing tm1 as the Cauchys bound
of the roots of the polynomial CP1,R (t)x0 ). It will be assumed that y1 (tm1 ) > 0,
otherwise tm = m , m N, may be chosen. Therefore, it is possible to find a large
enough instant t2 such as 2 < 1 and thus | yy21 (t(tmm )) | <
y2 (tm )
y1 (tm ) ) > 0,
2 x0
1 x0
2
1
64
deadlock, but Zeno solutions may exist in principle). Note that in the reset control
system (2.4), the reset action occurs at the instants t at which the output y(t) is zero.
The set MR is defined as in (2.5), that is, as the set of states that belong both to the
null space of C (and then the output is y = Cx = 0) and to the image space of AR
(they are the after reset states). In addition, the set M will be defined as in (2.6),
where the points that are the after-reset states are removed from the reset surface to
satisfy condition (A2), otherwise an infinite number of resets may be produced after
a reset action, and beating would be present.
Obase =
C
CA
..
.
(2.16)
CAn1
Proof Beating is avoided by defining the after-reset surface MR by (2.6). Thus, the
proof is centered around deadlock. In general, given any initial condition x0 D , the
reset surface M is first contacted at the instant t1 = 1 (x0 ). Then the reset instant
t1 is simply given by t1 = inf{t > 0|x(t, x0 ) M }, and thus is a solution of the
equation CeAt1 x0 = 0. In addition, it must be true that eAt1 x0
/ R(AR ) according
to the definitions of the after-reset and reset surfaces as given by (2.5) and (2.6),
respectively. For simplicity, consider first that the closed-loop state matrix A has
distinct eigenvalues, then the matrix exponential may be computed by using the
CaleyHamilton method, that is,
eAt1 = 0 I + 1 A + + n1 An1 ,
(2.17)
65
1 ,
e2 t1 = 0 + 1 2 + + n1 n1 ,
2
n t1
= 0 + 1 n + + n1 n1
e
n .
(2.18)
(2.19)
0 = T
C
CA
..
.
x0 = et1 U ()Obase x0 ,
(2.20)
CAn1
where U () = V ()1 . Now, if condition (2.16) is satisfied, then the right-hand
side of (2.20) is an analytical function (in fact, a sum of exponentials) that is not
zero for all t 0. As a result, it has no isolated zeros, and then the reset instant
t1 is lower-bounded. In other words, solutions of (2.20) may not be obtained for
arbitrarily small values of the reset instant t1 , and thus deadlock does not occur if
condition (2.16) is satisfied.
In the case in which the eigenvalues of A may be repeated, a similar argument may be applied. Note that eAt1 may be written as the infinite series
i t1i
Ai t1i
D(A) =
i=0 i! . Thus the polynomial D() =
i=0 i! can be factorized as
D() = Q()P () + R(), with R() = 0, or deg(R) < deg(P
) = n. jIn addition, R has degree no greater than n 1, and thus R() = n1
j =0 j . Since
the characteristic polynomial is zero for the eigenvalues of A, D(k ) = R(k ) for
n1
ik t1i
j
k t1 = R( ) =
k = 0, 1, . . . , n 1. And then D(k ) =
k
i=0 i! = e
j =0 j k
for k = 0, 1, . . . , n 1. This can be compactly expressed as V T () = et1 , and the
expression (2.19) is obtained.
Now, if A has r different eigenvalues with respective %
multiplicity order ni , and as
a consequence the characteristic polynomial is p() = ri=1 ( i )ni , then again
there exist unique polynomials Q and R such as D() = Q()P () + R() where
D() = et1 and R = 0 or deg(R) < deg(P ). Here R can be expressed as R() =
n1
i
i=0 i , where the coefficients are unique. Since p and its derivatives up to order
66
nr are zero at i ,
d j D(i ) d j R(i )
=
dj
dj
i = 1, 2, . . . , r, j = 0, 1, . . . , ni 1.
(2.21)
r n
i 1
ei ei t1 ej i ,
(2.22)
i=1 j =0
r n
i 1
j
V ()
W =
ei ej eTi
.
j
i
i=1 j =0
(2.23)
1 0
0
1 0 0
A = 0 1 1 ,
AR = 0 1 0 ,
C = 1 0 0 , (2.24)
0
1 1
0 0 0
where in addition the sets MR and M are defined according to (2.5) and (2.6) as
MR = R(AR ) N (C) = span (0, 1, 0)T ,
and
&
M = N (C) \ MR = span (0, 1, 0)T , (0, 0, 1)T span (0, 1, 0)T .
Note that this reset system cannot be realized as a control reset system with the
structure of Fig. 2.1.
67
In [12], it is correctly argued that for any initial condition x0 = (0, a, 0)T MR ,
the solution is ill-defined since the continuous dynamics makes the system instantly
evolve to the set MR and then the system instantly resets once the reset surface
is reached, making infinitely many resets without leaving the reset surface; in fact,
there is deadlock. Note that this is due to the fact that the after-reset surface MR is
a subset of the unobservable subspace of the linear base system, which is given in
this case by
C
1 0 0
(2.25)
N CA = N 1 0 0 = N (C) MR .
1 0 0
CA2
0
0
1
1 0 0
A = 1 0.2 1 ,
AR = 0 1 0 ,
C= 010
(2.26)
0 1 1
0 0 0
that has an unobservable mode corresponding to a stable polezero cancellation in
s+1
the linear base system, where the plant has a transfer function P (s) = s(s+0.2)
, and
1
the base compensator is C(s) = s+1
(corresponding to a first order reset element
FORE). In addition, the after-reset and reset surfaces are given by MR = R(AR )
N (C) = span{(1, 0, 0)T } and M = N (C) \ MR = span{(1, 0, 0)T , (0, 0, 1)T } \
span{(1, 0, 0)T }, respectively. In this case, the set MR is not a subset of the linear
base system unobservable subspace given by
C
0
1
0
0.2
1 = span (1, 0, 1)T .
N CA = N 1
0.2 0.96 0.2
CA2
(2.27)
As a result, Proposition 2.3 may be used to ensure that the system is well-posed.
Figure 2.2 shows the system solutions corresponding to two initial conditions.
68
Fig. 2.2 System solution for the well-posed reset system example
Proposition 2.4 The reset system (2.4)(2.6) does not have Zeno solutions if MR
N (Obase ) = {0}.
Proof The basic idea of the proof consists of showing that the reset system (2.4)
(2.6), with an initial condition in MR , can only have finite sequences of reset intervals k , k = 1, 2, . . . , m 1 such as m1 < m2 < < 1 = , for some
> 0 arbitrarily small but fixed, and some finite positive integer m. In fact, in the
following it will be shown that at most there will be sequences of length m 1, with
m being the dimension of the after-reset surface MR .
Without loss of generality, it is considered that the plant state equations (2.1) are
given in an observer canonical form, that is,
0 0 0
a0
b0
1 0 0
b1
a1
,
Bp = . ,
Cp = 0 0 . . . 1 ,
Ap = . . .
.
.
.
.
.
.
.
.
. .
.
. .
.
0 0 1 anp 1
bnp 1
(2.28)
then C = (0, 0, . . . , 1, 0, . . . , 0) and thus
0
0
0 0
1
0 ... 0
C
CA 0
0
0 1 anp 1 X . . . X
..
..
..
..
..
..
.. . .
..
..
. .
. .
.
.
.
.
.
Obase =
=.
,
CAnp 1 1 an 1 X X
X
X
X
p
..
..
..
.
.
.
.
.
..
..
..
.. . . . ..
. ..
.
.
.
(2.29)
where X stands for a non (necessarily) zero term.
69
For simplicity, the case of full reset is approached at first. Thus, an after-reset
state x MR is given by
x = (x1 , x2 , . . . , xnp 2 , xnp 1 , 0, 0, . . . , 0)T
(2.30)
for some values x1 , . . . , xnp 1 R, with np being the number of plant states. Thus,
m = np 1 in the case of full reset.
Let us start with the case m = 1, which corresponds to second order plants and
full reset compensators of arbitrary order. In this case, if the reset control system
is well-posed, it is well known that the reset is periodic after reaching the set MR
from any initial condition; thus starting from MR , the reset will be periodic and
Zeno solutions are not possible. In fact, there exists no initial condition in the set
MR that contacts M in an arbitrarily small time since reset intervals are constant.
The case m = 2 is analyzed in the following. Consider an initial condition x0 =
x1 MR , that is, x1 = (x1 , x2 , 0, 0, . . . , 0)T . If the solution x(t, 0, x1 ) contacts the
reset surface MR at time t1 = 1 , thus 1 = 1 , for some 1 > 0 arbitrarily small,
then
12
CA2 x1 + .
(2.31)
2
Since the control system (2.4)(2.6) is well-posed, the right-hand side of (2.31)
is not identically zero for any x1 MR . Now using the special structure given in
(2.29), one obtains
1
(2.32)
0 = x2 + x1 + O 12 ,
2
0 = CeA1 x1 = Cx1 + 1 CAx1 +
where the terms of order 12 and higher maybe be neglected, in principle. Note that
for (2.32) to be satisfied for an arbitrarily small 1 > 0 it must be true that x1 = 0
and x2 = 0.
In addition, the following after-reset state x2 is given by x2 = AR x(t1 , 0, x1 ) =
(x1 + O(12 ), x2 + 1 x1 + O(12 ), 0, 0, . . . , 0)T . Repeating the argument, the solution
x(t, t1 , x2 ) will again contact M at the instant t2 = t1 + 2 . If 2 = 2 1 for
some 2 > 0, then it is verified that
0 = x2 + 1 x1 +
2
x1 + O 12 ,
2
(2.33)
where the properties O(22 ) = O(12 ) for 2 1 and O(k) = O(), for a real constant k, have been used. Now, using (2.32) and (2.33), the result is that given some
1 > 0 arbitrarily small, 2 = 1 + O(12 ) < 0, which is absurd. Thus, by contradiction it is true that 2 > 1 , and thus any initial condition in the set MR that
produces a first reset interval 1 > 0 arbitrarily small, gives a larger second reset
interval 2 > 0. Thus Zeno solutions do not exist in this case either.
In the rest of the proof, the terms O(1m ) are directly neglected for simplicity. For
the case m = 3, consider an initial condition x0 = x1 = (x1 , x2 , x3 , 0, 0, . . . , 0)T
MR . Applying a similar argument, the result is now that if a sequence of decreasing
70
(2.34)
and now, eliminating x1 , x2 , and x3 , after some computation, 3 is given as a function of 1 and 2 by the second order equation
(1 + 2 )2 + (1 + 22 )3 + 32 = 0
(2.35)
having the solutions 3 = 2 < 0 and 3 = (1 + 2 ) < 0, which is a contradiction. Thus, no initial condition in MR can produce a sequence of resets intervals
that converge to zero, and again Zeno solutions do not exist for the case m = 3.
For the general case in which the dimension of MR is m, with initial state x0 =
x1 = (x1 , x2 , . . . , xm , 0, 0, . . . , 0)T , a similar reasoning results in the set of equations
m
k=1
m
i
i=1 k=1
1m1
xk = 0,
(m + 1 k)!
2mi 1ik
xk = 0,
(m + 1 i)!(i k)!
(2.36)
...
m
i
mi (1 + + m1 )ik
m
i=1 k=1
(m + 1 i)!(i k)!
xk = 0,
71
xp
x
T x = T x = xp = z,
x
x
that is,
0n np
T = Inp np
0n np
In n
0np n
0n n
0n n
0np n .
In n
(2.37)
(2.38)
Note that T is a square matrix, all of whose entries are 0 or 1, and in each row
and column of T there is precisely one 1. This means that T is a permutation matrix.
Clearly, such a matrix is unitary, hence orthogonal, so T T = T 1 . The nonsingular
matrix T allows us to rewrite the dynamical system via a similarity transformation
(congruence transformation):
if z(t)
/ M,
z (t) = Az(t)
z t + = A R z(t)
if z(t) M,
(2.39)
where A = T AT T , A R = T AR T T , and C = CT T , and in addition the reset surface is transformed into M = {z Rn : T T z M }. Note that C = CT T = enp +n
so that the output is not changed by the transformation, i.e., y(t) = znp +n (t) as
expected. Henceforth, (2.39) is in full-reset form. Since observability is invariant
under similarity transformations, it is clear that (2.4)(2.6) is well-posed if and only
if (2.39) is well-posed. Finally, to complete the proof it is necessary to show that the
observability matrix has the structure given in (2.29) (using state transformations
x
if needed). This is simply done by considering the substate z1 = xp . In general,
z
w
there always exists a state transformation of z = x1 to w = x1 such that the
state submatrix corresponding to z1 is in the observability staircase form, and thus
the observability matrix has the structure given in (2.29) once unobservable states
are eliminated. This concludes the proof.
0 1
,
1 1
Bp =
1
,
0
Cp = 0 1 ,
x1
x2
, is
(2.40)
72
Br =
1
,
1
x4
, is given by
1 0
,
0 0
(2.41)
that is, the reset control system has a partial reset compensator: it is a parallel connection of an integrator and a Clegg integrator, where only the state x4 is set to zero
at the reset instants.
The closed-loop system is defined by the matrices
1 0 0 0
0 1 1 1
0 1 0 0
1 1 0 0
A=
C= 0100 ,
AR =
0 0 1 0,
0 1 0 0 ,
0 0 0 0
0 1 0 0
(2.42)
and the closed-loop state x = ( x1 x2 x3 x4 )T .
This reset control system is well-posed, since
0
1
0
0 0
0
N (Obase ) = span
,
(2.43)
MR = span , ,
1
0
1
0
0
1
Ar =
0 0
,
0 0
x3
Cr = 1 1 ,
A =
and then MR N (Obase ) = {0}. Following the reasoning given in the proof of
Proposition 2.4, the closed-loop state x can be transformed into a state z in which
the observability matrix has the form (2.29). In this case, this is obtained with z =
( x3 x1 x2 x4 )T . Thus, the initial conditions that produce a crossing in an arbitrarily
small time > 0 are of the form z1 = ( 1 2 0 0 )T , or equivalently,
+T
*
x1 = 0 1 0 .
2
(2.44)
(2.45)
and, according to Proposition 2.4, x2 cannot produce a new crossing in a time less
than or equal to . This fact can be verified by computing solutions to the implicit
equation 0 = CeAt ( 0 1 0 )T for t, given R. The solution is shown in Fig. 2.3,
where t = 1 (( 0 1 0 )T ) is given.
Note that for t to have an arbitrarily small value, an initial condition x1 in the
after-reset surface must have the form (2.44), that is, = /2 in Fig. 2.3. Then, as
a result the state after the first reset x2 has the form (2.45), that is, = +/2 (see
also Fig. 2.3). And then the value of the second reset instant can be obtained from
Fig. 2.3. The result is that if the first reset instant is arbitrarily small, then the second
reset instant is arbitrarily close to 3.15.
73
(2.46)
74
where G1 is the spectral projector given by G1 = v1 wT1 , and v1 and w1 are the right
and the left eigenvectors, respectively. Thus, the set R can be simply computed as
C Re{G1 }
R = N
,
(2.47)
C Im{G1 }
where Re{G1 } and Im{G1 } stands for the real and imaginary parts of G1 , respec
tively. And thus the set of reduced states is given simply by Rn \ R.
The order of the reset control system, and in particular the dimension of the
after-reset surface, is key in the analysis of the reset instants, thus in the following
different cases corresponding to dimensions of the after-reset surface 1, 2, and 3
are treated separately.
M R) = 1
2.3.1 dim(M
In this case, since the after-reset surface has dimension 1, any initial condition
x0 MR can be generated by one vector u, that is, x0 = u for some R. Thus,
simply by using the homogeneity property of the functions tk = k (x0 ), k = 1, 2, . . . ,
it is clear that k (x0 ) = k (u) = k (u), that is, they are all constant functions over
the set MR , as long as (C, A, u) is reduced (otherwise there are no crossings). In
other words, starting from an initial condition in the after-reset surface, the reset
instants are periodic.
For initial conditions that are not elements of the after-reset surface, the first reset
instant is in general different. Using Proposition 2.2, if the reset system has a pair of
dominant complex eigenvalues then the first reset instant is uniformly bounded, that
is, 1 (x0 ) < , for some upper bound > 0 and for any initial condition x0 such
that (C, A, x0 ) is reduced. In the case in which (C, A, x0 ) is not reduced, 1 (x0 ) is
not necessarily bounded, that is, the initial condition may not cross the reset surface.
2.3.1.1 Example
Consider the reset control system (2.4)(2.6) where the plant, with state xp = xx12 ,
is given by the state space model
0 0
1
,
Bp =
,
Cp = 0 1 ,
(2.48)
Ap =
1 1
0
and the reset compensator, with state xr = x3 , is given by
Ar = 1,
Br = 1,
Cr = 1,
A = 0,
(2.49)
that is, the reset compensator is FORE, and the state x3 is set to zero at the reset
instants.
75
0
A = 1
0
0
1
1 0 ,
1 1
1 0 0
AR = 0 1 0 ,
0 0 0
C= 010 ,
(2.50)
(2.51)
Now, the subspace of not reduced states R is computed. The closed-loop state
matrix A has the eigenvalues 1 = 0.12 + j 0.74, 2 = 0.12 j 0.74, and 3 =
1.75. Thus, the reset control system has two complex dominant eigenvalues, with
their spectral projectors being
0.41
R = span 0.55 .
0.72
G2 = G1 .
(2.52)
, which in this example
(2.53)
Since, according to (2.51) and (2.53), every nonzero after-reset state is reduced,
it results in that the reset instants are periodic as discussed above. However, note that
if the initial condition is not an after-reset state, two types of solutions may occur
(see Figs. 2.42.5):
solutions with no crossings if
0.41
x0 = 0.55 ,
0.72
for some real number , and
solutions with a first crossing at a finite time and then an infinite number of periodic crossings.
76
M R) = 2
2.3.2 dim(M
The case of dim(MR ) = 2 is slightly more involved since, as it will be described
below, the resets instants are not periodic in general, and a large variety of solutions
may appear regarding the number and structure of crossings associated to a given
initial condition. By simplicity, the particular case of a third order plant and a full
reset compensator is considered. Without loss of generality, consider an observer
0 0 a0
b0
Ap = 1 0 a1 ,
Bp = b1 ,
0 1 a2
b2
77
Cp = 0 0 1 .
(2.54)
(2.55)
for some values xx12 R2 . Using again the fact that tk = k (x0 ), k = 1, 2, . . . , are
homogeneous, for the systems solutions with an initial condition in the after-reset
surface it is enough to check the reset instants
some subset of R2 , for example,
x1 in
cos
the unit circle. Using polar coordinates x2 = sin , for [0, ), [0, 2),
it is true that
cos
cos
sin
sin
(2.56)
k
= k 0
..
..
.
.
0
0
for k = 1, 2, . . . . As a result the reset instants are functions of the single parameter .
In general, the systems solutions may exhibit no crossings, a finite number of
crossings, or infinitely many crossings. If the initial condition is an after-reset state,
that is, x0 MR , no crossings or a finite number of crossings may occur if some of
the after-reset states are not reduced. Otherwise, an infinite number of crossings is
produced.
2.3.2.1 Example
Consider again the reset control system with a FORE compensator, with parameters
Ar = 1, Br = 1, and Cr = 1, and a plant given by the state space model
0 0 0.35
3
Ap = 1 0 2.40 ,
Bp = 1 ,
Cp = 0 0 1 .
(2.57)
0 1 4.35
0
The reset control system is well-posed since the base linear system is observable.
Using the parameter as in (2.56), the function 1 can be computed by solving the
implicit equation
cos
sin
CeAt1
(2.58)
0 =0
0
78
Fig. 2.6 Reset instants corresponding to after-reset states as a function of the parameter
for t1 = 1 ( ). The result is given in Fig. 2.6, where it can be seen that the mapping
1 has a discontinuity at = , and in addition it results in the reset instants being
uniformly bounded for initial conditions in the after-reset surface. This is due to the
fact that all the after-reset states are reduced, as it will be seen below.
The after-reset surface is given by
0
0 1
,
MR = span
(2.59)
0 0
0
0
and the subspace of not reduced states by
R = N
C Re{G1 }
C Im{G1 }
0.20
0.81
0.31
, 0.47 ,
= span
0.92 0.05
0.16
0.35
(2.60)
thus it can be easily checked that MR R = {}, which means that all the after-reset
states are reduced. In Figs. 2.7 and 2.8, several simulations corresponding to different initial conditions in the after-reset surface MR are shown, including closed-loop
outputs and control inputs. Note that since the after-reset states are reduced, there are
an infinite number of resets corresponding to each initial condition, and as indicated
above, a bound over the reset intervals may be found.
In the case in which the initial condition is not an after-reset surface state, it may
occur that no crossings are produced. This is, in fact, the case for initial conditions in
the set R given by (2.60). Otherwise, an infinite number of crossings are produced
79
Fig. 2.7 Closed-loop output for three initial conditions in the after-reset surface
Fig. 2.8 Control input for three initial conditions in the after-reset surface
in this example. Different simulations for not reduced initial conditions are shown
in Figs. 2.92.10
80
Fig. 2.9 Closed-loop output for three initial conditions (not reduced)
Fig. 2.10 Control input for three initial conditions (not reduced)
M R) 3
2.3.3 dim(M
For higher order reset control systems, the functions k , k = 1, 2, . . . , depend on
more than one parameter, and in general they can be described by the values of 1
over the unit ball in Rm , where m is the number of the after-reset states. In general,
the function 1 can exhibit an infinite number of discontinuities over that domain
81
even in the case m = 2, resulting in very complex patterns of the resets instants as a
function of the initial condition. A detailed analysis of these patterns is still an open
issue. In the following, an example corresponding to a second order partial reset
compensator and a third order plant is analyzed.
2.3.3.1 Example
In this example, a PI + CI compensator is used. It consists of a parallel connection
of a PI compensator and a Clegg integrator, and it is a partial reset compensator.
Consider the PI + CI compensator with a state space realization given by
0 0
1
Ar =
,
Br =
,
0 0
1
1 0
Dr = 0.2,
A =
(2.61)
Cr = 0.2 0.2 ,
0 0
and a third order plant given by
1
Bp = 0 ,
0
Cp = 0 0 1.33 .
(2.62)
1
0
0
0
A= 0
,
0
0
1.33
0
0
0
0
1.33
0
0
and
1
0
AR =
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
1
0
C = 0 0 1.33 0 0 , (2.63)
0
0
0
.
0
0
(2.64)
In this case, the base system is unobservable, with the unobservable subspace in
this case being given by
(2.65)
N (Obase ) = span 0 ;
82
Fig. 2.11 Closed-loop output for three initial conditions in the after reset surface
however, the reset control system is well-posed since the after-reset surface is
0
0
1
0 1 0
(2.66)
MR = span 0 , 0 , 0
1
0
0
0
0
0
and MR N (Obase ) = {0}. In addition, it can be checked that the after reset states
are not reduced, in other words, assuming initial conditions in the after-reset surface
there are always an infinite number of crossings.
In Fig. 2.11, closed-loop outputs corresponding to different initial conditions are
obtained. Also in Fig. 2.12 the different control inputs are shown. Note that in general, as the dimension of the after-reset surface increases, the structure of the reset
instants becomes not that simple, for example, in the interval [0, 15]s one initial condition does not produce a crossing while the others produce five reset actions. As a
conclusion, as the dimension of the after reset surface increases, the reset instants
occur with more complex patterns.
83
Fig. 2.12 Control input for three initial conditions in the after-reset surface
Consider the reset control system of Fig. 2.1, where the plant and the reset compensator are given by (2.1) and (2.2), respectively, and with a reference input r and
a disturbance input d generated by exosystems, with the state space models
1 (t) = A1 w1 (t),
w
w1 (0) = w10 ,
r(t) = C1 w1 (t),
(2.67)
2 (t) = A2 w2 (t),
w
w2 (0) = w20 ,
d(t) = C2 w2 (t),
(2.68)
with w2 Rm2 . These exosystems allow the generation of signals like steps, ramps,
sinusoids, . . . . Now, the feedback connection is given by e = r y and u = v + d,
and the base linear closed-loop system may be described by
x (t) = Ax(t) +
0
Br
r(t) +
0
Bp
d(t)
(2.69)
where x = xxpr , and xp and xr are the plant and compensator states. In addition,
the reset instants are defined as those instants t at which the closed-loop output
y(t) = Cx(t) is equal to the reference signal r(t) = C1 w1 (t). Define the augmented
state z as z = ( wT1 wT2 xT )T . Then, the reset map can be defined in the augmented
84
state space by
I
0
A R =
0
0
0
I
0
0
0
0
I
0
0
0
0
A
(2.70)
MR = R(A R ) N (C),
\ MR
M = N (C)
(2.71)
z (t) = Az(t)
if z(t)
/ M,
(2.72)
z(t + ) = A R z(t) if z(t) M
with the state space matrix
A1
0
A =
0
Br C1
0
A2
Bp C2
0
0
0
Ap
Br Cp
0
0
.
Bp Cr
Ar
(2.73)
In the augmented space state representation, Propositions 2.3 and 2.4 can be
directly applied, giving the next result.
Proposition 2.5 Consider the reset control system of Fig. 2.1, with the plant and
reset compensator given by (2.1) and (2.2), respectively, and with inputs r and d
generated by the exosystems (2.67)(2.68). If MR N (Obase ) = {0}, where Obase
is the observability matrix
C A
(2.74)
Obase =
..
.
C A n+m1 +m2 1
then
1. The reset system is well-posed.
2. The reset system does not have Zeno solutions.
Proof The augmented state can be partitioned as z = xz1r , and thus the state matrix
(2.73) is partitioned as
A p
B p Cr
A =
,
(2.75)
Br C p
Ar
85
where
A1
Ap = 0
0
0 u0
A2 u 0 ,
Bp C2 uAp
0
B p = 0 ,
Bp
C p = C1 0 Cp . (2.76)
Part 1 of the proposition is a direct application of Proposition 2.3 to the system (2.72)
A R , and C.
On the other hand, since the state matrix A
defined by the matrices A,
has the structure given in (2.75), it always possible to make a state transformation of
z1 in such a way that corresponding principal minor be in the observer form. Thus,
the reasoning used in the proof of Proposition 2.4 may be used to prove Part 2.
In general, a simple sufficient condition for the well-posedness of a reset control
system with exogenous inputs it that the augmented closed-loop system (2.72) be
observable, that is, the matrix Obase be full rank.
xr (t) = r(t),
xr (t + ) = 0,
v(t) = xr (t).
xr (t) r(t) = 0,
xr (t) r(t) = 0,
(2.77)
1 (t) =
w
w1 (t),
0
w1 (0) =
a sin
a cos ,
r(t) = ( 1 0 )w1 .
