SMC SarahSpurgeon
SMC SarahSpurgeon
SMC SarahSpurgeon
Control
THEORY
ANO APPLICATIONS
Christopher Edwards
and Sarah K. Spurgeon
Sliding Mode Control
Theory and Applications
Systems and Control Book Series
Editcd by John O'Reilly and Eric Rogers
Published
Forthcoming
Prom I'rocess Data to PID Controller Design, L. Wang and W.R. Cluett
Sliding Mode Control:
Theory and Applications
Christopher Edwards
and
Sarah K. Spurgeon
This book contains infonnat ion obtaine cl from a11thentic a ncl highly regarclecl so11r ces Re printe cl mate rial is <JIIOtecl with
permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish
reliable data and infonnation, but the author and the publisher cannot assume responsibility for the validity of ali materia ls
or for the consequences of their use.
No part of this book may be reprinted, reproduced, transmitted, or utilized in any fonn by any electronic, mechanical, or
other means, now known or hereafter invented, including photocopying, microfílming, and recording, or in any information
storage or retrieval system, without written permission from the publishers.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.
Series Introduction x ii
Preface xiv
Appendices
References 225
Index 233
Series Introduction
It was in response t,o these exciting ncw developrncnts that the present book series
of Systcm and Control was conceived. It publishes high-quality research texts and
reference works in the diverse areas which systems and control now includcs. In
addition to basic theory, experimental and/or application studies are wclcorne, as
an) cxpository texts where theory, verification and applications come togethcr to
provide a unifying coverage of a particular topic or topics.
E. Rogers
J. O'Reilly
Preface
Fully worked design examples are an addit ional tutoría! feature. Industrial case
studies, which present the results of sliding mode controller implementations, are
used to illustrate the successful practica! application of the theory.
The book is structured as follows. Chapter l introduces the concept of sliding
mode control and illustrates the attendant features of robustness and performance
specification using a straightforward example and graphical exposition. The neces-
sary background for more formal development is presented in Chapter 2 to enhance
readability.
Chapter 3 formulates the general multivariable sliding mode control problem. The
interpretation of the sliding surface design problem as a straightforward linear
state feedback problem for a subsystem is emphasised. Possible control strategies
to enforce sliding motion are identified and the problem of smoothing undesirable
discontinuous signals addressed. Controller design issues are considered in Chap-
ter 4. A number of methods for sliding surface design are described. The so-called
unit vector control structure is exploited as this lends itself to the development of
simple numeric design algorithms. Frameworks for the solution of both regulation
and tracking problems are presented. A design study from the flight control area
is used to illustrate the application of the proposed design methods.
As developed thus far, the controllers require full state information. In practice, it
may be impossible or impractical to measure all of the system states. The next three
chapters address this problem. Chapter 5 investigates the possibilities of obtaining
sliding mode schemes where the control law ancl decision rule are a function of
the measured outputs. The associated limitations relating to both nominal system
structure and attendant design freecloms are demonstrated. A design algorithm is
presented and illustrated using numerical examples. Chapters 6 and 7 consider the
possibilities of both state reconstruction and control using sliding mode techniques.
Again design fra.mcworks are cmphasiscd.
The book concludes with industrial studies describing both sliding mode controller
design and implementation. Chapter 8 considers two automotive studies. The
automotive environment is harsh; components are subject to large temperature
variations and shock, and parameters can be expected to vary significantly during
the lifetime of the vehicle. Robustness is thus a key requirement. Further, there
is the need to produce high levels of performance and safety from inexpensive
components. Sliding mode design will be seen to fulfil these requirements.
Chapter 9 considers sliding mode control of a gas-fired furnace. The furnace cly-
namics are highly nonlinear and any control strategy is required to yield robust
performance over a la.rge operating range. Tracking requirements a.re essential;
when firing ceramics, for example, it is essential that the furnace temperature
follows a prespecified profile exactly if the end product is to be of good quality. Ef-
ficiency of combustion is also a priority; large magnitude interacting control signals
result in significant gas wastage. It will be established that sliding mode control
can achieve the desired specifications.
In terms of the interdependence of the individual chapters, Chapters 1 and 2 are
not formally connected to the rest of the text. Chapter 1 only seeks to motívate
sorne of the ideas that are expressed rigorously in Chapter 3. Chapter 2 is meant
to inclicate the topics in systems theory that are required in subsequent chapters.
xvi PREFACE
Readers who have not encountered the state-space and Lyapunov theory presented
there are encouraged to refer to more specialist texts in these particular areas which
develop the ideas in a more measured and expert way. Chapter 3 is the corner-
stone of t he book; in particular Section 3.6 very much sets the tone of the control
laws which are discussed in Chapters 4, 5 and 7. Chapter 4 is quite strongly
linked with Chapter 3. Although its main thrust is to describe different hyper-
plane design strategies, it <loes introduce in a tutoría! way the integral action and
model-following approaches for reference tracking, which subsequently appear in
Chapters 5 and 7. Although Chapter 5 (by necessity) considers and develops differ-
ent control strategies because of the removal of the state availability assumption,
parallels are still drawn with the results of Chapter 3. Chapter 6, which is pri-
marily concerneci with the ciesign of ohservers, <loes 11tilise sorne of t.he icieas and
results from Chapter 5 because a similar canonical form is used as the fundamental
design framework. As a result, sorne of the proofs are more sketched but suitable.
cross-referencing has been made to the original development. (This is true of the
run of chapters from 3 through to 7.) Chapter 7 essentially reworks sorne of the
ideas of Chapter 4 (the reference tracking methodologies) and considers t he state
feedback controllers in conjunction with the state observers developed in Chap-
ter 6. The first actuator case study in Chapter 8 could (almost) be tackled once
the concepts from Chapter 1 have been understood. The idle speed control problem
in Section 8.3 and the furnace case study in Chapter 9 rely heavily on the integral
action controller exposition from Chapter 7.
Finally it is our plea.sure to thank our families, friends and colleagues who ha.ve
helped in the preparation of this book. Severa! deserve a special mention. First and
foremost we would like to thank Aamer Bhatti for his assistance in the preparation
of Chapter 8. We are especially grateful to him for allowing us to use, at the time of
writing, unpublished work pertaining to the control of engine speed. We also wish
to thank Russell Jones of Lucas Varity for his help with the automotive actuator
problem, which also appears in Chapter 8. The furnace application described in
Chapter 9 would not ha.ve taken place without the help and technical assistance of
Haydn Porch, Sean Goodhart, Rut h Davies and Patrick Holmes. We thank them for
their cooperation and patience during the trials initially at the i\ilidlands Research
Centre at Solihull and subsequently at the Gas Research Centre in Loughborough.
In terms of the preparation of the document we would like to thank Doug Pratt
for drawing many of the figures which appear throughout the text. We are also
very grateful to the brave souls who helped proofread and provide valuable feed-
back: we would like to thank Robert Cortez who suffered t he earliest draft and also
our graduate students Ashu Akoachere and Guido Hermman. We are particularly
grateful to Xinghuo Yu for his thorough reading of the first draft and for provid-
ing many pertinent and insightful comments which we have tried to incorporate.
Last but not least, we would like to thank Xiao-Yun Lu for helping us read and
correct the final proofs. Of course, despite our best efforts, we are sure that many
typographical errors still remain; for these we t ake ful] responsibility.
Chris Edwards
Sarah Spurgeon
March 1998
Chapter 1
1.1 IN TRODUCTION
Variable structure control systems (VSCS) evolved from the pionecring work in
Russia of Emel'yanov and Barbashin in the early 1960s. The ideas <lid not appear
outside of Russia urrtil the mid 1970s wherr a book by Itkis (1976) arrd a survey
paper by Utkirr (1977) were published irr Errglish . VSCS corrccpts havc subsequently
been utilised irr thc design of robust regulators, model-reference systcrns, adaptive
schemes, tracking systems, state observers and fault detection schemcs. Thc ideas
have successfully been applied to problems as diverse as automatic flight control,
control of electric motors, chemical processes, helicopter stability augmentation
systcms, space systems and robots. This chapter seeks to motívate and introduce
the concepts which will be considered more formally later in the book.
Variable structurc control systems, as the namc suggests, are a class of systems
whercby the 'control law' is deliberatcly changed during the control process ac-
cording to sorne dcfirred rules which dcpend on the state of the system. For the
purpose of illustration consider the double intcgrator given by
jj(t) = u(t) (1.1)
Initially consider the effect of using the feedback control law
u(t) =- ky(t) (1.2)
where k is a strictly positive scalar. One way of analysing thc rcsult ing closed-loop
motion is by means of a phase portrait, which esscntially is a plot of velocity against
position. Substituting for the control action in cquation (1.1), and multiplying the
resulting equation throughout by y, yields
yjj = -kyy (1.3)
Integrating this cxpression gives the following relationship between velocity and
position
i;2 + ky2 = e (1.4)
whcre e reprcsents a constant of integration resu lting from the initial conditions and
is strictly positive. Importantly, time <loes not appear explicitly in the expression
2 SLIDING MODE CONTROL
in (1.4). In the special case when k = l, equation (1.4) represents a circle with
centre at the origin and radius ,/c. More generally, a plot of iJ against y is an
ellipse which depends on the initial conditions as shown in Figure 1.1. In terms of
regulation - i.e. steering arbitrary initial conditions to the origin - a control law of
the form given in (1.2) is not appropriate since, as shown in Figure 1.1, the y a.nd
y variables do not move towa.rds the origin. 1
u(t)=-k1 y(t) 1t(t) =-k2 y (t)
(b) y (a) (b) y (a)
where O< k1 < 1 < k 2 . The phase plane (y,y) is partitioned by the switching
rule into four qua.drants separated by the axes as shown in Figure 1.1. The control
la.w u = -k2 y will be in effect in the quadrants of the phase plane la.belled (a).
In this region, the distance from the origin of the points in the pha.se portrait
decreases along the system trajectory. Likewise, in region (b) when the control law
u = - k 1 y is in operation, the distance from the origin of the points in the phase
portrait also decreases. This system clearly fits the description of a VSCS given
earlier. The phase portrait for the closed-loop system under the variable structure
control law u is obtained by splicing together the appropriate regions from the two
pha.se portraits in Figure 1.1. In this way the phase portrait must spiral in towa.rds
the origin a.nd an asympt.otically stable motion results, as shown in Figure 1.2.
This can be verified more forma.lly by considering the function
V (y, y) = y2 + y2 (1.6)
which, from Pythagoras' theorem, represents the square of the dista.nce from the
point (y, y) to the origin in the pha.se pla.ne a.nd ma.y be viewed as representing
the energy of the system. The time deriva.tive of V(y,y) along the closed-loop
trajectories is given by
2yy + 2yy
2yy(l -k1) if yy < o
2y(y +u)= {
2yy(l - k2) if yy > o
1 Since y and y remain bounded for all time the closed-loop is stable; it is not, however,
asymptotically stable - see Section 2.2 for details.
INTRODUCTION 3
This is always negative by the way the gains are constructed. Thus the distance
from the origin is always decreasing, which agrees with the intuitive observation
made earlier. By introducing a rule for switching between two control structures,
which independently do not provide stability, a stable closed-loop system has been
obtained.
A more significant example results from using the variable structure law given by
-1 if s (y, y) > O
u(t) ={ 1 if s(y,y) < O
{l. 7)
where mis a positive design scalar. The reaBon for the use of the term 'switching
function' is clear, since the function given in equation (1.8) is used to decide which
control structure is in use at any point (y, y) in the phase plane. The expression
in equation (1.7) is usually written more concisely aB
where sgn(·) is the .signum, or more colloquially, t he sign function. The signum
function exhibits the property that
.ssgn(s) = Jsl
This simple result will be exploited often in the analysis which follows.
The expression given in (1. 7) is used to control the double integrator. For large
values of iJ the phase portrait, obtained from joining the parabolic components of
the constituent laws, is shown in Figure 1.3. The dotted line in the figure represents
the set of points for which s(y, y) = O; in this case a straight line through the origin
of gradient -m. However, for values of y satisfying the inequality mlyl < 1 then
or equivalently
lim
s--;o+
s< O and lim
._,o-
s> O (1.11)
Consequently, when ml:ril < 1 the system trajectories on either side of the line
[, 5 = {(y,y) : s(y,y) = O} (1.12)
point towards the line. This is demonstrated in Figure 1.4, which shows different
phase portraits intercepting the same point on the line C, from different initial
conditions.
F igure 1.4: Phase portrait of the system under VSC near the origin
Intuitively, high frequency switching between the two different control structures
INTRODUCTION 5
will take place as the system trajectories repeatedly cross the line .C8 • This high
frequency motion is described as chattering. If infinite frequency switching were
possible, the motion would be trapped or constrained to remain on the line .Cs. The
motion when confined to the line .Cs satisfies the differential equation obtained from
rearranging s(y, y) === O, namely
This represents a first-order decay and the trajectories will 'slide' along the line .Cs
to the origin (Figure 1.5).
y
y
trajectory
ss < 0 (1.14)
which is referred to as the reachability condition. Thus, in terms of VSCS design,
the éhoice of the switching function, represented in this situation by the parameter
m, governs the performance response; whilst the control law itself is designed to
guarantee that the reachability condition (1.14) is satisfied. In this case, as argued
earlier, the reachability condition is only satisfied in a domain of the phase plane
This section explores in more detail the properties of the ideal sliding motion and
the control action necessary to maintain such a motion. Consider once again the
double integrator in equation (1.1) and the control law given in (1.7). The phase
portrait in Figure 1.6 is from a simulation of the closed-loop behaviour when m = 1,
and the initial conditions are given by y = 1 and y = O. The two-stage nature of
the dynamics is readily observecl: the initial (parabolic) motion towards the sliding
surface, followed by a motion along the line y = -y towards the origin.
The control action associated with this simulation is given in Figure 1.7. It can
be seen that sliding takes place after O. 732 second when high frequency switching
takes place.
§ 0.5
-~
<.)
-< o
ot,
u
"'o -0.5
-1
o 2 3 4 5 6 7 8 9 10
Time, sec
and thus since s(t) = Oforal! t ~ ts, it follows from (1.15) that a control law which
maintains the motion on Cs is
This control law is referred to as the equivalent control action. This is not the
control signa! which is actually applied to the plant but may be thought of as the
control signa! which is applied 'on average'. This can be demonstrated by passing
the discontinuous control signa] (in Figure 1.7) through the low pass filter
to obtain the low frequency component ua(t). Figure 1.8 shows u 0 (t), together with
the associated equivalent control when T = 0.04. It can be seen that the filtered
(or averaged) control signa! agrees with the equivalent control action defined in
equation (1.16). Of course agreement can only tal(e place once sliding is established
- in this case after O.732 second. Consequently, the dotted line, representing the
equivalent control, is only drawn from this time onwards, as shown in Figure 1.8.
e::
o
·.:,
()
<( 0.5
8
i:
o o
u
".':l~ -0.5
ti:
-1
o 2 3 4 5 6 7 8 9 10
Time, sec
The control signa! applied to the plant may be thought of as comprising 'low' and
'high' frequency components so that
u(t) = ..__.,
Ua(t) + (u(t) - u 0 (t))
'------v-------
lo-w frequency high frequ ency
8 SLIDING MODE CONTROL
For systems which may be modelled as strictly proper transfer functions, 2 for exam-
ple, the high frequency component may be considered to be beyond the bandwidth;
hence the control action that affects the dynamic response is ua(t), i.e. the equiv-
alent control. A more dramatic demonstration of the strength of the notion of an
equivalent control will be given later when the effects of plant/model mismatches
will be examined.
Suppose the double integrator in equation (1.1) is a linear approximation of the
real system on which the control law is to be implemented. Suppose that the real
system is in fact a pendulum, formed from a light rod and a heavy mass, as shown
in Figure 1.9.
íf\. T
.. ....:...... . Mg
The variable B represents t he angular displacement from the vertical and T repre-
sents a torque applied at the point of suspension, which will be considered to be
t he control input to the system. Ignoring the effects of friction, the system can be
represented mathematically as
·· l I
0(t) =- sin 0(t) + Ml2u(t ) (1.18)
9
where M represents the mass of the bob, l the length of the rod and g the acceler-
ation due to gravity. By appropriate scaling, the essential dynamics of the system
are captured by
y(t) = -a1 siny(t) + u(t) (1.19)
numerator polynomial is strict ly Iess than the denominator poly nomial and tbus G(jw) -+ O as
w-+ OO.
INTRODUCTION 9
-0.1
-0.2
-0.3
,>, -0.4
-0.5
-0.6
-0.7
The key result is that, in finite time, the phase portrait intercepts the sliding surface
Ls and is forced to remain there. The significance of this is that, once ideal sliding
is establíshed, the double-integrator system and the normalised pendulum behave
in an identical fashion, namely
An alternative interpretation is that the effect of the nonlinear term a 1 sin y(t),
which may be construed as a disturbance or uncertainty in the nominal double-
integrator system, has been completely rejected. As such, the closed-loop system
is said to be robust, i.e. it is insensitive to mismatches between the model used for
control law design and the plant on which it will be implemented.
The idea of completely rejecting a disturbance or uncertainty is decidedly different
from other control approaches such as H2 and H 00 . These linear methodologies
attempt to minimise in sorne sense the transfer functions relating the disturbances
to the outputs of interest. Indeed, at first, it is difficult to imagine how the effect
of an unknown disturbance can be cancelled. It is at this point that the concept
of the equivalent control can be used to provide insight. Arguing as before, once
an ideal sliding motion has been attained, s(t) = O and .s(t) = Ofor all subsequent
time. From equations (1.8) and (1. 19) it follows that
e: 0.5
.s:
ü
<(
o;:, o
e:
o
u -0.S
-1
o 2 3 4 5 6 7 8 9 10
Time, sec
However, passing the control action in Figure 1.11 through the low pass filter given
in ( 1.17) provides the average applied control signal shown in Figure 1.12. This is
quite different to the average control signal obtained previously and, as expected,
is visually identical to the equivalent control from (1.22), shown as a dotted line.
e:
o
·;:;
u 0.5
<(
g
e:
o o
u
"O
~ -0.5
.B
ti:
-1
o 2 3 4 5 6 7 8 9 10
Time, sec
Apart from the robustness properties exhibited by the sliding motion, another
benefit accruing from this situation is that the pendulum is forced to behave as a
first-order system. This guarantees that no overshoot will occur when attempting
to regulate the pendulum from an arbitrary initial displacement to the origin (pro-
vided sliding can be attained quickly enough). This can be seen from Figure 1.13,
which displays the exponential decay characteristic typical of a first-order system.
The two properties of an ideal sliding motion described previously, namely distur-
bance rejection and order reduction, are the key properties that have motivated
the study of controllers which induce sliding motions.
INTRODUCTION 11
...e
<!)
E 0.8
<!)
ü
"' 0.6
].
Q
.... 0.4
..::!
~ 0.2
e
<(
2 3 4 5 6 7 8 9 10
Time, sec
Consider the pendulum example from the previous section together with the control
law in (1.7) and consider the response of the system to large initial displacements.
Figure 1.14 represents a plot of the switching function with initial conditions y = 3
and y= O.