(2.78)
Since a disturbance signal is not considered in this example, Proposition 2.5 can
be used by eliminating the row and column blocks corresponding to the disturbance
A R , and C.
The result is
exosystem in the matrices A,
0
A =
0
1
0
0
0
0
0
0
1
0
0
,
1
0
C = 1 0 1 0 .
(2.79)
86
Fig. 2.13 Reference signal and closed-loop output corresponding to a well-posed reset control
system
1
0
1 0
0 1
Obase =
(1 + 2 )
0
1
0
0
(1 + 2 ) 0
1
(2.80)
which is full rank for any > 0 (and does not depend on a), and thus the system is well-posed for any sinusoidal reference input. Figures 2.13 and 2.14 show
a simulation of the reset control system for = 0.5 rad/s, with an initial condition
w1 (0) = (1 1)T for the exosystem, and with zero initial condition for the Clegg
integrator and the integrator.
87
1 0 0
0 0
C = 1 0 1 , (2.81)
A = 0 0 ,
A R = 0 1 0 ,
0 0 0
1
0 1
and the observability matrix of the augmented base system is given by
1
0 1
1
Obase = 1
2
1 1
(2.82)
which is full rank for any > 0. Although Proposition 2.5 cannot be used in this
case because the plant does not fit the model (2.2), the well-posedness can be assured
since Obase is full rank. However, the well-posedness argument cannot be used for
assessing the existence of Zeno solutions. In fact, this system has a Zeno solution
with sequences of reset instants {tk }, k = 1, 2, . . . , converging to t = n for every
integer n 1.
Figure 2.15 (top) shows the output response y(t) of the closed loop (and output of
the Clegg integrator) to the sinusoidal reference w(t) = sin t. It can be shown that
a Zeno behavior appears to the left of every t = k , k = 1, 2, 3, . . . . The solution
88
Fig. 2.15 (Top) Response y(t) to a sinusoidal reference of a feedback loop with a Clegg integrator;
(Bottom) detailed zoom of the three first visible resets close to t =
starting from the right of every y(k) = 0 escapes the Zeno behavior until it again
reaches another Zeno point at t = (k + 1) .
Figure 2.15 (bottom) plots a detailed zoom of the first three visible resets close to
t = . These reset times are t1 2.28, t2 2.92, and t3 3.12. In fact, there exists
an infinite sequence of resets {tk } converging to t = .
It is illustrative to see in Fig. 2.16 the time response of this system in a threedimensional plot that shows the time evolution of the state (w11 , w12 , xp ) =
(r, r , y) . The first two coordinates are from the exosystem, and the first one is
the reference r(t) to the closed loop, in this case r(t) = sin t and r (t) = cos t,
with = 1. Thus the trajectory lies entirely within the cylinder r 2 + r 2 = 1.
The three-dimensional plot reveals clearly that, in the state space (r, r , y),
there are actually two Zeno accumulation points, namely Z1 = (0, 1, 0) and Z2 =
(0, 1, 0). The first one corresponds in the time domain (Fig. 2.15) to the zero
crossings of the reference for t = 2k and the second one for t = (2k + 1) , with k
integer.
Figure 2.16 also shows two relevant planes: the null space N (C) of C =
(1, 0, 1), that is, the plane y = r that triggers the reset action, and the plane MR
given by y = 0 where the state is projected immediately after a reset y(tk+ ) = 0.
89
90
y(tk ) = 0.
Since tk tends to , a key simplifying assumption is that, in the limit, we can approximate sin t by (t ). This simplification is also helpful because the Poincar
maps give rise to implicit equations not solvable analytically. In this way, close to
t = , we have:
y = (t ) y(t),
y(tk ) = 0,
t tk .
To determine the next reset time tk+1 , we have to impose the condition y(tk+1 ) =
r(tk+1 ). Again we replace r(tk+1 ) = sin tk+1 by (tk+1 ). Thus the Poincar map
tk tk+1 is given implicitly by the condition
(tk+1 ) = (tk 1)e(tk+1 tk ) tk+1 + + 1.
Introduce the change of variables tk = dk , with dk 0 . Using the asymptotic approximation e(tk+1 tk ) = e(dk+1 dk ) 1 (dk+1 dk ) gives rise to
dk+1 = (dk 1)(1 dk+1 + dk ) dk+1 + 1,
from which we can solve explicitly dk+1 = dk2 /(dk 1). This law for dk 0
approaches dk+1 = dk2 , or equivalently, approaches the limit relation
|tk+1 | = |tk |2
which proves that there exists an infinite sequence of reset times tk tending to and
governed, in the limit, by a quadratic recurrence law. This explains why in Fig. 2.15
there are only a few visible resets: to detect the reset at tk = + dk , we should
implement a simulation step size smaller than |dk |, that, from the quadratic law
|dk+1 | = |dk |2 , decreases very fast with k, implying a strong computational cost.
References
1. Bainov, D.D., Simeonov, P.S.: Systems with Impulse Effect: Stability, Theory and Applications. Ellis Horwood Limited, Chichester (1989)
2. Baos, A., Mulero, J.I.: On the well-posedness of reset control systems. Technical Report
TR-DIS-1-2011, University of Murcia (2011)
3. Beker, O., Hollot, C.V., Chait, Y., Han, H.: Fundamental properties of reset control systems.
Automatica 40, 905915 (2004)
4. Bell, P.C., Delvenne, J.-C., Jungers, R.M., Blondel, V.D.: The continuous SkolemPisot
problem: on the complexity of reachability for linear ordinary differential equations (2009).
arXiv:0809.2189v2 [maths.DS]
References
91
5. Branicky, M.S., Borkar, V.S., Mitter, S.K.: A unified framework for hybrid control: model and
optimal control theory. IEEE Trans. Autom. Control 43(1), 3145 (1998)
6. Carrasco, J.: Stability of reset control systems. Ph.D. Thesis (2009) (in Spanish)
7. Grizzle, J.W., Abba, G., Plestan, F.: Asymptotically stable walking for biped robots. Analysis
via systems with impulse effects. IEEE Trans. Autom. Control 46, 5164 (2001)
8. Haddad, W.M., Chellaboina, V., Nersesov, S.G.: Impulsive and Hybrid Dynamical Systems:
Stability, Dissipativity, and Control. Princeton University Press, Princeton (2006)
9. Hainry, E.: Reachability in linear dynamical systems. In: Computability in Europe 2008. Lectures Notes in Computer Science, vol. 5028. Springer, Berlin (2008)
10. Lakshmikanthan, V., Bainov, D.D., Simeonov, P.S.: Theory of Impulsive Differential Equations. World Scientific, Singapore (1989)
11. Michel, A., Hu, B.: Towards a stability theory of general hybrid dynamical systems. Automatica 35, 371384 (1999)
12. Neic, D., Zaccarian, L., Teel, A.R.: Stability properties of reset systems. In: Proc. of 16th
IFAC World Congress, Prague, Czech Republic (2005)
13. Schumacher, J.M.: Time-scaling symmetry and Zeno solutions. Automatica 45, 12371242
(2009)
14. Yang, T.: Impulsive Control Theory. Lecture Notes in Control and Information Sciences,
vol. 272. Springer, Berlin (2001)
Chapter 3
3.1 Introduction
Stability is a main concern in control systems, and thus stability of reset control
systems is a key issue both from an analysis and a design point of view. In the
classical works of Clegg and Horowitz, the main stability tool was the describing
function that, although giving only approximate results, provides information about
the phase lead characteristic of reset compensators. In the seminal work [18], the stability problem is explicitly formulated, concluding that the design problem will be
solved more satisfactorily when stability criteria are developed for feedback loops
containing nonlinear elements such as FORE ( first order reset element).
In the last decade, the stability problem has been approached from different perspectives, and nowadays a number of results are available for reset control systems,
including systems with time delays. A basic question is about the relationship between the stability of the reset control system and its base linear system. It is important to emphasize that the stability of a reset control system is not always guaranteed
by the stability of the base system; in other words, reset actions can destabilize a stable base linear system, thus reset should be carefully used in practice. Conversely,
a proper reset control design may stabilize an unstable base linear system, thus stability of the base linear system is neither sufficient nor necessary for the stability of
the reset control system.
This chapter addresses the stability of reset control systems with finite-dimensional base systems. Section 3.2.1 deals with reset-times independent conditions,
that is, stability conditions that are not explicitly based on the reset instants, including the H condition developed in [9]. In Sect. 3.2.2, we study stability conditions
depending on the reset time intervals and based on the induced discrete-time system obtained by sampling the reset control system at the after-reset instants. This
approach is further extended in Sect. 3.2.3 to the stabilization problem, imposing
stabilizing upper bounds (resp., lower bounds) on the reset intervals, that is, forcing (resp., avoiding) reset actions. Section 3.3 considers inputoutput L2 stability
by using passivity-based techniques and providing a number of useful results on
the passivity and dissipativity of different reset systems. Finally, the chapter concludes with Sect. 3.4 where an analysis by the describing function is presented,
A. Baos, A. Barreiro, Reset Control Systems, Advances in Industrial Control,
DOI 10.1007/978-1-4471-2250-0_3, Springer-Verlag London Limited 2012
93
94
95
V
x
Ax < 0,
for x = 0,
(3.2)
and
V (x) := V (AR x) V (x) 0,
for x M .
(3.3)
(3.4)
(3.5)
and the following stability result is obtained, where the matrices C0 and B0 are
defined by
0np n
C0 = (Cp , 0n n P ),
B0 = 0n n .
(3.6)
In n
96
A P + P A < 0,
B0 P = C0
(3.7)
(3.8)
(3.9)
97
In spite of its usefulness, the H -condition has some limitations. Firstly, since
the quadratic function V (x) has a continuous derivative over Rn , it is clear that (3.2)
is equivalent to xT (AT P + P A)x < 0 for all x Rn and x = 0, and thus the base
linear system must be always asymptotically stable. Secondly, note that even in the
case in which the base linear system and the reset system are asymptotically stable,
it might happen that the H condition does not hold. This is, for example, the case if
the reset action induces small positive increments V (x) > 0 that are compensated
by large decrements between reset instants, due to V (x) < 0, so that the net effect is
a net decrement of V (x), and thus the reset system is stable.
As a result, the H -condition has two basic limitations: (i) it does not apply to
reset systems with unstable base system, and (ii) it is only a sufficient condition,
and thus there exist stable reset systems for which an SPR H does not exist. In
the following, two examples are developed: an example in which the H -condition
gives stability of a reset system; and a second example of a stable reset system with
a stable base system in which the H -condition does not apply.
Example: H is SPR Consider Example 6 in [10], which uses a reset control system with plant P (s) = 1/s and base linear controller Rb (s) = 1/(s + 1), connected
in negative feedback, without exogenous inputs. If the state vector is x = (x1 , x2 )
with x1 the plant state and x2 the (reset) controller state, then it results in a reset
system like that in (2.4)(2.6) with
1 0
0
1
,
C= 10 .
(3.10)
A=
,
AR =
0 0
1 1
Since the base linear system is observable, the reset control system is well-posed.
In addition, C0 and B0 are given by
0
B0 =
,
C0 = 1 ,
(3.11)
1
and thus it is easily obtained that (A, B0 ) is controllable and (A, C0 ) is observable.
In addition, from (3.9), H is simply given by (for this case n = 1 and then P = 1
without loss of generality):
s +1
1 0
,
H (s) = 1 (sI A)
= 2
1
s +s +1
and its frequency response is
H (j ) =
2
( 2 )
+
j
.
(1 2 )2 + 2
(1 2 )2 + 2
Finally, since H (s) is Hurwitz, Re{H (j )} > 0 for all finite , and furthermore, lim 2 Re{H (j )} = 1 > 0 for any value of R, we can conclude
that H is SPR, and as a consequence, the control system is quadratically stable and
thus asymptotically stable.
98
In this example, the after-reset surface MR and reset surface M are given by
MR = N (C) R(AR ) = {0},
,
1
M = N (C) \ MR = span
{0}.
0
(3.12)
Note that since the after-reset surface is the origin 0, the state evolution of this
system is very simple, the equilibrium point is reached after the first reset if a reset
occurs. In addition, sincethe matrix A does not have real dominant eigenvalues, its
eigenvalues are 12 j 23 , and (A, C, x0 ) is reduced for any initial condition x0 ,
so the reset surface is always reached in a finite time.
This is a very simple example of a reset control system that is asymptotically
stable, and this stability is perfectly captured by the H -condition. But it might
happen that the H -condition fails in detecting stability, as seen in the following
example.
Example: H is not SPR This example shows a case of a reset control system
with a stable base system in which the H -condition does not apply, and thus the
stability of the reset system cannot be asserted. As it will be shown in Sect. 3.2.2.1
(see the second example), this case corresponds indeed to a stable reset system.
Consider a reset control system with a base compensator Rb (again corresponding to a FORE compensator) and a plant P as given by the transfer functions:
Rb (s) =
3.2
,
s 1
P (s) =
s + 0.25
.
s 2 + 2s
(3.13)
1 0 0
2
0
1.6
0
0 ,
C = 1 0.25 0 ,
A= 1
AR = 0 1 0 ,
0 0 0
2 0.5 1
(3.14)
where A is Hurwitz (all its eigenvalues have strictly negative real part). In this example, the compensator is full reset, and so n = 0. In addition, np = 2 and n = 1.
Thus is a real number, and
C0 = Cp P = 0.25 1 ,
(3.15)
where P = 1 without loss of generality. In addition,
B0 = 0np 0n In = 0 0 1 .
(3.16)
5s 2 + (8 + 10)s + 2
.
5s 3 + 5s 2 + 6s + 4
(3.17)
99
In Fig. 3.1, it is shown that although the base system is stable (A is Hurwitz) the H condition is not satisfied since there does not exist any R such as Re{H (j )} >
0 for any > 0.
100
In this section, stability conditions will be developed by using an auxiliary timevariant discrete time system, where the dynamics of the auxiliary system will explicitly depend on reset time intervals. This approach is further extended in Sect. 3.2.3
to the stabilization problem, imposing restrictions over reset intervals to be in some
designed set. A simple observation of the reset control system equation (2.4), and
its solution as given in Proposition 2.1, shows that if it is sampled at the after-reset
instants tk+ , k = 0, 1, 2, . . . , the sampled system is given by
+
x tk+1
, x0 = AR eA(k+1 (x0 )k (x0 )) x tk+ , x0
(3.18)
(3.19)
As a result the stability of the reset control system is given by the stability of the
auxiliary time-invariant discrete time system, taking into account that the reset sur-
101
face is reached from any nonzero initial condition in the case in which the base
system is not stable. The following result easily follows.
Proposition 3.4 A regular reset control system with m = 1 is asymptotically stable
if the matrix AR eA is Schur stable, that is, all its eigenvalues are strictly inside the
unit circle. In addition, a reset control system with m = 1 is not stable if it is not
regular, or is regular and AR eA is not Schur stable.
In general, for the case m > 1, the reset intervals sequences may exhibit complex
patterns as has been discussed in Sects. 2.3.2 and 2.3.3. However, the stability problem of the reset control system can be reduced to solutions with initial conditions
in the after reset surface MR . In addition, if the base system is not stable, the regularity condition is needed to assure that MR is uniformly reachable in a finite time;
otherwise, the reset control system is not stable. On the other hand, the auxiliary
time-varying discrete time system will be given by
xd (k + 1) = AR eAk+1 xd (k),
xd (0) = x0
(3.20)
102
x(s, x0 ) ds,
(3.23)
tk
where the induced norm A is always bounded by some real number , that
is, A . Finally, using Gronwall inequality, x(t, x0 ) x(tk+ , x0 )e(ttk ) ,
showing that the solution x(t) is always bounded in every interval t (tk , tk+1 ], resulting in x(t, x0 ) e(ttk ) x0 , where and do not depend on the initial
condition x0 M . In addition, since the reset control system is regular, if the base
linear system is unstable, that is, A is not Hurwitz, the solution x(t, x0 ) satisfies
x(t, x0 ) ek x0 eM x0
(3.24)
for t 0, including reset instants. Thus, it is obtained that the system (2.4) is stable.
Note that in the case in which A is Hurwitz, reset intervals do not need to be upperbounded since the exponential is always bounded, and stability of (2.4) directly
follows.
Asymptotic stability follows by using similar arguments and the fact that
x(tk+ , x0 ) k x0 , k 0 for any x0 Rn , with 0 < 1, by uniformly
asymptotic stability of the discrete time system (3.19). Then, as t , it is true
that k since t (tk , tk+1 ] and tk+1 tk = k+1 (x0 ) > m for any x0 Rn ,
and in addition since the reset control system is regular,
x(t, x0 ) k ek x0 k eM x0 ,
thus x(t, x0 ) 0 for any x0 Rn as t .
(3.25)
Proposition 3.6 A regular reset control system is asymptotically stable if there exist
a positive definite matrices sequence Pk , k = 1, . . . , such that I Pk I and
k
eA
AR Pk+1 AR eAk Pk I
(3.26)
for some finite positive constants , , and , and for any reset interval sequence
(1 , 2 , . . . ), where k > m , k = 1, 2, . . . .
Proof It is a direct application of Proposition 3.5 and discrete time-varying systems
stability results as given, for example, in [27, Theorem 23.3].
In comparison with the reset times-independent results in the previous section,
notice that there the stability conditions are:
T T
x (A P + P A)x < 0
if x
/ M \ {0},
(3.27)
xT (ATR P AR P )x 0 if x M ,
which implies that the quadratic Lyapunov function V (x) = xT P x is decreasing
over the time intervals (tk , tk+1 ) and non-increasing at the reset instants tk , k =
103
1, 2, . . . . These stability conditions have also been previously published in [23] and
the seminal work [2], in the context of impulsive systems with impulses at fixed
instants.
It is clear that from the first condition of (3.27) the base linear system must always
be asymptotically stable. Note that even in the case in which the base linear system
is asymptotically stable, the second condition of (3.27) may be not compatible in
general with the solutions of the first condition.
In addition, in the previous section it has been shown that (3.27) is equivalent
to the H -condition. Thus, the main contribution of Proposition 3.5 (and Proposition 3.6) is to circumvent the H -condition by allowing that the continuous time
evolution between two consecutive reset instants, in (tk , tk+1 ), k = 1, 2, . . . , is not
necessarily associated with a Lyapunov function (quadratic or not) in those intervals,
thus results are less conservative: (i) the proposition may be applied to unstable base
linear systems, (ii) it may be applied to stable base linear systems in which condition
(3.27) cannot be applied, as we will see in the two following examples.
3.2.2.1 Examples
Unstable Base System Consider a reset control system as given by (2.4), where
A, AR , and C are given by
2 0 1
0 0,
A= 1
0 1 1
1
AR = 0
0
0
1
0
0
0,
0
C= 001 .
1
s(s+2)
(3.28)
and a FORE
y(t) = 0.42e0.57t cos(0.37t) + 0.37 e0.57t sin(0.37t) 0.42 e2.15t x01 .
Thus, after the first reset, reset instants are periodic with period = 2.3 (the first
zero of y(t) that is bigger than m = 0.1) for any initial condition. Note that AR eA
is Schur stable (its eigenvalues are strictly inside the unit circle), but that the reset
control system is not regular since the after-reset surface is not reached for initial
conditions x0 span{(0.898, 0.418, 0.133)T }. Thus, Proposition 3.4 can be used
to state that the reset control system is unstable. If the initial condition set were
restricted to MR [7] then the reset control system would be stable since AR eA is
Schur stable.
104
AR P AR eA P + I vs
Stable Base System This example is given to show how the H -condition can be
overtaken for reset control systems with stable base system. Consider a reset control
system as given by (3.21) with a dwell-time constant m = 0.1 and matrices
2 0 1.6
1 0 0
0 ,
A = 0.5 0
AR = 0 1 0 ,
C= 001
(3.29)
2 1 1
0 0 0
3.2
and P (s) = s+0.25
. It can be easily checked that, alcorresponding to Rb (s) = s1
s 2 +2s
though A is Hurwitz (and thus the reset control system is regular), the H -condition
is not satisfied, that is, there does not exist any R ( is scalar in this case) such
that Re{H (j )} = Re[ 0 1 ](j I A)1 [ 0 0 1 ]T > 0 (see the second example
in Sect. 3.2.1).
Now, it has been found that (see the search procedure given in next section) (3.26)
is feasible for constant matrices Pk = P , where P is given by
and for any k 0.03, k = 1, 2, . . . . Figure 3.2 shows the maximum eigenvalue of
eA AR P AR eA P + I vs , for = 0.001. It is always negative for 0.03.
As a result, Proposition 3.6 guarantees that the reset control system is asymptotically stable. As it will be developed in next section, for the case of stable base
105
Fig. 3.3 Base system (dashed line), and reset system (solid line)
systems this method can be used to compute the values of the dwell-time constant
m that stabilizes the reset control system. In this example, stability is assured for
any m 0.03. Figure 3.3 shows a time simulation of the closed-loop output for
the base and the reset control system.
eA
AR P AR eA P I
(3.31)
106
is satisfied for some P > 0 and > 0. Therefore, the reset control system (3.21) is
asymptotically stable for that m .
Proof Since A is Hurwitz and then the reset control system is regular, stability directly follows from Proposition 3.6, by taking Pk = P , for k = 1, 2, . . . . Note that
since A is Hurwitz, the matrix eA can be made arbitrarily small in norm, by choosing large enough. Thus, for any > 0 there always exists a m such for any
> m it is true that eA . Some valid explicit relations = f () could be
derived using some canonical form of A and eA (diagonal, Jordan form, etc.) and
bounding |ei |, where i are the (stable) eigenvalues of A. Then, for the sequence
P1 = P2 = = P for some P > 0, a large enough can be chosen ( > m ) so
that the left side of the inequality (3.31) approaches 0P = P , which is obviously
smaller than I for small > 0.
A Search Procedure for m Proposition 3.7 gives a condition that is easy to
check in practice by using available algorithms for solving LMIs. However, some
care has to be taken because as it is well known the norm eA is not necessarily
monotonic with respect to . For A Hurwitz, when it is true that eA 0,
but this does not imply monotonicity of eA as a function of .
Before the search procedure, a lower bound of the solution (LB ) can be found
by the map of the feasibility of the LMI (3.31) vs . This lower bound is defined by
LB R such that for all LB the LMI (3.31) is feasible for some P ().
A search procedure for m is proposed as follows:
1. Initial value: Let us start with a large initial value L so that the LMI (3.31) is
feasible for the single value = L with respect to some P = PL > 0, and also
for all > L the LMI (3.31) is feasible.
2. Feasible values of P : A set of feasible values of P is obtained simply by sampling the interval (0, L ] and solving the LMI (3.31) for P with a given sampled value of . More precisely, with some step size 1 , a set of reset intervals
{1,1 = 1 , 2,1 = 21 , 3,1 = 31 , . . . } (0, L ] is chosen. Now, for every
k,1 a matrix Pk > 0 is computed by solving the LMI (3.31). As a result, a set
of feasible pairs (k,1 , Pk ), k = 1, 2, . . . , is obtained. Let P = {P1 , P2 , . . .}.
3. Feasible values of : The next step is to compute the minimum value m,k
such that for any > m,k the LMI (3.31) is feasible for some Pk P. In
order to obtain that result, every matrix Pk P generates a matrix Qk () =
eA AR Pk AR eA Pk + I . Then, the interval (0, L ] is sampled with another
step size 2 such that the maximum eigenvalue of Qk (L i2 ), where i =
0, 1, . . . , is computed and i is increased until the maximum is non-negative for
some value of i = ik + 1; therefore, for all Pk P, a m,k = L (ik )2 is
found.
Again the LMI has to be checked with some accuracy 2 (not necessarily
equal to 1 ), for {1,2 = 2 , 2,2 = 22 , 3, = 32 , . . . }. As a result,
a set of pairs (k,m , Pk ), k = 1, 2, . . . , is now obtained.
4. Finally, the desired m is the minimum value of m,k for k = 1, 2, . . . , obtained
for some km , and Pkm is the associated matrix.
107
(s+1)(s+10)
P (s) = s 4 +4s
3 +6s 2 +5s98 , respectively, corresponding to a state-space realization
with matrices A and AR :
4
0.750 0.312 1.531
4
8
0
0
0
0
2
0
0
0
A= 0
,
0
0
4
0
0
0.25 0.656 1.875 0.391 1
(3.32)
1 0 0 0 0
0 1 0 0 0
AR =
0 0 1 0 0.
0 0 0 1 0
0 0 0 0 0
Note that the reset control system is regular since A is Hurwitz. It has been found
that the LMI (3.31) is feasible for 27.02 and some P > 0. As a result of the
search procedure a matrix Pm has been found:
4
(3.33)
Pm = 10 0.0395 0.0034 0.4300 0.0597 0.1110
108
AR P AR eA P + I vs
109
is very oscillatory since the reset intervals are close to the stability limit. Figure 3.7
shows a simulation for m = 30. In this case, the system behavior is much less
oscillatory due to the fact that reset intervals are far away from the stability limit.
110
111
A Low Order Base System Consider a reset control system with the base compensator Rb and the plant P given by
Rb (s) =
8
,
s +1
P (s) =
1
.
(s + 2)s
(3.36)
s3
8
.
+ 2s + 8
(3.37)
+ 3s 2
Ap =
Ar = 1,
2 0
,
1 0
Br = 4,
1
Bp =
,
0
Cr = 2,
Dr = 0,
Cp = 0 1 ,
2 0
2
1 0 0
0
0 ,
A= 1
AR = 0 1 0 ,
0 4 1
0 0 0
(3.38)
Dp = 0, (3.39)
C= 010 .
29.18 10.56
1.49
P = 10.56 176.23 19.39 ,
1.49 19.39 93.07
(3.40)
(3.41)
(3.42)
350
,
s +1
P (s) =
s5
+ 13s 4
s +3
+ 87s 3 + 305s 2 + 350s
(3.43)
that results in a unstable base control system, for example, the (unstable) closedloop transfer function from the reference input to the output is given by
112
AR P AR eA P + I vs
Fig. 3.9 Closed-loop output for the linear base control system (dashed line), and the reset control
system (solid line)
T (s) =
s6
+ 14s 5
+ 100s 4
350(s + 3)
.