2 3 4 5 6 7 8 9 10
Time, sec
It can be seen that at approximately 1.5 seconds the trajectory intercepts the
switching line .Cs but insufficient control energy is available to maintain a sliding
motion. The sign of the control law switches, the trajectory pierces the switching
line and moves away before intercepting the line again at approximately 3.5 seconds,
at which point sliding takes place. This result is not perhaps surprising since
the initial conditions correspond to releasing the pendulum from near the upward
vertical which will intuitively generate larger angular velocit ies. T his also agrees
with the theory presented earlier since even for the nominal double integrator, a
s!iding motion could only be guaranteed in the region of the phase plane for which
the angular velocity is less than 1/m. The result in this case is that the pendulum
crosses the downward vertical before a sliding motion can be established and thus
overshoot occurs.
12 SLIDING MODE C ONTROL
Since the key properties of robustness and order reduction are only obtained once
sliding is induced, from the perspective of control law design, the time taken to
induce sliding should be minimised, and the region in which sliding takes place
maximised. If the magnitude of the control signa! is limited to the range ±1 then
the bang-bang controller given in (1. 7) may well represent a legitimate control
scheme. Otherwise it is instructive to consider modifying the control law given in
(1.7), through the inclusion of a linear feedback component, to attempt to ensure
that once the trajectories reach the sliding surface they are forced to remain there.
A candidate control structure is given by
where li ,l2 and p represent scalars yet to be designed. The intention is to choose
the three parameters in (1.23) so that the inequality ss < O is always satisfied for
the normalised pendulum equation. From the definition of the switching function
given in (1.8) it follows that
3
e
.!2
t,
e:
:::,
2
IJ.
bl)
e
:2
2
-~
Cll
2 3 4 5 6 7 8 9 10
Time, sec
3
.,
"
.,E
g 2
~
i5....
:i"'
bl)
e
-<
2 3 4 5 6 7 8 9 JO
Time, sec
The most serious objection t o the control law given earlier is perhaps that choosing
p = l is rather conservative in the sense that it need only be greater than 0.2.:i for
the 17-reach ability condition to be satisfied. A lower value of p would reduce the
amplitude of the high frequency switching, which is advantageous from the point
of view of limiting wear and t ear on the actuators. Unfortunately, retaining t he
same linear component and reducing the sea.lar p results in even slower attaimnent
of the sliding surface. To overcome this difficulty, a j udicious modification to (1.27)
is to add the term -<I>s, where <I> is a p ositive design scalar, so that l 1 = - <I>m and
l2 = - (m + <I>) making
u(t) = -(m + <I>) y(t) - <I>my(t) - psgn(s(t)) (1.28)
Arguing as before, it can be established that with this control law the inequality
ss :5 - <I>s 2 - ryjsj (1. 29)
is satisfied. Consequently, since <I>s 2 ~ O, an 17-reachability condition has been
established and a sliding motion will t ake place. By ignoring t he nonlinear term in
(1.29) it follows that
d
dt is(t)J :5 - <I>Js(t)I
which implies
14 SLIDING MODE CONTROL
where ls(0)I represents the initial distance away from the sliding surface. The
parameter <I> can thus be seen to affect the rate at which the sliding surface is
attained. As a result of this modification, p can be chosen as small as possible (to
reduce the amplitude of the switching) and <I> can be chosen to determine the time
taken to attain sliding.
The following simulation results are obtained from using the control law in (1.28)
with <I> = 1 and p = 0.3 and assuming, as befare, the initial conditions y = O and
y = 3. Figure 1.17 represents the switching function as a function of time and
demonstrates that the time to reach the sliding surface has been greatly reduced.
3 .------,----,---T""-----r-- - . - - - . - - - - , - - - - - , - - - , -- - - ,
o~-~--~--~--~--~--~-~--~--~-~
O 2 3 4 5 6 7 8 9 10
Time, sec
To obtain a faster response, the value of <I> can be increased assuming that suffi-
cient control energy is available. As argued earlier, the reaching-time behaviour
displayed in Figure 1.17 is preferable to Figure 1.15 since ali the robustness and
arder reduction properties only occur once the surface has been attained. The cor-
responding angular displacement is given in Figure l. 18.
3
i::
V
~ 2
"'
~
Q
id
:l
"º
"'
<I'.
o
o 2 3 4 5 6 7 8 9 10
Time, sec
Here again the settling time is much improved when compared with the one ob-
tained in Figure 1.16. The big advantage of the controller given in (1.28) over (1.27)
is that the control signa! is much less aggressive in the sense that the amplitude of
the switching is now ±0.3 as shown in Figure 1.19.
INTRODUCTION 15
e:
o
·,s o
<(
2
uº
o -2
-4
o 2 3 4 5 6 7 8 9 10
Time, sec
In the nominal double-integrator case with p = Oin equation (1.28), it follows that
the closed-loop system satisfies
This represents a stable motion with poles at { - <I>, -m}. It was established earlier
that the pole at -<I> corresponds to the rate at which the sliding surface is attained.
The other pole, located at - m, corresponds to the pole of the sliding motion. It
can thus be argued that the linear part of the control action establishes a sliding
mode for the nominal system whilst the discontinuous component counteracts the
effects of the uncertainty or nonlinearity.
In certain problems, such as control of electric motors and power converters, the
control action is naturally discontinuous and sliding mode ideas can be used to
obtain extremely high performance. 3 Although the control signa! obtained from
Figure 1.19 is preferable to the earlier designs in terms of its chattering behaviour,
in many situations such a control signal would still not be considered acceptable.
A natural solution is to attempt to smooth the discontinuity in the signum func-
tion to obtain an arbitrarily close but continuous approximation. One possible
approximation is the sigmoid-like function
s
VJ(s) = (lsl + ó) ( 1.31)
where ó is a small positive scalar, which is shown in Figure 1.20. It can be visualised
that as ó ➔ O, the function VJ(·) tends pointwise to the signum function. The
variable ó can be used to trade off the requirement of maintaining ideal performance
with that of ensuring a smooth control action.
Assuming the same initial conditions as in the previous section and using the
control law given in (1.28) with VJ(s) replacing sgn(s) with ó = 0.005, the closed-
loop response given below is obtained.
3 For details see Utkin (1992) .
16 SLIDING MODE CONTROL
-
20ó
- 1
-
20ó
The angular displacement (Figure 1.21) is indistinguishable from Figure 1.18. The
key effect, however , is t hat the control action is srnooth, as shown in F igure 1.22.
3
~
e
<l)
8
<l)
o 2
"'
a.
ci"'....
""
~
<(
2 3 4 5 6 7 8 9 10
Time, sec
e
.Q
t, o
<(
2
E -2
uo
-4
o 1 2 3 4 5 6 7 8 9 10
Time, sec
should be stressed that ideal sliding no longer takes place: the continuous control
action only drives the states to a neighbourhood of the switching surface. However ,
arbitrarily close approximation to ideal sliding can be obtained by making 6 small.
In the literature this is often referred to as pseudo-sliding.
(1.32)
The switching function from (1.8) can also be conveniently expressed in matrix
t erms as
s(y,y) = Sx(t) (1.33)
where
S=[m 1]
The normalised pendulum can also be written in state-space form as
(1.34)
s(x) = Sx (1.36)
18 SLIDING MODE CONTROL
The earliest work in English on the properties of sliding modes appeared in the
late 1970s in the form of two books; Itkis (1976) and Utkin (1992). References to
the early work, pioneered mainly in t he former USSR, are given in Utkin (1992).
VSCS concepts have subsequent ly been utilised in the design of robust regulators,
model-reference systems, adaptive schemes, tracking systems, state observers and
fault detection strategies. Many of these ideas will be examined later in the book,
where more extensive rcferences will be given.
A recent survey paper which gives rnany references to the various application areas
in which sliding mode ideas have been u tilised is Hung et al. (1993).
The concept of equivalent control is attributed to Utkin (1977). A rigorous in-
terp retation of the r elationship between the equivalent control law and the low
frequency components of a discontinuous control action maintaining a sliding mo-
tion is described in Utkin (1992). The 17-reachability condit ion stems from the work
of Slotine (1984) .
T hP. r:ontrol law described in equat ion (1.28) is essentially the single-input sirnplifi-
cation ofthe control law proposed by R yan & Corless (1984) . This will be discussed
at length in Chapter 3.
All the simulations in the ch apter have been performed using the Runge-Kutta
integrat ion routines in the SIMULINK libraries of MATLAB.
T he example of the VSCS for t he double integrator in Section 1.1 is taken from
Utkin (1977). The pendulum example has been used by many authors to demon-
strate various nonlinear phenomenon; see for example Khalil (1992).
Chapter 2
2.1 IN TRODUCTION
This chaptcr provides thc necessary background in system tlwory and control topics
that is rcquired for thc sliding mode control dcveloprnents which are to follow. It
can be regardcd as optional rcading for those readcrs familiar with the concepts,
uscd as a reference whilst reading thc rcst of the tcxt, or considered a tutorial
introduction which is rcad in total before addressing subsequent chapters.
whcre x E fil" and the function f : IR" x IH+ ➔ JRn is such that a well-defined
solution to t.he differentia.l equation exists. Denote this solution at time tas x(t,:i: 0 )
wherc x 0 reprcscnts the initial conditions at t = O.
Suppose that f (O, t) = Oforal! t, i.c. the dynamical system dcfincd in (2.1) has an
equilibrium point at the origin. This imposes no undue restriction since if (2.1) has
an cquilibrium point at Xe thcn the change of coordinatcs x M x + Xe translatcs
th<•) equilibriurn point back to the origin.
D e finition 2 .1 The origin of ( 2.1) is said to be stable if given any E > O there
exists a 6 > O ,mch that if llxoll < ó f;hen llx(t,xo)II < E for all t > O.
Less formally thc definition mea,ns that by starting close enough to thc equilibrium
point, the solution will thcreafter always rcmain arbitrarily close.
T he distinction here is that for stability, although the solution x(t, x0 ) must remain
close to the origin, it does not have to return to the origin. For example, consider
20 SLIDING MODE CONTROL
= X2(t) (2.2)
2
-W X1 (t) (2.3)
where w > 1 and x 1(t) and x2(t) are thought of as components of the vector x(t)
in equation (2.1). Arguing as in Section 1.1, it follows that the solution to (2.2)
and (2.3) satisfies
(2.4)
for ali t > O where r is a positive constant depending on the initial conditions. This
represents an ellipse (as shown in Figure 1.1) and it follows that the pha.se portrait
líes between two concentric circles, with centres at the origin, of radius r and r/w.
Thus given any E > O, if the initial conditions x(O) satisfies llx(O)II < é/w then
and hence the simple harmonic oscillator satisfies the definition of stability. How-
ever llx(t)II is bounded away from the origin by a circle of radius r and hence llx(t) II
cannot tend to zero. Therefore the oscillator is stable but not asymptotically stable.
Before discussing in more detail the properties of general nonlinear systems, the
special case of linear systems will be examined.
Consider the special case where f(x, t) is a time invariant map and linear with
regard to the states so that
x(t) = Ax(t) (2.5)
where A E ]Rnxn _ In this situation equation (2.5) yields an analytic solution
(2.6)
This represents a natural extension of the scalar case. The luxury of an explicit
solution to equation (2.5) enables necessary and sufficient conditions for stability
to be established in terms of the eigenvalues of A. Suppose >. is a real eigenvalue of
A and suppose the initial conditions x 0 represent an associated right eigenvect or.
Using the properties of eigenvalues and eigenvectors
and therefore
(Jn , + -1 ", 2 t2 + ... + 1 ", n t n + ...) xo
e
At
xo = + ,..t 2 1n .
MULTIVARIABLE SYSTEMS THEORY 21
where V is a matrix whose columns consist of the linearly independent right eigen-
vectors corresponding to the distinct eigenvalues Ai, i = 1, . .. , n of the matrix A,
¡\ = diag(-\ 1 , ... , >-n) and the rows of the ma.trix WT consist of the correspond-
ing left eigenvectors, i.e. v- 1 = WT . Here x(O) represents a.n a.rbitrary ínitia.l
condítíon. Again the time doma.in cha.ra.cteristics of the system a.re determíned by
the eigenvalues of A. In a.ddition, the a.ssocia.ted ríght eigenvectors determine the
'sha.pe' of a given mode. As the solution depends upon a linear combination of
functions of t he form vie>.,t, a.ppropriate eigenvector entries enable the tra.nsient
e>-;t to contribute, or not, to a particular state varia.ble. In this wa.y it is seen that
the ent ire eigenstructure, and not just the eigenvalues, a.re effective in determining
the time response of a system.
The previous section considered the special case of linear systems. This enabled a.n
analytic expression for the solution to be obta.ined. For general nonlinea.r systems
22 SLIDING MODE CONTROL
(2.9)
D efinit ion 2 .3 The origin of the system (2.1 ) is said to be quadratically stable
if there exist symmetric positive definite matrices P, Q E ]Rn xn such that the
total time derivative satisfies
The inequality above implies JJa:(t) [I < e-at where o:= Amin(P- 1 Q) and hence the
origin is asymptotically stable. If f (x, t) = Ax(t) then it is well known that A has
stable eigenvalues if and only if, given any symmetric positive definite matrix Q,
there exists a unique symmetric positive definite matrix P satisfying the Lyapunov
equation
(2.11)
Consequently, any stable linear system is quadratically stable. A symmetric pos-
itive defin ite matrix P satisfying (2.11) will he rnferre<l to as a Lyapnnov matrix
for the matrix A.
Lyapunov theory may also be used as a means of examining the robustness of a
given linear system; suppose
where the matrix A is stable and t(·) is an imprecisely known function which
represents uncertainty in the system. Let the pair of positive definite matrices
(P , Q) satisfy the Lyapunov equation (2.11) and define
Amin(Q)
µ= (2.13)
Amax (P)
ancl suppose that the uncertain func tion satisfies
then the system in (2. 12) is stable. This can be established by using V= xT Px as
a Lyapunov function: thc dcrivativc along the trajectories satisfies
where t he Cauchy- Schwarz inequality (see Appenclix A.2.6) has been used to obtain
the last inequality. Now
where the Rayleigh principie has been usecl to obtain the middle inequality. Also
directly from the Rayleigh principie
T
-X Qx:::; ->.min(Q)Jlx JJ2 (2.17)
Thus from the inequality in (2.15) and using (2.16) and (2.17) it follows that
and therefore if t(·) satisfies (2.14) the Lyapunov derivative is always negative and
stability is proved.
In view of t he condition (2.14) , it is natural to attempt to choose Q in an effort to
maximise (2.13). Patel & Toda (1980) show t hat t he maximurn is given by
- 1
ti= Amax(P)
(2.18)
(2.20)
24 SLIDING MODE CONTROL
Let E e 1Rn be a bounded set, then the following definition can be made.
Definition 2.4 The solution x(·) to the uncertain system (2.20) is said to be
ultimately bounded with respect to the set [. if
• on any finite interval the solution remains bounded, i.e. if llx(to)II < ó
then llx(t)II < d(ó) for any t E [to, ti)
• in finite time the solution x(t) enters the bounded set[. and remains there
for all subsequent time.
The set [. is usually an acceptably small neighbourhood of the origin and the
concept is often termed practica[ stability.
As an example, consider the uncertain system
where the function ~(-) is unknown but bounded and is assumed to satisfy
For the analysis which follows, A and D will be used to represent the system and
uncertainty dist ribution matrices in equation (2.21). Let P denote the positive
definite solution to the Lyapunov equation
(2.23)
which will exist because the system matrix A is sta.ble. Consider the function
V(t) = x(tl Px(t); the derivative along the system trajectories is
V - x(PA+ATP)x+2xTPD~(t,x)
< -llxll2 + 2lxT PDll~(t,x)I (2.24)
✓xTPDDTPx
then from equation (2.26) it follows that for x f/. Dó , V < O. Therefore there exists
a to > O such that the states x(t) E Dó for ali t > t 0 . In other words, x(t) is
ultimately bounded with respect to Dó, which acts as the ellipsoid [..
MULTIVARIABLE SYSTEMS THEORY 25
Now consider the special case when ~(t,x) == sin(2t); this clearly satisfies (2.22).
The phase portrait shown in Figure 2.1 represents a simulation of (2.21) with the
initial condition x0 == [10 O]. The disc D6 is shown as the dotted circle where in
this case 6 == 1.2748. It can be seen that, as predicted by the theory, the states
enter the clise and remain there.
Note in this case the solution tends asymptotically to the solution x 1 (t) = cos(2t) -½
and so ultimately the phase portrait sweeps out an ellipse with majar and minar
axes 1 and ½respectively.
. .
. .
. .
:. :.
. .
.· ( ~
,,...
.
:·-.,
o .. '················(·· .. •. .- ·)' ... ,
. '·- ·•'
-5 .
F ig ure 2.1 : Phase portrait together with the ultimate boundedness set
The remainder of this chapter will confine its attention to the special case of linear
systems and will introduce sorne of the important concepts from linear systems
theory.
where A E IR.''xn, BE 1Rnxm and CE IR,px n_ The variables u(t) and y(t) will be
referred to as the inputs and outputs respectively. The matrices A, B and C will be
26 SLIDING MODE CONTROL
termed the system, input distribution and output distribution matrices respectively.
In these terms the general linear system given in (2.5) is considered to represent a
free motion, i.e. a motion which is independent of any control action or exogenous
force. For convenience the system in (2.28) and (2.29) will be referred to as a
system triple (A, B , C).
If TE JRn xn is nonsingular then the change of coordinates x ~ Tx induces a new
system representation with system matrix TA r - 1 , input distribution matrix T B
and output distribution matrix cr- 1 , i.e. the triple (TAT- 1 , TB, CT- 1 ) .
The third condition, which is often the most convenient method of establishing
controllability, is often referred to as the Popov-Belevitch-Hautus rank test or
PBH test for short.
Definition 2.6 The linear system is said to be complete/y observable if the
output function y(t) over some time interval [to, t 1 ] uniquely determines the
initial condition x(to) .
• (A, B) is stabilisable
TAT- 1 = [ AAu
21
O ]
A22
TE = [ B0 ] (2.30)
2
where the pair (A 22 , B 2 ) is completely controllable. Note t hat because of the special
structure of the canonical form in (2.30) it can readily be observed that
-X(A11) e -X(A + BF)
for FE IRm><n _ Consequently, the pair (A,B) is stabilisable if and only if Au is
stable.
By duality, a similar canonical form exists for pairs (A, C) which are not observable.
P(s) = [ sl CA -t ] (2.34)
This result defines a set of complex frequencies z which are associated with specific
directions x(O) and u(O) in the state and input spaces for which the output of
the system is zero. These elements are called invariant zeros. It is clear that
information regarding the existence of invariant zeros comes from examining the
rank of P(z) in equation (2.37). For example, in the case of a square system - i.e.
systems with equal numbers of inputs and outputs - in order for equation (2.37)
to have a nonzero solution for x(O) and u(O), det(P(z)) must be zero.
28 SLJDING MODE CONTROL
J
1
= 2 Jo
reo a:(t? Qx(t) + u(t? Ru(t) dt (2.41)
where Q and R are positive definite symmetric matrices which penalise the devia-
tion of the state from the origin and the magnitude of the control signa.\, respec-
tively. The optimal solution, for any initial state, is given by
(2.42)
where X is the unique positive semidefinite solution (X > O), of the algebraic
Riccati equation
(2.43)
pair (A, B) is controllable, the pair (A, C) is observable and the dimensionality
requirement
m+p+ I 2:n (2.47)
is satisfied. These requirements are referred to as the Kimum- Davison conditions.
It has been demonstrated that control solutions which depend upon the availability
of the state vector are straightforward to determine. However, as argued earlier,
for practica! purposes, only a subset of the state vector may be available for mea-
surement.. It, has been seen that if the control problem is solved upon the basis of
output information only, the solution set is restricted. An alternative method to
overcome problems with restricted measurements is to use a state feedback con-
troller in conjunction with a dynamical system whose purpose is to estímate the
state vector. Such a dynamical system is called an observer. For the observable
system
and hence the observer is seen to be a model of the plant, which is driven by the
mismatch between the plant and observer outputs. Define the difference between
the actual state x(t) and the estimated state x(t) to be
may be accurately measured and thus need not be estimated. An observer that
estimates fewer than n state variables is called a reduced-order observer. Assume
that x E IRn and y E JRP where the output vector y can be measured accurately.
Since the p output variables are linear combinations of the state variables, only n - p
state variables need to be estimated. In this case the reduced-order observer will
be an observer of order n - p. Without loss of generality, consider the partitioned
state equations
y(t) [o I J[xi(t)]
u(t)
(2.55)
The state equation for the unmeasured portion of the state is thus given by
x1(t) = A11x1(t) + A12y(t) + B1u(t) (2.56)
Rewriting the dynamic equation of the second subsystem as
i;(t) -A22Y(t) - B2ii(t) = A21X1(t) (2.57)
an 'output equation' for the subsystem (2.56) is obtained. If in equation (2.57)
the signal i;(t) - A 22 y(t) - B 2 u(t) is regarded as the available output vector and
A 12 y(t) + B 1 u(t) represents the externa! input applied to the subsystem, then
equation (2.50) may be used to define an observer for the system (2.56) so that
:h (t) = (A 11 + LA2i)x1 (t) + A12Y(t) + B iu(t) - L(y(t)-A22Y(t)- B2u(t)) (2.58)
The reduced-order observer defined above requires the signal iJ which is clearly
undesirable. Define a modified state vector of t he reduced-order observer by
ry(t) = x1 + Ly (2.59)
then equation (2.58) yields
ry('t) = (A.11 + LA21):í:1 (t) + (A12 + LA22)y(t) + (B1 + LB2)u(t)
= (A11 +LA21)rJ(t) + (A12+LA22-(A11 +LA21)L)y(t)+(B1 +LB2)u(t) (2.60)
The feedback controller
u(t) = -Kx(t) = -K1 x1 (t) - K2y(t) (2.61)
may now be implemented as
u(t) = -K1(11- Ly) - K2y(t) (2.62)
which depends only upon the state of the reduced-order observer (2.60) and the
output of the system.
Nonlinear systems and Lyapunov stabilíty theory are described in Khalil (1992) and
Slotine & Li (1991). Formal statements and Lyapnnov arguments associated with
ultimate boundedness are discussed in Ryan & Corless (1984) and Khalil (1992).
Further details of quadratic stability and its relationship with other stability theory
concepts, including H 00 ideas, are given in Corless (1994).
Chapter 3
3.1 INTRODUCTION
In this chapter the intuitive concepts of Chapter 1 will be fashioned into rigor-
ous theory for general multi-input multi-output systems. lnitially the problem of
regulating a finite-dimensional linear time invariant system subject to parameter
uncertainty or nonlinearities will be considered. A controller will be sought to
force the system states to reach, and subsequently remain on, a predefined surface
within the state space. The dynamical behaviour of the system when confined to
the surface is described as an ideal sliding motion. The advantages of obtaining
such a motion are twofold: firstly there is a reduction in order; and secondly the
sliding motion is insensitive to parameter variations implicit in the input channels.
The latter property of invariance towards so-called matched uncertainty makes the
methodology an attractive one for designing robust controllers for uncertain sys-
tems. The design approach comprises two components: the design of a surface
in the state space so that the reduced-order sliding motion satisfies the specifica-
tions imposed on the designer; and the synthesis of a control law, discontinuous
about the sliding surface, such that the trajectories of the closed-loop motion are
directed towards the surface. An alternative interpretation of the latter property is
that the discontinuous control action renders the sliding surface invariant and (at
least locally) attractive. From this description the closed-loop dynamical behaYiour
obtained from using a variable structure control law comprises two distinct types
of motion. The init ial phase, occurring whilst the states are being driven towards
the surface, and often referred to as the reaching phase, is in general affected by
any matched disturbances present. Only when the states reach the surface and
the sliding motion takes place <loes the system become insensitive to ali matched
uncertainty. Ideally the control system should be designed so that the initial (dis-
turbance affected) reaching phase is as short as possible.
The motivation for considering differential equations with discontinuous right-hand
sides is that they occur naturally in physical systems; for example, the resistance
force due to dry Coulomb friction takes values of opposite sign depending on the
direction of motion; for details see Utkin (1992) . Also, in the case of certain
electric motors and power converters, the control action is naturally discontinuous.
In other systems, especially mechanical ones, the introduction of discontinuous
control action will not induce an ideal sliding motion. Imperfections in the process
32 SLIDING MODE CONTROL
such as delays and hysteresis will conspire to induce a high frequency motion known
as chattering. This is characterised by the states repeatedly crossing rather than
remaining on the surface. Such a motion is highly undesirable in practice and
will result in unnecessary wear and tear on the actuator components. It is usual
in such situations to modify the discontinuous control action so that, rather than
forcing the states to lie on the sliding surface, they are forced to remain within an
(arbitr arily) small boundary !ayer about the surface. In the literature this is often
referred to as a pseudo-sliding motion. The total invariance properties associated
with ideal sliding will be lost. However, an arbitrarily close approximation to ideal
sliding can usually be obtained.
Initially a nominal linear system will be considered and a methodology will be pre-
sented for designing a hyperplane which guarantees that the reduced-order sliding
motion is stable. The properties of such a sliding motion will then be examined - in
particular the invariance property with regard to matched uncertainty. Sufficient
conditions will be established which guarantee the existence of an ideal sliding mo-
tion. A variety of different control laws will be presented which induce and maintain
sliding despite the presence of matched uncertainty. This leads to a more detailed
analysis of a general multivariable problem in which the effects of both matched
and unrnatched uncertainty are considered. Finally, the notion of a pseudo-sliding
motion will be discussed which will be motivated as the motion resulting from ap-
proximating the discontinuous control laws, required for ideal sliding, by arbitrarily
close continuous ones.
Consider the uncertain linear time invariant system with m inputs given by
x(t) = Ax(t) + Bu(t) + f(t,x,u) (3.1)
where A E lR"xn and B E IR"xm with 1 ::; m < n . Without loss of generality it
can be assumed that the input distribution rnatrix B has full rank. The function
f : IR x IRn x lR.m ....+ IR.n is assumed to be unknown but bounded by sorne known
functions of the state. Different restrictions can be placed on this function, which
represents the para.meter uncertainty or nonlinearities present in the systern. In
general terrns, the problern to be addressed is that of synthesising a control law
which forces the states back to the origin frorn sorne a.rbitrary initial displacernent,
i.e. a regulation problern. In la.ter chapters a more general (output) tracking
problem will be considered. This will be achieved by using an integral action or a
rnodel-following methodology to transform the problem into one of regulation.
Let s : IRn -t IRm be a linear function represented as
s(x) = Sx (3.2)
For reasons that will become clear la.ter, this function will be termed t he switching
function.
SJ,Jl)JNG MODE CONTROL 33
D efinition 3.1 Suppose there e:r:ists a finite time ts s1u:h that the solution to
( 3. 1) represented by x(t) satisfies
then an ideal sliding motion is said to be tnking place for ali t > ts.
R cmarks
• In this definition the switching function has been thought of as the function
of time defined by s(t) = Sx(t).
• It must be stressed that tcchnically there is no rcason for S to be restricted
to be a hyperplane - more general (possibly time-varying) surfaces rnay be
considered; see for example DeCarlo et al. (1988).
H is now possihlc to be more specific about. the direction of this chapter. The
first problem that. will be addressed is that of designing a swit.ching function so
t.hat the motion of t he dynamical system when confined to thc hyperplane S is
stablc. Secondly, it considcrs t hc problem of designing variable structure control
laws so that in finite time the states are forced onto and subsequently rcmain on
the hyperplane S.
In or<ler to meet the finite time requirement, it proves necessary to use a con-
trol action which is discontinuous about the surface S. Potentially this presents
difficulties from a mathematical viewpoint.
T his concept is discussed at greater length in Utkin (1992). A more recent ap-
proach using the theory of differential inclusions is taken by R.yan (1988). The
most intuitively appealing approach, however, is the method of equivalent control
proposed by Utkin (1977) . Broadly speaking, the equivalent control is t he control
action necessary to maintain an ideal sliding motion on S.
In clescribing the methocl of equivalent control it will initially be assumecl that the
uncertain function in equation (3.1) is identically zero, i.e.
x(t) = Ax(t) + Bu(t) (3.5)
Suppose at time t 5 the systems states lie on the surface S defined in (3.3) and an
ideal sliding motion takes place. rviathematically this can be expressed as Sx(t) = O
ancl s(t) = Sx(t) = O for all t 2'. t 5 • Substituting for x(t) from (3.5) gives
Sx(t) = SAx(t) + SBu(t) = O for all t 2'. t. (3.6)
Suppose the rnatrix Sis such that the square matrix SE is nonsingular. This cloes
not present any serious clifficulty since by assumption B is full rank and S is a
design pa.rameter.
D efinition 3.2 The equivalent control associated with the nominal system
( 3.5), written as U eq, is defined to be the unique solution to the algebraic equa-
tion ( 3. 6), namely
(3.7)
SLIDING MODE CONTROL 35
Define
(3.9)
then P. is a pro_jection operator and satisfies two simple yet very important (from
the viewpoint of sliding mode control) equations:
These can be readily established by direct evaluation. Using these tools the next
section investigates the properties of the sliding motion and provides a mechanism
for hyperplane design.
Initially in this section the linear time invariant system given in (3.5) will be con-
sidered. As in the previous section it will be assumed that a controller exists which
induces an ideal sliding motion on the surface S given in equation (3.3). The first
result confirms the sliding motion is of reduced order and provides insight into its
associated eigenstructure.
Proposition 3.1 The sliding motion given in equation (3.8) is of reduced order
and the eigenvectors associated with any nonzero eigenvalues of the system matrix
(3.11)
From the proposition above, it is clear that the matrix Aeq in (3.11) can have at
most n - rn nonzero eigenvalues. Suppose by choice of S these are distinct and that
V E m,nx (n-m) is a matrix formed from the (right) eigenvectors associated with
the n - m nonzero eigenvalues. The matrix V must then satisfy the two conditions
in the following proposition.
Proposition 3.2 The matrix of right eigenvectors V satisfies
SV =O and rank [ V B ] =n
Proof
The first result follows immediately from Proposition 3.1. To prove t,he secon<l
result, consider solving
(3.12)
where r¡1 and r¡2 are vectors of appropriate dimension. Multiplying (3.12) by S and
using the previously established fact that SV = O implies SBr¡2 = O. This in turn
implies r¡ 2 = O since by assurnption the square matrix SE is nonsingular. Equation
(3. 12) thus becomes V r¡1 = O, which implies r¡1 = O since the eigenvectors that
comprise V are linearly independent as the associated eigenvalues are distinct.
Thus the only solution to (3.12) is r¡1 = O and r¡2 = O, which implies that t he
columns of the matrix [ V B] are linearly independent. ■
From equation (3. 7) the equivalent control can be considered to be the linear state
feedback component necessary to maintain the reduced-order motion. Because
of the straightforward and explicit representation in equation (3.7) it would be
tempting to think of using the signa!
u(t) = Kx(t) (3.13)
This equivalent control action is now dependent on the unknown exogenous signa!
and therefore cannot be realised in practice. It should therefore be stressed that
the notion of the equivalent control is best thought of as a tool for the analysis of
the motion obtained from constraining the system states to remain on S.
Using the ideas introduced thus far, it is already possible to prove the key prop-
erty of the reduced-order sliding motion - namely its invariance towards matched
uncertainty.
Theorem 3 .1 The ideal sliding motion is totally insensitive to the uncertain junc-
tion ((t,x) in equation (3.14) ifR(D) C R(B) .
Proof
Substituting the e4uivalenL conLrol law (3.15) i11Lu the uncertain system represen-
tation (3.14), it follows that the sliding motion satisfies
x(t) =P 8 Ax(t) + PsD((x, t) for ali t 2 ts and Sx(ts) =O (3.16)
where P 5 is the projection operator defined in (3.9). Now suppose R(D) e R(B),
then there exists a matrix of elementary column operations RE IRm x l such t hat
D = BR. As a result it follows that PsD = Ps(BR) = (P.B)R = O by the
projection property described in equation (3.10). The reduced-order motion as
given in (3.16) reduces to
for ali t 2 ts and Sx(t,) =O (3.17)
which does not depend on the exogenous signa!. ■
Using this definition, Theorern 3.1 may be paraphrased as: whilst sliding, the
reduced-order motion is completely insensitive to matched uncertainty. This in-
variance property makes VSCS a powerful too! for controlling uncertain systems
and is a strong motivation for the continuing research interest in the area.
From the preceding analysis it is clear that, when matched uncertainty alone is
present, it is sufficient to consider the nominal linear system representation when
designing the switching function. However , whilst it can be seen from equation
(3.8) that the sliding motion depends on the choice of sliding surface, the precise
effect is not readily apparent. A convenient way to shed light on the problem is to
first t r ansform the system into a suitable canonical form. In this form the system
is decomposed into two connected subsystems, one acting in R(B) and the other
in N(S) . Since by assumption rank(B) = m, there exists an invertible matrix of
elementary row operations Tr E IRnxn such that
(3.18)
where B 2 E IRmxm and is nonsingular. This matrix can readily be found using
Gaussian elimination, which reduces the E matrix to row echelon form. 2 However,
2 For details see for example Strang (1988).
38 SLIDING MODE CONTROL
there exists an orthogonal matrix Tr satisfying (3.18) which can be computed via
' QR' decomposition (see Appendix A.2.5). From a purely computational point of
view it is more convenient and numerically stable to deal with orthogonal matrices,
since inverses can be obtained by straightforward transposition. A further use-
ful property is that orthogonal coordinate transformations preserve t he Euclidean
norm (see Appendix A.2.6). These computational issues will be discussed in more
detail in the next chapter. By using the coordinate transformation x ++ Trx if
necessary, it can be assumed without loss of generality that the distribution matrix
has the form of equation (3.18). If the states are partitioned so that
(3.19)
where x 1 E m,n-m and X2 E m,m then the nominal linear system (3.5) can be
written as
This representation is referred to as regular form. Equat ion (3.20) describes the
null space dynamícs and equation (3.21) describes the range space dynamics. If the
switching function matrix in this coordinate system is partitioned compatibly as
(3.22)
(3.25)
The ideal sliding motion is given therefore from the combination of equations (3.25)
and (3.24). These equations give a much clearer picture than the previous descrip-
tion of sliding given in (3.8) . In particular, two points are worth noting:
• The matrix S2 has no direct effect on the dynamics of the sliding motion and
acts only as a sca.ling factor for the switching function.
-A12
rank [ zI- A B]
=
rank
l zI-A11
-A21
rank [ z I -An
zI- A 22
A .12 ] +m for all z E C
This implies
and from the Popov- Belevitch- Hautus (PBH) rank test 3 it follows that (A, B) is
controllable if and only if the pair (A 11 , A 12 ) is controllable. ■
Using t his proposition, provided the original pair (A, B) is controllable, the pair
(A 11 , A 12 ) is controllable and any robust linear state feedback method can be ap-
plied to designing M. Severa! approaches have been proposed, including quadratic
minimisation, direct eigenvalue placement, eigenvalue placement within a region
and eigenstructure assignment methods. These will be discussed in sorne depth in
the next chapter.
Evaluating the expression for sliding motion given in (3.8) in the regular-form
coordinates, it can easily be verified by direct computation that
l An
- MA11
(3.26)
and furthermore
l Au
- MA11 ] [ - IM (3.27)
feedback law defined in (3.13). From this standpoint, the problem appears closely
related to the so-called output zeroing problem for the (in this case fictitious)
input/output system defined by ( A, B, S). In fact, it will be shown that the poles
of the sliding motion are given by the invariant zeros of the triple (A, B, S). At this
stage, this is no more than a.n observation. However, when the problem of designing
sliding mode controllers ha.sed upon output information only is considered, this
interpretation will provide good insight into the class of systems for which the
theory is applicable.
Proposition 3.4 The poles of the sliding motion are given by the invariant zeros
of the system triple (A, B , S) .
Proof
By definition, the invariant zeros are given by
{z E C : P(z) loses normal rank}
where Rosenbrock's system matrix
P(z) = [ z I_S A !]
In this case (A, B, S) has the sa.me number of inputs as outputs a.nd therefore P(z)
loses rank if and only if det P(z) = O. Substituting for the triple (A, B, S) from
equations (3.20) a.nd (3.21)
P(z) = [
z l -A11
- A21
-S1
-A12
z l -A22
-S2
Ü
B2
0
l
Since by construction B2 is nonsingular
Because S 2 is nonsingula.r it can be shown by direct evaluation that for all values
of z E <D
where Ah and M are the matrices which appear in equations (3.24) and (3.25).
Since the inner and outer matrices on the right-ha.nd side of t he above equivalence
are independent of z, and both have determinant equal to unity, it follows that
det [ zl - Au -A12 ] det [ zl - Af1 O ]
-S1 -S2 O - S2
det(zl - Af 1) det(-S2)
Therefore since S 2 is nonsingular
Up to this point the concept of sliding modes and sliding surfaces has been intro-
duced in a rigorous way for multi-input linear systems. The relationship between
the choice of sliding surface and the resulting sliding motion has been examined
and an explicit relationship has been identified which is convenient from the design
viewpoint. The key notion of matched uncertainty has been introduced and the
fundamental property that the sliding motion is completely insensitive to matched
uncertainty has been proved. But two issues of importance, however, have not yet
been addressed:
The effect of unmatched uncertainty will be explored in due course but first suffi-
cient conditions for the existence of a sliding mode will be derived. This is vital to
the design of any VSCS since the robustness properties outlined above will exist
only if a sliding motion can be reached and maintained.
lim
s-+O+
s< O and lim
s-+o-
s> O (3.28)
D = S n !1 = {x E !1 : s(x) = O}
The expression given in (3.28) is often replaced by the equivalent, but more succinct
criterion
ss < 0 (3.29)
The expressions in (3.28) and (3.29) are termed reachability conditions. An equa-
tion of this type was encountered previously in Chapter 1 for the double-integrator
example
fj(t) = u(t) (3.30)
42 SLlDING MODE CONTROL
where
s(y,y) =my+y (m > O) (3.32)
it was argued that the switching function satisfies the reachability condition (3.29)
in the domain
n = {(y,y) : ml:zil < l}
Remark
In general, if the reachability condition (3.29) is satisfied globally, i.e. O = IR",
then since
1 d 2 .
-s =ss
2
dt
it follows that the function
V(s) = ½s2 (3.33)
is a Lyapunov function for the state s. This observation will provicle the key to
extending these ideas to the multi-input situation.