+ 392s 3 + 655s 2 + 700s + 1050
(3.44)
113
AR P AR eA P + I vs
In this case, it has been found that for the positive-definite matrix P
16.22 1.16
2.14
2.42
0.79
3.08
1.16 14.68 1.46
5.44
1.93
1.74
2.14
1.46
9.77
2.92
10.78
2.26
P =
2.42
5.44
2.92
9.22
10.45
0.68
0.79
1.93 10.78 10.45 28.11 9.55
3.08
1.74
2.26 0.68 9.55 27.35
(3.45)
(3.46)
114
Fig. 3.11 Closed-loop output for the linear base control system (dashed line), and the reset control
system (solid line)
The work [24] approaches the problem for compensators in which its output has
the same sign as its input, and the zero reference case is considered. In addition,
in [1], L2 -stability conditions for the case of nonzero references are given. The
conservatism given by the H -condition is improved for that kind of system.
On the other hand, dissipative systems theory was developed in [33], where the
passive system concept which was developed in electric circuit and mechanical systems was extended to abstract systems. That theory is widely used in interconnected
systems due to the fact that the feedback interconnection of two passive systems is
a passive system. In [29], passivity techniques are shown to be a powerful tool for
nonlinear control. Dissipative systems theory has been developed for a hybrid system in [17], where dissipative hybrid systems definitions are developed for classic
concepts such as supply rate. A few works have dealt with passive systems theory
and hybrid systems [35], [11], [25], [8]. In spite of the fact that a single inputoutput
approach to passive systems theory cannot be sufficient for more general hybrid systems, this section will show how several passive properties can be obtained for reset
compensators.
The goals of this section are: (a) to obtain stability conditions that are applicable
to feedback interconnections of reset compensators and nonlinear plants and (b) to
find passive reset compensators that can be used in passive control techniques. The
passivity approach will be used for these purposes. Easily checkable passivity conditions on the base linear compensator will be presented. In general, the feedback
system given by Fig. 3.12 will be considered, where w and d are the reference and
disturbance inputs, respectively. R is a single-input single-output (SISO) reset compensator to be defined later on, and P is an SISO plant.
115
(3.47)
The feedback system of Fig. 3.12 is L2 -stable (with respect to inputs w and d)
if for every input signals w L2 and d L2 , the outputs u L2 and y L2 . In
addition, it is finite-gain stable if there exists a positive constant > 0 such that
y2 + u2 (w2 + d2 ).
Here the plant P is represented by the state space model
xp = f (xp , u) if u R,
(P )
(3.48)
y = g(xp , u)
if y R,
where np is the dimension of the state xp , f : Rnp R Rnp is locally Lipschitz, g : Rnp R R is continuous, f (0, 0) = 0, and g(0, 0) = 0. In addition,
following the notation given in [17], the reset compensator dynamic is described by
three elements: (a) a continuous-time dynamical equation, (b) a difference equation,
and (c) a resetting law. Whenever the resetting law is not held, the system flows
continuously; otherwise, when the resetting law is satisfied, the system jumps in a
discrete way. An LTI base compensator will be considered in this work, then the
reset compensator R is given by the impulsive differential equation (IDE):
xr = Ar xr + Br e if e = 0,
if e = 0,
xr+ = A xr
(R)
(3.49)
u = Cr xr + Dr e,
where nr is the dimension of the state xr , A is a diagonal matrix with zeros in
the states to be reset and ones in the rest of compensator states, n is defined as
116
the dimension of the reset subspace, and n is defined as the dimension of the nonreset subspace (n + n = nr ). When A = 0, R will be referred to as a full reset
compensator; otherwise, it will be referred to as a partial reset compensator.
The first equation in (3.49) describes the continuous compensator dynamic at the
non-resetting instants, and the second equation gives the reset operation as a jump
of the compensator state at the reset instants. Note that reset instants are produced
when the compensator input is zero. The base compensator is simply obtained by
elimination of the reset actions in (3.49) and thus has the transfer function Rbl (s) =
Cr (sI Ar )1 Br + Dr , where it is assumed that (Ar , Br , Cr , Dr ) is a minimal
realization.
It is assumed that the feedback system is well-posed and thus Zeno solutions are
avoided. Alternatively, time regularization may be imposed (see (2.7)), where the
resetting law is deactivated during a time m after every reset action. As a result, it
is assumed that there exists a finite number of resets over every finite time interval,
thus the reset actions will be given by the countable set {t1 , t2 , . . . , tk , . . . } where
tk+1 tk > m for all k = 1, 2, . . . and limt tk = .
A system H : L2,e L2,e , with input u and output y = H u is said to be passive
if there exists a constant 0 such that
u (t)y(t) dt ,
T 0, u L2 .
(3.50)
u (t)y(t) dt +
0
u (t)u(t) dt +
y (t)y(t) dt,
(3.51)
for all functions u, and all T 0, then the system is input strictly passive (ISP)
if > 0, output strictly passive (OSP) if > 0, and very strictly passive (VSP) if
> 0 and > 0. In particular, for a linear time-invariant system with a transfer
function given by H (s) = C(sI A)1 + D, with A Hurwitz and the pair (A, B)
controllable, it is true that [21]
1. The system is passive if and only if Re[H (j )] 0 for all .
2. The system is ISP if and only if there is a > 0 such that Re[H (j )] > 0
for all .
3. The system is OSP if and only if there is an such that Re[H (j )] |H (j )|2
for all .
Note that these three properties are verifiable in a Nyquist diagram: if H (j ) is in
the closed right half-plane then the system is passive; if H (j ) is in Re[H (j )]
1
> 0, the system is ISP; and if H (j ) is inside the circle with center at s = 2
and
1
radius r = 2 , then the system is OSP.
When the system is in state space representation, the passivity of the inputoutput
maps as before is replaced by the following notions from dissipative systems theory.
117
(3.52)
where f and g have the same properties as in (3.48). H is said to be dissipative respect to a supply rate w(t) = w(u(t), y(t)) if there exists a storage function
V (x) 0 such that the following dissipation inequality holds
T
w(u(t), y(t)) dt; u, x(0), T 0.
(3.53)
V (x) V (x(0)) +
0
Following [29], the relationship between dissipative and passive systems is given
by the choice of a particular supply rate:
1. The system H is passive if it is dissipative with respect to supply rate wp = u y
and V (0) = 0.
2. The system H is input strictly passive (ISP) if it is dissipative with respect to
supply rate wi = u y u u, for some > 0.
3. The system H is output strictly passive (OSP) if it is dissipative with respect to
supply rate wo = u y y y, for some > 0.
4. The system H is very strictly passive (VSP) if it is dissipative with respect to
supply rate wv = u y u u y y, for some > 0 and > 0.
Comparing (3.50) and (3.53), since V (x) 0, the relationship between and the
initial condition is clear, namely, = V (x(0)). An important result is the Passivity
Theorem (see [29]): the feedback interconnection of Fig. 3.12 is finite-gain stable if
R + P > 0,
P + R > 0,
where R , R , P , and P are such that they satisfy (3.51) for the systems R and P ,
respectively.
118
compensators given by (3.49). This is due to the fact that the system loses all of its
memory at every reset time, i.e., the base compensator is restarted with zero initial
condition. Therefore, a space states description is not necessary because the output
compensator just depends on the input since the last reset time. This is the key point
in the next proposition.
Proposition 3.10 A full reset compensator R (given by (3.49) with A = 0) is passive, ISP, OSP, or VSP if the base compensator is passive, ISP, OSP, or VSP, respectively.
Proof The result is proved for the VSP case, the rest of the cases follow similar
arguments. If the base compensator is VSP, then there exist > 0 and > 0 such
that (3.51) is satisfied for all T 0 and for all u L2 . Since, by assumption, in
every finite time interval there are a finite number of resets instants {0 < t1 , t2 , . . . }
with 0 < t1 < t2 < and limk tk = , there exists a finite integer k such that
for every T (tk , tk+1 ] the integration interval [0, T ] can be divided into a finite
number of subintervals
u (t)y(t) dt =
t1
u (t)y(t) dt +
i=k1
ti+1
i=1
T
u (t)y(t) dt
ti
u (t)y(t) dt.
(3.54)
tk
In addition, since the base system is time invariant and VSP by assumption, it is true
that
t1
t1
t1
u (t)y(t) dt +
u (t)u(t) dt +
y (t)y(t) dt,
(3.55)
0
ti+1
u (t)y(t) dt
ti
and
0
ti+1
u (t)u(t) dt +
ti
T
tk
u (t)y(t) dt
ti+1
y (t)y(t) dt,
(3.56)
ti
T
tk
u (t)u(t) dt +
y (t)y(t) dt
(3.57)
tk
Thus, the reset compensator is a VSP system with the same constants , , and as
its base compensator.
119
Remark 3.1 Note that the formulation of passivity, ISP, OSP, and VSP given in [29]
(Sects. 2.2 and 3.1) is completely general, and thus does not involve any assumption
regarding the continuity of system trajectories of the reset system R as given by
(3.49). In particular, the state trajectories may be discontinuous at the reset instants,
and passivity definitions and the passivity theorem can be directly applied to reset
systems in order to obtain inputoutput stability properties.
Remark 3.2 Note that in the proof of the previous proposition, linearity of the base
compensator has been used simply to state that it has null initial conditions at ti+ for
all i = 1, 2, . . . . Thus, it holds for any full-reset compensator, not necessarily linear,
as long as = 0 or, using the dissipativity framework to be elaborated in the next
subsection, the storage function of the compensator satisfies V (0) = 0.
Some examples of reset compensators are analyzed in the following.
k
s+b
(3.59)
where it will be assumed that k, b > 0. The real part of its frequency response is
given by
kb
,
b2 + 2
and thus, applying Proposition 3.10, it is directly obtained that
Re[FORE(j )] =
(3.60)
1+Ts
1+Ts
(3.61)
120
1 + (T )2
,
1 + ( T )2
(3.62)
1 + (T )2
.
1 + ( T )2
(3.63)
1 + Ti s 1 + Td s
1 + Ti s 1 + Td s
(3.64)
121
dissipative with respect to supply rate wp , wi , wo , or wv , respectively, and a storage function V (x) that satisfies
V (A x) V (x)
(3.65)
(3.66)
For a given input u(t), by assumption there is a finite number of reset times ti , i =
1, 2, . . . , k, in the interval [0, T ], with T (tk , tk+1 ]. Thus, it is true that
T
V (x(T )) V x tk+ +
wv (t) dt
(3.67)
tk
since in the interval [tk+ , T ] the reset compensator behaves like its base compensator
with x(tk+ ) the initial condition. Applying a similar argument, it is true that
ti
+
wv (t) dt
(3.68)
V (x(ti )) V x ti1 +
ti1
V (x(t1 )) V (x(0)) +
ti
(3.69)
wv (t) dt.
0
Then, applying condition (3.65) at the reset instants, it is obtained that V (x(ti+ )) =
V (A x(ti )), i = 1, . . . , k. Using this fact and (3.67)(3.69), the result is
+
V x(T ) V x tk1
+
+
= V x tk1
+
+
V x tk2
+
tk
wv (t) ds +
tk1
wv (t) dt
tk
wv (t) dt
tk1
wv (t) dt ,
tk2
(3.70)
122
u (t)y(t) dt +
u (t)u(t) dt +
y (t)y(t) dt,
(3.71)
Remark 3.3 A necessary and sufficient condition for the satisfaction of (3.65) can
be given as follows for a convex storage function V . Note that if V is convex then
V satisfies (3.65) for all x Rnr if and only if V
x (A x) ker A (see [26]). In the
linear case (quadratic V ) and A being given as the projection on the first vector
component, this is equivalent to the decoupled quadratic function V given by:
0
Q11
V (x) = x
x
(3.72)
0
Q22
where Q11 and Q22 are positive-definite matrices such that Q11 Rn n and
Q22 Rn n . It is clear that V satisfies (3.65) since
0
0
V (A x) V (x) = x
x 0, x Rnr .
0 Q22
Remark 3.4 A well-known condition for an LTI system with state-space representation (A, B, C, D) to be SPR is that
QA A Q QB C
>0
(3.73)
B Q C
D + D
for some Q > 0. In addition, the system (A, B, C, D) is dissipative with respect
to the storage function V (x) = x Qx (see [21]). Note that if the LMI (3.73) is
satisfied for a block diagonal Q = Q011 Q022 with the structure given by (3.72), then
Proposition 2 can be applied and as a result the partial reset compensator with base
system (A, B, C, D) will be VSP.
Some examples of passive partial reset compensators are given in the following.
1 + 1 Ti s 1 + 2 Tr s
1 + T i s 1 + Tr s
(3.74)
where 1 (0, 1) and 2 (0, 1). Now, the transfer function is not enough to define
the reset compensator because, with the same Ti , Tr , 1 , and 2 , infinitely many
123
reset compensators can be defined. Thus, a state space description has to be given.
In particular, consider the base compensator
PII R (s) =
1 + 2s 1 + 0.012s
1 + 3s 1 + s
(3.75)
given by the block diagram of Fig. 3.13, where the left block gives the state x1 = u1
to be reset.
Then, the compensator is defined by its state space matrices given by
0.333 0.247
0.003
Ar =
,
Br =
,
0
1
1
Dr = 0.008,
Cr = 0.444 0.658 ,
and
A =
1
0
0
.
0
(3.76)
Note that by simply checking the Nyquist plot of the system (Ar , Br , Cr , Dr ) it
is straightforward to show that it is SPR (see Fig. 3.14). In addition, a decoupled
storage function has been found that satisfies (3.72), which is given by
1.20
0
V (x) = x
x.
(3.77)
0
8.94
As a result, an application of Proposition 3.11 (see Remark 3.4) guarantees that
the partial reset compensator is VSP.
124
125
3.3.4 Example
In this example, Propositions 3.113.12 will be used to investigate the inputoutput
stability of a reset control system in which the plant is nonlinear; in particular, a reset
compensator for a pH control system will be designed based on the above described
passivity properties. Firstly, consider a stream (with flow rate F ) of distilled water
that is mixed with a titrating stream (with flow rate u) of a strong acid in a continuous
stirred tank reactor (CSTR) [34] with the following assumptions (see Fig. 3.15):
(a) the mixture is perfect and instantaneous, (b) the pH sensor in placed very near
to acid injection, and (c) the flow rate of titrating stream is negligible compared
with the flow rate of the process stream, i.e., F + u F . Under these restrictions,
the process can be modeled as a linear plant in cascade with a nonlinear passive
memoryless system (see Fig. 3.16).
The tank will be considered as the part of the pipe where acid is injected. Due
to conservation of charge, the concentration of the charge sum (W ) is taken as the
state of the linear subsystem, i.e., W = [H+ ] [OH ]. The dynamics of the CSTR
is given by
V W = F (W Wdw ) + uWa
(3.78)
where V is the volume of the tank, F is the flow rate of the process stream, u is the
flow rate of the titrating stream and Wa is the concentration of charge in the titrating
stream. In particular, for distilled water Wdw = 0, but also a stream process of strong
base (Wb < 0) can be studied if Wb " Wa . Furthermore, the output of the linear
system equals the state, yL = W . The relationship between this concentration and
126
the pH is given by
pH(W ) = log
W+
W 2 + 4 1014
.
2
(3.79)
Taking an operation point, with a flow rate constant uss , an equilibrium point is
obtained as F Wss + uss Wa = 0. Therefore, Wss = uFss Wa . Working with increments with respect to the equilibrium point, W = W Wss , the new dynamics are
given by
= FW + uWa
V W
(3.80)
where u = u uss .
The nonlinear memoryless function is defined by pH = pH (Wss ) pH (W ).
The system was inverted in order for an increment in the input to result in an increment in the output. Finally, the model was recast as the block interconnection of
Fig. 3.16, where L is an LTI system given by:
x = ax + bu,
(L)
(3.81)
y = x,
where a =
F
V
and b =
Wa
V .
(3.82)
from which it follows that x(x) > 0 for all x = 0. It can be easily shown (see, for
example, [19] or [21]) that the system of Fig. 3.16 with L and given by (3.81)
(3.82) is passive.
A numerical example is developed for the next data: V = 0.1 l, F = 5 102 l/s,
Wa = 1 mol/l, uss = 0.5 104 l/s. The linear plant from (3.81) is
L(s) =
and the static nonlinearity is given by
(x) = 3 log10
20
,
2s + 1
x+
x 2 + 4 1014
2
(3.83)
.
(3.84)
A reset-PII R compensator has been also designed for tracking some pH references, consisting in deviations about the equilibrium point. By using Proposition 3.12, finite-gain L2 -stability of the feedback system can be directly guaranteed since reset-PII R is VSP (see Sect. 3.3.2.1). A simulation of the control system
127
is shown in Fig. 3.17, where in addition an important characteristic of reset compensation is clearly shown: the possibility of obtaining very fast responses without
overshooting, overcoming fundamental limitations of LTI compensation.
128
zero error crossing as the resetting law) the describing function is only a function of
the input frequency (it does not depend on the input amplitude) makes the analysis
very simple, and thus is a good tool for a first stability analysis.
In the following, the describing functions of several basic reset compensators,
such as the Clegg integrator and FORE, are obtained.
E
K
,
s + a s 2 + 2
(3.85)
and the time response v(t) will consist of two terms: one corresponding to the transient response given by the compensator modes, and the other corresponding to the
steady response given by the sinusoidal input modes. Since the reset instants will
be given by the crossings of the sinusoidal input with zero, and thus they will be
periodic with fundamental period 2/ (see Fig. 3.18), it turns out that v(t) will
also be periodic with that period.
To compute the FORE describing function, a simple procedure is to add to the
base frequency response the contribution of the transitory terms vt (t) in v(t) over
a period, for example [0, 2
]. By symmetry (see Fig. 3.18), it will be computed by
129
KE at
e .
a 2 + 2
(3.86)
The result is
1
/
0
vt (t)ej t dt
E
2j
j 2K2
1 + ea ,
2
2
(a + j )(a + )
(3.87)
and finally adding to the frequency response of the base system, the describing function of FORE results as
K
22 (1 + ea )
FORE() =
1+j
.
a + j
(a 2 + 2 )
(3.88)
Note the simple form of the describing function that only depends on the frequency. This allows us to make direct comparisons of the FORE describing function and the frequency response of its base system. The Clegg integrator describing
function can be simply obtained by setting K = 1 and a = 0 in (3.88):
4
1
1+j
,
CI() =
j
(3.89)
and since 1 + j 4 = 1.62ej 51.85 , it is clear that Clegg integrator gives a phase lead
of almost 52 over an integrator (it also increases the gain with a factor of 1.62
approximately). Figure 3.19 shows this fundamental property of the Clegg integrator
and FORE, that is, the achievement of a phase lead up to 52 with respect their base
linear compensator.
130
Reset control systems with a reset band are feedback control systems (see
Fig. 2.1) where the compensator R is given by the impulsive differential equation
/ B ,
if (e(t), e(t))
B ,
xr (t + ) = A xr (t)
(R)
(3.90)
131
whereas R0 (or simply R) is the standard reset compensator without reset band.
Note that Rb stands for the base reset compensator. For simplicity, the describing
function (DF) of a FORE compensator is considered first.
3.4.2.1 FORE
To obtain the describing function with a reset band, again the sinusoidal response
needs to be calculated. In this case, note that since FORE is time-invariant, the
FORE response to e(t) = E sin(t) over the period [t , 2
t ] is the same
(except for a translation) as its response to e(t t ) over the period [0, 2
] (see
Fig. 3.21). Thus, FORE DF can be obtained as the contribution of the transitory
terms in the response to e(t t ) = E cos(t ) sin(t) E sin(t ) cos(t), plus
the base frequency response. But note that the transitory terms consist of two terms:
one term v1t corresponding to the response to E cos(t ) sin(t), and a second term
vt2 with the response to E sin(t ) cos(t).
) at
By using (3.86), the term vt1 is directly given by vt1 = KEa 2cos(t
e , and thus
+2
using again symmetry arguments, its harmonic balance (corresponding to the fun-
132
vt1 (t)ej t dt
0
2/
1
e(t t )ej t dt
2 0
1
KE2 cos(t )
(1 + ea )
(a+j )(a 2 +2 )
.
E j t
2j e
(3.91)
sin(t ) at
On the other hand, the second transitory term is vt2 = KEa
e , and its
a 2 +2
harmonic balance is now
KEa sin(t )
1 /
(1 + ea )
vt2 (t)ej t dt
(a+j )(a 2 +2 )
0
.
(3.92)
=
2/
E j t
1
e(t t )ej t dt
2j e
2 0
Finally, the FORE describing function is obtained by adding to (3.91) and (3.92)
the base frequency response. After some simple manipulations the result is
.
2
FORE (E, )
j 2(1 + ea )
j sin1 ( )
E , (3.93)
=1+
e
1
+a
FOREb (j )
E
E
(a 2 + 2 )
K
, and the identities sin(t ) = E and cos(t ) =
where FOREb (j ) = j +a
/
2
1 E have been used. Note that in contrast to (3.89), FORE (E, ) depends
both on the magnitude and the frequency of the input. In general, expression (3.93)
is valid for 0 E 1. Obviously, for > E, no reset action is performed, and thus
the describing function will directly be the base frequency response.
Note that the FORE introduces extra phase lead for frequencies over a1 , in comparison with FORE for some values of (see Fig. 3.22). Since in addition for increasing values of a significant phase lag is introduced, it seems that a proper
selection of should be in the range 0 < /E < 0.8.
Again, note that the reset band introduces extra phase lead for values 0 < /E <
0.8 and at every frequency (see Fig 3.23).
133
different FORE compensators, and thus the calculation of its describing function is
straightforward. In addition, partial reset may be implemented by considering the
base frequency response of the non-reset element.
On the other hand, although the derivation is somehow involved (see Appendix A
of [6]), a closed-form expression of the describing function can be obtained for the
134
general case of full reset with a reset band. The result is:
R (E, ) = Rbase (j )
+
Ar
1
1
1
Cr e Ar sin 1( E ) G
, I + eAr e( j I )(sin ( E )) Br ,
E
(3.95)
where G is given by
1
1
G(x, ) = (j I Ar )1 ej sin (x) + (j I + Ar )1 ej sin (x) (j I Ar )1 .
(3.96)
Using (3.95)(3.96) with = 0, a general closed-form expression can also be
obtained for the describing function R0 (j ) of a full reset compensator without a
reset band. It is given by
R0 (E, ) = Rb (j ) +
Ar
1
1
Cr e Ar F (0, ) I + eAr e( j I )() Br (3.97)
+ (j I + Ar )1 (j I Ar )1 eAr + I Br .
R0 (j ) = Rb (j ) +
(3.98)
Note that in general R0 (j ) will depend only on frequency , and R (E, ) both
on the amplitude E and the frequency . Obviously, the expression for FORE as
K
given in (3.93) is a particular case of (3.95) for Rb (j ) = j +a
, Ar = a, Br = 1,
Cr = K. In addition, the CI FD expression in (3.94) is a particular case for K = 1
and a = 0.
135
line to the graph of the error signal at that point. Therefore, the following expression
is obtained
tan =
de(t) e(t)
=
,
dt
h
(3.99)
where h is the time delay process. From this equation, the new variable reset condition is expressed as
h
de(t)
+ e(t) = 0.
dt
(3.100)
As it can be seen, no parameter has to be fixed in this new condition, since the
time delay h can be estimated from plant tests (usually a worst-case estimate is a
good option). With this new reset condition (3.100), the reset controller (3.90) is
expressed now in the state-space as
/ Bhv ,
if (e(t), e(t))
Bhv ,
xr (t + ) = A xr (t)
(R)
(3.101)
R2 | he(t)
+ e(t) = 0 .
(3.102)
In this case, the reset surface is a continuous function of the error signal, as it
can be seen in Fig. 3.24. When there is no time delay, h = 0, the reset band surface
(3.102) is reduced to the reset compensator without reset band.
In the frequency domain, the describing function of the Clegg integrator with a
variable reset band may be computed, resulting in
j2
1
1
CI h () =
1+
1 + ej 2 tan (h) ,
(3.103)
j
136
where h is the time delay of the process. Note that in the case of the variable reset band the describing function only depends on the input frequency and that, in
contrast to CI, CI h gives a frequency dependent phase lead. Figure 3.25 shows the
phase of CI h for different values of h. It is clear from this figure that CI h gives an
1
137
Fig. 3.26 Time simulation of feedback system with FORE and first order plant (note that e = y)
functions depend on the parameter /E, where /E (0, 1) defines the reset band
in relation to the compensator input amplitude, if a limit cycle exist for some then
there always exist limit cycles for any value of E such that /E (0, 1). Thus,
a normalized reset band /E and a describing function with amplitude E = 1 will
be used.
As a result, the critical locus 1/R/E (1, ) will be used to evaluate the crossings, and then values of /E and will be obtained instead. It turns out that for each
given frequency , critical loci 1/R/E (1, ) are circle segments in the Nyquist
plane as /E goes from 0 to 1.
3.4.3.1 FORE with a First Order Plant
A simple analysis shows that there are no limit cycles in the case corresponding to
a feedback control system (see Fig. 2.1) with a FORE reset compensator having a
stable base compensator and a first order stable plant P (s). The reason is that the
after-reset state of the closed-loop system will always correspond to a zero compensator output, and an alternating value of the error , , . . . is obtained. The only
case in which a limit cycle may exist is the case in which the error signal reaches
the reset band from each one of the two values and , and this is only possible if
the base feedback system is unstable.
In this example, a FORE compensator with an unstable base compensator
2
1
FOREb (s) = s1
, = 0.1, and a first order plant P (s) = s+0.5
is considered. A simulation of the feedback system reveals (see Fig. 3.26) that a limit cycle exists with
amplitude E = 0.20 and frequency = 1 rad/s.
138
Fig. 3.27 Crossings of critical loci 1/FORE/E (1, ) for = 0.5, 1, 1.5 rad/s with Nyquist plot
of the first order plant
In Fig. 3.27, the crossings of critical loci 1/FORE/E (1, ) for = 0.5, 1,
1.5 rad/s with P (j ) are shown. Then, it is possible to evaluate numerically those
crossings corresponding to a limit cycle, and in addition values of /E and . The
result is that there exist limit cycles for any value of E such as /E = 0.68 and for
= 1 rad/s. In this example, the amplitude will be E = /0.68 = 0.1/0.68 = 0.15.