Unfortunately, although equations (3.28) ancl (3.29) are commonly encountered in
the literature, they do not guarantee the existence of an ideal sliding motion as
given in Definition 3.1. Essentially these conclitions only guarantee that the sliding
surface is reached asymptotically; for example, consicler once more the clouble-
integrator in equation (3.30), this time with the linear state feedback control law
where <I> is a positive design scalar. It was shown in (1.30) that using the control
law (3.34) results in a stable closed-loop motion with poles at ( - m, -<I> ). More
importantly, direct computation reveals
s = -<I>s (3.35)
which implies
ss = - <I>s2 <O
and hence the reachability condition (3.29) is satisfied. From equation (3.35) it
follows t hat
s(t) = s(O)e-<I>t
and therefore if s(O) f; O, i.e. initially the states do not líe on the sliding surface,
then s(t) :j; O for a ll t > O, although of course s(t) --+ O as t--+ oo. This means that
the slicling surface is only reached asymptotically rather than in finite time.
A stronger condition, guaranteeing an ideal sliding motion, is the 77-reachability
condition given by
ss S, -771s1 (3.36)
where 77 is a small positive constant. By rewriting equation (3.36) as
SLlDING MODE CONTROL 43
The r¡-reachability condition has also been encountered in Sectíon 1.3, where the
control law
u(t) = -my(t) - psgn(s(t))
was used to control the double-integrator. It was shown there that
ss = -pjsj
and hence the r¡-reachability condition is satisfied. Given these sufficient conditions
for reachability, the next section considers the synthesis of control laws which induce
ideal sliding motions.
The control structures that are most commonly employed comprise a linear state
feedback component and a nonlinear or discontinuous component. Consider an nth
order single-input uncertain system represented by
where the nominal pair (A, b) is controllable and Aper(t) is a t ime-varying uncer-
tain matrix, which belongs to R(B) and represents any parameter uncertainty or
nonlinearities in the system. In an appropriate coordinate system the pair (A,b)
can always be written as
o 1 o o o
o 1
A= b= (3.38)
o
o o 1 o
-ai -a2 - an 1
which is often referred to as the controllability canonical form; see for example Ero-
gan (1991). One of the advantages of this canonical form is that the characteristic
equation for the matrix A is given by
(A 11 ,A12 ) from (3.20), which plays the pivota.l role in the design of the switching
surface, is gi ven by
o 1 o o o
o 1
A11 = o A12 =
o 1 o
o o 1
It can easily be verified that the pair (A11 , A12 ), as predicted from Proposition 3.3,
is controllable. Let
(3.39)
be a row vector M of order n - l. Because the pair (A 11 , A12 ) is itself in con-
trollability canonical form, the system matrix governing the reduced-order sliding
motion Ah = An - A 12 M has the coefficients of M as its last row. Therefore the
characteristic equation of Af¡ is given by
An-l + mn-1An- Z + ... + m2A + m1 = O
Thus, provided the coefficients m 1 , ... , m,,_ 1 constitute a Hurwitz polynomial, 4
it follows that the eigenvalues of Af 1 will be stable. Consequently, the linear
functional
n-1
It will be assumed that, for ali t, the perturbations ~ ;(t), which constitute the
matched uncertainty, satisfy
fori=l,2 ... ,n (3.43)
4
A polynomial is said to be Hurwitz if a li its roots have negative real parts.
SLIDING MODE CONTROL 45
where u 1(t) is a state feedback law - often the nominal equivalent control - and un
is a discontinuous or switched component. Differentiating equation (3.40) gives
n- 1 n
s(t) = ¿ miXi+1 (t) - ¿ (ai + 6.;(t)) x;(t) + u(t)
i= l i=l
n n _
-a1x1 (t) + ¿ (mi- 1 - a;)x¡(t) - L 6.i(t)xi(t) + u(t)
i =l
Two different choices of un(t) will now be explored. Firstly define a scalar function
p(t,x) such t hat
n
p(t,x) 2: I ¿6.¡(t)x;(t) I +17 for ali t, x (3.47)
i=.l
where 77 is a small positive design scalar. Under the assumptions on the 6.;'s
given in equations (3.43) this is always possible. 5 A potential expression for the
discontinuous component is
where
ki = { k¡ if sx; > 0 (3.50)
k; if SX; < o
and again r¡ is a small positive scalar. Using t he nonlinear control component above
in the expression (3.46) it can be established that
n
ss =¿ sx;(t)(k; - 6 ; (t)) - r,lsl :S -171s1
i=l
since by construction
sx;(t)(k; - 6 ;(t)) < O fori= 1, 2 ... ,n
Once again an r¡-reachability condition has been established and therefore ideal
sliding in finite time is guaranteed.
Remarks
• It should be noted that many different control structures appear in the liter-
ature although the ones discussed above are probably the most cornmon. A
slightly different control structure will be given in the actuator case study in
Chapter 8 and others can be found in Utkin (1977).
• It should be pointed out t hat the linear control structure defined in equation
(3.45) is in fact the 'nominal equivalent control' since it can easily be shown
that n
a1x1(t) +¿ (a; - m;- 1)x¡(t) = -(Sb)- 1 SAx(t)
i=2
The next subsection seeks to broaden the scope of the ideas presented above to a
more general multivariable situation.
Consider once more the pendulum arrangement in Figure 1.9. Now assume that a
friction term exists which results in a new normalised pendulum equation:
where, as befare, the control input u(t) represents a torque applied at the point of
suspension. Define a scalar function
= x2(t)
= - a2x2(t) - a 1 ~(t)x1 (t) + u(t) (3.53)
SLIDING MODE CONTROL 47
In obtaining the description in (3.53) the fact that t he term sin(y) can be written
as y ~(y) has been used. T his is a different uncertainty formulation than that used
in Chapter 1, where essentially the whole term sin(y) constituted the uncertainty.
The advantage of this approach is that in (3.53) the uncertainty bound is smaller.
Choosing the switching function
"'o 0.8
·.:;
·;;;
o 0.6
"-
....
"'
-; 0.4 .
ofJ
e
~ 0.2
o
o 2 3 4 5 6 7 8 9 10
Time, sec
T he solid line in Figure 3.2 shows a simulation of the closed-Ioop system obtained
from using the control law in equations (3.54) and (3.55) and the nominal values
of a 1 and a2 wit h initial conditions y = l and y = O. The dashed and dotted lines
represent the responses when the parameter a 1 is pert urbed to the values 0.2 and
0.3 respectively. The corresponding switching functions are given in Figure 3.3.
e::
.:2 0.8
tí
e::
:::,
u. 0.6
bJl
e::
E 0.4
.':l
-~ 0.2
(/)
2 3 4 5 6 7 8 9 10
Time, sec
extension of the single-input problem described earlier will be considered, i.e. the
control of the uncertain system
x(t) = (A + A per(t)) x(t) + Bu(t) (3.56)
where A E IRn x n and B E IRnxm and Ape,-(t) represents any uncertainties or
nonlinearities which, as before, are assumed to be matched so that
Aper(t) = Bb,.(t) (3.57)
where b,.(t) is sorne bounded time-varying matrix. In Section 3.2 the hyperplane
S was defined as the null space of a design matrix SE IRmxn _ An alternative and
equivalent definition of the same hyperplane is
m
S = n
i=l
S;
where Á.(t) = A + Aper (t) and bi represents the ith column of the input distribution
matrix. Once ideal sliding has been established on S1 , the reduced-order motion is
given by
(3.60)
SL!DING MODE CONTROL 49
where A eql represents the equivalent dynamics obtained from assuming that sliding
takes place on S1 . The second control component is then designed to induce sliding
on S2 by considering equation (3.60) as a single-input equation in u 2 . Since it is
assumed that sliding takes place on S 1 for any control signals u 2 , ... , um, it follows
that sliding takes place on S1 n S2 • Using this inductive argument, it follows that
sliding will eventually take place on íl~ 1 Si = S.
Although this approach is conceptually appealing, in practice difficulties often oc-
cur. This method is obviously dependent on the order in which sliding t akes place
on the constituent surfaces S1, .. . , Sm. In general there is no requirement that S 1
be t he surface on which t he system slides first. This explains why the method is
often referred to in the literature as a fixed-order approach. In practice the order-
ing of the constituent sliding surfaces is pararnount in avoiding unnecessarily large
control signals.
A more recent and perhaps more practica! method is the so-called diagonalisation
approach. It was noted in Section 3.3 that in order for a unique equivalent control
to exist, the matrix S must be chosen to ensure that det(SB) f= O. In addition it
has been demonstrated that the sliding motion is invariant to nonsingular scaling
of the switching function. As a result, it can be assumed that, by design,
SB=A (3.61)
(3.62)
where .ói(t) is t he ith row of .ó(t) and J\.i is the ith diagonal element of A. Thus,
for example, if
(un)i = - p;(x, t)sgn(s;)
where
p;(t,x) > l.ó;(t)x(t) I
then, using t he arguments given in the previous section, a sliding mode is induced
on each equation given in (3.61) . In summary this approach reduces thc problem
of multivariable controller design to m single-input problems. The advantage of
the diagonalisation approach just described over the hierarchical approach is that
it is not necessary to define a priori the order in which sliding occurs on t he
constituent subsurfaces . In both approaches, however , controller switching takes
place at points in the state space ot her than those constituting S, whenever the
50 SLIDING MODE CONTROL
A more recent and more convenient control structure for multivariable systems is
the so-called unit VP.ctor approach. Consicler the uncertain system
x(t) = Ax(t) + Bu(t) + fu(t,x) + f m(t,x,11) (3.65)
where the functions f u : 1R x 1Rn ➔ R (B))1- and fm : 1R x 1Rn x 1R111 ➔ R(B) are
unknown but bounded and satisfy
(3.68)
and a(·) is a known function. Here, by definition, the functions fu (·) and f m (-)
represent the unmatched and matched uncertainty components respectively. With-
out loss of generality it can be assumed that equation (3.65) is already in regular
form and so
Aux1(t) + A12X2(t) + fu(t ,x) (3.69)
A21 X1 (t) + A22X2(t) + B 2u(t) + Ím(t, x, tt) (3.70)
where J,, and f m represent projections of fu and f m into the coordinates of the
subspaces N(S) and R(B) respectively, which is possible since N(S) e,R(B) = 1Rn.
It can be assumed that the transformation used to obtain the canonical form was
an orthogonal one. Hence the Euclidean norm is preserved, which implies
llfu(t,x)II < k1 llxll + k2 (3.71)
llfm(t, X, u)II < k3jluJJ+ a(t,x) (3.72)
and B T B = Bi' B2 . Consequently
and furthermore
(3.74)
where the first equality is a well-known property of singular values and the second
follows from t he definition of the spectral norm. Combining equations (3.73) and
(3.74) it follows that inequality (3.68) can be written as
(3.75)
SLJOING MODE CONTROL 51
(3.76)
(3.78)
and K-(A) represents the condition number. This is always possible since, for exam-
ple, if A is a sca.!ed identity matrix then K-(A) = 1 and the above equation reduces
to (3.75), which is satisfied by assumption.
Define a linear change of coordinates by
(3.79)
(3.80)
then
(3.84)
where q> E IR,mxm is any stable design matrix. The nonlinear component is defined
to be
- (t )A- 1 ~\IP2s(t)II
Un (t) - - p 'X for s(t) -:f. O (3.85)
where P 2 E IR,mxm is a symmetric positive definite matrix satisfying the Lyapunov
equation
(3.86)
and the scalar function p(t,x), which depends only on the magnitude ofthe uncer-
tainty, is any function satisfying
(3.87)
52 SLIDJNG MODE CONTROL
_fu_
IIP2sll = sgn ( 8 ) (se/ O)
and hence the control structure (3.85) becomes a scaled relay structure as described
in equation (3.48).
The next subsection demonstrates that the control law (3.83) induces a sliding
motion on S in finite time despite the presence of uncertainty.
Substituting equations (3.83) to (3.85) into (3.82) and simplifying, produces the
system
±1(t) (3.88)
s(t) (3.89)
where the fact that sT P2P2s = IIP2sll 2 has been used in establishing the first in-
equality. Since S2 = AB2 1 it follows that IIS2II < IIB2 1 IIIIAII- Using this inequality
and rearranging (3.87) it is straightforward to show that
Cornbining inequality (3.90) and (3.93) it follows that the derivative of the Lya-
punov function satisfies
(3.94)
SLIDING MODE CONTROL 53
This inequality will be used to show that sliding on S takes place in finite time.
From the Rayleigh principie
(3.97)
where s0 represents t he initial value of s(t) at t = O. This is a natural multivari-
able analogue of the single-input case: a Lyapunov function has been found which
depends on the switching function states and which satisfies (3.96), in a sense the
natural extension of the r¡-reachability condition. The next subsection investigates
the dynamical behaviour when confined to the hyperplane S.
The equation representing the motion when confined to the sliding surface is ob-
tained by substituting s = O into equation (3.88), giving
(3.98)
(3.99)
then V(x 1 ) < O and the system (3.98) is quadratically stable. As discussed in
Section 2.2.2 the optima! choice of Q 1 to maximise the value of the scalar µ is to
let Q 1 = I , giving
The expression (3.100) is clearly not satisfied unless k 2 = O. In such a situation the
best that can be guaranteed is that the states xi(t) are ultimately bounded. As
discussed in Section 2.2.2, ultimate boundedness ensures that, in finite time, the
solution enters a set E (usually an acceptably small neighbourhood of the origin),
and remains there for ali subsequent time.
During the sliding motion x2 = -M1x1 and therefore llxll ~ (JI + IIMll2 )llx111 -
Consequently, the bound on the unmatched uncertainty (3.71) can be written as
(3.101)
where k1 k1 J I + IIM.112. Then, arguing as before and assurning 2k1 < P,, it
follows that
V(x1) < -llx1IIAmax(P1) (P,llx1II - 2ll!ull)
< - llx1IIAmax(Pi) (P,llx1II - 2k1llx1II - 2k2)
and consequently ll(xi) < O if x 1 (/_ E1 where
E1 = {,i:1 E :IR,(n-m) : llx1II < 2k2/(µ - 2k1) + E} (3.102)
for sorne small scalar 1: > O. Therefore the states x 1 ( ·) are ultimately bounded with
respect to the ellipsoid E1.
Consider the problem of controlling the angular position of the shaft in a DC motor.
T he differential equation governing the mechanical part of the system, modelled as
a disc rotating about the shaft, is given by
where i 0 (t) is the armature current. This in turn is assurned to satisfy Kirchhoff's
voltage law
d .
L-d ia(t) + Ri 0 (t) = v 0 (t) - Ke0(t) (3.108)
t
where R and L are the armature resistance and inductance respectively and the
last term represents the electromotive force. The applied armature voltage v0 is
taken to be the control variable.
For illustration assume that the inductance L is not known precisely; specifically
suppose that the error ratio
(3.109)
satisfies l~LI < 0.1, where Lo is sorne nominal inductance value (or in other words
the value of L is known to an accuracy of 90%). Sirnilarly, suppose the inertia
coefficient (or the motor gain) is not known perfectly, so that
(3.110)
where Jo is sorne nominal value and l~J 1< 0.5. Finally, assume that the friction
effects are negligible, i.e. b = O. Using the variables 0,0 and ia as states, the
nominal system from equations (3.106) to (3.108) can be written in state-space
formas
±(t) = Ax(t) + Bu(t) (3.111)
where
o
~
1
l
o
A= [
o
_ &_
Lo
&
Jo
R
- Lo
and B=
[ o
1
Lo l (3.112)
It can be seen that the chosen coordinates constitute a regular form and therefore,
in the notation of Section 3.6, the matrices
(3.114)
(3.115)
Therefore the parameters ( and Wn represent the damping ratio and natural fre-
quency respectively. The sliding motion governed by (3. 115) can thus be assigned
an appropriate dynamics by choice of ( and Wn. From the choice of M given in
(3.114), the switching function, obtained from selecting S 2 = 1, is given by
s(t) =[ :Jtw;. 2:Jt(wn 1 ] x(t)
(3.116)
where x 1 , x 2 and x 3 are the componeuts of the state-space vector,6 and the·matched
uncertainty component
k _ _ l_
3
- l OLo
and the bounding function
(3.118)
(3. 119)
6 This should not be confused with the notation of Section 3.6 where xi and x2 represent
partitions of the state-space vector not its components.
SLIDING MODE CONTROL 57
which implies
llfu(t, x)II ::; ½IIMll llx(t)II (3.120)
since l~JI < 0.5. Consequently, comparing this with the general description of the
uncertainty class in equation (3.66), appropriate values for the coefficients are
k1 = ½IIMII and
Because of the special structure of the uncertainty in this particular case, provided
~J is such that Kt/ J and Kt/ 10 have the same sign, the reduced-order sliding
motion will be stable. This can easily be established from the Routh criteria; for
details see for example Brogan (1991).
An appropriate sea.lar function to premultiply the unit vector component can now
be described. Since A= 1/Lo and x:(A) = 1 it follows that
A21=A(3,1:2);
A22=A(3,3);
L=0. 046;
Aper=[O 1 O; O -b/J Kt/J; O - Ke/L -R/L];
Bper= [zeros(2 , 1); 1/L) ;
Cper=eye(3);
X=[1; O; O);
gamma2=0 . 01;
In the simulation of Figure 3.4 the sliding motion has been assigned the properties
( = 0.95 and Wn = 2.5. The design parameter if>, governing the rate of attainment
of the sliding motion, has been chosen to be - 20. This reflects an attempt to
ensure that the control action lies in the region of (-5, 1).
The solid Jine represents a simulation of the closed-loop system with init ial condi-
tions 0 = 1, é = O and Ía = O.
o 0.8
·::,
·¡;;
p..
o 0.6
~
;i 0.4
Cf)
"'
<( 0.2
The dashed line represents the closed-loop response when the inductance is given
as L = 0.046 and the motor gain / inertia is perturbed to I<t/ J = 6.0. Both of these
perturba.tions satisfy the bounds on ~J and ~~ - Because the system just considered
is single-input and of low order it is amenable to the sort of analysis described
SLIDING MODE CONTROL 59
above. For more general systems and in particular for multi-input systems, the
hyperplane design problem possesses more degrees of freedom which can be prof-
itably exploited. A general discussion exploring these possibilities is described in
Chapter 4.
The guaranteed attainment of an ideal sliding mode in finite time requires that
the control action is discontinuous across S. As noted earlier, in the control of
certain electric motors the control action is naturally discontinuous. However,
for mcchanical systems the implementation of such a control law would produce
a chattering motion in a boundary of the surface S rather than an ideal sliding
motion. Such high frequency switching might excite unmodelled dynamics and
impose undue wear and tear on the actuators, so the control law would not be
considered acceptable. Severa! modifications have been proposed to overcome this
diffi.culty and they are described in the next section.
In this section a more general definition of a sliding motion (a.5 opposed to ideal
sliding) will be given. In the literature this is often referred to a.5 pseudo-sliding. It
will obviate the need for the control law to be discontinuous and as a result presents
a more practica! solution from the viewpoint of implementation. The notion of
ideal sliding will t hus represent in a sense the limiting ca.se or the (idealised) best
performance which could be obtained.
In this section the essential difference is that, rather than requiring the states to be
forced to remain on the surface S, the states will only be required to remain arbi-
trarily close to the surface. It should be noted that the total invariance properties
with respect to matched uncertainty will be lost, although very clase approximation
to the ideal sliding case can usually be achieved. In terms of designing a realisable
control law for practica! application, a trade-off must be made between achievable
performance and the smoothing of the discontinuous control action.
(1) V(t,x,s) is positive definite with respect tos, i.e. V(t,x,s) > O if s -1- O for
all x E n, and on the spheres 11s11 = r for all x E n
• inf V(t,x,s)=hr
ll•ll=r
and hr > O f or r -1- O
• sup V(t,x,s)=Hr>O
ll•ll=r
where h,. and H,. depend on r.
(2) The total time derivative of V(t, x, s) has a negative supremum for every
x E n except at points on the switching surface where the control action may
be undefined and hence the derivative of V(t,x, s) does not exist .
This is clearly positive definite with respect to S and satisfies the supremurn and
infimurn conditions. The total time derivative is given by
m
V= 2¿ s;s¡
i= l
This condition is the one most often cited in the literature when demonstrating the
existence of a sliding mode and it was encountered earlier in Section 3.5. It should
be noted that the state feedback controller given in (3.34), namely
u(t) = -(m + 4>) y - 4>my(t)
satisfies the notion of sliding given above. In Section 3.5 it was shown that using
this feedback law to control the double-integrator (3.30) resulted in the switching
function, thought of as a function of time, satisfying
ls(t) I = ls(O)le-<I>t
Sliding in the sense of Definition 3.4 now occurs because, given any ó > O, there
exists a time t. (given explicitly in this case by [log(ls(O) l/8)]/4> such that for all
t > t. the switching function satisfies ls(t)I < ó.
The most common approach to control law design is to 'soften' the discontinuity
in the control laws. T his can easily be accomplished in the case of relays and unit
vector strategies. Probably the earliest and most intuitive approach is to utilise a
boundary layer approach, whereby the discontinuous component is replaced by the
continuous nonlinear approximation
-1
The size of ó defines the size of the boundary !ayer in which a high gain linear
feedback control law operates. In the single-input case this can be visualised with
the unit vector component (shown as a dotted line) being replaced by the function
(shown as a solid line) in Figure 3.6. An ideal sliding motion will no longer take
place and the system will not be totally invariant to matched uncertainty. However ,
the switching function state s will be uniformly ultimately bounded with respect
to the ellipsoid
(3.124)
Although no ideal sliding takes place, the states x 1 (·) remain ultimately bounded
with respect to a neighbourhood of the origin (albeit a larger one than before).
62 SLIDING MODE CONTROL
IIP2s(t)II > ó
o< IIP2s(t)II ~ ó (3.125)
Pzs(t) =O
where the design scalar q E [O, 1) and as before ó is a small positive scalar. Por
the single-input case, the smoothing of the discontinuous relay function in a neigh-
bourhood of the origin is shown in Figure 3. 7 for different values of q.
- ó-
-ó-
Again the dotted line represents the signum function. As q -7 O the slopes of the
functions in the neighbourhood of the origin become steeper and the approxima-
tions become more precise. (The solid line represents q == 0.05 and the chain dotted
line q = 0.5.)
An alternative differentiable approximation is
- 20/i
, - - - - - - - -- ---- -- --
1
'
- 1
- 20li
As shown in Figure 3.8, the scalar ó does not define the boundary !ayer (at Jlsll == ó
SLJDING MODE CONTROL 63
the nonlinear control action is half of its maximum value), but ultimate bounded-
ness results can be demonstrated just as for the other parallel boundary techniques.
The motion resulting from using the fractional approximation (3.126) is studied in
detail in Burton & Zinober (1986) where, in particular, a formal exa.mination of its
chatter elimination properties is made.
An experimental comparison of these different approaches to eliminate chattering
(technically for the case of a tracking controller) is made in deJager (1992), which
compares the steady-state errors that result from using the continuous approxima-
tions in the presence of a constant disturbance. It is reported that no significant
difference arises between the various methods.
3.8 SUMMARY
The key ideas of VSCS with a sliding mode have been presented. The notions of
a sliding surface and an ideal sliding mode have been introduced, together with
sufficient conditions for their existence. The exposition has been restricted to un-
certain linear systems, and only hyperplanes have been considered as potential
sliding surfaces. Implicitly it has been assumed that all t he states are available to
the controller. This restriction is often cited as a limitation of the technique since in
practice only certain 'outputs' may be directly measurable. In later chapters VSCS
concepts will be examined for uncertain systems for which only outputs are avail-
able. In this situation the class of hyperplanes and controllers considered must be
restricted to those requiring only the system output information, or else estimates
of the unavailable interna! states must be generated for use by the control law.
Both these possibilities are investigated. For the time being, the assumption that
all the states are available will be retained. The next chapter considers different
approaches for the design of the switching surface.
Because of the nature of this chapter, it is difficult not to have been influenced
either directly or subconsciously by the work of other authors. Severa! excellent
introductory tutorial papers exist in the literature, most notably those of Utkin
(1977), DeCarlo et al. (1988) and Hung et al. (1993). Introductory material also
appears in t he early chapters of two essentially research-orientated collections by
Zinober (1990,1994) .
For those interested in the rigorous fundamental ideas concerning the existence of
solution and the Filippov construction, the reader is directed to the seminal work
of Utkin (1992).
The 77-reachability condition a.nd the estimation of the time to reach the sliding sur-
face appears in the work of Slotine & Li (1991) as does the use of VSCS controllers
for electric motors and power converters. This a.rea is also discussed in more detail
by Sira-Ramírez & Lischinsky-Arenas (1991) and the references contained within.
This chapter considers only systems with nominal linear parts and only hyperplanes
as potential sliding surfaces. It should be stressed that in principie these ideas can
64 SLIDING MODE CONTROL
be used for more general nonlinear systems and sliding surfaces. More information
is given in DeCarlo et al. (1988) and Utkin (1992).
The original formulation of the invariance principie given in Theorem 3.1 was stated
and proved by Drazenovié (1969). The projection operator approach is introduced
by El-Ghezawi et al. (1983) and the use of 'QR' decomposition to separate the
original system to expose the null and range space dynamics first appears in Dorling
& Zinober (1986). This material also appears in the early chapters of Zinober
(1990).
The use of the controllability canonical form as a platform for designing sliding sur-
faces and switched controllers for single-input systems with matched uncertainty is
widespread throughout the early literature, for example Utkin (1977) . The pitfalls
of using such an approach when unmatched uncertainty is known to be present
is analysed in Spurgeon (1991). A worked example relating to the design of a
controller using the hierarchical approach is given in DeCarlo et al. (1988).
The relationship between the poles of the reduced-order sliding motion and invari-
ant zeros is considered in Young et al. (1977) and by El-Ghezawi et al. (1983) .
The 'unit vector' approach described in Section 3.6 was introduced by Ryan &
Corless (1984). In many ways this work is a hybrid of VSCS concepts and the
uncertain systems approach of Leitmann. Furt her details and relevant references
are given in the original paper of Ryan & Corless (1984). The exposition in Section
3.6 <loes not explicitly consider the issue of existence of solution. This is treated
in careful detail by Ryan & Corless (1984): the uncertain functions fm(·) and
f,, (-) are assumed to be Carathéodory functions, w hich guarantee the existence of
solution away from the sliding surface; they also prove, using the Brouwer fixed
point theorem, that s(t) = O is a feasible solution to the differential inclusion
representing the dynamics on the hyperplane.
The parameter values for the DC motor example in Section 3.6.4 are taken from
the case study in Bélanger (1995).
The more general definition of sliding given in Section 3.7 is discussed in detail
in the work of Utkin (1992) and in the tutorial paper by DeCarlo et al. (1988)
from which Figure 3.5 is taken. Other more complex approximations to smooth
discontinuous unit vector controllers are proposed by Zhou & Fisher (1992) and
Luo & Feng (1989) . An analysis of the smoothed unit vector controller described in
Section 3. 7 is considered in Ryan & Corless (1984) and more recently by Spurgeon
& Davies (1993).
Chapter 4
It has becn established in Chapter 3 that the switching surface design problem
may he interpreted as a static state feedback selcction problem for a particular
subsystem. This chaptcr will consider the developmcnt of appropriatc methods to
solve this problem. The ernphasis will be on numerical tractability and MATLAB
rnfiles will be provided to implement thc design procedures. The first algorithm
is based upon robust eigcnstructure assignment as this offers an effoctive method
of minimising t h e effects of unmatchcd parameter variations on the sliding mode
dynamics. Thc second will be based upon optima! control theory. A final switching
surface design procedurc will be synthesised whereby a desircd modal structure is
specified for the ideal sliding mode dynan1ics. This mcthod is particnlarly cffectivc
for the solut ion of aerospace and vibrntion control problems.
The exposition t hus far has considcrcd rcgulation problems. This chapter will re-
view methods for incorporating a refcrcncc signa! within the sliding mode design
procedure. This is clearly a nontrivia1 problem in that it has been dcmonstratcd
that a sliding system completely reject.s a class of forcing functions. lt is thus
imperative that the scheme is configured in such a way as to ensurc the reference
signa! is not rejectcd. Two methods for incorporating a tracking rcquirement wiJl
be considered. The first involves a rnodel-reference approach and the second in-
volves t hc addition of integral action. The chapter will conclude with a dcsign
study concerning the d_evclopment of a fiight control law. This will illustrate the
application of many of thc techniques introduced in the chapter.
In Chapt.er 3 a particular canonical form, the so-called regular form, was shown
to yield a convenient interpretation of the reduced-order sliding mode dynamics.
Consider the nominal linear model of an uncertain systcm, given by
x(t) = Ax(t;) + Bu(t) (4.1)
where rank(B) =m and (A, B) is a controllable pair. Define an associated switch-
ing function
s(t) = Sx(t) (4.2)
66 SLIDING MODE CONTROL
(4.5)
(4.6)
The matrix sub-blocks in (4.3) and (4.4) can be obtained in terms of the original
pair (A, B) from
T. ATT _ [ An (4.7)
r r - A21
(4.8)
This transformation may be effectively accomplished for any given control\able pair
(A,B), where Bis full rank, using the MATLAB commands described below.
mfile: changes coordinates to bring about regular form
During sliding motion, the switching function s(t) will be identically equal to zero.
Equation (4.5) thus gives
(4.9)
SLIDING MO DE DESIGN APPROACHES 67
In this chapter, techniques for switching surface selection are developed and it is
thus reasonable to make the a priori assumption that S will be selected so that
the matrix product S B is nonsingular. It has been seen in Chapter 3 that this
assumption implies that the matrix S 2 is nonsingular. Condition (4.9), which
defines t he sliding mode, may therefore be re-expressed as
z2(t) - s2-1S1z1(t)
-Mz1(t) (4.10)
where M E lRmx(n-m) is defined to be
The development of the z2 partition in the sliding mode is thus seen to be lin-
early related to the z 1 partition. The sliding mode is governed by equations (4.3)
and (4.10):
z1 (t) A11z1(t) + A12z2(t) (4.12)
z2(t) - M z 1(t) (4.13)
This is an (n - m)th order system in which z 2 has the role of a linear full-state
feedback contr ol signal. Closing the loop in (4.12) with the feedback from (4.13)
(4.14)
The mínimum design requirement is that asymptotically sta.ble dynamics must
be ensured during sliding so that z 1 ➔ O as t ➔ oo. The existence problem
beco mes that of fixing M = S 21 S1 to gi ve (n - m) negati ve poles to the closed-
loop system (4.14). Any additional degrees of freedom may be used to furt her tailor
the performance during the sliding mode. In particular , it will be seen that making
the system (4.14) maximally robust to system uncertainty which occurs wi thin
this subsystem has the effect of rendering the sliding mode dynamics minimally
sensitive to the unmatched uncertainty in the system. Recall that a sliding system
can be made totally insensitive to the matched uncertainty in the problem but
will be effected by any unmatched uncertainty contributions. It is t hus desirable
to minimise the effects of any unmatched signals. The existence problem may be
solved by using a modified form of a ny standard design procedure which prescribes a
linear full-state feedback controller for a linear dynamical syst em. The two methods
which will be considered here are based upon a robust pole assignment approach
a nd also the minimisation of a.n integral cost functional with qua.dratic integrand.
However, it is straightforward to formula.te the switching surface design problem
in t erms of any preferred state feedback design procedure. It should be noted tha.t,
whichever scheme is chosen, fixing M does not uniquely det ermine S since there
a.re m 2 degrees of freedom in the rela.tionship
(4.15)
In the work which follows, the hyperplane matrix S will be determined from M by
letting S2 = Im, giving
ST; = [ M Im ] (4.16)
This a.pproach minimises the calculation in proceeding from M to S and so reduces
the possibility of numerical errors.
68 SLIDING MODE CONTROL
The eigenvalues of the system during sliding will consist of the set of (n - m)
stable eigenvalues assigned to the spectrum of A11 - A 12 M from equation (4.14),
plus the value zero repeated m times. If the state is to remain within the null
space of S, the effect of the parameter variat,ions which do not occur in the range
space of B , Le. any unmatched parameter variations, must be minimised. This will
be achieved by developing a robust control design approach to solve the existence
problem; the aim will be to make the nonzero sliding mode eigenvalues insensitive
to perturbations using robust eigenstructure assignment. This will minimise the
effect of parameter variations outside the range space of B. Consider the closed-loop
system equations (4.12) and (4.13) to be replaced by the perturbed representation
(4.17)
where the second term on the right-hand side of the equation represents a perturba-
tion of the nominal system due to parameter variations. This term will usually be
both time-varying and arbitrary, and the robust assignment approach will seek to
minimise its effect. The nominal sliding mode system defined by equations (4.12)
and (4.13) is of order (n - m), with m control inputs. The control distribution
matrix A12 E lR(n-m)xm must have rank l, which satisfies
1 ~ l ~ min{n - m , m}
where A121 has l linearly independent columns while the matrix A122 is zero. Let
(4.18)
where
(4.19)
The effect of perturbations can be minimised by sett ing the matrix M 2 equal to
zero. The remaining feedback matrix partition M 1 is then determined by applying a
robust control system design procedure to the matrix pair (A 11 , A121 ) . The matrix
M can t hen be calculated from a simple permutation of the rows of the partítioned
feedback matrix using equation (4.18).
An appropriate methodology for determining M 1 is described below. It has been
shown by Wonham (1967) that a full-state feedback matrix can be fou nd to assign
SLIDING MODE DESIGN APPROACHES 69
(4.21)
(4.22)
where the columns of K(-\;) E q::(n+m)xm forma basis for the null space of G(-\;).
Using this notation it is possible to present the following well-known result, which
will be quoted without proof.
Theorem 4.1 The necessary and sufficient conditions to find a real matrix F
satisfying (4.20) are
A desirable property of any closed-loop system is that the eigenvalues are rendered
insensitive to perturbations in the coefficient matrices of the system equations; such
a well-conditioned solution will be termed robust. It has been seen in Theorem 4.1
that the solution to the problem of designing a state feedback controller using
eigenvalue placement is generally underdetermined with ma.ny degrees of freedom.
The objective of the design process under consideration here will be to determine
a robust control system by restricting the degrees of freedom of the assignment
problem. Let v; and w; be the right and Jeft eigenvectors of the closed-loop
system matrix corresponding to an eigenvalue A;. Therefore
C; = (4.25)
70 SLIDING MODE CONTROL
(4.26)
where 11",(V) denotes the condition number of the matrix V = [v1 , ... , Vn ] of right
eigenvectors. Essentially the condition number is a measure of orthogonality of the
eigenvectors v;. The closer the eigenvectors of a matrix are to being orthogonal,
the smaller is the associated condition number and the greater is the robustness
of the eigenvalue locations to changes in the elements of the matrix. The feedback
matrix is obtained by assigning a set of linearly independent right eigenvectors
corrcsponding to the required eígenvalues such that the matríx of eigenvectors is
as well conditioned as possible. The assigned eigenvalues are then as insensitive to
perturbations as possible. The robust eigenvalue assignment problem may therefore
be stated as follows:
Given A, B and a set of desired eigenvalues, find a real matrix F and a nonsingular
matrix V satisfying
(A- BF)V = VA (4.27)
[ Q; q.; ] [ ~] (4.30)
(4.31)
This vector has mínimum misalignment angle to q; and follows directly from the
formula (A.37) in Appendix A.2.6 using t he fa.et that Q(>..;)TQ(>..;) = Im. The
sea.lar term IIQ(>..;? q;II ensures that vrew is normalised. The procedure is contin-
ued iteratively until there is either no improvement in the condition number or a
prespecified maximum number of iterations has been exceeded.
Having determined an appropriate set of eigenvectors, it is relatively straightfor-
ward to recover the state feedback matrix F which prescribes the desired modal
structure to the closed-loop system (4.20) . Equation (4.20) may be rewritten as
N(>..;) ]
[ Ail - A B l[ M(>..;) k; =o (4.33)
as K(>..i) is a basis for the null space of S(>..;). Expanding equation (4.33) yields
(4.34)
The vector k¡ effectively determines where in the allowable subspace V¡ exists and
is thus known for given v;, as seen in equation (4.31). It follows from (4.36) that
the feedback matrix may be determined from
(4.37)
where t he k¡ are those which have been used to determine the corresponding robust,
allowable eígenvector V; at the end of the iteration procedure. In this way the
feedback matrix M in (4.14) and hence the swítching surfa.ce a.re specified. MATLAB
commands to establish a robust switching surface selection for a given (A, B) pair
are given below.
mfile: hyperplane design using a robust pole placement method
M=vplace(A11,A12,p);
Remark
The MATLAB command place from the control system toolbox implements the
robust eigenstructure assignment approach described in this section. However, in
the mfile fra.gment above, a modification of the place command, termed vplace, has
been used in preference; for details see Appendix B.
(4.38)
where Q is both symmetric and positive definite and t 5 is the time at which the
sliding motion commences. The aim is to minimise equation (4.38) subject to the
system equation (4.1). It is assumed that the state of the system at time ts, x(ts),
SLJD!NG MODE DESIGN APPROACHES 73
is a lmown initial condition and is such that x(t) ➔ O as t ➔ oo. The matrix Q
from equation (4.38) is transformed and partitioned compatibly with z :
(4.39)
(4.40)
i n order to solve this problem it is desirable to express it in the form of the standard
LQR problem met in Chapter 2, where z1 , which determines the system dynamics
in the ideal sliding mode, has the role of the state and the effective control input is
a function of z2 . To this end, it is necessary to eliminate the cross term 2zf Q 12 z2
from equation (4.40). Noting that the last two terms in equation (4.40) may be
factored to yield
2z¡ Q12z2 + z'[ Q22z2 = (z2 + Q2lQ21z1? Q22(z2 + Q 2{Q21zi) - z¡QJ1Q 2,}Q21z1
Define
Q Qn - 2
Q12Q }Q21 (4.42)
V Zz + Q:;{Q21Z1 (4.43)
J =2 11"°
t,
z 1T Qz1
- + v T Q22V dt (4.44)
(4.45)
Eliminating the z2 contribution using equation (4.43), the modified constraint equa-
tion becomes
(4.46)
where
(4.47)
The problem thus becomes that of minimising the functional (4.44) subject to the
system (4.46) and has thus been interpreted as a standard linear quadratic optima!
state regulator problem (see Section 2.3.3). The positive definiteness of Q ensures
that Q 22 > O, so that Q 221 exists, and also that Q > O (see Appendix A.2.7).