Note that the describing function method gives a very accurate prediction of the
limit cycle. In addition, a perturbation method can be used to analyze the (local)
stability of the limit cycle. It can be shown that for the limit cycle to be stable a
sufficient condition is
R(E, ) I (E, ) R(E, ) I (E, )
>0
E
(3.104)
139
Fig. 3.28 Crossings of critical locus 1/CI b (E, ) with Nyquist plot of the plant
cycle for any value of E such as /E = 0.82 and = 0.92 rad/s. For this example,
the amplitude is E = /0.82 = 0.3. Again, the prediction of the limit cycle is very
accurate. The deviation of the amplitude is consistent with the fact that signal v
in Fig. 3.28 may have significant higher order harmonics. In addition, condition
(3.104) is satisfied for (E0 , 0 ) = (0.3, 0.92), and thus the limit cycle is also stable.
140
Fig. 3.29 Crossings of critical locus 1/CI b (E, ) with Nyquist plot of the second order plant
fundamental period 2/ (see Fig. 3.21). In addition, it is also symmetric and can
be expressed as
v(t) =
a 2 +2
if t [0, ],
y(t )
if t [ , 2
],
(3.105)
K
|.
where M() and () are the magnitude and phase of | j +a
Now, computation of high order harmonics is relatively simple in this case since
by symmetry only odds harmonics are different from zero. Defining C(E, , ) =
KE( cos(t )+a sin(t ))
(1 + ea/ ), higher harmonics have a particularly simple
(a 2 +2 )
form. After some computation, spectral coefficients of the Fourier series v(t) =
ak =
C(E,,)
a+j k
for k = 3, 5, . . . ,
for k = 2, 4, . . . .
(3.106)
141
Fig. 3.30 Sinusoidal response of FORE for = 1 rad/s, K = 1, and a = 1, and approximations
with the first plus third harmonics, and with 20 harmonics
Therefore, by using the first n odd harmonics of FORE response as an approximation, a measure of the approximation can be defined as the L2 [0, 2
]-norm
of the residual term. Then, v(t) as given by (3.105) can be expressed by v(t) =
vn (t)+ vn (t) where vn (t) = nk=n a2k1 ej (2k1)t , and vn (t) = v(t)vn (t). Then
vn 2 = 2
k=n+1
|a2k1 |2 = 2
k=n+1
a2
C
,
+ (2k 1)2 2
(3.107)
where C stands for C(E, , ) (since C does not depend on k, the arguments have
been dropped for simplicity). By using the function (also known as the digamma
function), the norm can be computed in a closed-form as
+ ia
ia
C
+ ia
C
n+
n+
=
Im n+
vn =
i2a
2
2
a
2
(3.108)
for a = 0. Figure 3.31 shows a measure of how low order harmonics represent the
vn
sinusoidal FORE response. Note that in general the ratio v
only depends on the
2
142
vn
v )
Im{ (n +
1
2
a
+ i 2
)}
a
Im{ ( 12 + i 2
)}
1
2
(3.109)
n
+
vn 2 =
n
+
n
+
=
,
2
2
2
i22
2
(3.110)
vn
is constant (independent of the frequency ), and is
and in this case the ratio v
given by (it has the same values as that of FORE for |a/| = 1, see Fig. 3.32)
1
1
2
1
(1
+
i)
=
0.8331
Im
n
+
.
1
2
Im{ ( 2 (1 + i))}
(3.111)
For the example of FORE plus a first order plant, the low-pass condition can
be now evaluated by computing how the plant filters the harmonics of the FORE
sinusoidal response. In Fig. 3.33, it is shown how for different values of |a/| the
relative weight of the first harmonic is much more significant, and it is a good approximation of the filtered FORE response. For the example of CI with a second
order plant, the results are analogous.
Note that the harmonics structure of the FORE or the CI response is always
given as shown in Fig. 3.31, then the low-pass condition can be evaluated for a
vn
=
v
Im{ (n + 12 (1 + i))}
143
Fig. 3.32 Sinusoidal response of CI for = 0.92 rad/s, and approximations with fundamental
component, with the first plus third harmonics, and with 20 harmonics
vn
v )
144
given plant by computing how the harmonics are filtered, and then evaluating how
the plant modifies the harmonics structure, for example, as shown in Fig. 3.33. This
figure shows that the relative error norm of the first harmonic approximating the
FORE response is around 10%, and that the relative error considering both the
first and third harmonics is around 5%, thus it may be expected that the describing
function method works well in many cases.
References
1. Aangenent, W.H.T.M., Witvoet, G., Heemels, W.P.M.H., van de Molengraft, M.J.G., Steinbuch, M.: An LMI-based L2 gain performance analysis for reset control systems. In: Proceedings of the American Control Conference (2008)
2. Bainov, D.D., Simeonov, P.S.: Systems with Impulse Effect: Stability, Theory and Applications. Ellis Horwood, Chichester (1989)
3. Baos, A., Carrasco, J., Barreiro, A.: Reset times-dependent stability of reset control system.
In: European Control Conference, Kos, Greece (2007)
4. Baos, A., Carrasco, A., Barreiro, A.: Reset times dependent stability of reset systems with unstable base system. In: Proc. IEEE International Symposium on Industrial Electronics, Spain
(2007)
5. Baos, A., Dormido, S., Barreiro, A.: Stability analysis of reset control systems with reset
band. In: 3rd IFAC Conference on Analysis and Design of Hybrid Systems, Zaragoza, Spain
(2009)
6. Baos, A., Dormido, S., Barreiro, A.: Limit cycles analysis in reset systems with reset band.
Nonlinear Anal. Hybrid Syst. (2010). doi:10.1016/j.nahs.2010.07.004
7. Baos, A.: Carrasco, J., Barreiro, A.: Reset times-dependent stability of reset control systems.
IEEE Trans. Autom. Control 56(1), 217223 (2011)
8. Bemporad, A., Bianchini, G., Brogi, F.: Passivity analysis and passification of discrete-time
hybrid systems. IEEE Trans. Autom. Control 53(4), 10041009 (2008)
9. Beker, O.: Analysis of reset control systems. Ph.D. Thesis, University of Massachusetts,
Amherst (2001)
10. Beker, O., Hollot, C.V., Chait, Y., Han, H.: Fundamental properties of reset control systems.
Automatica 40, 905915 (2004)
11. Camlibel, M.K., Heemels, W.P.M.H., Schumacher, J.M.: On linear passive complementary
systems. Eur. J. Control 8(3), 220237 (2002)
12. Carrasco, J., Baos, A., van der Schaft, A.: A passivity approach to reset control of nonlinear
systems. In: 34th Annual Conference of the IEEE Industrial Electronics Society, Orlando,
Florida, USA (2008)
13. Carrasco, J., Baos, A., van der Schaft, A.: A passivity-based approach to reset control systems
stability. Syst. Control Lett. 59(1), 1824 (2010)
14. Bergen, A.R., Franks, R.L.: Justification of the describing function method. SIAM J. Control
9, 568589 (1971)
15. Fernndez, A., Barreiro, A., Baos, A., Carrasco, J.: Reset control for passive bilateral teleoperation. IEEE Trans. Ind. Electron. (2010). doi:10.1109/TIE.2010.2077610
16. Guo, Y., Wang, Y., Xie, L., Zheng, J.: Stability analysis and design of reset systems: theory
and an application. Automatica 45(2), 492497 (2009)
17. Haddad, W.M., Chellaboina, V., Nersesov, S.G.: Impulsive and Hybrid Dynamical Systems:
Stability, Dissipativity, and Control. Princeton University Press, Princeton (2006)
18. Horowitz, I.M., Rosenbaum, P.: Nonlinear design for cost of feedback reduction in systems
with large parameter uncertainty. Int. J. Control 24(6), 9771001 (1975)
19. Khalil, H.K.: Nonlinear Systems, 2nd edn. Prentice Hall, New York (2002)
References
145
20. Krishman, K.R., Horowitz, I.M.: Synthesis of a nonlinear feedback system with significant
plant-ignorance for prescribed system tolerances. Int. J. Control 19(4), 689706 (1974)
21. Lozano, R., Brogliato, B., Egeland, O., Maschke, B.: Dissipative Systems Analysis and Control. Springer, London (2000)
22. Mees, A., Bergen, A.: Describing functions revisited. IEEE Trans. Autom. Control 20(4),
473478 (1975)
23. Michel, A., Hu, B.: Towards a stability theory of general hybrid dynamical systems. Automatica 35, 371384 (1999)
24. Neic, D., Zaccarian, L., Teel, A.R.: Stability properties of reset systems. Automatica 44(8),
20192026 (2008)
25. Pogromski, A., Jirstrand, M., Spangeus, P.: On stability and passivity of a class of hybrid systems. In: Proceedings of the 37th IEEE Conference on Decision and Control, Tampa, Florida,
USA (1998)
26. Rockafellar, R.T., Wets, J.-B.: Variational Analysis. Series of Comprehensive Studies in Mathematics, vol. 317. Springer, Berlin (1998)
27. Rugh, W.J.: Linear System Theory, 2nd edn. Prentice Hall, New York (1996)
28. Sanders, S.R.: On limit cycles and the describing function method in periodically switched
circuits. IEEE Trans. Circuits Syst. I 20(4), 473478 (1975)
29. Van der Schaft, A.: L2 -Gain and Passivity Techniques in Nonlinear Control. Springer, Berlin
(2000)
30. Vidal, A., Baos, A.: QFT-based design of PI + CI compensators: application in process control. In: IEEE Mediterranean Conference on Control and Automation, pp. 806811 (2008)
31. Vidal, A., Baos, A.: Reset compensation for temperature control: experimental application
on heat exchangers. Chem. Eng. J. 159(13), 170181 (2010)
32. Vidyasagar, M.: Nonlinear Systems Stability. Prentice-Hall, London (1993)
33. Willems, J.C.: Dissipative dynamical systemsPart I: General theory. Arch. Ration. Mech.
Anal. 45, 321351 (1972)
34. Wright, R.A., Kravaris, C.: On-line identification and nonlinear control of an industrial pH
process. J. Process Control 11, 361374 (2001)
35. Zefran, M., Bullo, F., Stein, M.: A notion of passivity for hybrid systems. In: Proceedings of
the 40th IEEE Conference on Decision and Control, Orlando, FL, USA (2001)
Chapter 4
4.1 Introduction
Since the early proposals for reset control, including the Clegg integrator [13] and
the first-order reset element (FORE) [20], [15], the underlying idea was to achieve
fast and robust control solutions for problems under linear limitations. The simple
nonlinear effect of resetting was proved to be useful for overcoming linear fundamental limitations.
As it is well known in the literature on fundamental limitations [1] (see also
Chap. 1 of this monograph), linear plants that have time-delays (and also plants
having right half-plane poles or zeros) are specially hard to control: if the delay is
significant, it is impossible to obtain a linear controller providing simultaneously a
fast and a robust behavior. One has to choose between a fast response but with poor
stability margin, or a robust stability margin, but with slow response. As a consequence, since reset control is a simple nonlinear method for overcoming fundamental limitations, and since time-delay is one of the main sources of such limitations,
it is of great interest to study the problem of delayed reset systems.
A design idea to obtain a reset controller for a plant with time-delay can be
taken from the design approach proposed in [15] and [20]: first, tune the base LTI
controller so that the base LTI feedback loop becomes stable and satisfies part of
the performance specifications (typically, fast transient and large bandwidth). Then,
in a second step, a reset action is included to achieve the other part of the desired
objectives (typically, bounded overshoot and large stability margins).
However, this method should be carefully applied since it is well known [8] that
the reset action may destabilize a well designed LTI base control system, and this
fact is one of the main drawbacks for the practical application of reset control.
A solution to this problem has been reported in [8, 9] for finite-dimensional LTI
plants. The result is a frequency domain condition, the H -condition (see the previous Sect. 3.2.1.1 of this monograph), that guarantees closed-loop stability of the
reset system.
Since then, a number of useful results have appeared addressing the analysis and
design of reset systems [6], [17], [18], [25]. However, none of them considered the
A. Baos, A. Barreiro, Reset Control Systems, Advances in Industrial Control,
DOI 10.1007/978-1-4471-2250-0_4, Springer-Verlag London Limited 2012
147
148
case of time-delay plants, so it is then of great interest to study delayed reset systems.
It should be emphasized that this is extremely difficult: the combination of reset and
delay moves the problem to the field of hybrid and distributed systems. Even the
issue of existence and uniqueness of solutions may cause trouble unless special care
is taken, for example, by means of the time-regularization procedure [25].
It should be noted that stability of delay impulsive systems is a very active
research area in different areas of mathematics and systems theory, two classical monographs are [2] and [21]. In general, as reported in [2] (see the previous
Sect. 1.5.1 of Chap. 1 in this monograph), impulsive systems may be classified into
(i) systems with impulses at fixed instants, (ii) systems with impulse effect at variable instants, and (iii) autonomous systems with impulse effects. Systems (ii) and
(iii) have state-dependent impulses, but there is an important difference: the reset
surface is time independent in the case (iii). Traditionally, most of the research effort has been dedicated to cases (i) and (ii), and thus case (iii) that includes reset
control systems is less developed. In addition, in the control field, the main effort
has been directed at systems without time delays [28], [18]. For the case of impulsive delay systems, available results in the literature are concentrated in the case (i)
(see [22], [14], [24], and the references therein), and case (ii) (see [23]).
The material in this chapter is based on the journal papers [5] and [7], and on
the preliminary conference papers [3] and [4]. Some of the results in these references can be regarded as generalizations of the H -condition for LTI plants with
time-delays. In [5], the case of delay-independent stability was considered, a standard LyapunovKrasovskii functional was used, and the instrumental technique for
passing between time and frequency domains was the KalmanYakubovichPopov
(KYP) lemma. In [7], the delay-dependent case was addressed. As the combination
of delay-dependent LMIs with the KYP lemma may give rise to extremely large matrix inequalities, which do not show much insight into the results, special care was
taken in the choice of appropriate Lyapunov functionals and passivity/frequency
domain techniques.
There are several LyapunovKrasovskii (LK) functionals that can be used, so
that the delay h appears explicitly in the stability conditions. Among these delaydependent results, the complete LK functional (Theorem 5.19 in [16]) has the advantage that it gives a sufficient but also necessary condition for stability; however,
it is a distributed functional and requires some kind of discretization for practical
computation.
Another useful result is given by Proposition 5.17 in [16], with good LMI computability and which is not too conservative, in general. This functional was studied
in [4] for application to reset systems. Finally, for interpretability in terms of uncertainty, the LK functional in [26] has the advantage that it allows for an easy
interpretability (detailed in [29]) so that the stability conditions can be understood
in terms of scaled small-gain for an uncertain plant. In the delay-dependent case,
for passing from the time-domain to the frequency-domain, in place of the involved
KYP techniques in [5], a novel interpretation was presented in [7] based on passivity
and impulsive control. In this way, all the stability conditions can also be interpreted
in terms of robust stability against certain uncertainties.
149
The organization of the chapter is as follows. After introducing formally the problem and basic results in Sect. 4.2, the time-domain delay-independent stability of reset systems is studied in Sect. 4.3 with the equivalent frequency domain conditions
and interpretations presented in Sect. 4.4, followed by an example in Sect. 4.5. Then,
the delay-dependent stability results are considered in Sect. 4.6 with the frequency
domain counterpart in Sect. 4.7, including a passivity analysis of reset systems and
robustness interpretations. The chapter concludes with an illustrative example of the
delay-dependent case in Sect. 4.8.
x r = Ar xr + Br e,
xr (t + ) = A xr (t),
(4.2)
(R)
v = Cr xr + Dr e,
where the second equation of the controller represents the impulsive or reset actions,
applied when a certain reset condition, to be specified later, holds. If the connection
from P0 to R (or from R to P0 ) is affected by a delay h1 (or h2 ), then it is easy to
see that the total delay h = h1 + h2 can be moved to the plant input or output.
For instance, move the delay to the plant input, as shown in Fig 4.1. As the
main interest is stability, let us suppose that the exogenous loop signals are zero
(r = d = n = 0), so that the closed loop connections are:
u(t) = v(t h),
e(t) = y(t),
(4.3)
(4.4)
150
Since the results in this chapter are independent of the precise structure of A, Ad ,
let us assume in what follows, for the sake of generality, that A, Ad are arbitrary
matrices, not necessarily coming from a control loop like the one in Fig. 4.1.
Now, to complete the system equations, it remains to introduce AR and M . Let
us suppose that only the second part of the controller states are reset to zero, that is,
A = diag(1, . . . , 1, 0, . . . , 0). Then AR takes the form [9]:
AR = diag(In12 , 0n3 ),
(4.6)
so, the reset action acts over the last n3 states of the state x Rn . These dimensions
are defined by:
n = n1 + n2 + n3 = n1 + n2 +n3 ,
n23
(4.7)
n12
where n1 (n23 ) is the number of states of the plant (controller). Accordingly, partition the state vector:
(4.8)
x = x
1 , x2 , x3 .
x
p
x
r
In principle, it is considered that the reset is applied at the time t when x(t) hits
the hypersurface
(4.9)
M = x Rn : Cx = 0 .
As it is desired that 0 = Cx defines a hypersurface, C is a row vector, C R1,n .
From the state partition (4.8), the reset condition matrix C can be partitioned as:
C = (C1 , C2 , C3 ).
There are several possibilities for C. If the IDDE system comes from a control
loop like the one in Fig. 4.1, and the reset condition is activated when the error
crosses zero, then C = (Cp , 0, 0). The most general case is an arbitrary row matrix C, but it causes some difficulties if C3 = 0, that is, if the part x3 of the controller
states affected by reset have influence via C3 on the reset condition. For this reason,
several results in this chapter assume the structure C = (C1 , C2 , 0) for C.
151
In this way, (4.5), (4.6), and (4.9) define, in principle, the system that we are
going to study. However, it is known that the IDDE (4.5) may exhibit beating and
have Zeno solutions as has been remarked in [25] for the case h = 0. A certain
strategy for avoiding these pathological solutions has been analyzed in Chap. 2 of
this monograph. This problem of well-posedness of solutions affects all hybrid systems, and in particular reset systems. Note that the well-posedness issue has been
addressed in Chap. 2 for finite-dimensional reset systems. However, the existence of
the time-delay complicates the situation further because it makes the system infinitedimensional. In this chapter, the useful idea of time regularization (TR) will be used,
and thus the IDDE (4.5) is recast in the form (IDDE-TR):
= 1,
+ = 0,
if (x(t)
/ M ) ( ),
if (x(t) M ) ( > ),
(4.10)
where is a finite positive scalar, arbitrarily small. In this way, every reset in (4.10)
will take place at least seconds after the previous reset instant. That is, the reset
instants tk (k = 1, 2, . . .) satisfy tk+1 tk > (k = 1, 2, . . .). Choosing a sufficiently
small time-step the behavior the IDDE and IDDE-TR systems (4.5) and (4.10) is
almost equal for most practical systems, and in this way the regularized system is
formally free of dynamic pathologies.
Now, it is important to distinguish between the (lumped) instantaneous state
x(t) Rn and the true (distributed) state given by some piecewise continuous function xt : [h, 0] Rn , where xt ( ) = x(t + ), [h, 0]. In the following,
C([h, 0], Rn ) will stand for the set of piecewise continuous functions mapping
[h, 0] to Rn , with a finite number of discontinuity points.
For the sake of concreteness, the piecewise continuous functions in
C([h, 0], Rn ) will be assumed continuous from the left at the discontinuity points,
that is, x(tk ) = x(tk ) and, after a reset, x(tk+ ) = AR x(tk ).
Then, the existence and uniqueness of solutions to (4.10) is simply guaranteed
in a constructive way. First, as the counter starts at = 0, there exists a unique
solution to the following initial value problem:
x = Ax(t) + Ad x(t h) for t 0,
(4.11)
for [h, 0],
x( ) = x0 ( )
where x0 () is the initial condition. The starting trajectory of the linear system (4.11)
is then well defined and is not affected by the first reset until some instant t1 > .
Now, in a recursive way, if the trajectory is well defined until some reset instant
t = tk , then introduce the reset jump x(tk+ ) = AR x(tk ) and after the reset, solve a
new linear problem similar to (4.11). In a constructive way, the solution is
x(t) = e
A(ttk )
AR x(tk ) +
eA(ttk ) Ad x(tk + h) d,
(4.12)
and this solution is valid for all t (tk , tk+1 ], up to reaching the new reset time
tk+1 > tk + .
152
For the stability analysis of (4.10), let us rewrite the system in a form that makes
explicit only the state x and implicit the time counter :
x (t) = Ax(t) + Ad x(t h),
(4.13)
(IDDE-TRx )
x(t + ) = AR x(t) only if x(t) M .
This should be understood as equivalent to (4.10) but we only pay attention to
x(t), driven by the continuous dynamics most of the time, and subject to reset only
if x(t) M (note that the iff condition is not used, due to the fact that < may
disable the reset).
Now, let us introduce a basic result to address the stability of time-delay reset
systems. It is common practice in stability analysis of time-delay systems [16] to
start with a LyapunovKrasovskii functional V (xt ) and force its derivative to be
negative, V (xt ) < 0, along the solutions of the system.
If the system also contains jumps or resettings, then it is natural to impose that
the variation of the functional at the reset instants t = tk has to be nonpositive,
V (t) = V (xt + ) V (xt ) 0.
There are other possibilities, different from imposing V < 0 and V 0 as basic
stability conditions. These requirements may be conservative, in general, and within
the area of impulsive systems there is a vast amount of literature offering alternative
techniques. But, at the same time, notice that many reset design approaches (see,
e.g., [15], [20], [9], [18]) are based on the useful idea of tuning the linear system
to be fast transient, underdamped but stable, and introducing reset for adding extra damping and decreasing overshoot. Within this idea, both the continuous and
the impulsive mode must be stable, and so the basic stability principle (V < 0 and
V 0) is very well suited and adequate.
This principle is formalized in the following proposition taken from [7] and based
on the results in [9] for systems without time-delays.
Proposition 4.1 Let the functional V (xt ) : C([h, 0], Rn ) R be continuously
differentiable, positive-definite, radially unbounded and such that along the solutions of the reset system
d
V (xt ) < 0,
dt
xt = 0,
V = V (xt + ) V (xt ) 0,
x(t) M .
(4.14)
(4.15)
153
for some P , Q > 0. It is clear that V (t) 0 and that V (t) = 0 if and only if
xt ( ) = 0 Rn , for each [h, 0]. Thus the functional is positive-definite, and
also differentiable and radially unbounded. It is clear as well that the zero state
xt = 0 is a fixed point, or equilibrium point, of the reset system. The question now
is whether this equilibrium is stable or not. First, let us formalize a little the sense
of stability to be used in relation to the aforementioned quadratic functional.
Definition 4.1 The IDDE reset control system (4.5) is said to be quadratically
(asymptotically) stable when there exist P , Q > 0 such that the functional V in
(4.16) satisfies the conditions (4.14) and (4.15).
These two inequalities ensure that there exists a common Lyapunov function for
the system (4.5) without reset (the base LTI system), using (4.14), and for the reset
system xt + (0) = AR xt (0) using (4.15).
Proposition 4.2 The system (4.5) is quadratically stable if and only if there exist
matrices P , Q > 0, of size n n, satisfying the following conditions:
A P + P A + Q P Ad
< 0,
(4.17)
A
Q
dP
and
A
R P AR P 0
(4.18)
154
Proof Condition (4.17) is commonly used in delay-independent stability of timedelay systems [16], and easily follows from the fact that (4.16) can be written as
V (xt ) = x(t) P x(t) +
x (t + )Qx(t + ) d,
x(t)
x(t h)
A P + P A + Q
A
dP
P Ad
Q
x(t)
,
x(t h)
from which (4.17) is finally derived. It remains to prove that condition (4.18) follows
from (4.15). Since the reset action is only active when xt (0) M , and thus does
not affect the delay buffer x(t + ), for any (h, 0), the integral part of the
Lyapunov functional (4.16) does not contribute to the jump, thus the reset jump in
the Lyapunov functional V in (4.15) results in that
V = x
t (0) AR P AR P xt (0) 0,
for every xt (0) M = Ker C, from which (4.18) directly follows.
Note that the reset action adds condition (4.18) to the standard condition of delayindependent stability given by the linear matrix inequality (LMI) (4.17). Since the
matrix AR has a very particular structure, this new condition can be considered as
an extra LMI, or simply interpreted as restriction over the structure of the matrix P
that is a solution of the LMI.
Let us study the structure of condition (4.18). First, notice that from (4.6) AR has
the form
AR = diag(In , On ),
(4.19)
n = n n = n1 + n2 ,
so that n is the number of (controller) states affected by reset and n is the number
of (plant and controller) states not affected by reset. Now, partitioning P according
to (4.19), we have:
P1 P2
P1 0
,
(4.20)
A
P
A
=
A
=
A
R
R
R
R P P
0 0
3
2
155
(4.21)
where = (1 2 ). Now, assume here that C has the form C = (C1 , 0) with C1
of size (1 n ), as typically the last n states do not influence the reset condition.
Notice that in this case x = (x
1 , x2 ) belongs to Ker C = Ker(C1 , 0) if and only if
C1 x1 = 0. Furthermore, (4.21) is nonpositive when x
1 P2 x2 = 0 for all x1 , x2 such
that C1 x1 = 0. This amounts to P2 = C1 M for some matrix M , and hence the
next proposition easily follows.
Proposition 4.3 The system (4.5), with C = (C1 , 0) and C1 a row of length n n ,
is quadratically stable if and only if there exist matrices P , Q > 0, with
C1 M
P1
(4.22)
P=
M C1
P3
that satisfy (4.17), where M is some (column) matrix of size n 1.
156
(4.25)
(4.28)
for some column matrix M of dimensions n 1, and some P3 > 0. This restriction
can be finally recast as
B0 P = C0
for B0 = [0n In ] and C0 = [M C1 P3 ].
(4.29)
157
(4.30)
(4.31)
M0
where A = A + I . So, using KYP, since (A, B0 ) is controllable, > 0 such that
(j I A )1 B0
(j I A )1 B0
0 for all R {}.
M0
I
I
(4.32)
Notice that from the last equation (statement (i) in KYP) we recover (4.31) (statement (ii) in KYP), but with P = P in place of P > 0. However, since A is Hurwitz, the existence of a symmetric P such that A P + P A = Q < 0 implies [19]
that such P is positive-definite, i.e., P > 0. Hence, the last two equations (4.31) and
(4.32) are completely equivalent.
From the last condition, putting H (j ) = C0 (j I A)1 B0 and noticing that
j I A = j I (A + I ) = (j )I A, the positive real condition required for H (s ), for some small > 0, is obtained. In other words (see [19]), it
amounts to the strictly positive real (SPR) requirement on the system H (j ) given
by the triple (A, B0 , C0 ), which is precisely the so-called H -condition in [9].