Furthermore, the controllabilíty of the original (A, B) pair ensures that the pair
74 SLIDING MODE CONTROL
(A11 , A12 ) is controllable (Proposition 3.3) and this in turn is sufficient to ensure
the controllability of (Á., A12). It follows that a unique positive definite solution
A is guarant eed for the algebraic matrix Riccati equation which is associated with
the problem defined by equations (4.44) and (4.46)
• .T 1 T •
P1A + A Pi - P1A12 Q 22 A 12 Pi + Q =O (4.48)
T he optimal v minimising equation (4.44) is given by
v = -Q2/A{2Piz1 (4.49)
Comparing this with (4.13) yields the expression M = Q 221 (A[2 Pi + Q 21 ). The
MATLAB mfile fragment below illustrates this design procedure.
mfile: hyperplane design using a linear quadratic cost function
Qt=Tr*O*Tr' ;
Form reduced order system descript ion and associated weighting matrix
'l.
Qhat=Q11 - Q12*inv(Q22)•Q21;
Ahat=A11-A12*inv(Q22)*Q21;
M=inv(Q22)*(A12'*P1+Q21);
S2=eye(mm); 'l. For simplicity
S=S2*[M eye(mm))*Tr;
eigenvectors. For the multiple-input case, it has been shown in Proposition 3.3 that
if the matrix pair ( A, B) defining the system (4.1) is controllable, then the matrix
pair (A11 ,A12 ) of equation (4.3) is aJso controllable. The direct consequence of
this result is that the eigenvalues of the sliding mode system (4.14) can be placed
arbitrarily in the complex plane by suitable choice of the matrix M. The remain-
ing degrees of freedom available in the assignment problem will be used to modally
shape the system res ponse using a judicious choice of eigenvector form. For this
application of the t echnique, the objective will be to modally shape the system re-
s ponse during the sliding mode . Assuming t hat sliding motion has commenced on
the null space of S, the dynamics exhibited by the system can now be represented
by
x(t) = (A - BK)x(t) (4.51)
where
( 4.52)
This describes the equivalent system motion developed in Section 3.4. During
sliding motion, t he system state must lie within the null space of S by definition
and so
Sx =O <=} S(A - BK)x =O
<=} R(A - BK) ~ N(S) ( 4.53)
Let Ai, i = 1, ... , n be the eigenvalues of t he equivalent system, A - BK, with
corresp onding right eigenvectors Vi, i = 1, ... , n. Then, from equation (4.53):
S(A - BK)v¡ = AiSvi = O ( 4.54)
which shows that either Ai = O or Vi E N(S). It has been shown in the discussion
of the equivalent system motion in Section 3.4 that (A - BK) = Aeq has precisely
m zero eigenvalues. Assuming that Ai, ·i =
1, .. . , n - m a re the nonzero <li:stim:t
eigenvalues of t he equivalent system motion, t he corresponding eigenvectors Vi,
i = 1, ... , n - m will determine the null space of S, since dim [N(S)] = n - m .
It has been shown in t he previous section that the eigenvector corresponding to a
given eigenvalue must lie in an allowable subspace which is determined by t he plant
matrix A , the input matrix B and the eigenvalue itself. If information is known
about a desirable weighting of t he system sta.tes for ea.ch mode, it is possible to
choose a desired eigenvector specification. This will not necessarily be achievable
because it may not lie within the prescribed allowable subspace. In the theoreti-
cal discussion b elow it will be assumed for simplicity that the desired eigenvalues
are real. The modifications required to cover the complex case are outlined in
Appendix B. Assume that q ~ n components are specified. For q ~ m it may be
possible to assign an allowable eigenvector with these components specified exactly.
For q > m an allowable eigenvector which is in sorne sense 'closest' to the desired
eigenvector must be selected. This t echnique of shaping the modal structure will
be referred to in the work which follows as direct eigenstructure assignment. For a
given problem only a few of the components of a desired eigenvector form vf are
usually specified (Harvey & Stein , 1978) ; t he rest may be arbitrary. To account for
t his, a row reordering operation is computed such that
- d V- s· ]
Rvi = [ iíf (4.55)
76 SLIDING MODE CONTROL
(4.56)
where N 1 E IRqxm. The required ó; is thus determined from the linear equation
(4.57)
Whether this equation has a solution will depend on the size of q relative to m and
on the rank of N 1. It can be verified that if q :S m and N 1 is full rank, the equation
always has an exact solution and thus the q ::; m components of the eigenvector
can be assigned exactly. If N 1 is rank deficient for the chosen q :S m components,
the designer should consider interchanging sorne of the specified desired nonzero
eigenvector entries with the unspecified components to render N 1 full rank. In this
way q :S m components can again be specified exactly. In the case when q > m,
equation (4.57) <loes not have an exact solution unless
(4.59)
It should be noted that S is not uniquely determined from the selected eigenvectors
since the equation
has m 2 degrees of freedom. If the eigenvector matrix of equation (4.60) results from
a modal shaping of the system response, the null space dynamics determined by
S will exhibit the required response shaping. In the mfile fragment which follows
the null space of G(>.i), which in turn determines N(>.i), is obtained numerically
using the singular value decomposition approach described in Appendix A.2.4. The
hyperplane equation (4.60) is solved in a similar way to obtain the matrix S .
Once a controller is found to drive the system state and thereafter maintain it
within the null space of S, the technique of VSCS is seen to have the potentia! to
force a nonlinear or perturbed system to exhibit a required modal structure when
in the sliding mode. The MATLAB commands given below allow the designer to
assign desired real eigenvalues and eigenvectors to the ideal sliding mode dynam-
ics. Appendix B contains MATLAB commands which enable the assignment of an
eigenstructure that includes complex conjugate assignment.
SLIDING MODE DESIGN APPROACHES 77
% For the system pair (A,B), the vector lambda defines the n-m desired
% sliding mode poles. The columns of the nx(n- m) matrix specpos are
% used to distinguish vector entries which are arbitrary (denoted by
%O in the relevant matrix entry) from entries which are to be
% specified (denoted by 1 in the relevant matrix entry). The columns
% of the nx(n-m) matrix specent specify any desired eigenvector entries.
[nn,mm]=si ze(B);
end
Up to now only regulation problems have been considered where the objective is to
drive the statcs to zero. The possible ways in which a tracking rcquircment can be
incorporated in the design framework will now be explored. It should be recalled
that when in the sliding mode a system completely rejects any signals which satisfy
the matching conditions. The inclusion of a tracking requirement is thus nontrivial
and must always be achieved in such a way as to ensure that the command is not
wholly or partly rejected by t he sliding system.
78 SLIDING MODE CONTROL
for sorne arbitrary gain rnatrix G. lf the first derivative of the error is to be zero
then, using equations (4.68) and (4.62), the equation
(4.70)
It is thus seen that direct use of the rnodel states w through the control loop has no
effect on the condition for rnodel-following. Noting equation (4.67), equation (4.70)
is satisfied for al! x, w and r if
lf equations (4.71) to (4.72) are to hold for an arbitrary time t, all higher-order
derivatives of the error will also be zero. Consider a control law with the structure
(4. 73)
where
-Ke(t) (4.74)
E t (Am - A)x(t) + Et Emr(t) (4.75)
Substituting the control law (4. 73) into equation (4.66) and assuming conditions
(4.71) and (4.72) hold, then
HD=E (4.77)
it follows from Theorem 4.2 that there exist compatibly dimensioned matrices F
and G such that
BF Am - A (4.80)
BG == Bm (4.81)
It follows that the control law defined by equations (4.73) and (4.74) with u 2
expressed as
u2(t) = Fx(t) + Gr(t) (4.82)
will a.lso achieve perfect following. However, it should be noted that equation (4.75)
will use mínimum control effort dueto the use of the pseudo-inverse.
The design of variable structure model-following control systems will now be con-
sidered. The required system performance is usually specified via the model, and
the controller synthesis is to minimise the error between the model states and the
controlled plant. The associated control component (4.74) must provide robustness
to parameter variations and other externa! uncer tainties which occur in the plant.
Consider the time invariant multivariable plant described by equation (4.61). The
plant output is required to follow the output of the model (4.62). The dynam-
ics of the model-following error system can be obtained in a suitable form from
equation (4.65) by the addition and subtraction of a term A mx, yielding
( 4.85)
During the sliding mode the error state will satisfy the equation
Se(t) =O ( 4.86)
Differentiating this equation with respect to time and substituting from equa-
tion (4.83) yields
Substituting this equivalent control into the model-following error system (4.83)
gives
(4.88)
In the slicling mode, the error system is formally equivalent to a system of order
n - m and the response will be determined by the n - m nonzero eigenvalues of
the system (4.89). Given a matrix pair (A,,., B) which is stabilisable, a hyperplane
matrix S can be found such that the error tends to zero with increasing time.
This hyperplane matrix can be designed by any of the techniques described in
Section 4.2. Sufficiently fast error decay when sliding can be ensured by placing
the n - m eigenvalues sufficiently far enough into the left-hand half of the complex
plane. Having chosen a suitable hyperplane matrix, a unit vector nonlinearity
control structure (3.83) to (3.85) can be determined to drive the error vector into
the null space of S and thereafter maintain it within this sliding subspace. The m
range space eigenvalues, associated with the design matrix <I>, are chosen to ensure
that the error state approaches the sliding surface in a suitably short time. Let
A E m,mxm denote a symmetric positive definite matrix satisfying the Lyapunov
equation
(4.90)
82 SLID!NG MODE CONTROL
Consider initially the development of a tracking control law for the nominal linear
system
x(t) Ax(t) + Bu(t) (4.94)
y(t) Cx(t) (4.95)
which is assumed to be square. In addition, for convenience, assume the matrix pair
(A, B) is in regular form. The control law described here utilises an integral action
methodology. Consider the introduction of additional states Xr E }RP satisfying
x- = [ Xr ] (4.98)
X
(4.99)
where x 1 E IRn and xz E IRm . The (augmented) nominal system can then be
conveniently written in the form
where
(4.102)
(4.103)
A A (4.105)
S = [ S1 S2 ]
and assume S2 = AB2 1 where A is a nonsingular diagonal design matrix. If a
controller exists which induces an ideal sliding motion on S then the ideal sliding
motion is given by
(4.106)
P(z) = [ zI_~A ~ ]
Utilising the PBH rank test, the pair (Án, .412 ) is controllable if and only if
rank [ z Iv
0
C1
zl-Au
-C2]
A =n for ali z E {: (4.107)
12
If z = O then
C1 - C2 C1
rank [ zlp
0 zl-An A12 ]=n {::} det [
-An
- C2 ]
A12
i O
{::}
det [ Z: C2 ] i- O
A12
Otherwise z i- O and
zlv C1
rank [ O zl -An rank [ zl -A11 A12 ] = n- p
However, from Proposition 3.3, (A, B) is controllable if and only if (An, A12 ) is
controllable and therefore by assumption
from the PBH rank test applied to (A 11 ,A 12). Therefore assertion (4.107) is true
and (Á 11 , .41 2) is controllable.
■
ThP. <lP.velopment that follows mirrors the approach in Section 3.6. Specifically,
define
T- [I" º]
S1 S2
s -
(4.108)
ªT s
[ X¡ ]
Xz
then with respect to these new coordinates the nominal system can be written as
where <P is any stable design matrix. The overall control law is then given by
(4.112)
SLIDING MODE DESIGN APPROACHES 85
(4.114)
The positive scalar function which multiplies the unit vector component can be
obtained from arguments similar to those in Section 3.6, It follows that, in terms
of the original coordinates
(4.115)
(4,119)
The configuration of the aircraft has changed very little in principie since the begin-
ning of aviation; almost al! aircraft have a standard arrangement of rudder, aileron
and elevator control surfaces. However, it can easily be shown (Huenecke, 1987)
that this design has considerable disadvantages for particular aircraft manoeuvres.
Consider the situation ill ustrated in Figure 4, l.
: -1----~~
1
1
1
1
1
1
1
1
1
1
1 1
An aircraft is flying on a straight track. The pilot wants to point the aircraft
towards the point P. Note that there is a longitudinal distance separating the air-
craft from P as well as a vertical separation. The aircraft must either increase its
86 SLIDING MODE CONTROL
altitude or pitch by the angle "f. A conventional aircraft will first of ali use the ele-
vator, after which a rotation about the lateral axis will be initiated. The required
movement in the vertical direction can now be accomplished; the path alteration
will therefore take the form of an oscillation.
If the combined rotational and translational motion could be successfully decou-
pled, the pilot would have a much more straightforward flying task. Consider the
following two configurations for decoupling of the longitudinal motions. Firstly, the
pilot commands the pitch angle 0 whilst maintaining zero perturbation of the flight
path angle 'Y- This configuration is illustrated in Figure 4.2 and will be referred
to as pitch-pointing control. For the second configuration, the pilot commands the
flight path angle I whilst maintaining zero perturbation of the pitch angle 0. This
so-called vertical trans/atíon control is illustrated in Figure 4.3.
Such advanced longitudinal control modes can be realised by using flaps or canard
control surfaces in conjunction with the elevator. In this way, for example, a change
of flight path without rotation can be achieved by directly influencing the lift act-
ing upon the vehicle. A computer-based automatic flight control system is required
for the coordinated actuation of the control elemcnts. Smith & Anderson (1982)
applied optima! control theory for the design of such decoupled controllers. How-
ever , it is extremely difficult to translate specifications for piloted vehicles into the
required form for an optima! control problem; for example, the required damping,
frequency and decoupling are not easily incorporated into a quadratic performance
index, which results in quadratic weight adjustments by tria! and error. This can
SLIDING MODE DESIGN APPROACHES 87
l
where
pitch angle (rad)
pitch rate (rad/sec)
x(t) = ªu angle of attack (rad) (4.121)
Br¡:
O 1 O o
O - 1.99 - 13.41 - 18.95
A= O 1.00 - 1.74 -0.08 (4.123)
[ o o o - 20
o o o o
and
BT =[0 O O 20 O ] (4.124)
O O O O 20
The eigenvalues of the open-loop system are given by
The performance of the flight path angle is of direct interest and so in the state-
space representation the pitch angle state is replaced by the flight path angle using
the relationship
(4.125)
The state-space representation in equations (4.120) to (4.124) is therefore modified.
T he state vector (4.121) becornes
(4.126)
88 SLIDI NG MODE CONTROL
The remaining rows of the A matrix are unchanged, as are the E and u formula-
tions.
Here a design method using direct eigenstructure assignment to obtain the required
decoupling, damping and rise time of the system response together with a set of
feed-forward gains which are computed to ensure steady-state tracking of a step
command from the pilot will be used to prescribe an ideal model. A variable
structure model-following control system will then be derived.
In order to provide rapid manoeuvring, precision tracking and precise flight path
control, the following desired closed-loop eigenvalues are chosen
The desired short period frequency and damping are thus chosen to be ( = 0.8
and wn ;e:: 7 rad/sec. The desired eigenvectors are chosen to decouple the short
period and fiight path modes. The desired eigenvectors corresponding to the short
period mode are thus selected to have no contribution from the first state, which
corresponds to the flight path angle, but are allowed contributions from the pitching
motions of the aircraft. The desired eigenvector corresponding to the flight path
mode is chosen to have the opposite characteristics. These choices should prevent
a pitch attitude command from causing significant change in the flight path angle;
the performance specification is thus to minimise the coupling between pitch rate
and flight path angle. The desired and achievable eigenvectors are
= [ vf V~ v{ vf vf]
o o 1 V14 V15
l
1 + V22j l - v22j o V24 V25
= V31 + j V31 - j o V34 V35 (4.128)
V41 +V42Í V41 - V42Í V43 1 V45
V51 + V52j V51 - V52j V53 V54 1
l
Vij
Vª V~ l
o o 0.67 0.01
1.00 - 9.50j 1.00 + 9.50j - 0.33 0.97 - 0.03
0.11
-0.93+ j -0.93 - j - 0.33 -0.06 0.03 (4.129)
- 1.82- 2.24j - 1.82 + 2.24j 0.29 1.00 -0.11
2.97 - 2.62j 2.97 + 2.62j - 0.18 - 0.21 1.00
SLIDING MODE DESIGN APPROACHES 89
(t) = [ ,(t)
8(t) ] 10 1 0 0 ]
y [ 1 O O O O x(t)
= Hx(t) (4.131)
The inverse of the steady-state gain for the triple (A+ BF, B, H) will provide an
appropriate feed-forward injection signal
(4.132)
T he eigenvalues of the range space dynamics, linked with the design of the lin-
ear control component (4.91), are chosen to be - 20 and -20. This ensures rapid
approach to the sliding subspace whilst accommodating the bandwidths of the ele-
vator and flaps. The linear component may be expressed in terms of the physically
meaningful error coordinate system as
Le(t)
- 1.2285 -0.1858 -2.1624 0.0782 0.0460 ] ( ) (4 141)
= [ 1.8631 0.2827 3.0804 -0.1299 -0.0660 e t ·
The parameters of the nonlinear control component (4.92) are given by
0.025 O ]
[ (4.142)
O 0.025
where the gain p is set equal to unity. This control scheme is augmented to form
the full model-following control law of equation (4.93). The step response of the
aircraft system when pitch-pointing is shown in Figure 4.4. Excellent decoupling
of the pitch and flight path channels is exhibited.
't;l
e
"'s
E
o 0.5
u
.Q
¿:
o
o 2 3 4 5 6 7
Time, sec
Here the pilot's command is incorporated using the alternative integral action ap-
proach. The output equation is taken to be defined by the matrix H in equa-
tion (4.131). The design matrix r in equation (4.97) is given by
and the poles of the ideal sliding mode dynamics have been chosen to be
{ -5.6 ± 4.2j, -1, -0.4, - 0.7}
The corresponding desired eigenvectors which define the required decoupling are
selected as
SLIDING MODE DESIGN APPROACHES 91
From equations (4.115) to (4. 118) the controller gains are given by
where the eigenvalues of the range space dynamics are taken to be - 20 and - 20.
The step response of the aircraft system when undergoing a vertical translation
manoeuvre is shown in Figure 4.5. Excellent decoupling of the pitch and flight
path channels is again exhibited.
1.5
¡:¡
E
"'u
o."'
0.5
i5"'
~ - ..
,, ~.
-~
t:: o
>"'
-0.5
o 2 3 4 5 6 7
Time, sec
Fig ure 4 .5: Vertical translation manoeuvre with integral action VSCS
4 .6 SUMMARY
This chapter has developed design algorithms appropriate for the construction of
sliding mode control schemes for uncertain systems where ali of the states are mea-
surable. One particular algorithm has been presented which seeks to minimise
the effects of unmatched parameter variations, and in this way it has been shown
92 SLIDING MODE CONTROL
that the sliding system cannot only reject a class of uncertainty completcly but
can also be made rnaximally robust to the remaining unmatched uncertainty. Ap-
propriate methodologies have been describcd to solve servomechanism and other
tracking problems. Much of the work has been illustrated in the development of an
advanced aircra.ft flight control system. In later chaptcrs industrial application of
thesc results will be explored and the further development of thc design algorithrns
to encompass the output feedback case will be considered.
The regular form used for sliding mode dcRign is doscly related to the controllability
cancnical form for a multivariable linear system dueto Kwakernaak & Sivan (1972).
The robust eigenstructure assignrnent procedure used in Section 4.2.1 is based
upon the work of Kautsky, Nichols and co-workers (Kautsky et al., 1985; Kautsky
& Nichols, 1983; Nichols, 1987). The quadratic Dinimisation approach of Utkin
& Young (1978) is also described in Zinober & Dorling (1990). The latter also
considcrs eigenstructure assignment procedures for hyperplane design. F\irther
approaches for solving the existence problem are available in the contributions of
Wocdham & Zinober (1993) and Zinober (1994), which describe procedures that
force t he rcduced-order eigenvalues to lie in specified regions of the complex plane.