Remark 4.1 It can be seen that the previous SPR condition, characterized in [19], is
equivalent to the Output Strictly Passive (OSP) condition as defined in [10]. More
precisely, since A is Hurwitz and H is strictly proper, the SPR condition amounts,
following [19], to H (j ) + H (j ) > 0 for all finite , and for to
lim 2 H (j ) + H (j ) > 0.
On the other hand, [10] characterizes the OSP condition, with A Hurwitz and
(A, B0 ) controllable, by
H (j ) + H (j ) |H (j )|2
for some > 0 small enough. By expanding H (s) = H1 /s + H2 /s 2 + and
putting s = j with , the equivalence between the two previous conditions,
that is, the equivalence between the SPR and OSP requirements for a strictly proper
H (s) can be verified.
We will find again the OSP condition later, in Sect. 4.7, where the passage from
time to frequency domain, instead of using KYP, will be solved by an approach
based on approximating the resets by suitable Dirac delta pulses.
158
Q
M M = C0
0
C0
0
0
0
Q + I
(4.33)
where
C0 = [M C1 P3 ].
Proof The treatment of the reset-delay conditions in (4.17) and (4.18) can follow a
completely parallel manipulation to conditions in Proposition 4.5. First, notice from
Proposition 4.3 that (4.18) is equivalent to (4.22). This condition has already been
reformulated in the form
(0 I ) P = (M C1 P3 ),
B0
(4.34)
C0
159
A P + P A + Q
P Ad P B0 C0
0, (4.36)
P > 0, > 0,
A
Q + I
0
dP
B0 P C0
0
0
where A = A + I . Now, reproducing (4.30) (4.31), and simultaneously interchanging the second and third row-blocks and column blocks, we obtain that
> 0:
B1
Q
C0
0
A P + P A P (B , A )
0
d
C0
+
(4.37)
0
0
0.
B
0
0
0
Q + I
P
0
A
d
M
The last equation is ready for KYP, reaching to the frequency condition for
quadratic stability: > 0,
(j I A )1 B1
(j I A )1 B1
0 for all R {},
M
I
I
(4.38)
where M , B1 are given in (4.37). This concludes the proof.
The matrix M could be diagonalized in the form M = F J F , with J diagonal with entries 1 or 0. Thus, depending on the signature of M (the number of
positive, zero, and negative eigenvalues), condition (4.38) might be presented more
suitably as a small-gain frequency condition, or as a positive-real frequency condition.
Bd = Ad ,
d Dh ,
(4.39)
for all .
(4.40)
160
The impulsive or reset action is implemented by the lower loop which is zero most
of the time. It only works at the reset instants when the reset states (the last n
states) are switched to zero. This resetting is easily shown to be equivalent to the
application of Dirac impulses () in the following form. Consider the reset system
Rr from yr to ur defined by:
(Rr ) ur (t) =
(t tk )P31 yr (t),
(4.41)
k
with
Br = (0, In ) ,
Cr = (M C1 , P3 ),
t (tk , tk + ), < .
tk+
tk
ur = Br P31 yr (tk )
0
=
P31 (M C1 , P3 )x(tk ).
I
Now, since at the reset instants
0 = Cx(tk ) = (C1 , 0)
x12 (tk )
x3 (tk )
= C1 x12 (tk )
where the state has been partitioned making explicit the last n coordinates x3 (the
coordinates affected by reset),
0
0
=
,
x = x tk+ x tk =
x3 (tk )
P31 P3 x3 (tk )
161
which is the precise change that produces the required reset. Thus so far we have
proved that the IDDE system can be realized in the block diagram form given by
Fig. 4.2. To complete the first step, notice that the resetting subsystem Rr : yr
ur satisfies
u
yr (tk )P31 yr (tk ) 0,
(4.42)
r ( )yr ( ) d =
0
so, as shown in Fig. 4.2, the resetting block can be treated as a particular case,
r Rr of a general uncertainty r : yr ur satisfying the positivity condition:
ur = r yr
u
(4.43)
r ( )yr ( ) d 0.
0
In the second step, the interpretation of the upper loop condition is provided. The
quadratic stability of the continuous dynamics amounts to (4.14), V < 0, and to the
first LMI in (4.35). This LMI can be treated within the KYP lemma, with M =
diag(Q, Q ), for Q = Q I , and it is easy to show that it is equivalent to
Gd (j )
- 1/2
1
- < 1 for all ,
-Q (j I A ) Ad Q1/2
(4.44)
where Gd (s) is the open-loop nominal (d = 0) transfer function of the upper loop
in Fig. 4.2 (from ud to yd ). In this way, (4.44) is a Q-scaled strict small-gain condition on the uncertain upper loop. The small-gain condition is strict due to the
presence of the small > 0 in A = A + I and Q = Q I .
It is important to notice that stability is achieved for all values of the delay 0
h < . This means, in particular, stability for h = 0 (so that A + Ad is stable)
and for h (so that A stable). Notice also that, from (4.35), A P + P A
Q 2P < 0, so that x P x is a Lyapunov function for the system x = Ax.
In the third and final step, an interpretation of the lower loop condition is given.
As the lower loop performs the resettings, the required condition (4.15) is V 0
at the reset instants. It is equivalent to the second condition in (4.35), rewritten now
as Br P = Cr . Furthermore, it amounts (given A P + P A < 0) to:
P Br Cr
A P + P A
0.
(4.45)
> 0,
Br P Cr
0
Recasting the previous LMI via the KYP lemma, it easily follows
> 0,
0 Gr (j ) + Gr (j ),
(4.46)
where Gr (s) = Cr (j I A)1 Br is the open-loop nominal (r = 0) transfer function of the lower loop in Fig. 4.2 (from ur to yr ). In this way, (4.45) is a strictly
positive real (SPR) condition on the uncertain lower loop.
Summarizing, the LMIs (4.37) and (4.38), expressed in terms of large matrices,
have been interpreted here in terms of structural properties of the IDDE system.
162
A=
2
0
,
0 0.9
Ad =
0.25
0
,
0.25 0.25
(4.47)
0.1
0.05
.
0.05
0.1
Q
M = C0
0
C0
0
0
0
Q + I
163
0.1
0.05
0
0
0
0.05
0.1
0.1
0
0
0.1
0
0
0
= 0
,
0
0
0
0.09 0.05
0
0
0
0.05 0.09
and finally, the frequency response matrix H (j ) resulting from the KYP lemma
(4.38) is
H (j ) = G (j )MG(j ),
where
G(s) =
(j I A )1 B1
I
164
Fig. 4.4 Simulation of the delay-independent example: x1 (t) (dashed) and x2 (t) (solid)
Notice the behavior as in Fig. 4.3: G(j ) tends to (023 , I33 ) , and
thus H (j ) tends to the 3 3 submatrix of M formed by its last three rows and
columns. This submatrix has a zero eigenvalue, and this explains the high frequency
limit in Fig. 4.3.
Figure 4.4 shows a simulation of the example with a delay h = 2 s, and with
the initial conditions x1 ( ) = x2 ( ) = 0 for [2, 0) and x1 (0) = 10, x2 (0) = 0.
It can be seen that initially (t < h = 2) the state x1 (t) tends to zero with the stable dynamics given by A = diag(2, 0.9). But when t = h = 2 s, the influence
of Ad x(t h) appears and the trajectories x(t) are affected, giving rise to several
resets at the times t1 2.1, t2 4.6, t3 7.3, . . .. These reset times are defined by
x1 (tk ) = 0. The simulation of this reset system for h = 2 s shows a stable behavior.
Since we have proven that this system is stable independently of the delay h, we can
as well predict stability for all fixed h 0.
165
of the basic Proposition 4.1, the derived stability conditions depend explicitly on the
delay h. In this way, the stability conditions may hold only for some values of h. This
is more natural since many practical systems are typically stable for small values of
the delay, but unstable for large delays.
In order to gain generality, we assume here for C the structure C = (C1 , C2 , 0),
based on the state partition (4.8) x = (x
1 , x2 , x3 ) and the corresponding dimensions (4.7) n = n1 +n2 +n3 . This structure is equivalent to the one used in the delayindependent case, with the notation equivalence n = n3 and n = nn = n1 +n2 .
The proposed structure C = (C1 , C2 , 0) is more general than C = (C 1 , 0, 0), the
one adopted in [9]. For comparison with this last case, let us state the fact that there
exists an invertible transformation T , T 1 Rnn such that
C
C
T
(C1 , C2 , 0) = (C 1 , 0, 0) diag(T12 , In3 ) .
(4.48)
One possible solution is C 1 = (1, 0, . . . , 0) and T12 with its first row equal to
(C1 , C2 ) and the remaining rows completed for full rank.
Now, the key point to obtain delay-dependent stability conditions is to choose
an appropriate LyapunovKrasovskii (LK) functional. There are several LK functionals that can be used for generating delay-dependent conditions. The complete
LK functional (Theorem 5.19 in [16]) has the advantage that it gives a sufficient
but also necessary condition for stability, however, it is a distributed functional and
requires discretization for practical computation. Another useful result is given by
Proposition 5.17 in [16], with good LMI computability and which is generally not
too conservative. Finally, the LK functional in [26] has the advantage that the stability conditions are interpretable in a robust stability framework, in terms of a certain
small-gain loop [29]. This LK functional takes the integral-quadratic form given by:
V (xt ) = V1 (xt ) + V2 (xt ) + V3 (xt ),
(4.49)
with
1
V1 (xt ) = x (t)P x(t),
2
0
x (t + )Qx(t + ) d,
V2 (xt ) =
h
and
V3 (xt ) =
G(xt+ ) A
d XAd G(xt+ ) d d,
(4.50)
(4.51)
(4.52)
where
G(xt+ ) = Ax(t + ) + Ad x(t + h).
(4.53)
166
has been already treated in [26], and it is finally expressed in an LMI form. The
requirement that V (xt ) 0 when x(t) M can be addressed in a special form,
related to the one in Theorem 5 of [9]. By combining these two approaches, we
obtain the following proposition.
Proposition 4.7 The IDDE reset system (4.5) is globally and asymptotically stable,
if there exist Q > 0, V > 0, W , and
P = P = (Pij )i,j =1,2,3 > 0,
(4.54)
(4.55)
with
W Ad
Q
0
A A
dV
A
d Ad V
V
0
h(W + 12 P )
0
0
V
< 0,
(4.56)
where
N=
1
(A + Ad ) P + P (A + Ad ) + W Ad + A
d W + Q,
2
(4.57)
(4.58)
Then, it can be seen that, after a reset jump x(t) AR x(t), the LK functional
V = V1 + V2 + V3 only changes in its first contribution V1 because the integrals in
V2 and V3 are not affected. In this way, the condition V = V1 0 for x M =
T 1 Imag K amounts to the negative semidefiniteness, P 0, of
P = K T (AR P AR P ) T 1 K,
(P )
(4.59)
167
(4.60)
0
(P ) = 0
P
13
0
0
P
23
P13
P23 .
P33
(4.61)
0
P = K (P )K = 0
K
P13
1
0
0
P23
K1 P13
P23 .
P33
(4.62)
P23 = 0,
and M :
P13
= M C 1 ,
(4.63)
as Imag K1 = Ker C 1 , for some column matrix M of appropriate size. Then, using
T P T = P , we can say that
(0, 0, In3 )P = P13
(4.64)
, P23 T12 , P33 = (M C 1 , 0)T12 , P33 ,
and using (4.48) and P33 = P33 , one immediately obtains (4.55) = (4.64), proving
the fact that (4.55) holds iff P 0 iff V 0 for x M , which concludes the
proof.
T = In , and P = P ,
Remark 4.2 Notice that when C2 = 0, one can take C = C,
which is the case in [9]. In this way, the result here extends the previous results not
only to the delayed case, but also to a more general form of the resetting hypersurface M , with C2 = 0.
Remark 4.3 Usually, the term V3 in (4.52) is expressed by replacing Ax(t + ) +
Ad x(t + h) by x (t + ) [26]. Here, this is possible only in the continuous mode
of operation. At the reset instants, the relation is no longer valid, so that the correct
expression for V3 is the one in (4.52).
Remark 4.4 Delay margin. If the delay h is not given, it is important to estimate the
set of values h Ireset for which (4.54) and (4.56) are feasible. Notice that stability
of the reset system is more restrictive than stability of the linear system, so that
Ireset Ilin . Determining those h Ilin for which x = Ax + Ad x(t h) is stable
168
is well known. It is known that h can have an effect of conditional stability: Ilin =
[0, 1 ) (2 , 3 ) (m1 , m ), with the k easily computable [16].
With this is mind, as Ireset is bounded by Ilin , one can always define a fine mesh
of values {hk }N
k=1 covering this set and check feasibility of the N related LMIs.
Furthermore, (4.56) can be written as F + hG < 0, and one can always force negative definiteness by checking Fk + hk Gk < k I , for some k > 0. Then, it is
easy to show that we have a stable
0 range (hk k , hk + k ) =: Ik provided that
k = k /Gk . Thus, the union k Ik =: Inum provides a numerically guaranteed
stable range Inum Ireset that can be made arbitrarily close to Ireset by increasing
the number of LMIs tested.
(4.65)
with
B = (O, O, In3 ),
H = (M C1 , M C2 , P33 ).
(4.66)
(4.67)
(4.68)
Proposition 4.8 If the conditions (4.55) and (4.56) hold, then the base-linear system BLSio is strictly passive with respect to the storage functional V (xt ) (4.49).
169
Proof The arguments are sketched as follows. Recall that the system BLSio is
strictly passive with respect to V (xt ), by definition, when
y (t)u(t) > VBLSio (xt ),
xt = 0,
(4.69)
for all u(t), y(t), xt solutions to the system. By rebuilding the time derivative of the
LyapunovKrasovskii functional, it can be seen that
1
1
VBLSio = VBLS + x (t)P Bu(t) + u (t)B P x(t),
2
2
(4.70)
where the subindex of the V denotes the system where the time derivative of V is
computed. The last expression is easy to deduce and shows that the only influence
of u in BLSio affecting V = V1 + V2 + V3 is through the term V1 . The derivatives of
the remaining integral terms V2 , V3 can be written [26] as:
V2 = x (t)Qx(t) x (t h)Qx(t h)
and
V3 = hF (t)
F (t + ) d,
with
F (t) = (Ax(t) + Ad x(t h)) A
d XAd (Ax(t) + Ad x(t h)) .
And hence V2 , V3 depend only on past values x(t + ), 0 of the state and not
on the input u(t), so that (4.70) is true. Then, the passivity condition amounts to
y (t)u(t) = x (t)H u(t) > VBLS (xt ) + x (t)P Bu(t).
(4.71)
Now, the only possibility for the previous relation to hold is (setting u = 0) that
VBLS (xt ) < 0 (ensured by (4.56)) and (as u(t) is completely free) that (4.65) is satisfied (equivalent to (4.55)). Notice that (4.56) is only a sufficient condition for
VBLS (xt ) < 0, according to [26]. This is the reason why the only if direction cannot be proved here.
Now, it is useful to show that the overall IDDE (4.5) is passive (and stable) by
proving that it is the negative feedback connection between BLSio and certain suitable subsystem R that performs the resetting actions. To show this, consider the
ideal resetting system R : y(t) u(t) =: v(t) with input y(t) = H x(t) and output u(t) = v(t) given by
1
(R) v(t) =
(t tk )P33
y(t),
(4.72)
k
170
Proposition 4.9 The IDDE system (4.5) is equivalent to the negative feedback interconnection between the base linear system with inputs and outputs BLSio and the
ideal resetting subsystem R.
Proof During the continuous mode, both configurations share the same autonomous
dynamics. At the reset times tk , the IDDE system resets to zero the component x3
of the controller state x
23 = (x2 , x3 ) (the last n3 states). At the same instants tk , the
system R behaves as follows. Notice that
y(tk ) = H x(tk ) = M C1 x1 (tk ) + M C2 x2 (tk ) + P33 x3 (tk ),
(4.73)
(4.74)
so from (4.72) u(t) = v(t) contains, locally around t = tk , a Dirac delta with
weight
1
u(t) (t tk )P33
P33 x3 (tk ) = (t tk )x3 (tk ).
This is the precise input u(t) that, applied to the BLS with input matrix B, per
forms the zero reset of x3 , the last n3 components of the state.
Figure 4.5 illustrates the main idea of the previous proposition. The evolution
of certain reset state x3 (t) is shown in Fig. 4.5 (top). It obeys the equation x3 =
f3 (x(t), x(t h)) and resets to x3 (tk+ ) = 0 at reset instants given by Cx(tk ) = 0.
Then, this behavior is equivalently
modeled by x3 = f3 (x(t), x(t h)) + u(t) with
the feedback law u = k (t tk )x3 (t) with corrective ideal Dirac impulses.
Indeed, this effect can be approximated
by rectangular pulses (t) with width
and height 1/ in the form u = k (t tk )x3 (tk ). Figure 4.5 (bottom) shows the
required train of finite pulses, for = 0.1. These ideas are exploited and formalized
in the following.
Now the question is: Could the stability of the IDDE system be proved with passivity techniques? From Proposition 4.7 (internal stability), one can always derive
an inputoutput frequency domain result, using the KYP lemma, as in Sect. 4.4.
However, it would be useful to find an independent approach based only on passivity. The main problem on the I/O stability of the loop {BLSio , R} is that the Dirac
impulses are not signals in L2 . Then, the only possibility is to consider L2 approximations of ideal impulses. So, consider the system R , y(t) u(t) =: v(t), with
input y(t) = H x(t) and output u(t) = v(t) given by
1
(t tk )P33
|y(tk )| sign(y(t)),
(4.75)
(R ) v(t) =
k
171
Fig.
4.5 (Top) A reset state x3 (t). (Bottom) Equivalent resetting impulses u =
k (t tk )x3 (tk ) ( = 0.1)
because y(t) may vary fast enough in the interval [tk , tk + ), changing its sign.
Connecting R to BLSio , from (4.66) and (4.68), and for t [tk , tk + ), we have
y(t)
= H x (t) = H Ax(t) + H Ad x(t h) +H Bu(t),
=:yX (t)
=:yX (t)
|y(tk )|
sign(y(t)).
(4.76)
To better see how R works, consider the scalar case with y(tk ) =: yk > 0, then
y(t)
= yX (t)
yk
sign(y(t)).
(4.77)
So, if " min(yk )/ max(|yX |), then the second term in (4.77) dominates the first
one, which is the objective. Thus, we make here two assumptions:
172
(i) the state x(t) is restricted to evolve in {x : x(t) < Xmax }, and
(ii) the resetting is applied only when |y(tk )| > ymin .
For the first assumption, the bound Xmax can be taken arbitrarily large, covering
the physically meaningful values of the states. This assumption also implies that the
stability results will have a local character, and ensures that yX (t) is bounded.
For the second assumption, ymin can be taken arbitrarily small, equal to the resolution level of the quantization in digital implementation. Both assumptions are
necessary, otherwise the dominance of yk / and the resetting effect will be lost. If
assumptions (i) and (ii) hold, then (4.77) for yk > 0 and 0 approaches
y(t)
yk
for t [tk , tk + ).
Integrating this equation from y(tk ) = yk gives rise to y(tk + ) = 0, that is, we
reset the output y(t) at the end of the pulse width. Since we are arbitrarily close
to the reset instant t [tk , tk + ), with 0, from (4.73) and (4.74) it follows
that y(t) P33 x3 (t) for t [tk , tk + ). Thus we are resetting as well the required
part of the state x3 at the end of the pulse: x3 (tk + ) = 0 when 0. In this
way, R approaches the ideal reset R (4.72) when 0, and the resetting action
in Proposition 4.9 is correctly applied because from (4.77) when 0 the value
yk = P33 x3 (tk ) decreases linearly to y(tk + ) = 0 in an arbitrarily fast way.
Now, the system R given by (4.75) maps L2e to L2e . However, R is not
bounded in L2 since it is possible to find signals in L2 such that their images
under R are not in L2 . A similar problem has been discussed in [11] in the framework of sampled-data systems. The solution adopted there to overcome the problem will be also used here. Basically, the system R is restricted to operate in the
space F L2 of filtered L2 signals, F being an LTI system with a strictly proper
transfer function and arbitrarily large bandwidth. More precisely, R is restricted to
operate on F L2 = {y F L2 : |y|
M} where the bound M on y can always
be imposed, for any fixed , from the discussion after (4.77). Thus, the system
L L is bounded, that is, R (F L ) L . This is a necessary
R : F L2e
2e
2e
2
2
condition for R to be output strictly passive.
L L
Proposition 4.10 The system R : F L2e
2e
2e (that approximates the
ideal reset R by finite rectangular pulses, and approaches R when 0) is an
output strictly passive (OSP) system for all > 0.
y (t)v(t) dt + i
y (t)y(t) dt + o
v (t)v(t) dt
(4.78)
173
Let us suppose here, for simplicity, the scalar case n3 = 1 and P33 = 1. The
multivariable case requires a more careful treatment and will be reported elsewhere.
First, the L2 norm of the output is
R y =
2
2
(t tk )|y(tk )| sign(y(t)) dt,
|y(tk )|2 .
k
|y(tk )|
tk +
|y(t)| dt.
tk
, |y|
Now, since y is a filtered L2 signal, that is, y F L2e
is bounded, and one
gets
tk +
|y(t)| dt ()|y(tk )|,
tk
and thus, from (4.78), R is output strictly passive (OSP) for some such that 0 <
.
Remark 4.5 An important consequence of the last result is that the stability condition for the feedback system {BLSio , R } can be translated into a frequency condition on BLSio . To see this, recall that a negative feedback system {G1 , G2 } is
finite-gain stable if any of (i)(iii) holds [10]: (i) G1 , G2 are ISP, (ii) G1 , G2 are
OSP, (iii) G1 is SP and G2 is passive (see also Sect. 3.3 for details).
174
Clearly, we are interested in case (ii), so if BLSio is OSP, then closed-loop stability (finite-gain) is guaranteed. Recall also that a stable and controllable LTI system
G(s) is OSP if there exists some o > 0 such that G(j ) + G (j ) o G(j )2
(see [10]). In the scalar case, this means that G(j ) lies inside a circle with center
s = 1/(2o ) and radius r = 1/(2o ). In summary, the required stability condition is
H (j ) + H (j ) H (j )2 , > 0, that is, an OSP condition on
1
H (j ) = H j I A Ad ej h B.
(4.79)
175
This figure contains two sets of blocks: the white fixed blocks and the gray uncertain blocks. The set of fixed blocks is derived from the base linear system and from
some matrices appearing in the LMI stability conditions, including B, H in (4.66).
There are three uncertain blocks, the block is the delay system ehs . The block Z
is the zero-order holder system (ehs 1)/(hs). The third block R is the impulsive
resetting subsystem introduced above.
The proof of the equivalence is omitted for brevity, but it can be checked starting
from (13) in [29], and with some mild notation changes, as M = I + X 1 W , with
X = 12 P . The interesting point of the relation between (4.13) and Fig. 4.6 is not only
the equivalence when Z, and R are as described above, but also the important
conclusion that stability is preserved when the first two blocks are replaced by any
pair of small-gain blocks (1 , 2 ) satisfying i 1, i = 1, 2. And the third
bock R can be replaced as well by any other OSP block with a stable strictly proper
transfer function 3 with Re[3 (i )] 0 for some > 0.
Thus, Fig. 4.6 shows explicitly the degree of robustness achieved by the LK
stability conditions, that is, which amount of uncertainty, and in what form, can
be tolerated. This offers useful insights into theoretical and practical aspects of the
problem.
Remark 4.7 (Using alternative LK functionals and LMIs) Within the field of linear time-delay stability there is a large literature, and thus, there are many other
LK functionals alternative to (4.49). In [4], a functional taken from [16] was used.
Similarly, many other LK functionals could be combined to the passivity analysis
presented here, although it is not always possible to derive robust interpretations as
in the previous remark.
s +1
,
s(s + 0.2)
and the feedback compensator R is a FORE compensator with base LTI system
given by
1
.
s +1
It can be shown that the base LTI control system is stable for delays h < 0.2. In
the following, the passivity result derived in Sect. 4.7 will be used to show that the
reset control system is also stable for that range of delays. Thus, the reset action does
not destabilize the control system. This is simply made by checking if the system
given by (4.79) is OSP. A delay h = 0.15 will be chosen.
FOREb (s) =
176
In this simple case, the plant has two states and the reset compensator only
one state, thus n1 = 2, n2 = 0, and n3 = 1. Using the realizations of P0 (s) and
FOREb (s):
0.2 0
1
Ap =
,
Bp =
,
Cp = 1 1 ,
1
0
0
and Ar = 1, Br = 1, Cr = 1, Dr = 0; the matrices A, Ad
0.2 0
0
0
0
0 ,
A= 1
Ad = 0
1 1 1
0
are
0 1
0 0.
0 0
177
Fig. 4.9 Time response y(t) (top) and u(t) (bottom) with reset (solid) and without reset (dotted)
In addition, here M = (), P33 are scalar, in particular P33 = 1 may be chosen
without lost of generality. As a result, matrices H and B are given by (4.66):
B = 0 0 1 .
H = Cp P33 = 1 ,
With the above values, system (4.79) is checked to be OSP. A simple exploration
(see Fig. 4.7) shows that in fact it is OSP for values of between 0.29 and 0.56.
Figure 4.7 shows the minimum value of Re{H (j )} over the -axis for different
values of . In particular, Fig. 4.8 shows the positive value of Re(H (j )) (logarithmic scale) vs for = 0.4.
As a result, the reset control system is guaranteed to be also stable, in this case
for h = 0.15. An identical procedure could be used to show that the reset action
never destabilizes a stable base LTI control system, and its stability limit is h = 0.2,
the same limit that the one for the base LTI control system. Figure 4.9 plots the
step response, showing the plant output y(t) and plant input u(t). Notice that the
reset instants appear actually when the error is zero, r(t) y(t) = 0, but the corrective effects (discontinuities in u(t) and derivative discontinuities in y(t)) appear
in the plant h = 0.15 later, due to the delay between the controller and the plant
(see Fig. 4.1). Notice also the damping provided by reset that improves the trade-off
between fast rise time and small overshoot.
178
An important question is whether the reset control system might be stable even
in the case in which the base LTI system is unstable, in this case for h > 0.2. The
proposed method cannot answer this question because the stability of the base LTI
system is a necessary condition. A way to overcome this problem could be to develop stability conditions based on the reset-times, as in Sect. 3.2.2. But those ideas
should be extended to time-delay, infinite-dimensional reset systems. This is a more
involved problem that requires further work.