Alternative methods for achieving robust modcl-refercnce control in the presence
of uncertainty and disturbances use the hyperstability concept or Lyapunov tech-
niqucs. Further dctails can be found in the following rcferences: Lanclau (1974),
Landau & Courtoil (1974), Landau (1979), Butchart & Shakcloth (1966) and also
Monopoli et al. (1968).
Fm· further detaib on the conditions for perfect model-following, see the contribu-
tions of Erzberger (1968), Chan (1973) and Corless et al. (1985) .
VSCS model-following control schemes are considercd in thc papers by Young
(1977), Zinober et al. (1982), Balcstrino et al. (1984), Dorling & Zinober (1985)
and Zinober (1984).
Thc integral action control law in this chapter 1s based on the work of Davics &
Spurgeon (1993).
Broussard's method (O'Bricn & Droussard, 1978; Broussard & O'Brien, 1980) can
be useful to compute thc feed-forward control effort that is required to ensure
steady-state tracking of more cornplcx signals in modcl-reference frameworks.
Further details of thc variable structure rnodel-following aircraft control systern can
be found in the paper by Mudge & Patton (1988).
Chapter 5
5.1 INTRODUCTION
In the previous chapter, arguably the most stringent constraint related to the
assumption that ali the states of the system to be regulated were available to the
controller. In most practica] situations this is not the case. In sorne circumstances
it is impossible or prohibitively expensive to measure ali of the process variables.
Alternatively, the system may be so complex that a system identification approach
must be adopted to obtain a reasonable model. In this case the states will usually
have no physical meaning and thus cannot be measured. This chapter considers
t he problem of designing the sliding surface and the variable structure control law
in such a way that only output information is required.
The majority of this chapter is devoted to developing a framework for designing a
regulator for uncertain systems where only output information is available to t he
controller. A particular canonical forro, which may be viewed as a special case of
the regular form introduced in Section 3.4, will be shown to be important in the
design procedure. At the end of the chapter, it will be shown that these ideas can
be incorporated into a model-reference structure to generate an output tracking
controller. The control law that will be developed is similar to the unit vector
struct ure employed in the previous chapter except that only output information is
used.
where x E lR", u E lRm and y E ffiP with m :S p < n. Assume that the nominal
linear system (A, E, C) is known and that the input and output matrices E and
C are both of full rank. The unknown function f: lR+ x lR" x lRm--+ lR", which
represents the system nonlinearities plus any model uncertainties in the system, is
94 SLIDING MODE CONTROL
for some known function 0: : IR+ x IRP ➔ IR+ and positive constant k1 < l.
The intention is to develop a control law which induces an ideal sliding motion on
the surface
S={xEIRn :FCx=O} (5.4)
for sorne selected matrix FE IRmxp _ A control law of the form
Consider the special case when the number of inputs equals the number of outputs:
such plants will be referred to as square. Consider first the choice of hyperplane to
guarantee a stable reduced-order motion.
From Section 3.3, in order for a unique equivalent control to exist, det(FC B) :f. O.
Beca.use by assumption m = p, it follows that F is a square matrix and therefore
the expansion of C(s1 - A) - 1 B; for furth.er details see for example Chen (1984).
USING OUTPUT INFORMATION 95
• det(CB) f= O
• invariant zeros of (A, B, C) are in ([!_ .
All the preceding analysis of course depends on the existence of a control law which
induces sliding. A natural starting point is to examine the effect that considering
outputs only has on the control structure developed in Section 3.6. T he nonlinear
cornponent in (3.85) depends only on the switching function and thus poses no
problem. The linear component, however, was shown to be
(5.7)
It is clear that unless by sorne good fortune CA= MC for sorne ME JR,'nxm then
(5.8) cannot be realised. In the remainder of this section a different approach will
be explored.
Assume that (5.1) is expressed in regular form and the triple (A, B , C) is partitioned
as
(5.9)
- [ J (5.10)
T= C1
Since CB = C 2 B 2 and both det(CB) =I= O and det(B2 ) f:. O it follows detC2 f:. O
and the transformation matrix in (5.10) is nonsingular. In the new coordinate
system
FC = [ O F] (5.11)
where Au = Au - A12C2 1 C1 -
Lemma 5.1 The invariant zeros of (A., B, C) are the eigenvalues of A11 .
Proof
This can be shown using a slight modification to the arguments in the proof of
Proposition 3.4.
■
p (5.12)
96 SL!DING MODE CONTROL
where the symmetric positive definite sub-block P 2 is a design matrix and the
symmetric positive definite sub-block Pi satisfies the Lyapunov equation
(5.13)
F~B'f P2 (5.14)
(5. 15)
and define
(5.18)
This scalar is well defined since the matrix on the right is symmetric and therefore
has no complex eigenvalues.
Lemma 5. 2 The symmetric matrix L(1') ~ P A 0 + AJ P where A 0 = A - 'YBFC is
negative definite if and only if 'Y > 'Yo.
Proof
Using the definition of Ao and the structural constraint (5.15) it follows that
Using the partitions of A,FC and P from equations (5.11) and {5.12) respectively
and the definitions of Q2 and Q3 it follows that
Using the properties of symmetric matrices (see Appendix A.2.7), and the fact that
Q 1 is positive definite, it follows that
where 12 is a positive design scalar which will be shown to define the region in
which sliding takes place. In the new coordinate system, the uncertain system
(5.1) can b e written as
Proposition 5.1 The variable structure control law above quadratically stabilises
the uncertain system given in equation ( 5.22) .
Proof
Consider as a candidate Lyapunov function the positive definite expression
(5.23)
Taking derivatives along the system trajectory, and using the structural constraint
from equation (5.15) gives
But by definition
p(t,y) == k1p(t,y)+k1rllFYll+a(t,y)+,2
> k1(ll11yll+,i1Fyll)+a(t,y)+,2
> k1 llull + a(t, y)+ ,2 (5.24)
if x i= O and , > ,o
and therefore the system is quadratically stable.
■
98 SL!DING MODE CONTROL
Corollary 5.1 An ideal sliding motion takes place on the surface S in the domain
=
D {x E lRn : IIB21 Atxll < ')'2 - r¡} where the matrix At represents the last m
rows of the closed-loop matrix Ao and r¡ is a small scalar satisfying O < r¡ < ')'2 .
Proof
Let the switching function s(x) = FCx then from equation (5.22) it follows that
Using the fact that FT = AB2 it follows that (F- 1 )T P2 F- 1 FCA 0 = B 21 A5 and
arguing as in Proposition 5.1 it can be verified that
V,,= 2sT B 21 A~x + 2s'1' ({ - v) ::; 2llslJIIB2 1 Atxll - 272llsll < -2r¡llsll (5.25)
if x E fl. From Proposition 5.1 the states x(t) are quadratically stable and so there
exists a t 0 such that x(t) E fl for ali t > t 0 . Consequently (5.25) holds for ali t > t 0 ,
and arguing as in Section 3.6.1, a sliding motion will be attained in finite time.
■
Remark
From the statement of Corollary 5.1 it is clear that the parameter ~¡2 defines the
region fl in which sliding takes place. As a result, outside of this region the system
trajectories may pierce the sliding surface without sliding taking place and the
control law will behave in a similar fashion to a relay. Because quadratic stability
has been demonstrated, eventually the trajectories will enter D and sliding will
take place.
A solution to the special case of square systems has been presented. The next
section aims to provide a framework for the general case when the number of
outputs is greater than (or equal to) the number of inputs.
Consider the system in (5.1) and assume that p 2: m and rank(CB) = rn. The
reason for imposing this rank restriction at the outset is that for a unique equivalent
control to exist, the matrix FCB E IRmxm must have ful] rank. It is well known
that
rank(FCB)::; min{rank(F), rank(CB)}
and so in order for F BC to have ful! rank both F and C B must have rank m . The
matrix F is a design parameter and therefore by choice can be chosen to be of full
rank. A necessary condition therefore for the matrix F BC to be foil rank is that
rank(CB) = m .
The first problem which must be considered is how to choose F so that the associ-
ated sliding motion is stable. A control law analogous to (5.19) to (5.21) will then
be used to guarantee the existence of a sliding motion.
USING OUTPUT INFO!lMAT!ON 99
In view of the fact that the outputs will be considered, it is convenient to introduce
a coordinate transformation to make the last p states of the system the outputs.
Define
Te -- [ N¡e ] (5.26)
where Ne E lRnx(n-p) and the columns span the null space of C. The coordinate
transformation x e-+ Tcx is nonsingular by construction and, as a result, in the new
coordinate system
From this starting point a special case of regular form will be established: suppose
in the new coordinate system
(5.27)
is nonsingular, and with respect to the new coordinates the triple (A, B, C) is in
the form
A = B C = [O T ] (5.29)
where A 11 E JR,(n- m)x(n- m) and the remaining sub-blocks in the system matrix
are partitioned accordingly.
Let
pBm ~
[F1 F2] =FT
where T is the matrix from equat ion (5.27). As a result
(5.30)
where
(5.31)
100 SLIDING MODE CONTROL
(5.32)
A 11 _ [ Auu (5.34)
- A1121
where A1111 E JR(n-p)x(n- p) and suppose the matrix pair (A1111 , A1121) is observ-
able.
l t fotlows that
l
[ , I - Am, A1112
rank [
zl-Au
C1 ] = rank A1~21 z l -An22
I p- m
k [ z l -Aun ]+(p- m)
= ran Aº
1121
for ali z E (C
and hence from the PBH rank test, and using the fact that (A1 m, A 1121) is ob-
servable, it follows that
rank [ z l C1A11 ] =n - m
T.ob. A l l l l r-1 _
obs -
[ At1
O and
where Aº11 E m,rxr , Aº21 E IR(p- m)x(n- p- r) , the pair (Aº22, Aº21 ) is completely
obser vable and r ~ Orepresents the number ofunobservable states of (A1111 , A1121) .
2 For further details see Chen (1984).
USING OUTPUT INFORMATION 101
T _ [ Tobs O ] (5.35)
ª- O Ip
which, when used in conjunction with Te and Tb from equations (5.26) and (5.28),
generates the canonical form described below.
Lemma 5.3 Let (A, B, C) be a linear system with p > m and rank(CB) = m.
Then a change of coordinates exists so that the system triple with respect to the
new coordinates has the following structure:
= (5.37)
where A11 E m,rxr, A~z E m,(n-p- r)x(n-p-r) and A~1 E m,(p-m)x(n-p-r) for
some r 2: O and the pair (A~2 , A~i) is completely observable.
(5.38)
C=[O T] (5.39)
Ac=Tc•A•inv(Tc);
Bc=Tc•B;
Cc=C*inv(Tc);
A1111=Aa(1:nn-pp,1:nn- pp);
A1121=Aa(nn- pp+1:nn-mm,1:nn-pp);
Af=Ta•Aa•inv(Ta);
Bf=Ta•Ba ;
Cf=Ca•inv(Ta);
In the case where r > O, the intention is to construct a new system (..411 , .81 , é\ )
which is both controllable and observable with the property that
and (5.40)
where A 122 E IR(n - m - r)xm and A t 2 E lR(n-p-r)x(p-m) and form a new sub-
system represented by the triple (..411 , A122 , é\) where
t,.
(5.41}
USING OUTPUT INFORMATION 103
Proof
Using the partition in equation (5.37), by definition
Af1 A\'2
Af, = A11 -Ai2 KC1
[ o A22 An] - [ Ü(n-m) x(n-p) A12K ]
o A21 A~
[ AJ
1 [A\'2 Ai1iJ
Á11 ]- [ ~ Ai21K ]
Ai22K
[ AJi
Af'.?, j Ai1i - A1~1K
A11 - A122KC1
] ]
Therefore ,\(Afi) = ,\(Af1 ) U ,\(Án - A122KC'i) as claimed.
■
Lemma 5.5 The spectrum o/ Aí'1 represents the invariant zeros o/ (A, B, C).
Proof
Let P(z ) denote Rosenbrock's system matrix
{::>
zl -Au -A12 ]
( [O -Ti] -Tz loses rank
since det(B2) -f. O where T1 E IR(p- m)xm and T 2 E IR.mxm representa partition of
thc matrix T. Substituting for An from (5.37) and repartitioning gives
- A\'2
z l - A11 zl - A 22
[ [O -Ti] Azi
o
104 SLIDING MODE CONTROL
where * represents a matrix sub-block which plays no part in the analysis. Because
T is full rank, the expression on the right, and hence P(z), loses rank if and only if
-Af2
zI - k 22
-A~1
l loses rank
By construction the pair (A~2 , A~1 ) is completely observable and therefore from
the PBH observability rank test
zI -A~2 ]
= n- p- r for ali z E
rank [ -
Aº21 (C
Therefore
(5.42)
where ÍJ¡ E IR(n-m-,·)xm' and is of full rank. If Km' = T;;J K and Km' is parti-
tioned compatibly as
tm'
tm-m'
then
A.¡¡ - A122KC1 = Á.11 - [ B¡ O ] Km1 C1 = A.u - B1KJ':1
and (Á.11, A122 , C1) is stabilisable by output feedback if and only if (Á. 11 , B1 , C1) is
stabilisable by output feedback. The reason for the somewhat tortuous argument
to establish the subsystem (Á.11 , B1, Ci) is that in order to use standard output
feedback results, the triple must be controllable, observable and satisfy the Kimura-
Davison conditions (see Section 2.3.4), which in this case amount to
Lemma 5.6 The pair (Au, B1) is cornpletely controllable and (A.11 , C1) is corn-
pletely observable.
P roof
Because the pair (A, B) is in the canonical form of Lemma 5.3, which is a special
3The phrase stabilisable with respect to output feedback will indicate that for the linear system
(A, B, C) there exists a fixed gain K such that the matrix A - BKC is stable.
USING OUTPUT INFORMATION 105
case of the 'regular form', the pair (A, B ) is completely controllable if and only if
t,he pair (A 11 , A 12 ) is completely controllable. Therefore from the PBH rank test
This implies
l
pair (A.n,C1)
rank [
z l - A.11
-
C1
= rank
¡z l - A~2
-A~1
O
- Ai."h
zl - A~
l p-m
= rank [ z l_-
A21
¿~ 2
] + (p - m) for ali z E (C
zl - A~2 ]
rank [ -Aº21 = n-p-r for ali z E (C
and therefore
rank [ zl - A1 ] =n - m- r for ali z E (C
01
hence (A.u, Ci) is completely observable.
■
The rest of this section considers the problem of constructing an appropriate control
law to induce sliding.
Assume there exist s a K1 E IRm'x(p-m) such that Á 11 - .B1K 1C1 is stable. Let
(5.44)
t = (5.45)
and C1 is defined in (5.31). In this new coordinate system the system triple
(Á, B, FC) has the property that
A = [ ;!11 (5.46)
A21
where Án = A11 - A12KC1 and is therefore stable. This is the same structure
as considered in (5 .11) of Section 5.3. An alternative description is that by an
appropriate choice of F a new square system (A, E, FC) has been synthesised
which is mínimum phase and relative degree l. Consequently, the control structure
described in Section 5.3 can be utilised to induce sliding on S. Let t he control law
be given by
u(t) = - "(Fy(t) - vy (5.47)
where
...!::1!.S!l_
p (t, y ) IJFy(t)II if Fy =/ O
(5.48)
o otherwise
and
(5.49)
where the Q 1 , Q 2 and Q3 satisfy expressions (5.13), (5. 16) and (5.17) and p(t, y) is
defined in (5.21). Arguing as in Section 5.3 the uncertain system is quadratically
stable andan ideal sliding motion is inducecl on S.
The results of this section can be summarised as follows.
Proposition 5.2 There exists a matrix F defining a surface S which prov·ides a
stable sliding motion with a unique equivalent control if and only if
• the triple (A11, .81, C1) is stabilisable with respect to output feedback.
5 .4.3 Example 1
A=[~ -1
~i
?3 - 1 B=[~] 1
(5.50)
USING OUTP UT INFORMATION 107
l
In the canonical form of Lemma 5.3 the system becomes
- 1.5816 0.0192
1.4071 0.3845
0.2953 0.3400
0.1457
-1.7080
0.1971
B-
º
O
[ -3.9016 l
and
o 0.3417 - 0.9398 "]
o 0.9398 0.3417
It can be verified that the ene-dimensional matrix B 2 = -3.9016, the orthogonal
matrix
T = [ 0.3417 -0.9398 ]
0.9398 0.3417
Here r = Ohence the original system <loes not possess any invariant zeros. Arbitrary
placement of the pales of Á 11 - Íh K 1 C1 is not possible since only a single scalar is
available as design freedom. For the single-input single-output system (Á 11 , .81 , C1 )
the variation in the poles of Á 11 - .B1 K 1 C1 with respect to K 1 can be examined by
root locus techniques; see for example Ogata (1997). In this case if the gain matrix
K = K 1 = - 1.0556 then >.(A11 - Í3iKC1 ) = {- 1, - 2}, from which
F F2 [ K 1 ] TT
F2 [ - 1.3005 -0.6503 ] (5.51)
where F2 is a nonzero scalar which will be computed later. Transforming the system
into the canonical form using f defined in (5.45) generates
p
1
= [ 0.3368 0.1891 ]
0.1891 0.5401
F = [ 5.0741 2.5370 ]
The following closed-loop simulation represents the regulation of the initial states
[ 1 O O] to the origin. Figure 5.1 represents a plot of the switching function versus
time. The hyperplane is not globally attractive since at approximately 0.3 second
it is pierced and a sliding motion cannot be maintained. Only after approximately
1 sec is sliding established. Figure 5.2 shows the decay of the states to the origin.
108 SLIDING MODE CONTROL
"' 0.5
%1
úí
o
§
"'
;>-.
(/)
-0.5
4 5 6 7
T ime, sec
In the analysis encapsulated in Proposition 5.2, it was assumed that the triple
(Á 11 , .81 , C\) was stabilisable with respect to output feedback. This property can
be guaranteed if the so-called Kimura-Davison condition holds. 4 If it is not possible
to synthesise a K 1 to stabilise the triple (Á 11 ,.81 ,é\) then it is natural to explore
the effect of introducing a compensator - i.e. a dynamical system driven by the
output of the plant - to introduce extra dynamics to provide additional degrees of
freedom.
Consider the uncertain system from equation (5.1) together with a compensator
given by
Xc(t) = Hxc(t) + Dy(t) (5.52)
where the matrices H E IRq xq and D E IRqxp are to be determined. Define a new
hyperplane in the augmented state space, formed from the plant and compensator
state spaces, as
(5.53)
4It is quite possible of course for the triple (A.11, .81 , C1) to be output feedback stabilisable if
the Kimura- Davison constraint <loes not hold, since it represents only a sufficient condition; for
instance, see Sect ion 5.4.3.