Notice also from the time response in Fig. 4.9 that if we want to avoid the delay
of h in the reset corrective effect on the plant output y(t), then we could arrange
it by anticipating the reset actions in some way. This anticipative reset can be implemented, for example, by using the concept of reset band introduced in Sect. 3.4.
Reset band is proposed in the next chapter as a design improvement and applied in
different cases as reported in Chap. 6.
References
1. strm, K.J.: Limitations on control system performance. Eur. J. Control 6, 220 (2000)
2. Bainov, D.D., Simeonov, P.S.: Systems with Impulse Effect: Stability, Theory and Applications. Ellis Horwood, Chichester (1989)
3. Baos, A., Barreiro, A.: Delay-independent stability of reset systems. In: 32nd IECON Conference, Paris, France (2006)
4. Baos, A., Barreiro, A.: Delay-dependent stability of reset control systems. In: Proceedings
of the American Control Conference. ACC, New York (2007)
5. Baos, A., Barreiro, A.: Delay-independent stability of reset systems. IEEE Trans. Autom.
Control 54(2), 341346 (2009)
6. Baos, A., Carrasco, J., Barreiro, A.: Reset-times dependent stability of reset control with
unstable base systems. In: Proceedings of the IEEE Int. Symp. on Industrial Electronics (ISIE),
Vigo, Spain (2007)
7. Barreiro, A., Baos, A.: Delay-dependent stability of reset systems. Automatica 46, 216221
(2010)
8. Beker, O.: Analysis of reset control systems. Ph.D. Thesis, University of Massachusetts,
Amherst (2001)
9. Beker, O., Hollot, C.V., Chait, Y., Han, H.: Fundamental properties of reset control systems.
Automatica 40, 905915 (2004)
10. Carrasco, J., Baos, A., van der Schaft, A.: A passivity based approach to reset control systems
stability. Syst. Control Lett. 59, 1824 (2010)
11. Chen, T., Francis, B.A.: Input-output stability of sampled data systems. IEEE Trans. Autom.
Control 36, 1 (1991)
12. Chen, J., Latchman, H.A.: Frequency sweeping tests for stability independent of the delay.
IEEE Trans. Autom. Control 40, 16401645 (1995)
13. Clegg, J.C.: A nonlinear integrator for servomechanism. Trans. AIEE, Part II 77, 4142 (1958)
14. De la Sen, M., Luo, N.: A note on the stability of linear time-delay systems with impulsive
inputs. IEEE Trans. Circuits Syst. 50(1), 149152 (2003)
15. Horowitz, I.M., Rosenbaum, P.: Nonlinear design for cost of feedback reduction in systems
with large parameter uncertainty. Int. J. Control 24(6), 9771001 (1975)
16. Gu, K., Kharitonov, V.L., Chen, J.: Stability of Time-Delay Systems. Birkhuser, Boston
(2003)
17. Guo, Y., Wang, Y., Zheng, J., Xie, L.: Stability analysis, design and application of reset control
systems. In: Proceedings of the IEEE Int. Conf. on Control and Automation, Guangzhou,
China (2007)
References
179
18. Haddad, W.M., Nersesov, S.G., Chellaboina, V.S.: Energy-based control for hybrid portcontrolled Hamiltonian systems. Automatica 39, 14251435 (2003)
19. Khalil, H.J.: Nonlinear Systems, 2nd edn. Prentice Hall, Upper Saddle River (1996)
20. Krishman, K.R., Horowitz, I.M.: Synthesis of a nonlinear feedback system with significant
plant-ignorance for prescribed system tolerances. Int. J. Control 19(4), 689706 (1974)
21. Lakshmikantham, V., Bainov, D.D., Simeonov, P.S.: Theory of Impulsive Differential Equations. World Scientific, Singapore (1989)
22. Liu, X., Ballinger, G.: Uniform asymptotic stability of impulsive delay differential equations.
Comput. Math. Appl. 41, 903915 (2001)
23. Liu, X., Wang, Q.: Stability of nontrivial solution of delay differential equations with statedependent impulses. Appl. Math. Comput. 174, 271288 (2006)
24. Liu, X., Shen, X., Zhang, Y., Wang, Q.: Stability criteria for impulsive systems with time delay
and unstable system matrices. IEEE Trans. Circuits Syst. 54, 10 (2007)
25. Neic, D., Zaccarian, L., Teel, A.R.: Stability properties of reset systems. Automatica 44, 8
(2008)
26. Park, P.: A delay-dependent stability criterion for systems with uncertain time-invariant delays.
IEEE Trans. Autom. Control 44, 876877 (1999)
27. Rantzer, A.: On the KalmanYakubovichPopov lemma. Syst. Control Lett. 28, 710 (1996)
28. Yang, T.: Impulsive Control Theory. Lecture Notes in Control and Information Sciences,
vol. 272. Springer, Berlin (2001)
29. Zhang, J., Knospe, C.R., Tsiotras, P.: Stability of time-delay systems: equivalence between
Lyapunov and scaled small-gain conditions. IEEE Trans. Autom. Control 46, 482486 (2001)
Chapter 5
181
182
x(t)
= e(t),
(PIbase )
v(t) = kp x(t) + kp e(t).
i
(5.2)
Finally, the PI + CI compensator is defined by adding in parallel a Clegg integrator (CI) to a PI controller (see Fig. 5.1). As a result, the PI + CI compensator
will have three terms: a proportional term P , an integral term I, and a Clegg integrator (CI) term. Note that a CI term by itself does not have the characteristic of
eliminating the steady-state error in response to step disturbances by itself, thus the
integral term I is used for that purpose. The main benefit of using the CI term will
be to improve the transient response of the system, being possible in principle to get
a better response without excessive overshooting in comparison with its base PIbase .
As it will be seen later, if a well-tuned PIbase does not have a closed-loop response
with enough overshoot then the compensator PIbase may be detuned to obtain a
faster response, and then reset is performed to improve settling time and reduce
overshooting. In addition, it is expected that a PI + CI compensator will obtain better
performance in comparison with a well-tuned PI compensator in some cases. It may
be argued that the comparison should be also done with a PID compensator, but it
should be noted that in general a PID compensation would have a significatively
higher cost of feedback (sensitivity with respect to sensor noise) even in the case
183
in which the term D is filtered, thus for a fair comparison only a PI compensator is
considered.
The block diagram structure of a PI + CI controller is shown in Fig. 5.1, where
the input e is the error signal, the output v is the control signal, kp is the proportional gain, and i is the integral time. The new parameter preset , the reset ratio,
is a dimensionless constant with values preset [0, 1] that accounts for the relative
weight of the CI term over the I term. Note that a PI + CI compensator is reduced
to a PI compensator if reset is eliminated, that is, preset = 0. And on the other hand,
a P + CI compensator would be obtained if preset = 1. However, as it was discussed
above, in general the reset should not be applied on the whole of the integral term
because the fundamental asymptotic property of the integral term would be lost, for
example, the steady-state error in the closed-loop step response could not disappear.
In the state space, a PI + CI compensator can be expressed by using two states,
one corresponding to the I term and the other corresponding to the CI term. Thus
the state will be xr = (xi , xir ) , where xi will correspond to the I-term state and xir
to the CI-term state. A PI + CI state space model is given simply by (4.2), where the
state space matrices are given by
0 0
,
(5.3)
Ar =
0 0
1
,
(5.4)
Br =
1
kp
1 preset preset ,
(5.5)
Cr =
i
Dr = kp ,
and the matrix A R22 selects the state to be reset, xir , and is given by
1 0
.
A =
0 0
(5.6)
(5.7)
kp
v(t) = i xpi+ci (t) + kp e(t),
where the previous model is obtained in the particular case of
rk =
184
with tk , k = 1, 2, . . . , being the reset instants, and xi and xir the states of the integrator and the Clegg integrator, respectively, in the two states model. In general, several
different values of rk may be used. As a result, this representation gives a particular
type of reset control system with variable reset that need a specific treatment ([17]
is an initial work in this direction).
On the other hand, as it is well known, the describing function gives good results
in practice if some low-pass filtering is assumed in the plant. A very important characteristic of the CI term is that its describing function, previously given by (3.89),
does not depend on the input amplitude, only on the frequency. This property is also
inherited by the PI + CI compensator whose describing function is simply given by
1 preset preset 1.62 j 38.1
(PI + CI)(j ) = kp 1 +
e
+
j i
i
j (i + 4 preset ) + 1
= kp
.
(5.9)
j i
Note that the PI + CI describing function reduces to the PI frequency response
function whenever reset is eliminated, that is, preset = 0. In Fig. 5.2, the describing
function of a PI + CI compensator is shown for several values of the reset ratio
preset . Note that the reset action may introduce a phase lead of up to 50 degrees and
a relatively small increment of the magnitude for frequencies lower than 1/i . This
is a basic property of the PI + CI compensator that, in comparison with its base
PI compensator, allows achieving both a bigger phase margin and a crossover gain
frequency. This means that a better performance both in terms of speed of response
and relative stability may be potentially obtained by means of PI + CI compensation,
overcoming fundamental limitations of PI compensation.
In addition, the simple form of the describing function makes it very appropriate
from a design perspective, and in fact it will serve as a basis of the tuning rules to
be developed in next sections.
185
Consider the standard feedback reset system of Fig. 2.1, where the reset controller R is a PI + CI compensator, and now the plant P is described by
(P )
xp (t0 ) = xp0 ,
y(t) = xp (t),
(5.10)
where xp is the plant state and xp0 the plant initial state. In this case, the time
constant and the gain k are scalars. After some simple computation, the closedloop system (2.4) is given by the matrices
1+kp k
A = 1
1
kp k
i (1 preset )
0
0
0
0
C= 100 ,
1
AR = 0
0
kp k
i preset
0
1
0
0
0 .
0
(5.11)
(5.12)
(5.13)
186
(5.14)
for t (tk , tk+1 ], where the state is now x = (xp , xi , xir ) , and A is given by (5.11).
In addition, the closed-loop output y is given by y = Cx = (1, 0, 0)x = xp , and by
definition after each reset time both the plant state xp (t) and the reset integral state
xir (t) are zero. Thus, taking a state after reset x = (0, xik , 0) at t = tk+ , it can be
shown after some computation that the output y(t) in the interval (tk , tk+1 ] is given
by
y(t) =
2 + 2 (ttk )
e
sin((t tk ))(1 preset )xik ,
(5.15)
where j are the nonzero eigenvalues of the matrix A (4.2), with and
being respectively
=
and
1 + kp k
2
.
kp k
.
= + 2
i
(5.16)
(5.17)
k k
Note that 2 < pi is assumed, otherwise the output y would not be oscillatory.
Using (5.15), the reset instants tk+1 , k = 1, 2, . . . , are given by
0 = y(tk+1 ) =
2 + 2 (tk+1 tk )
sin((tk+1 tk ))(1 preset )xik ,
e
(5.18)
and then it is not difficult to show that reset instants {t1 , t2 , . . . } are periodic with
fundamental period reset given by
reset =
(5.19)
On the other hand, the integral term state, xi (t), is not necessarily zero after the
reset instants. Its time evolution for t (tk , tk+1 ], k = 1, 2, . . . , is given by
1 preset (ttk )
xi (t) = preset +
e
sin (t tk ) + cos (t tk ) xik ,
(5.20)
and thus the values of xi (t) at the reset instants are simply given by
xi,k+1 = x(tk+1 ) = preset (1 preset )e xik .
(5.21)
187
0
0
e
AR eAreset = 0
(5.22)
1 (1 preset )(1 + e ) ,
0
0
0
where stands for a nonzero term. Note that the nonzero eigenvalues of
AR eAreset are 1 = e and 2 = 1 (1 preset )(1 1 ). Thus, AR eAreset
is Schur if and only if the nonzero eigenvalues of A are strictly in the left half-plane
(and thus a minimal realization of the base system is Hurwitz) independently of the
reset ratio preset , since |1 | < 1 if and only if > 0, and 2 [1 , 1) for values
preset [0, 1).
Thus, the closed-loop reset system is stable if and only if the closed-loop base
linear system is stable, which is a very convenient result. Unfortunately, this result
is no longer valid for high order plants and plants with deadtime, and some of the
stability results given in Chaps. 3 and 4 could be used.
1+e
(5.23)
which will relate the reset ratio, preset , with the integral time constant of the PI
controller, i , and its proportional gain, kp , through the parameters and . Note
that in general the integral term state cannot be forced to be zero at the first reset
instant t1 since (5.21) is only valid when xik , k = 1, 2, . . . , is an after-reset state.
It would be only possible if the initial state were an after-reset state, but this is a
188
trivial case since it corresponds to a zero initial state, and thus the (autonomous)
reset system state would not depart from the origin.
For a closed-loop step response that reaches the steady state at the second crossing, the tuning rules of the PI + CI controller for a first order plant can be summarized as:
1. Design the PIbase compensator to obtain kp and i for a desired oscillatory system
1+kk
kk
response (( 2 p )2 < pi ) by using any tuning rules.
2. Compute and by using (5.16) and (5.17).
3. Finally, calculate the reset ratio, preset , by using (5.23).
5.1.1.4 Example
Consider the following plant with a unit step as a reference signal:
xp (t) = 0.5xp (t) + 1.5u(t),
(P )
y(t) = xp (t).
(5.24)
Fig. 5.3 System responses and control signals for PI and PI + CI controllers
190
Internal model control (IMC) will be used to tune the PIbase compensator. The
IMC method consists of considering a known model of the system in order to get
a desired closed-loop response. For that purpose, the compensator parameters will
depend on the system parameters, and on the time constant of the desired closedloop response, . Specifically, when the system is modeled as an FOPDT system
P (s) =
k
ehs ,
s + 1
(5.25)
,
k( + h)
i = .
(5.26)
(5.27)
The value of the desired closed-loop time constant can be chosen freely, but
from (5.27) it has to satisfy h < < to get a positive and nonzero controller
gain. In general, the optimal value of is determined by a trade-off between:
1. Fast response and good disturbance rejection (small value of ), and
2. Stability and robustness (large value of ).
Among the different ways of choosing the value of that have been studied in
the literature (see [1214]), the one given by S. Skogestad [14] will be used in the
following. Thus, the value is chosen to be equal to the time delay = h. This
gives a reasonably fast response with good robustness margins. In this way, the PI
proportional gain (5.26) is given by
kp =
.
2kh
(5.28)
In the following, this method will be used as a basis to tune the PI + CI compensator, but the tuning procedure will depend on whether the delay is dominant or
not.
(5.29)
191
Disturbance
ITAE
(s2 )
Stability margins
IAE (s)
ITAE
(s2 )
m ()
Am (dB)
PI-IMC
2.20
5.05
5.85
237.45
61.3
9.92
PI-SIMC
3.09
14.07
2.33
67.2
52.5
9.64
PI + CI-SIMC
2.30
6.47
2.33
67.2
51
9.19
1
es .
20s + 1
(5.30)
Direct application of (5.26) and (5.27) gives a PI-IMC compensator with parameters
kp = 10 and i = 20 s. On the other hand, by applying the SIMC method, a PISIMC compensator is obtained with the same proportional gain, kp = 10, but with
a smaller integral time constant, i = 8 s. Finally, these parameters will be used to
tune the PI + CI controller, PI + CI-SIMC, with reset ratio value preset = 0.55.
In Fig. 5.4, closed-loop responses and control signals are compared for the PIIMC, PI-SIMC, and PI + CI-SIMC compensators. It can be seen that the PI-SIMC
gives a disturbance response faster than the one given by the PI-IMC controller.
On the contrary, when the reference signal is changed, the PI-SIMC controller increases the overshoot in comparison with the PI controller tuned by IMC, making
its transitory response longer. The PI + CI controller will be used to reduce this transitory response by decreasing the overshoot, and without modifying its disturbance
response, as it is shown in Fig. 5.4. As a result, it can be concluded that the PI + CI
compensator will give a good trade-off between reference tracking and disturbance
rejection, without increasing the cost of feedback (as a general rule, the cost of feedback for a PI + CI compensator is similar to the cost of its base compensator).
In order to get a better comparison between the different compensators, the values
of performance indexes such as the integral of absolute error (IAE) and the integral
time absolute error (ITAE) are presented in Table 5.1.
From these indexes, it can be concluded that the PI + CI controller makes the
disturbance response faster, and therefore much better, in comparison with the PIMC (IAEPI+CI-SIMC = 2.33 s versus IAEPI-IMC = 5.85 s), just by making the reference response slightly worse (IAEPI+CI-SIMC = 2.30 s versus IAEPI-IMC = 2.20 s).
In addition, the robustness indexes, phase and gain margins, are also computed in
Table 5.1 (by using the describing function in the case of PI + CI), and it can be
expected that the robustness properties of PI + CI are close to the values obtained
by its base compensator.
As a result, for FOPDT systems with dominant lag the PI + CI compensator
gives a good balance between disturbance rejection and reference tracking without
Fig. 5.4 System responses and control signals for PI-IMC, PI-SIMC, and PI + CI-SIMC controllers
192
5 Design of Reset Control Systems
193
sacrificing robustness. After extensive simulation, it has been found that a value
of preset [0.3, 0.6] gives a good response having better rejection/worse reference
tracking characteristic for increasing values of preset . Rules for tuning a PI + CI
compensator for a lag dominant system are now summarized:
1. Tune the base PI controller, PIbase , by using
kp =
2kh
and i = 8h.
2. Choose the reset ratio preset in the interval [0.3, 0.6], balancing between performance and robustness.
1
es .
2s + 1
(5.31)
For this system, a PI-IMC compensator is tuned by using the IMC method (see
(5.27)(5.28)). The following parameters are obtained: kp = 1 and i = 2 s. At first,
a PI + CI-IMC is tuned by using as base compensator the PI-IMC and a small value
of preset , for example, preset = 0.1.
In Fig. 5.5, the closed-loop responses and the control signals are shown. It can be
seen that the reset action makes the PI + CI behavior worse for a reference change
since overshoot does not decrease and undershoot increases considerably. On the
other hand, when an input load disturbance is considered, the PI + CI controller
behaves in the same way as the PI one since the response does not cross the reference
signal, and thus there are no resets.
In general, reset action does not improve the linear response due to the fact that
the base PI compensator, tuned by the IMC method, gives a response with a low
overshoot; in such a way, the control signal reaches the steady state very fast and
with practically no oscillations. Then, when the reset is performed, the control signal
is close to its steady-state value and any reset ratio, even small valued, makes the
control signal move away from its steady state-value giving a worse response.
To solve this problem, a detuning of the PIbase compensator is performed. This
detuning consists of making the closed-loop response faster, and thus more oscillatory, by fixing a smaller value of the time constant of the desired closed-loop
response . Note that the value had been fixed previously to be equal to the time
delay, h. Thus, in order to make the response faster, now will be fixed to be two
thirds of the time delay, that is, = 23 h.
With this new value, the integral time constant of the base PI controller (5.27)
does not vary, but the proportional gain (5.28) does. By replacing = 23 h in (5.26),
Fig. 5.5 System responses and control signals for PI-IMC and PI + CI-IMC controllers
194
5 Design of Reset Control Systems
195
Disturbance
IAE (s)
ITAE
PI-IMC
2.17
PI-des
2.10
PI + CI-des
2.15
(s2 )
Stability margins
IAE (s)
ITAE
5.05
2.00
5.01
1.67
5.03
1.67
(s2 )
m ()
Am (dB)
47.89
61.3
9.94
39.46
55.6
8.36
39.46
56
8.23
3
.
2kh
(5.32)
With this modification, the detuned PI (PI-des) base parameters are now kp = 1.2
and i = 2 s. In addition, the detuned compensator is the base for the reset compensator PI + CI-des, where a small reset ratio is also considered, for example,
preset = 0.1. In Fig. 5.6, closed-loop responses and the control signals are compared
for the PI-IMC, PI-des, and PI + CI-des compensators. In addition, for a better comparison, the values of the performance indexes are presented in Table 5.2.
It can be seen that PI-des, in spite of having bigger overshoots and undershoots,
by sacrificing some robustness, behaves better than the PI-IMC since, when the
reference signal changes, the system response is faster, and smaller IAE and ITAE
values are obtained. When input disturbances are considered, it is seen that PI-des
also makes the system response better, by obtaining a faster response with lower
performance indexes.
However, comparing PI + CI with its base compensator PI-des, it is deduced that
the PI + CI response is slightly worse with similar robustness values. Thus, it has to
be concluded that without any modification PI + CI compensation is not appropriate
for FOPDT systems with dominant delay. This worse behavior of the reset system is
due to the detrimental effect of the dominant time delay over the reset action. In the
next section, we will discuss how the use of a reset band can improve this behavior.
Fig. 5.6 System responses and control signals for PI-IMC, PI-des and PI + CI-des controllers
196
5 Design of Reset Control Systems
197
k
ehs ,
s + 1
(5.34)
where the model time constant is calculated with the following expression:
= 10 +
20
,
2
(5.35)
20
i0 .
+
2
(5.36)
i3
Finally, if the general system (5.33) has zeros, the numerator term (T0 s + 1) is
canceled by a neighboring denominator term (0 s + 1), where both T0 and 0 are
positive and real, using the following approximations
T0
if T0 0 h,
T0
T0 s + 1 h if T0 h 0 ,
(5.37)
0 s + 1
1
if h T0 0 ,
T0 if T 5h.
0
If there is more than one positive numerator time constant, then one T0 should be
approximated at a time, starting with the largest T0 .
Once the general system is approximated to an FOPDT model, the tuning rules
developed in the above section can be directly applied in order to tune the PI + CI
compensator. For example, consider the following system:
P (s) =
(15s + 1)
es .
(25s + 1)(20s + 1)(s + 1)
(5.38)
Firstly, its effective time delay is computed by using (5.36), and the value
h = 1.5 s is obtained. Secondly, the zero (15s + 1) is canceled against the neighboring pole (20s + 1) with (5.37). Since the condition 0 T0 5h is fulfilled, the
zero and pole are approximated as:
15s + 1 15
= 0.75.
20s + 1 20
(5.39)
198
Disturbance
ITAE
(s2 )
IAE (s)
Stability margins
ITAE
(s2 )
m ()
Am (dB)
PI-IMC
4.52
25.42
6.29
457.98
50.6
10.31
PI-des
5.55
38.00
3.18
200.30
42.6
9.77
PI + CI-des
4.81
26.70
3.18
200.30
40.8
9.25
From this model it is seen that its time constant, = 25.5 s, is much bigger than
its time delay, h = 1.5 s, thus the model (5.40) is a lag dominant system. As a result,
it may be expected that reset compensation will be effective in this case to improve
performance without sacrificing robustness and cost of feedback. Firstly, a base PISIMC compensator is tuned by using (5.28) and (5.29); the result is kp = 11.3 and
i = 12 s. In addition, for the PI + CI-SIMC compensator, an initial value of the
reset ratio preset = 0.55 is chosen.
For comparison purposes, a PI-IMC compensator will be also used. Using (5.27)
and (5.28), the parameters are kp = 11.3 and i = 25.5 s. In Fig. 5.7, closed-loop
responses and control signals are shown for PI-IMC, PI-SIMC, and PI + CI-SIMC
compensators. Both PI-SIMC and PI + CI-SIMC give a faster response than PIIMC, overall when input disturbances are considered. However, when the reference
signal changes, the system responses with PI-SIMC and PI + CI-SIMC exhibit more
overshoot, and therefore worse performance indexes, as it can be seen in Table 5.3.
When PI + CI is used, the PI-SIMC overshoot is reduced and better performance indexes for references are obtained. In spite of this overshoot reduction,
the PI + CI response is slightly worse than the PI-IMC response when reference
changes are considered (IAEPI+CI-SIMC = 4.81 s versus IAEPI-IMC = 4.52 s), but it
is much better when input disturbances are considered (IAEPI+CI-SIMC = 3.18 s versus IAEPI-IMC = 6.29 s). As a conclusion, and without surprise, it can be said that
PI + CI compensation gives the best balance between performance and robustness
for higher order systems with deadtime that can be well approximated by FOPDT
systems with dominant lag.
Fig. 5.7 System responses and control signals for PI-IMC, PI-SIMC, and PI + CI-SIMC controllers
200
Disturbance
ITAE
(s2 )
IAE (s)
Stability margins
ITAE
(s2 )
m ()
Am (dB)
PI
6.43
62.80
4.48
328.3
48.6
24.28
PI + CI
4.17
20.44
4.48
328.3
48.4
23.9
integrator, this is no longer true, and thus the reset ratio can be set at its maximum
value preset = 1.
An FOPDT system can be approximated to an integrating system when its time
constant is really large, :
P (s) =
k
ehs
s + 1
k hs k hs
e
= e .
s
s
(5.41)
2kh
and i = 8h.
(0.17s + 1)2
,
s(s + 1)2 (0.028s + 1)
(5.42)
(5.43)
With this model, the PI + CI compensator is tuned by using the tuning rules
given above. Firstly, the PIbase parameters are computed, and the following values
are obtained: kp = 0.3 and i = 13.5 s. Obviously, these parameters are used in the
PI + CI compensator together with a reset ratio, in this case, preset = 1.
In Fig. 5.8, the closed-loop responses and control signals are drawn for the PI
and PI + CI compensators (with the plant given by (5.41)). In addition, the values
of the performance indexes and stability margins are shown in Table 5.4.
From Fig. 5.8 and Table 5.4, it can be deduced that the reset action improves
considerably the linear system behavior since the PI + CI controller not only gives
a response with less overshoot but also reduces the settling time without modifying
Fig. 5.8 System responses and control signals for PI and PI + CI controllers
202
the response speed. On the other hand, when input disturbances are considered, the
reset control system behaves in the same way as the linear one since the response
does not cross the reference and thus the reset action does not occur.
It is worthy to note that in this case the stability margins, see Table 5.4, are hardly
modified with the reset action. Therefore, it can be concluded that the PI + CI controller makes the response much better without modifying its robustness properties.
203
kp
< pi
/
1+kk
kk
= 2 p , = pi 2
preset =
2kh
8h
[0.3, 0.6]
3
2kh
i
1+kkp 2
2 )
kk
preset
e
1+e
8h
Regarding the reset law, a first simple modification will consist of using a reset
band, as introduced in Sect. 3.4.2, that is, reset actions are produced when the error
signal is entering a predefined reset band. In a second stage, this reset band is considered to be time varying depending on the error signal and its time derivative; in
this way, reset actions may be more effective, for example, if different setpoints are
defined in the control problem, and moreover, not only in the transitory but also in
the steady state, adapting the bandwidth to the magnitude of the error and its time
derivative. In general, these changes over the reset law are specially effective when
significant time-delays are present, for example, in applications like process control
or teleoperation.
On the other hand, to improve robustness especially in the case in which large
plant parametric uncertainty exists, a frequency domain robust control technique
like Quantitative Feedback Theory (QFT) is useful for guaranteeing some prescribed
degree of robustness in stability and performance. Thus, the base linear compensator
can be designed using QFT with some robustness specifications. Finally, considering the case of PI + CI design, it has been found useful in some cases to use a
time varying reset percentage; in general, this is related with the use of partial reset
compensators with a time varying reset dynamics.
204
Fig. 5.9 System responses and control signals for PI and PI + CI controllers
.
r
tr
(5.44)
For example, consider the integrating system given in Sect. 5.1.4. In this case,
the rise time has a value tr 2.5 s and the delay is h = 1.69 s. Thus, the value of the
reset band is = 0.5. Figures 5.9 and 5.10 show the step response and the control
signal for the PI + CI compensator with a reset band of = 0.5, respectively. Note
that in comparison with the PI + CI compensator, the step response has improved in
terms of the overshoot and undershoot.
In addition, using the describing function analysis of the Clegg integrator with
reset band given in Sect. 3.4.2, it also possible to come up with a value of /r
for obtaining a maximum phase lead around the frequency 1/ h. In Chap. 3, the
describing function of the Clegg integrator with a fixed reset band was computed
and given by
/
1 ( E )2
j
4
1
j sin1 ( E )
DCI (E, ) =
e
,
(5.45)
1+
j
1
1
DPI+CI
(E, ) = kp 1 +
1+
ej sin ( E ) . (5.46)
j i
205
Fig. 5.10 System responses and control signals for PI and PI + CI controllers
For comparison purposes, in Fig. 5.11, the Bode plot is shown for a PI compensator (kp = i = 1) and for a PI + CI compensator (kp = i = 1 and preset = 0.50)
with several fixed reset bands ( E = 0.25, 0.50, 0.75, and 1.00). It can be seen that
the controller phase is increased with the fixed reset band for some values. Specifically, in this case, when preset = 0.50, the higher controller phase is obtained in
a range E [0.25, 0.50]. This phase increment should make the delay effect less
influential over the responses, by giving smaller overshoots and undershoots, and
hence, smaller IAE and ITAE values.
206
Fig. 5.11 Bode plots for PI + CI controller with several fixed reset bands
Fig. 5.12 System responses and control signals for PI and PI + CI controllers
207
Fig. 5.13 System responses and control signals for PI and PI + CI controllers
of prediction, such as the rate of change of the error signal. Specifically, the reset
percentage will be changed as a function of the error derivative through the tuning
of a new parameter, the derivative time constant, d . In addition, a low-pass filter
will be used with the pure derivative term to avoid the amplification of the sensor
noise,
F (s) =
1
,
f s + 1
(5.47)
where the filter time constant, f , will be another parameter to be tuned. In this way,
the variable reset percentage, p reset , can be given through the following expression:
p reset (t) = p reset d
deF
,
dt
(5.48)
where p reset is a constant base value of reset percentage and eF is the error signal
filtered by using F . As it can be seen, from (5.47)(5.48), three parameters must be
fixed to tune a PI + CI controller with variable reset percentage from a PI + CI one
with fixed reset percentage. To choose the values of these parameters, it is important
to observe that the variable reset percentage must make the reset action act strongly
over the transitory response, but weakly when the system reaches the steady-state.
In this way, the base reset percentage, p reset , should have a varying value in order
to avoid a high influence on the system response when the steady-state is reached.
On the other hand, the reset action should have strong influence over the transitory
response, so the derivative time constant, d , is fixed for that purpose. Finally, the
filter time constant, f , should be tuned to avoid the undesirable noise amplification.
In Chap. 6, an application example of the variable reset percentage technique will
be developed (see Sect. 6.3)
208
References
209
open loop gain function is described by L(s) = G(s)P (s), where the frequency
response of the feedback compensator is directly given by the describing function
of the PI + CI compensator (5.9), that is, G(j ) = (PI + CI)(j ). Alternatively, it
can be used to design the base compensator, guaranteeing that the response has the
desired oscillatory characteristic for the reset to be effective, and in the presence of
large plant uncertainty.
In Chap. 6, a number of examples and case studies will be developed by using
QFT.
References
1. strm, K.J., Hgglund, T.: The future of PID control. Control Eng. Pract. 9, 11631175
(2001)
2. Bakkeheim, J., Smogeli, O.N., Johansen, T.A., Sorensen, A.J.: Improved transient performance by Lyapunov-based reset of PI thruster control in extreme seas. In: Proc. IEEE Conference on Decision and Control, pp. 40524057 (2006)
3. Baos, A., Vidal, A.: Definition and tuning of a PI + CI reset controller. In: Proc. European
Control Conference, Kos, Greece (2007)
4. Baos, A., Vidal, A.: Design of PI + CI reset compensators for second order plants. In: Proc.
IEEE International Symposium on Industrial Electronic, Vigo, Spain (2007)
5. Baos, A., Carrasco, J., Barreiro, A.: Reset-times dependent stability of reset systems. In:
Proc. European Control Conference, Kos, Greece (2007)
6. Beker, O., Hollot, C.V., Chait, Y.: Plant with integrator: an example of reset control overcoming limitations of feedback. IEEE Trans. Autom. Control 46(11), 17971799 (2001)
7. Clegg, J.C.: A nonlinear integrator for servomechanisms. Trans. AIEE, Part II 77, 4142
(1958)
8. Fernndez, A.F., Barreiro, A., Baos, A., Carrasco, J.: Reset control for passive bilateral teleoperation. IEEE Trans. Ind. Electron. (2010). doi:10.1109/TIE.2010.2077610
9. Horowitz, I.M.: Quantitative Feedback Theory. QFT Press, Boulder (1992)
10. Horowitz, I.M., Rosenbaum, P.: Nonlinear design for cost of feedback reduction in systems
with large parameter uncertainty. Int. J. Control 24(6), 9771001 (1975)
11. Krishman, K.R., Horowitz, I.M.: Synthesis of a nonlinear feedback system with significant
plant-ignorance for prescribed system tolerances. Int. J. Control 19(4), 689706 (1976)
12. Lee, Y., Park, S., Lee, M., Brosilow, C.: PID controller tuning for desired closed-loop responses for SI/SO systems. AIChE J. 44(1), 106115 (1998)
13. Rivera, D.E., Morari, M., Skogestad, S.: Internal model control 4, PID controller design. Ind.
Eng. Chem. Process Des. Dev. 25(1), 252265 (1986)
14. Skogestad, S.: Simple analytic rules for model reduction and PID controller tuning. J. Process
Control 13(4), 291309 (2003)
15. Smith, C.L., Corripio, A.B., Martin, J.: Controller tuning from simple process models. Instrum. Technol. 22(12), 3944 (1975)
16. Vidal, A., Baos, A.: QFT-based design of PI + CI reset compensator: applications in process
control. In: 16th IEEE Mediterranean Conference on Control and Automation, Ajaccio, France
(2008)
17. Vidal, A., Baos, A.: Stability of reset control systems with variable reset: application to
PI + CI compensation. In: European Control Conference, Budapest, Hungary (2009)
18. Vidal, A., Baos, A.: Reset compensation applied on industrial heat exchangers. In: 14th IEEE
International Conference on Emerging Technologies and Factory Automation, Mallorca, Spain
(2009)
210
19. Vidal, A., Baos, A., Moreno, J.C., Berenguel, M.: PI + CI compensation with variable reset: application on solar collector fields. In: 34th Annual Conference of the IEEE Industrial
Electronics Society, Orlando, Florida, USA (2008)
20. Wu, D., Guo, G., Wang, Y.: Reset Integral-Derivative Control for HDD servo systems. IEEE
Trans. Control Syst. Technol. 15(1), 161167 (2007)
21. Yang, T.: Impulsive Control Theory. Lecture Notes in Control and Information Sciences,
vol. 272. Springer, Berlin (2001)
22. Zheng, Y., Chait, Y., Hollot, C.V., Steinbuch, M., Norg, M.: Experimental demonstration of
reset control design. Control Eng. Pract. 8(2), 113120 (2000)
23. Zheng, J., Guo, Y., Fu, M., Wang, Y., Xie, L.: Development of an extended reset controller
and its experimental demonstration. IET Control Theory Appl. 2, 866874 (2008)
Chapter 6
Application Cases
211
212
6 Application Cases
is tuned by using QFT and applied to the temperature control of an industrial heat
exchanger. In Sect. 6.1.1, a dynamic model of the heat exchanger is obtained for
different operation points based on experimental data. In Sect. 6.1.2, several PI + CI
compensators are designed. First, a PI + CI compensator is tuned by using QFT, and
then several design improvements such as the use of a fixed reset band and a variable
reset band, are analyzed. The different reset compensators are experimentally tested
and compared with a well designed PI compensator.
213
Shell diameter
104 mm
76.2 mm
Type of flow
parallel
Material
stainless steel
Tubes length
3000 mm
are placed at both the inlet and the outlet of the inner tube of the heat exchanger
for temperature measurements. The product is heated in the inner tube from the saturated steam condensation at the shell side. The rate of heat transfer depends on
the temperature difference between the cold and hot fluids in such a way that the
higher this difference is, the faster the heat is transferred. Therefore, the rate of heat
transfer, q in kcal/h, is expressed through the following equation:
q = U AT ,
(6.1)
214
6 Application Cases
the control system provides the corresponding electropneumatic globe valve aperture, u in %, in order to reach the desired reference temperature.
The industrial heat exchanger is modeled by using water as the product with a
flow rate of 250 L/h. The identification is done through pseudorandom binary sequence (PRBS) experiments and step test application. On the one hand, the heat
exchanger models are obtained by first order approximation describing the response
of the temperature along the heat exchanger to several step changes of valve aperture. Specifically, this identification was done by opening the control valve from 10
up to 25% in 5% steps and then by closing the control valve up to 10% by using
again 5% steps as can be seen in Fig. 6.4.
In addition, the PRBS experiments were also done to model the heat exchanger.
These experiments [30] consist of doing different tests by using the control signal
215
Gain ka b%
P1015%
0.32
87
88
P1520%
0.35
82
83
P2025%
0.40
80
65
P2520%
0.40
89
50
P2015%
0.35
95
76
P1510%
0.31
114
84
as the input, which will have a spectrum with significant components in some regions of interest. Then, the different response data are used for getting an uncertain
parametric model. The range [10, 25]% for the control valve aperture is identified
as the working range for this study. One example of a PRBS experiment is shown in
Fig. 6.5 for a valve aperture between 10 and 15%.
With these modeling experiments, the heat exchanger can be represented by a set
of first order plus time delay (FOPTD) transfer functions given by
Pab% (s) =
kab%
ehab% s .
ab% s + 1
(6.2)
For each valve aperture change, from a% to b%, a heat exchanger model is obtained, as given in Table 6.2. The heat exchanger uncertainty is expressed in parametric form as kab% [0.31, 0.40], ab% [80, 114], and hab% [50, 88]. Note
that every parameter combination is not possible, in fact, the gain is increasing for
increasing valve aperture values, while the time lag and the delay are decreasing.
216
6 Application Cases
6.1.2 PI + CI Design
In the following, a PI + CI compensator is designed by using the results given in
Chap. 5. Different control experiments using the industrial heat exchanger will be
carried out with different control systems including a base PI compensator and different reset PI + CI compensators (with/without reset band, . . . ). The experiment
will consist of setting a fixed product flow of 250 L/h, as considered in the modeling
procedure, and opening the control valve at 10%. When the steady-state temperature
is reached at this control valve aperture, the reference temperature is increased and
decreased by step changes with a magnitude of 2 C.
,
k( + h)
i = .
(6.3)
(6.4)
The value of the desired closed-loop time constant can be chosen freely, but due
to (6.3) it has to satisfy h < < to get a positive and nonzero controller gain.
In general, the optimal value of should be determined by a trade-off between:
1. Fast response and good disturbance rejection (small value of ) and
2. Stability and robustness (large value of ).
Between the different ways of fixing the value of which have been studied in
the literature [31, 35, 36], the one given in [36] is used in this work. So, the value
is chosen to be equal to the time delay = h. This gives a reasonably fast response
with good robustness margins. In this way, the PI proportional gain (6.3) is given as
kp =
.
2kh
(6.5)
For the industrial heat exchanger modeled by (6.2) and with parameters as in
Table 6.2, a nominal model has to be considered to tune the PI controller via IMC.
In this work, the chosen model is the one in which the control valve is opened from
15% up to 20%
P1520% =
0.35 83s
e
.
82s + 1
(6.6)
217
For this model, by using (6.4) and (6.5), the following PI parameters are obtained:
kp = 1.41 and i = 82 s. From now on, this controller is going to be named as PIIMC. The experiment explained previously has been carried out with this controller.
That is, the control valve is opened at 10% until the steady-state is reached. The
difference between the outlet and inlet temperatures, T = Tout Tin , is taken as
0 C. Then 2 C step changes are introduced into the reference temperature from
0 C up to 6 C and from this value down to 0 C again.
In Fig. 6.6, the temperature increments in the heat exchanger are shown together
with the control signal, that is, the control valve aperture. In both signals, the noise
introduced by the thermocouples and the perturbations due to the pressure drops
occurring at the electric boiler are visible. Note that the higher the control valve
is opened, the more oscillatory the response at the steady-state is. Regarding the
control system, it can be seen that the PI-IMC controller is robust for all the reference changes. In addition, fast response is obtained (rise time between 98 and 233
seconds) with a significant overshoot in some cases (between 4.5 and 26%).
218
6 Application Cases
(QFT) will be used to design PIbase , by using design specifications about closedloop stability and tracking performance. In a second stage, this PI + CI compensator
will be designed choosing a proper reset ratio preset .
PIbase Tuning The PIbase compensator will be tuned with QFT by following the
design steps (briefly) explained in Chap. 5. First of all, plant templates are computed
taking into account parametric uncertainty as given in (6.2) and Table 6.2, with the
nominal plant being
P1520% (s) =
0.35 83s
e
.
82s + 1
(6.7)
In this case, the templates are obtained for a wide set of working frequencies,
from 0.002 up to 0.02 rad/s.
Secondly, the boundaries are computed for each frequency from design specifications by using templates. Design specifications will be a minimum (robust) phase
margin of 24 and desired tracking performance, such as a rise time between 45 and
180 seconds and a maximum overshoot between 7 and 75%.
Note that a relatively small phase margin has been specified for the base PI control system. In addition, note also that the response is specified to be faster and with
more overshoot than the response obtained with the PI-IMC controller. These design specifications are considered due to the fact that the reset compensation will
improve the minimum phase margin and will reduce the overshoot of the PI compensator without reducing the response speed.
The next step consists of shaping the nominal open loop. Here a PI controller is
tuned by making the open-loop gain fit both the stability and tracking boundaries.
A shaping is shown in Fig. 6.7, corresponding to a proportional gain of kp = 1.6
and an integral time constant of i = 60 s.
Finally, once the base PI compensator is tuned, a closed-loop system analysis is
performed to ensure that the system fits the given stability and tracking specifications. This is usually referred as to compensator validation in QFT. This analysis
has to be done for a wider set of frequencies than the one used in the templates computation, and in this case, this set goes from 0.00001 up to 100 rad/s with a much
bigger grid. In Fig. 6.8, the stability analysis is shown, whereas the tracking analysis
is presented in Fig. 6.9.
As it can be seen, the designed closed-loop system satisfies both stability and
tracking specifications for every frequency. Therefore, the base PI compensator design is finished.
PI + CI Tuning Once PIbase has been designed, the obtained parameters are chosen as a part of the PI + CI parameters. But in addition, one more parameter, the reset
ratio preset , must be tuned. The expected result is an increment of the phase margin,
and thus a better transient response for the industrial heat exchanger. Several values
of preset are considered: 0.1, 0.2, 0.3, and 0.4.
These PIbase and PI + CI compensators are used to control the already described
heat exchanger. For this purpose, the same experiment carried out with the PI-IMC
219
controller is repeated. Figure 6.10 shows the temperature increment and the valve
aperture when the PIbase and PI + CI compensations are used.
220
6 Application Cases
As it can be seen, the PIbase and the PI + CI responses are equal until the first
reset action is performed. After this reset, the PI + CI response has the same or
smaller first overshoot while the other overshoots and undershoots are increased
with respect to the base PI system, even for small values of preset . This behavior
is compatible with the results given in Chap. 5, where PI + CI is shown to behave
worse than its base compensator for delay dominant systems.
221
IAE (s)
ITAE (s2 )
Phase margin ()
Base PI
3.31 103
12.45 106
41
PI + CI preset = 0.1
3.75 103
14.14 106
44
PI + CI preset = 0.2
3.28 103
12.09 106
46
PI + CI preset = 0.3
3.62 103
13.12 106
47
PI + CI preset = 0.4
3.87 103
14.54 106
47
To evaluate the tracking performance, the integral absolute error (IAE) and the
integral time absolute error (ITAE) are computed:
|e(t)| dt,
(6.8)
IAE =
0
ITAE =
t|e(t)| dt
(6.9)
222
6 Application Cases
Fig. 6.11 Responses and control signals of PI-IMC and PI + CI controllers with a fixed reset band
Table 6.4 Performance indexes for the PI and PI + CI control systems with a fixed reset band
Controller
IAE (s)
2.59 103
9.56 106
= 0.6
2.16 103
8.17 106
= 0.7
2.82 103
10.86 106
PI-IMC
PI + CI preset = 0.10 and
PI + CI preset = 0.35 and
ITAE (s 2 )
error signal and zero. To specify when the reset action must exactly be done, two
techniques are considered in this study: a fixed reset band and a variable reset band.
Fixed Reset Band The experiment is carried out again, but now for two PI +
CI compensators with fixed reset band. In this case, two different values of the
reset ratio are set and their corresponding fixed reset bands are computed. In one
controller, a low reset ratio is chosen and for this value, preset = 0.1, a fixed reset
band has been computed with the goal of having the maximum phase margin in the
control system. In this case, the maximum phase margin of 47 has been obtained
for a fixed reset band of E = 0.6. On the other hand, for a second PI + CI controller
with a fixed reset band, a higher reset ratio value is chosen, preset = 0.35. In this
case, the control system reaches the maximum phase margin of 60 when E = 0.7.
In Fig. 6.11, the heat exchanger responses and the control signals for the PI + CI
compensator and for the PI-IMC compensator are shown. In addition, their performance indexes are compared in Table 6.4.
From Fig. 6.11 and Table 6.4, it can be said that the PI + CI compensator with
preset = 0.35 and E = 0.7, in spite of having a higher phase margin, has a response
that it is not as good as that given by the PI + CI compensator with preset = 0.1 and
E = 0.6. This last PI + CI controller not only has lower IAE and ITAE values than
223
Table 6.5 Performance indexes for the PI and PI + CI control systems with a variable reset band
ITAE (s 2 )
Controller
IAE (s)
PI-IMC
2.59 103
9.56 106
2.13 103
8.09 106
2.92 103
10.67 106
the other PI + CI controller, but also has less overshoots and undershoots in such a
way that it gives the better response. In fact, this PI + CI controller response is also
much better than the PI-IMC response since it is faster and it has similar overshoots
and undershoots with lower performance indexes.
However, as it can be seen from the control signals, the control system just undergoes the reset action once in each reference. Therefore, after the first and the only
reset instant, the control system behaves as the lineal base PI controller. In order to
improve further the system performance, a variable reset band will be considered in
the next section.
Variable Reset Band Here a variable reset band is considered to overcome the
influence of the dominant time delay over the reset action. For this purpose, the
approach described in Chap. 5 will be followed. In this case, the reset conditions
are also given when the error signal is approaching zero, but now the value of the
error when the reset action is done will change constantly from one reset instant to
another.
For a variable reset band to be applied, only a nominal value of the time delay has
to be fixed in the variable reset band. For the industrial heat exchanger studied in this
work, the time delay is uncertain, and several values of time delay can be chosen.
Theoretically, the reset instants are going to take place h time units before the time
the error signal crosses to zero. Therefore, in order not to degrade the response to
any reference change, the nominal time delay is taken as the minimum one, that is,
h = 50 s. In addition, due to the presence of sensor noise, the derivative part of this
reset band will be filtered by a low-pass filter with a constant time 10 times smaller
than the nominal time delay.
For comparison purposes, the same experiment explained previously is carried
out under the same operation conditions. In this case, the PI-IMC controller response
is compared to the PI + CI controller with variable reset band and two different reset
ratios, preset = 0.1 and preset = 0.4. In Fig. 6.12, both the responses and the control
signals are drawn for these three controllers. In addition, the IAE and ITAE values
are reported in Table 6.5.
From Fig. 6.12 and Table 6.5, it is deduced that the PI + CI controller with a
variable reset band and reset ratio equal to preset = 0.1 is the best option to control
the temperature increments in an industrial heat exchanger since, in comparison
with the PI controller tuned with IMC, a faster response is obtained with a similar
overshoot and undershoot. In this case, the PI + CI controller does several resets
rather than just one as with the fixed reset band. Therefore, it can be concluded that
224
6 Application Cases
Fig. 6.12 Responses and control signals of PI-IMC and PI + CI controllers with a variable reset
band
the variable reset band can be used to significantly improve the reset performance
and the linear IMC response when dominant time delays are present in the system
dynamics.
225
tools has a number of advantages: they provide a multi-domain and modular approach to modeling, and transparent control principles based on energy and power
concepts. The passivity ideas can be adapted to the teleoperation framework, where
a human operator (actuating a master device) manipulates a remote machine (slave
device) which is possibly interacting with the environment. If some time delay exists, a communication channel transmitting velocity and force (which provides the
operator with a feeling of telepresence that is essential in applications [26, 29]) is no
longer passive. As an alternative, wave variableswhich are obtained from velocities and forces via the scattering transformation, see [4, 34]should be transmitted,
thus making the delayed line behave like an analog lossless LC line. This is called
line-passivation.
It turns out that after line-passivation, and with passive controllers at the master and slave sides, the overall teleoperation system becomes passive. Since it is
assumed that the external elementshuman operator and environmentbehave in
a passive way, this means that the system is stable for all possible constant values
of the time delay. This delay-independent stability property can also be extended
to time-varying delays [18], and therefore it is possible to use the Internet as the
communication channel, as has already been shown in several applications [37, 44].
Delay-independent stability can be interpreted partly as an advantage and partly
as a drawback: on the one hand, it is an advantage because the designer does not
have to worry about the delay. On the other hand, it is a drawback because, if the
passive controllers are stable for very large delay values, they tend to provide worse
performance than other (non-passive) controllers, which are unstable for large delays but stable for small nominal delays.
Realizing that the source of the problem is the time-delay, which introduces a
linear trade-off between speed and robustness, it becomes clear that using the resetcontrol principle in teleoperation might enable us to combine the robustness of the
passive solutions with the fast position tracking performance provided by a proper
design of the reset action.
The section is organized as follows. Section 6.2.1 provides an overview of the
system, and the principles of passive teleoperation. Then, Sect. 6.2.2 shows how
the use of reset is capable of overcoming the fundamental limitations of linear controllers, namely when time delays are present. The proposed reset strategy is presented and justified and the stability is proven from passivity properties. Finally, two
examples are presented in Sect. 6.2.3: a simulated simple robot and a real application with a laboratory gantry crane teleoperated with a haptic device.
226
6 Application Cases
the communication channel, a slave impedance controller (Ks , Bs ) and a slave device in contact with the environment. The interchanged variables are either power
variables (forces fi and velocities xi , with i {h, m, s, e} and with the additional
subindex d for delayed variables); or wave variables (um , us , wm , ws ) which result
from the scattering transformation of the former.
The communication channel contains the time delays which will be considered
constant and known throughout this study. This is a frequent assumption when deriving theoretical results for the first time; these may be extended for time-varying
delays later. Here we restrict ourselves to constant time delays; only in the last example of Sect. 6.2.3, and as a preliminary experimental validation, we apply the
method to unknown, time-varying delays. It is well-known that a time delay provokes a loss of passivity, which can be arranged using the scattering theory [4] or
the equivalent wave-variables [34] formulation given by the equations:
1/2
2B
I
f
um
m
=
(6.10)
xmd
wm
B 1
2B 1/2
and
ws
xsd
=
2B 1/2
B 1
I1/2
2B
fs
us
,
(6.11)
where the matrix B > 0 is the line impedance. The actually transmitted variables
are the wave variables:
us (t) = um (t T ),
wm (t) = ws (t T ),
(6.12)
(6.14)
227
This relation is true for all T , so, from the point of view of passivity, we do not
have to be concerned about its actual value. The interconnection of passive subsystems leads to an overall system that is itself passive [39], and hence stable behavior
is guaranteed.
As wave reflections can occur at both sides of a teleoperation scheme, there is a
need for impedance matching [34], which can be achieved if both sites are placed
under velocity control. This is one of the motivations for using the two symmetric impedance controllers mentioned before, which are placed at each end of the
transmission channel. Thus, each side is receiving force information and providing
velocity signals.
The master side of the setup consists of a master device, its impedance controller,
and the corresponding scattering transformation. The master device interacts with
the human operator, allowing him to specify the desired movement of the plant
(velocity command), while receiving some force feedback representing information
about the operating conditions. Similarly, the slave side consists of the slave device,
its impedance controller, and the scattering transformation.
The master and slave devices will be considered as mechanical systems whose
equations can be written in the port-Hamiltonian notation as:
z = [J (z) R(z)]
x = g T (z)
H (z)
,
z
H (z)
+ g(z)f,
z
(6.15)
where z are the states, x the output (velocities), f the input (forces), H the Hamiltonian function, and J, R the interconnection and dissipation structures, respectively.
Reproducing the previous equations with subindices m, s, we obtain the description
of the blocks Master device and Slave device in Fig. 6.13, additively including
the forces fh , fe .
Finally, the equations of the impedance controllers can be written as
fm = Km (xm xmd ) dt + Bm (xm xmd ),
(6.16)
fs = Ks (xsd xs ) dt + Bs (xsd xs ),
which are two Proportional-Integral (PI) controllers with transfer functions
Km
+ Bm em , em = xm xmd ,
fm =
s
(6.17)
Ks
+ Bs es ,
es = xsd xs .
fs =
s
In this way, we are not only transmitting energy through the strings Km , Ks , but
also dissipating some of it in the dampers Bm , Bs . Consequently, artificial damping
is injected, which is a requisite to ensure a passive behavior when it is possible that
the plant interacts with the environment. The parameters of this controller are the
228
6 Application Cases
stiffness constants of the springs, Km and Ks , and the viscous frictions, Bm and Bs .
If the latter are chosen to be equal to the scattering parameter of the transmission
line, B, no wave reflections will take place [34].
xI = xCI = xm xmd
if c = 0,
x+ x
I
I
if c = 0,
(6.18)
+ = 0
xCI
xI
fm = Km [(1 p) p] xCI + Bm em ,
with c the condition determining the reset instants (discussed later in (6.24)),
(xI , xCI ) the states of the linear and Clegg integrators, respectively, and (1 p, p)
the parameters that multiply the integrator states.
The PI + CI controller can be implemented as two parallel integrators (see
Fig. 5.1, with kp = Bm , Km = kp /i ), one of them being a Clegg integrator and
the other a normal one. The outputs of these two integrators are multiplied by p and
1 p respectively, for 0 < p < 1. Thus, the modification of parameter p provides a
simple way of tuning the controller.
The reason for performing only partial reset is that a totally reset integrator
loses the fundamental property of eliminating the steady-state error. Since feedback
control is performed on the velocity signal, full reset leads to a loss of information
about the slave position that cannot be recovered. When this happens, position drift
appears, a well-known problem in bilateral teleoperation. However, when partial
reset is used, position information is being kept in the linear integrator (the nonreset integrator placed in parallel to the reset one).
This can be more formally explained as follows. A steady state is achieved if
all the forces and velocities in Fig. 6.13 are zero. According to [17] and references
therein, the steady-state tracking-position error, referred to as position drift, is
zero if:
t
(xm xmd ) = 0,
(6.19)
xm =
0
where t is some time at steady-state. Let us assume that the error is initially zero, that
is, xm (0) = 0. Now consider the master impedance controller in Fig. 6.13 replaced
229
Bm
Km ,
xi = xI , and
(6.20)
y = fm = Bm (xm xmd ) + Km (xI (1 p) + xCI p),
t
where xI = 0 (xm xmd ) dt and xCI = tk (xm xmd ) dt, with tk being the last reset
time. Now, when in steady state, the controller has zero input (velocities) and zero
output (force). The problem is whether there may exist nonzero steady-state values
xI , xCI compatible with zero input and output. We will discuss three possible cases:
PI, CI, and PI + CI.
When a PI is used as a matching controller, the fact that the output is zero under
zero input yields (6.19) directly because xm corresponds to the state of the PI
controller. On the other hand, if a CI controller (this is, a PI + CI with p = 1) is used,
xm can be non-null under zero output and zero input due to the fact
t that every reset
action is deleting the controllers memory; null reset state entails tk (xm xmd ) = 0,
where tk is the instant of the last reset action. The advantage of using a PI + CI with
p < 1 is that (6.19) is guaranteed. This is based on the fact that the reset state (xCI )
cannot maintain a non-null value under zero input. In principle, there may exist a
combination (xCI , xI ) of non-null states in (6.18) that yield zero output (fm ) with
zero input. However, a sustained zero input entails that the reset law holds and that
the reset state is made zero. Therefore, at steady-state, the linear state (xI ) has to be
zero, and since xm = xI , (6.19) holds and the steady-state tracking-position error
is zero, as in the linear (PI) case.
Hence, the PI + CI manages to improve the transient response while retaining
the zero steady-state error property. This entails a trade-off between transient and
steady state performance, and the parameter p offers an intuitive way of adjusting
it. This can be done by a simple trial-and-error procedure, decreasing its value from
p = 1 until the steady state performance is good enough, that is, with an overshoot
and settling time which are both small enough to be acceptable.
The option of resetting the master side controller and not the slave one is supported by several sets of simulations and experiments. In them, the effect of resetting
the master has been found to be much more important, while reset of the slave has
had little influence. The explanation is related to the fact that the signal round trip
(from the velocity specified by the operator xm to force feedback fm in Fig. 6.13)
is closed by the master controller forming an external loop around the plant (slave)
that contains the time-delays. On the other hand, the slave PI forms an inner loop not
affected by delay. As the expected performance improvement due to reset is more
effective under fundamental limitations (time-delays), the master controller, and not
the slave, is the one that should be reset. The value of p should be chosen so that it
is close enough to 1 to reduce overshoot, but not so high as to affect the capability
of eliminating the steady state position error.
t
230
6 Application Cases
and a CI compensator (see Fig. 5.1), its passivity results from the passivity of both
subsystems as follows. Without loss of generality, gains have been restructured in
order to show a simpler and easier to understand proof. First, passivity of a PI can
be proven with the classical result for linear systems [28]: a system with a transfer function H (j ) is passive if Real[H (j )] 0 for all R. The PI transfer
function is given by:
1
,
(6.21)
HPI (j ) = k 1 +
T j
where T =
T
1p
(6.22)
Therefore, the PI compensator is passive. On the other hand, using Proposition 3.10 (see also [14]), passivity of a full reset compensator results from the passivity of its base linear compensator. The CI base linear compensator is a linear inblc (j ) = 1 . Since Real[H blc (j )] = 0 0
tegrator and its transfer function is HCI
CI
j
for all R, the linear integrator is passive, and therefore, using the above mentioned result, the full reset compensator CI is also passive.
Hence, it is immediately seen that the PI + CI compensator defined by (6.18) is
passive.
6.2.2.2 Reset Instants
The aforementioned passivity property is independent of the reset instants, provided
that there are a finite number of resets in any finite interval. This enables the designer
to choose the reset instants in the most convenient way for each application. In the
case of teleoperation with a known, constant delay, it is possible to profit from this
freedom by anticipating to this delay, in the following sense: the reset instants can be
chosen so that they correspond to the times when the velocity error is approaching
zero. The resulting strategy can be visualized in Fig. 6.14, where the evolution of the
velocity error, ev (t) = xm (t) xmd (t), is plotted. By monitoring the velocity error
and its variation, the reset instants can be chosen as those satisfying the following
condition:
ev (t1 )
ev (t1 ) + ev (t1 )T = 0.
(6.23)
ev (t1 ) =
T
Hence, (6.23) defines the reset condition c = 0 in (6.18), which amounts to applying a PD control on the error signal ev (t). That is, with a slight abuse of notation,
the signal c(t) that triggers resetting in (6.18) is fixed to
c(t) = (1 + sT ) ev (t).
(6.24)
Thus, instead of waiting for the error to be zero, this situation is predicted and
reset is carried out in advance. This mechanism also holds if the slave interacts with
the environment: in case of a contact, the slaves velocity is modified and the reset
instants are changed accordingly.
231
6.2.3 Examples
6.2.3.1 Example 1: Simple Robot
In order to show that the reset control action is useful in different applications, two
different teleoperated systems with two different approaches are chosen. First, we
choose a simple model of a one-dimensional robot used in many simulations of
bilateral teleoperation, consisting of a mass with friction. The Master Device and
Slave Device blocks depicted in Fig. 6.13 are then given by:
1
(fh fm ),
mr s + br
1
xs =
(fs fe ).
mr s + br
xm =
(6.25)
232
6 Application Cases
Fig. 6.15 Teleoperation in free space: (solid lines) master device, (dashed) slave; (top) without
reset, (middle) with partial reset (p = 0.5), (bottom) full reset
233
position, xm (0) = xs (0) = 0 m; but now there is a wall in the slave location at xs =
0.5 m. A sinusoidal force with a frequency of 1 rad/s is applied by the simulated
operator at the master side (fh ), with an amplitude of 2 N during the first cycle
and of 5 N during the second and third. After 15 seconds, this force is made zero,
simulating that the operator stops holding the master device. The coefficients of the
impedance controllers are now Km = Ks = 15 kg/s2 , Bm = Bs = 20 kg/s, and the
scattering factor is also set to b = 20 kg/s. The evolution of the system is shown in
Fig. 6.16, for the no-reset, full-reset, and partial-reset cases (from top to bottom). If
no reset is performed, the plant exhibits considerable overshoot.
A reset strategy can be applied to solve this issue. Since the overall delay is
T = 1, the reset instants satisfy ev + ev = 0. In this case, a full reset strategy (see
Fig. 6.16, middle) and a partial reset strategy with p = 0.85 (Fig. 6.16, bottom)
achieve similar results as far as overshoot reduction is concerned: at the time of 3
seconds, the overshoot is eliminated; at the time of 6 seconds, it is reduced by
more than 50%; and at the time of 13 seconds, by more than 60%. In both cases,
when moving in free space, the overshoot is reduced and, due to the fact that we
are anticipating the time delay, the slave position tracks the master position more
accurately, so the shapes of the trajectories are more similar. The difference in performance between both strategies is that with full reset there is an error in steady
state, which does not appear with partial reset: in this latter case, after the master device is released by the operator, this is pushed by the feedback force until it matches
the position of the slave device, thus eliminating the steady state position error. Obtaining simultaneously these two features (good tracking with reduced overshoot,
and no error in steady state) is possible due to the use of partial reset.
In Fig. 6.17, the force feedback to the user (fm ) is plotted for the no-reset (dashed
line) and reset cases (solid line). As expected, the use of reset decreases the magnitude of the force at the reset instants, resembling the physical equivalent of a charged
spring whose deformation is suddenly reduced.
234
6 Application Cases
Fig. 6.16 Manoeuvre with contact: (solid lines) master device, (dashed) slave; (top) no reset,
(middle) full reset, (bottom) partial reset (p = 0.85)
235
Fig. 6.17 Feedback forces with partial reset and without reset
Fig. 6.18 Gantry crane, with and without partial reset (p = 0.95)
236
6 Application Cases
Fig. 6.20 Teleoperating a crane over the Internet: (solid lines) master device, (dashed) slave
(gantry crane); (top) no reset, (bottom) partial reset (p = 0.95)
The proposed method resets the state of the master impedance controller and benefits from choosing the reset instants so that the system anticipates the communications channel time delay.
237
The simulated and experimental results presented demonstrate the good performance of the proposed scheme for different applications. The solution has shown its
usefulness even in the Internet-based teleoperation with time-varying delays.
An interesting research direction for future work is to consider the possibility of
using a time-varying parameter p, and its implications for passivity. An extension
for unknown, time-varying delays, is another natural line of research.
238
6 Application Cases
239
one located around 0.02 rad/s) [11]. Thus, the controller tuning should be careful to
avoid exciting these antiresonance modes.
6.3.2 PI + CI Design
Since the main goal of this work is to design a robust control system that guarantees
both stable and optimal operation for the ACUREX field, QFT will be used to design
the base LTI control system. Closed-loop stability and tracking specifications are
considered. In a second stage, this base compensator will be used to design the final
reset control system.
0.296
s2
0.0084 s + 0.00842 )
e20s .
s2
0.82
s2
)(1 + 0.01 s + 0.012 )
0.00452
0.83(1 +
(1 +
1.6
0.0045 s
(6.26)
Here the PI controller is tuned by making the open-loop gain fit both the stability and tracking boundaries. A shaping is shown in Fig. 6.23, corresponding to a
proportional gain of kp = 1.80 and an integral time constant of i = 150 s.
Finally, once the base PI compensator is tuned, a closed-loop system analysis is
performed to ensure that the system fits the given stability and tracking specifications. This is usually referred as to compensator validation in QFT. This analysis
has to be done for a wider set of frequencies than the one used in the templates
computation, and in this case, this set goes from 0.0001 up to 0.1 rad/s with a larger
240
6 Application Cases
grid. In Fig. 6.24, the stability analysis is shown, whereas the tracking analysis is
depicted in Fig. 6.25.
As it can be seen, the designed closed-loop system satisfies both stability and
tracking specifications, for every frequency, thus the base PI compensator design is
finished.
6.3.2.2 PI + CI Tuning
Once the base PI compensator has been designed, the obtained PI parameters are
chosen as part of the PI + CI parameters. But, in addition, another parameter, the
reset percentage (preset ), must be tuned. For a proper selection of this parameter, it
is important to note that for high values of preset a worse closed-loop response can
be expected. Here, the tuning rules given in Chap. 5 (see also [6, 7]) are used for
tuning this parameter for the nominal loop. The result is an increment of the phase
margin and thus a better transient response for the nominal loop, and also for rest of
possible loops derived from the uncertainty due to the extra phase lead given by reset
compensation. A value preset = 0.35 is obtained. For comparison purposes, Bode
plots of the nominal loops with the plant (6.26) and both compensators: the base
PI compensator (kp = 1.80, i = 150), and the PI + CI compensator (kp = 1.80,
i = 150, and preset = 0.35) are shown in Fig. 6.26. The extra phase lead given by
241
242
6 Application Cases
243
reset action can be very sensitive to these changes. In fact, in the PI + CI design,
the reset percentage has been tuned by fixing a low value, preset = 0.2, but from
244
6 Application Cases
Fig. 6.29 Field responses of PI + CI with both fixed and variable reset
Fig. 6.27, it can be seen that this fixed value of reset percentage has only improved
the field response at some reset times. To extend this response improvement over all
the reset times, the reset percentage should be variable.
In the following, the reset percentage is modified at each reset time using (5.47)
and (5.48). Consider the system responses given in Fig. 6.27, that is, the reference
temperature will experience the same step changes along the simulation. The control will be done by the PI + CI controller tuned in previous section, where its parameters were kp = 0.77, i = 120, and preset = 0.2. In addition, another PI + CI
controller is used with the same parameters, but with a variable reset percentage as
in (5.47)(5.48), where the chosen values are p reset = 0.05, d = 0.55, and f = 2.
The systems responses are shown in Fig. 6.29.
Comparing closed-loop responses in Fig. 6.29, it can be said that closed-loop
response is considerably improved by making the reset percentage variable since
both overshoots and undershoots are strongly reduced without decreasing the speed
of response. Finally, for comparison purposes, in Fig. 6.30, closed-loop responses
are plotted for the base PI controller and the already tuned PI + CI controller with
variable reset percentage.
As can be seen in Fig. 6.30, the PI + CI controller with a variable reset percentage makes the response to have less overshoot in comparison to the PI, without
modifying the response speed. Therefore, it can be concluded that a variable reset
percentage improves considerably the PI response.
References
245
References
1. Aguilar, R., Martnez, S.A., Rodrguez, M.G., Soto, G.: Process analysis for treatment of
industrial plating wastewater: simulation and control approach. Chem. Eng. J. 105, 139145
(2005)
2. Alpbaz, M., Karacan, S., Cabbar, Y., Hapoglu, H.: Application of model predictive control
and dynamic analysis to a pilot distillation column and experimental verification. Chem. Eng.
J. 88, 163174 (2002)
3. Altinten, A., Ketevanlioglu, F., Erdogan, S., Hapoglu, H., Alpbaz, M.: Self-tuning PID control
of jacketed batch polystyrene reactor using genetic algorithm. Chem. Eng. J. 138, 490497
(2008)
4. Anderson, R.J., Spong, M.W.: Bilateral control of teleoperators with time delay. IEEE Trans.
Autom. Control 34(5), 494501 (1989)
5. Arbaoui, M.A., Vernires-Hassimi, L., Seguin, D., Abdelghani-Idrissi, M.A.: Counter-current
tubular heat exchanger: modeling and adaptive predictive functional control. Appl. Therm.
Eng. 27, 23322338 (2007)
6. Baos, A., Vidal, A.: Design of PI + CI reset compensators for second order plants. In: IEEE
International Symposium on Industrial Electronics, Vigo, Spain (2007)
7. Baos, A., Vidal, A.: Definition and tuning of a PI + CI reset controller. In: Proceedings of
the European Control Conference, Kos, Greece (2007)
8. Baos, A., Yaniv, O., Montoya, F.J.: Nonlinear QFT synthesis by local linearization. Int. J.
Control 76(5), 429436 (2003)
9. Baos, A., Garcia, P., Checa, L.: Robust control of thermal treatments in can industry. In: Proceedings of the 4th IFAC/CIGR WorkshopControl Applications in Post-Harvest and Processing Technology (CAPPT 2006), Potsdam, Germany (2006)
10. Berenguel, M., Camacho, E.F., Rubio, F.R.: Simulation software package for the ACUREX
field. Internal Report, Departamento de Ingeniera y Automtica, ESI Sevilla, Spain (1994)
11. Camacho, E.F., Berenguel, M., Rubio, F.R.: Advanced Control of Solar Plants. Springer,
Berlin (1997)
246
6 Application Cases
12. Camacho, E.F., Rubio, F.R., Berenguel, M., Valenzuela, L.: A survey on control schemes for
distributed solar collector fields. Part I: Modeling and basic control approaches. Sol. Energy
81, 12401251 (2007)
13. Camacho, E.F., Rubio, F.R., Berenguel, M., Valenzuela, L.: A survey on control schemes for
distributed solar collector fields. Part II: Advanced control approaches. Sol. Energy 81, 1252
1272 (2007)
14. Carrasco, J., Baos, A., van der Schaft, A.J.: A passivity-based approach to reset control systems stability. Syst. Control Lett. 59(1), 1824 (2010)
15. Carrizales-Martnez, G., Fermat, R., Gonzlez-Alvarez, V.: Temperature control via robust
compensation of heat generation: Isoparaffin/olefin alkylation. Chem. Eng. J. 125, 8998
(2006)
16. Cervera, J.: Ajuste automtico de controladores en QFT mediante estructuras fraccionales.
Ph.D. Thesis, University of Murcia, Spain (2006)
17. Chopra, N., Spong, M.W., Ortega, R., Barabanov, N.E.: On tracking performance in bilateral
teleoperation. IEEE Trans. Robot. 22(4), 861866 (2006)
18. Chopra, N., Berestesky, P., Spong, M.W.: Bilateral teleoperation over unreliable communication networks. IEEE Trans. Control Syst. Technol. 16(2), 304313 (2008)
19. Cirre, C.M., Moreno, J.C., Berenguel, M.: Robust QFT control of a solar collectors field. In:
Martnez, D. (ed.) Improving Human Potential ProgrammeResearch Results at PSA Within
the Year 2002 Access Campaign. Serie Ponencias CIEMAT, pp. 2733 (2003)
20. Fernndez, A., Barreiro, A., Raimndez, C.: Digital passive teleoperation of a gantry crane.
In: Proc IEEE Int. Symp. Ind. Electron., Vigo, Spain (2007)
21. Fernndez, A.F., Barreiro, A., Baos, A., Carrasco, J.: Reset control for passive teleoperation
applications in process control. In: 34th Annual Conference of the IEEE Industrial Electronics
Society, Orlando, Florida, USA (2008)
22. Fernndez, A.F., Barreiro, A., Baos, A., Carrasco, J.: Reset control for passive bilateral teleoperation. IEEE Trans. Ind. Electron. (2010). doi:10.1109/TIE.2010.2077610
23. Fischer, M., Nelles, O., Isermann, R.: Adaptive predictive control of a heat exchanger based
on a fuzzy model. Control Eng. Pract. 6, 259269 (1998)
24. Hokayem, P., Spong, M.W.: Bilateral teleoperation: an historical survey. Automatica 42(12),
20352057 (2006)
25. Horowitz, I.M.: Quantitative Feedback Theory. QFT Press, Boulder (1992)
26. Hyodo, S., Soeda, Y., Ohnishi, K.: Verification of flexible actuator from position and force
transfer characteristic and its application to bilateral teleoperation system. IEEE Trans. Ind.
Electron. 56(1), 3642 (2009)
27. Karacan, S., Hapoglu, H., Alpbaz, M.: Application of optimal adaptive generalized predictive
control to a packed distillation column. Chem. Eng. J. 84, 389396 (2001)
28. Khalil, H.K.: Nonlinear Systems. Prentice Hall, New York (2002)
29. Khan, S., Sabanovic, A., Nergiz, A.O.: Scaled bilateral teleoperation using discrete-time
sliding-mode control. IEEE Trans. Ind. Electron. 56(9), 36093618 (2009)
30. Landau, I.D.: System Identification and Control Design. Prentice Hall, New York (1990)
31. Lee, Y., Park, S., Lee, M., Brosilow, C.: PID controller tuning for desired closed-loop responses for SI/SO systems. AIChE J. 44(1), 106115 (2004)
32. Loulch, Z., Cabassud, M., Le Lann, M.V.: A new strategy for temperature control of batch
reactors: experimental application. Chem. Eng. J. 75, 1120 (1999)
33. Maidi, A., Diaf, M., Corriou, J.P.: Optimal linear PI fuzzy controller design of a heat exchanger. Chem. Eng. Process. 47, 938945 (2008)
34. Niemeyer, G., Slotine, J.-J.: Stable adaptive teleoperation. Int. J. Ocean Eng. 16(1), 152162
(1991)
35. Rivera, D.E., Morari, M., Skogestad, S.: Internal model control. 4. PID controller design. Ind.
Eng. Chem. Process Des. Dev. 25(1), 252265 (1986)
36. Skogestad, S.: Simple analytic rules for model reduction and PID controller tuning. J. Process
Control 13(4), 291309 (2003)
References
247
37. Slama, T., Trevisani, A., Aubry, D., Oboe, R., Kratz, F.: Experimental analysis of an internetbased bilateral teleoperation system with motion and force scaling using a model predictive
controller. IEEE Trans. Ind. Electron. 55(9), 32903299 (2008)
38. Smith, C.L., Corripio, A.B., Martin, J.: Controller tuning from simple process models. Instrum. Technol. 22(12), 3944 (1975)
39. van der Schaft, A.J.: L2-Gain and Passivity Techniques in Nonlinear Control. Springer, London (2000)
40. Vidal, A., Baos, A.: QFT-based design of PI + CI reset compensator: applications in process
control. In: 16th IEEE Mediterranean Conference on Control and Automation, Ajaccio, France
(2008)
41. Vidal, A., Baos, A.: Reset compensation for temperature control: experimental applications
on heat exchangers. Chem. Eng. J. 159(13), 170181 (2010)
42. Vidal, A., Baos, A., Moreno, J.C., Berenguel, M.: PI + CI compensation with variable reset: application on solar collector fields. In: 34th Annual Conference of the IEEE Industrial
Electronics Society, Orlando, Florida, USA (2008)
43. Vidal, A., Vidal, A., Baos, A.: Reset compensation applied on industrial heat exchangers.
In: 14th IEEE International Conference on Emerging Technologies and Factory Automation,
Mallorca, Spain (2009)
44. Wirz, R., Marin, R., Ferre, M., Barrio, J., Claver, J.M., Ortego, J.: Bidirectional transport
protocol for teleoperated robots. IEEE Trans. Ind. Electron. 56(9), 37723781 (2009)
Index
A
ACUREX field, 238
After-reset surface, 58, 73
Anticipating reset, 178, 230
B
Beating, 63
C
Clegg integrator
definition, 6
describing function, 17, 129
Horowitzs design procedure, 20
Clegg integrator with advanced reset/variable
reset band
definition, 17
describing function, 18, 135
Clegg integrator with reset band
describing function, 132
Crane, 233
D
Deadlock, 63
Delay margin, 167
Delay-dependent stability
frequency domain, 168
time domain, 164
Delay-independent stability
frequency domain, 155
time domain, 153
Delayed differential equation, 150
Delta (Dirac) function, 160
Detuned PI (PI-des), 195
Dirac impulse (delta), 169
Dominant eigenvalue, 61
Dwell time, 105
E
Existence and uniqueness of solutions, 151
F
Feedback force, 233
Feedforward, 238
First order plants, 184
First order plus deadtime (FOPDT), 188, 215
lag dominant, 190
time-delay dominant, 193
FORE (first order reset element)
definition, 1
describing function, 129
Horowitzs design procedure, 23
passivity, 119
FORE with reset band, 131
describing function, 132
G
General reset control system, 26
full reset, 26, 117
partial reset, 26, 120
with reset band, 130
General reset controller, 7
describing function, 134
with reset band
describing function, 132
H
H-beta condition, 27, 156
generalized, 158
Half-rule, 195
Haptic device, 235
Heat exchanger, 212
High-order systems, 195
Hybrid systems, 37
249
250
I
IAE (integral absolute error), 221
Impedance controllers, 227
Impulsive control system, 32
Impulsive dissipative system, 35
Impulsive systems, 27
autonomous systems with impulses, 30
impulses at fixed times, 29
impulses at variable times, 29
Input strictly passive (ISP), 116
Inputoutput stability, 113
Integrating systems, 198
Internal model control (IMC), 190, 216
Skogestad IMC (SIMC), 190
ITAE (integral time absolute error), 221
J
Jordan form, 61
K
KalmanYakubovichPopov (KYP) lemma,
155
L
Limit cycles, 136
Linear base system, 59
Linear fundamental limitations, 4
frequency domain, 11
time domain, 9
Linear matrix inequality (LMI), 154, 166
Lyapunov stability
definitions, 94
H-beta condition, 95
reset-times dependent conditions, 99
reset-times independent conditions, 94
LyapunovKrasovskii (LK) functionals, 153,
165, 175
O
Output strictly passive (OSP), 116, 157, 174
P
Passive system, 116
pH control, 125
PI+CI, 181
design for a heat exchanger, 216
design for solar collector field, 239
design for teleoperation, 228
summary of tuning rules, 202
variable reset band/advanced reset, 205,
223
with fixed reset band, 203, 222
Port-Hamiltonian systems, 227
Index
Position drift, 228
PRBS (pseudo random binary signal), 215
Q
Quantitative Feedback Theory (QFT), 208,
218, 237
templates, boundaries, and loop shaping,
208
templates, boundaries and loop shaping,
239
R
Rectangular pulses, 172
Reduced initial condition, 61, 100
Regular reset system, 100
Reset band, 129
Reset lag compensator, 119
Reset percentage (ratio), 183, 240
variable, 206, 242
Reset PID compensator, 120
Reset PII R compensator, 122
Reset surface, 58
Reset times, 2
Robustness, 174
S
SCADA, 213
Scattering, 226
Small gain, 159, 174
Solar collector fields, 237
Spectral projector, 61
Stabilization by reset, 105
stable base system, 105
unstable base system, 110
Storage functional, 168
Strictly positive real (SPR), 157, 174
T
Teleoperation, 224
passive, 225
Temperature control, 211
Time regularization, 59, 151
V
Variable reset, 184
Variable reset band, 134
Very strictly passive (VSP), 116
W
Well-posedness, 63
Z
Zeno solutions, 63, 67
Wastewater Systems
Carlos Ocampo-Martinez
Tandem Cold Metal Rolling Mill Control
John Pitter and Marwan A. Simaan
Lebih dari sekadar dokumen.
Temukan segala yang ditawarkan Scribd, termasuk buku dan buku audio dari penerbit-penerbit terkemuka.
Batalkan kapan saja.