USING OUTPUT I NFORMATION 109
where Fe E lRmxq and F E lRmxp. As in Section 5.4, assume that the nominal
linear system (A, B, C) is in the canonical form of Lemma 5.3 and partition the
matrix FT, where T is the orthogonal matrix from the output distribution matrix,
as
p~r ~
[ F1 F2 ] = FT
In an analogous way define D1 E lRqx(p-m) and D2 E lRqxm as
(5.54)
If the states of the uncer tain system are partitioned as
X= [ X¡ ] tn-m (5.55)
X2 tm
then the compensator can be written as
(5.56)
where C 1 is defined in equation (5.31). Assume that a control action exists which
forces and maintains motion on the hyperplane S e given in (5.53). As usual, in order
for a unique equivalent control to exist, the square matrix Fz must be invertible. By
writing K = F 2- 1 F 1 and defining K c = F2- 1 Fe then the system matrix governing
the reduced-order sliding motion, obtained by eliminating the coordinates x 2 , can
be written as
[Af2 1 A~ 1 - A121K]
.A11 - A122KC1
(D1 - D2K)C1
As in the uncompensated case, it is necessary for the eigenvalues of Af1 to have
negative real parts. The design problem becomes one of selecting a compensator,
represented by t he matrices D 1 , D 2 and H , and a hyperplane represented by the
matrices K and K e so that the matrix
(5.59)
is stable. Again if there is ra.nk deficiency in the matrix A122 the problem is
over-parametrised. As in Section 5.4, suppose rank(A112 ) = m' < m and Jet
Tm' E lRm xm be a matrix of elementary column operations such that
110 SLIOJNG MODE CONTROL
where Í31 E rn,(n-m-r)xm' and is of full rank. Define partitions of the transformed
hyperplane matrices as
= [ ~~~ ]
tm'
T~,1 K and
tm-m'
(5.60)
As before, the matrix given in (5.60) will be written as the result of an output
feedback problem for a certain system triple. Unfortunately, a degree of over-
parametrisation is still present in (5.60), which for simplicity will be removed by
defining
(5.61)
This is comparable to t he situation which occurred in the uncompensated case
where K 2 was found to have no effect on Án - É 1 KC1 . The key observation is
that equation (5.60) can now be written as
[Au b~t~
<1C1 -B}{cl] = [A31~] - [11 -t] [i: I~r] [~1 f]
Thus by defining
A _ [ Au o ]
q - O Üq x q
then the parameters K1,Kc1, D1 and fI can be obtained from output feedback
pole placement of the triple (Aq, Bq , Cq). In order to use standard output feed-
back results it is necessary for the triple ( Aq, Bq, Cq) to be both controllable and
observable.
Lemma 5. 7 The matrix pairs (Aq, Bq) and (Aq, C9 ) are controllable and observable
respectively.
Proof
From the definition of (A9 ,Bq) it follows that
rank [ zl - A 9 Bq ] =: rank [ z l - Au ÍJ 1 ] +q
for all z E (C. From Lemma 5.6 the pair (A11 , ÍJ1 ) is controllable and therefore from
the PBH rank test (Aq, B 9 ) is controllable, as claimed. Using the fact that from
Lemma 5.6 pair (A 11 , 6 1 ) is observable, a similar argument proves that (A9 , Cq) is
observable. ■
The Kimura- Davison conditions for the triple (Aq, Bq, Cq) amount to requiring
that
m' + q + p 2 n - r + I (5.62)
Thus for a large enough q, the Kimura- Davison conditions can always be satisfied
and the static output feedback method can be employed. 5
5 An example of such an approach is given in Bag et al. (1997).
USING OUTPUT INFORMA'.l'ION 111
Ra:ther than formulating the hyperplane design problem as a static output feed-
back problem, this section explores an observer-based methodology. Consider the
compensator defined in (5.52) then, as in the previous section, eliminating any in-
variant zeros, the assignable dynamics of the sliding motion are given by the system
matrix
(5.63)
(5.64)
and the pair (A~2 , A~1 ) is observable. Recall that the structure of the output
distribution matrix above was used as a sta.rting point for the development of a
reduced-order Luenberger observer in Section 2.3.5. Consequently, if the input
dist ribution matrix is partitioned conformably so that
(5.74)
where
(5.75)
It can easily be verified that t he closed-loop system formed from (5.66) and (5.67)
is given by
(5.76)
and from the sepa.ration principie the closed-loop poles are ,\(H) U .\(Án -A 122 K).
The system matrix associated with (5.76) is identical to the system matrix of the
reduced-order sliding motion given in (5.59). Therefore the choice of compensator
mat rices in (5.69) to (5.71) and the hyperplane matrices (5.72) and (5.73) give rise
to a stable sliding mode.
MATLAB commands to generate the compensator matrices are given below.
mfile: generates a compensator based on a reduced-order observer
A11o=Af(1 :r,1:r) ;
A12o=Af(1:r,r+1:nn-pp);
A22o=Af(r+1:nn-pp,r+1:nn-pp);
A21o=Af(nn- pp+1:nn-mm,r+1:nn-pp);
A121m=Af(1:r,nn-pp+1:nn- mm) ; 'l. Extracts the various sub-blocks
A122m=Af(r+1:nn-pp,nn- pp+1 :nn-mm); 'l. f rom the system matri x A under
A22m=Af (nn-pp+1 : nn- mm,nn- pp+1:nn-mm); 'l. t he as sumption that a coordinate
A121=Af(1:r,nn-mm+1:nn); 'l. transformation has been used to
A122=Af(r+1:nn-rnm,nn-rnm+1:nn); 'l. obtain the canonical f orm of
A211=Af(nn-mm+1:nn,1:r) ; 'l. Lemma 5.3
A212=Af(nn-rnm+1:nn,r+1:nn-pp);
A213=Af(nn-rnm+1:nn,nn- pp+1 :nn-rnm);
A22=Af(nn-mm+1: nn, nn-mm+1:nn) ;
H=A22o+Lo*A21o;
US!NG Ol/TPUT INFORMAT!ON 113
D1=A122m+Lo*A22m-H•Lo;
D2=A1221+Lo*A1222;
K=K2-K1*Lo;
Kc=K1;
Phi=diag(p3);
Lambda=S*Bf;
L=-inv(Lambda)*S*Ahat + inv(Lambda)*Phi*S
P=lyap(Phi',eye(mm));
The control laws, based on the plant outputs described earlier in this chapter, tend
to produce 'high gain' controllers. The next section considers a control law which
utilises the plant outputs and the compensator states and is more akin to the state
feedback controllers of Section 3.6.
Assume that there are r (stable) invariant zeros and part ition the state vector x 1
as in (5.55) so that
X¡ = [ ::~ ]
X12
~:-¡,-r
lv- m
As a result, the (original) compensator can be written as
(5.77)
Define a new dynamical system by
and
Using the partitions (5.36), (5.37), (5.40) and (5.65), the original dynamics can be
written as
(5.80)
114 SLIDING MODE CONTROL
(5.89)
Note that the triple (Á., B , C) can be obtained from the canonical forro (A , B , C)
via a similarity transformation. Thus the original system together with the com-
pensator can be written as
é(t) = fI e(t) (5.92)
i(t) Áx(t) -Áee(t) + B(u(t) + ((t,x,u)) (5.93)
Note also that the sliding surface S e can be written as
{x E lR,n : Sx =O}
USING OUTPUT INFORMAT!ON 115
where
S = A [ Omxr I<c K Im ] (5.94)
Define a switching function
s(t) = Sx(t) (5.95)
and define a linear feedback component
(5.96)
where A = S B and <I> E lRmxm is a stable design matrix. Let P be t he unique
positive definite solution to the Lyapunov equation
(5.97)
(5.104)
where the equality on the right comes from rearranging (5.100). Therefore from
inequalities (5.103) and (5.104) it follows that
(5.105)
Because e(t) satisfies the unforced stable system in (5.92) , e(t) enters n in finite
time and remains there. Consequently, the inequality in (5.105) demonstrates that
an ideal sliding motion is induced on Se in finite time.
■
Remarks
The control law in (5.96) to (5.100) is effectively a state feedback controller since
the components Zr and Xc are estimates of the true states Xr and x 11 (up to a
coordinate transformation). Indeed, viewed in this way, the control law can be
thought of as being equivalent to the state feedback controller in Section 3.6 and
not 'high gain' in nature.
-2 1
A=
[
o o
O 1 e-[
-
ºO oO o1 1]
O (5.106)
1 - 6
This system is already in the canonical form of Lemma 5.3 and thus
. _ [ Af1
A 11 - o [!]
In terms of the compensator design, the triple of interest is given by
-
An =[ O 4 ]
1 O (5.107)
and so the triple (..411 , A122 , C1 ) is not stabilisable by static output feedback and a
compensator-based approach must be employed. It follows from (5.107) that
and so from equations (5.69) to (5.71) an appropriate parametrisation for the com-
pensator is
H = Lº D1 = 4 - (L 0 )2 D2 = 1
USING OUTPUT INFORMATION 117
0.5
e:
.g 0.4
u
e:
::,
u. 0.3
OJ)
e:
:.ag
-~
V)
0.1
2 3 4 5 6 7 8 9 10
Time
where L º is any negative scalar which will appear as one of the eigenvalues of (5.59).
In the simulation which follows Lº = - 2.5 and A(A 11 - A 122 1C) = {-1 , -1.5}.
Since the system has an invariant zero at -2, the sliding motion will have poles at
{- 1, -1.5, - 2}. T he pole represented by 4> which governs the range space dynamics
has been chosen to be - 5. For simplicity the scaling factor for the sliding surface
is F2 = l. All the available degrees of freedom have now been assigned. Figure 5.3
is a plot of the switching function against time; it can be seen that sliding occurs
after approximately 1 second. Figure 5.4 shows the evolution of the states against
time. Initially the states of the compensator have been set to zero. The states of
the system have a nonzero initial condition which needs to be regulated to zero.
o .. ·· :-
"'
Q)
lo
cií -1
E
.':l
"'
>-,
V)
-3
2 3 4 5 6 7 8 9
Time
Figures 5.5 and 5.6 show the evolution of the error states ec and er , Initially ec
is nonzero since the state x 11 was given a nonzero initial condition. As indicated
in equation (5.88), this error system is completely decoupled and decays away to
zero at a rate governed by the invariant zero (Figure 5.5). The error states er,
shown in Figure 5.6, although initially zero, ar e coupled to the state ec as shown
in equation (5.87) . However, this too decays asymptotically to zero in accordance
with the theory. Notice from Figure 5.3 that, although the states initially lie on
the sliding surface, a sliding motion is not maintained; this is due to the fact that
the error term e is initially too large. A sliding motion occurs after approximately
1 second, by which time the error e has decayed sufficiently.
118 SLIDING MODE CONTROL
o
u' -0.2
~
"' -0 .4
2:l
~
CI'.)
....
s
-1
o 2 3 4 5 6 7 8 9 LO
Time
1-0.os
"'
<)
~ -0.1
CI'.)
....
o
J3 -0.15
2 3 4 5 6 7 8 9 10
Time
Consider the inverted pendulum with a cart shown in Figure 5.7. Assume that the
pendulum rotates in the vertical plane and the cart is to be manipulatecl so that
t he pendulum remains in an upright position. The cart is linked by a transmission
belt to a drive wheel which is driven by a DC motor.
- X
where the values of the physical parameters used are given in Table 5.1.
A linearisation of the system in (5.108) and (5.109) has been made about the
equilibrium point at the origin. Using x, 0, x and 0 as system states, and assuming
that only 0, x and x are available as measured outputs, results in the triple
1
[ oo oo o
A
O -1.9333
O 36.9771
-1.9872
6.2589 -0.1738
OOtl j B = [ 03t5]
-1.0095
o
~l
1 o
e
[o o 1
o
o
1
(5.110)
1t can be easily verified that the Markov parameter CE is foil rank and in the
canonical forro of Lemma 5.3
-o.1is2 O 30.8881
o o
-0.4572
1.0000
j B= o
O j (5.111)
A=
[
1.0000 o o 3.1496
[
-0.0091 O 1.9333 -2.0158 -0.3~05
An =
[ -0.1452
O
1.0000
o
o
30.8881
o
o
l
not possess invariant zeros. As a consequence, in this example,
_ º A122 =[
-0.4572
1.0000
3.1496
l
The theory requires a state feedback gain K. to be chosen so that A.11 - A 122K.
is stable. In the absence of any initial mismatch between the compensator states
and the states of the system, ec(t) == O for ali time, and the recluced-order sliding
motion will be governed by t his system matrix. If the canonical forro in (5.111)
is viewed as regular form then the quadratic minimisation method described in
Section 4.2.2 can be used to optimally select K.
120 SLIDING MODE CONTROL
If, in the notation of Section 4.2.2, the positive definite matrix Q in the cost
function is chosen to be diag(lO, 1, 1, 0.1), then the matrices Q = diag(0.052, 10, 1)
and Q 22 = 1.9920 can be identified. Solving the Riccati equation yields
This gives a reduced-order motion wit h potes {- 4.3241 ± 1.7852j, - 3.1623}. The
range space dynamic matrix cJ! has been chosen as - 6 and the compensator gain
matrix Lº = -9.8548, which results in H = - 10 and
D= [O -67.6603 31.4960 ]
0.5
e
o
·.;¡ 0.4
<.)
e
"
l,l. 0.3
bJ)
e
:i:: 0.2
.B
-~ 0.1
"'
2 2 .5 3 3.5 4 4.5 5
Time, sec
0.4 - - - -- - - - - - -- - ~ - - - -- - - - - -- ~
"'
'5
&
::,
o
Figure 5.8 is a plot of the switching function against time and shows that within
0.25 second a sliding motion is est ablished. It a!so demonstrates the advantage
of setting the initial conditions of the compensator appropriately since a sliding
USING ou·rPUT INFOR.MATION 121
motion is sustained at the first instant when s(t) = O; this should be compared
with Figure 5.3 in the previous example. Figure 5.9 plots the evolution of the
system outputs. The dashed line represents the angular position 0 and shows that
the settling time is approximately 2 seconds.
The following results are obtained from using the controller on the nonlinear equa-
tions of motion. In this situation both matched and unmatched nonlinearities are
present and so the theory described earlier is not technically valid. In practice,
however, a good response is obtained. The following simulations use the same ini-
tial conditions as Figures 5.8 and 5.9.
0.5
g 0.4
'Be
:::l
¡,,. 0.3
bJl
e:
:2 0.2
B
'§: 0.1
Cll
Figure 5.10 is a plot of t he switching function against time and shows that within
0.25 second a sliding motion is established. The output responses shown in Fig-
ure 5.11 are identical to those of Figure 5.9. Again the dashed line represents
angular position.
0.4
.':!l
0.2 >
::,
B-
::,
o o
-0.2
o 0.5 1.5 2 2.5 3 3.5 4 4.5 5
Time, sec
5 . 7 .1 Aircraft Example
Consider the following fifth-order linear system representing the lateral axis model
of an L-1011 fixed-wing aircraft with the actuator dynamics removed. The state
vector is represented by
e/> bank angle (rad)
r yaw rate (rad/sec)
X p roll rate (rad/sec)
(3 sideslip angle (racl)
X5 washed-out filter state
USING OUTPUT I NFORMATION 123
r
0.03086 -0.9960 - 0.00003 -0.1170
0.5000 o -O L]
o o o o
l
1
-0.7440 -0.0320
e o o 1
B = 0.3370 - 1.1200
o o o
0.0200 o 1 o o
o o
The inputs are
= [ t]
rudder deflection (rad)
u aileron deflection (rad)
and the outputs
An ideal model for this system has been developed using the eigenstructure assign-
ment method described in Section 4.5.1. The state gain matrix
-0.3131 3.3211 -0.1386 - 0.7379 4.1180 ]
(5.119)
Lx = [ 3.9524 5.6616 2.2906 -65.6425 -90.7262
provides an ideal model system matrix Arn = A+ BLx with eigenvalues at
{- 0.05, -2 ± l.5j, - 1.5 ± 1.5j}
This model essentially decouples the rolling and yawing motions of the aircraft.
Clearly, since the system has only two inputs, ali four out puts cannot be controlled
independently. For this reason a matrix of controlled outputs
10000] (5.120)
Cm= [ O O O 1 O
has been introduced which identifies bank angle <fa and sideslip angle /3 as the
outputs of interest. (For control purposes the four signals in y will be fed back.)
The feed-forward matrix Lr has been calculated as in Section 4.5.1 to ensure that
t he steady-state gain from the reference signal to the controlled outputs via the
triple (Am,BLr,Cm) is unity.
It can easily be checked that the Markov parameter CB is of full rank and a
realisation of (A, B , C) in the canonical form of Lemma 5.3 yields
-0.0133 -0.0007 -0.0172 0.2838 0.6266
o o - 0.0008 -0.9110 0.4125
[ Au
A21
A12
A22 ] 0.7071
0.1227
0.0386 -0.0858
0.0004
0.4167
4.1018 - 0.8028
0.9210
0.4471
(5.121)
l
Since r = O it follows that Á11 = A11 and A122 = A12 , hence
- 0.0133 - 0.0007 -0.0172
o O -0.0008
0.7071 0.0386 -0.0858
and
1221
[ A 1222
A
] =[
0.2838 0.6266
- 0.9110 0.4125
0.4167 0.9210
l
Here the gain matrix Lº has been chosen so that H = - 5, which results in
D = [ 5.5198 0.0243 - 34.8353 - 0.2729 ]
Because in this example Á11 = A n and A122 = .412 , the state feedback gain lC
can be chosen by applying the quadratic minimisation method of Section 4.2.2 as
detailed in Section 5.6.3. In the notation of Section 4.2.2, the symmetric positive
definite matrix Q in the cost function has been chosen to be diag(5, 1, 1, 5, 5), which
results in a switching function matrix
and sliding mode poles at {- 0.4470, - 2.5061 , - 2.2372}. The range space dynamic
matrix tI! has been chosen as diag(- 5, -5) and in the simulation which follows
p = 1.
0.3
0.2
Zl
::,
e-::, 0. 1
o
o
-0.1
o 20 40 60 80 100 120 140
Time, sec
Figure 5.12 shows the response of the closed-loop system to a step change on bank
angle. The bank angle plant response is identical to that of the ideal model. The
plant exhibits an in it ial sideslip transient before at taining the perfect decoupling
of the ideal model. In order to make the error system apparent, the initial state
of the compensator has been chosen to be nonzero. Figure 5.13 shows the first 10
seconds of the simulation so that the plant and ideal model discrepancies can be
observed. Figure 5.14 shows that a sliding motion in the error system occurs after
approximately 1 second.
USING OUTP UT INFORMATION 125
0.06
~ 0.04
8
ul
"
B-
::,
o
-0.02
o 2 3 4 5 6 7 8 9 10
Time, sec
Figure 5.13: Evolution of the errors between plant and ideal model outputs ey
0. 1
e
.s
ü
e 0.05
&:
00
=
:E
B o ._,
·~
(/)
-0.05
o 2 3 4 5 6 7 8 9 10
Time, sec
5.8 SUMMARY
Design procedures have been presented to synthesise robust output feedback con-
trollers for uncertain systems. The class of systems to which the results apply
has been identified, and includes the requirement that the nominal linear system
is minimum phase. Emphasis has been placed on the tractability of the associ-
ated design procedure. Two design methods have been employed for the design
of the sliding surface: one utilises established static output feedback pole place-
ment results, the other a reduced-order Luenberger observer approach. It has been
demonstrated that ali the assumptions imposed on the system pertain to the design
of the sliding surface. The proposed controller which guarantees attainment of a
sliding mode despite the presence of uncertainty, requires no additional assump-
tions. The last section demonstrated how the previous output feedback results can
be viewed within the context of a model-reference approach.
For linear systems with no uncertainty the output feedback problem has been inves-
tigated by White (1990) and by El-Khazali & DeCarlo (1991,1992). For uncertain
linear systems an approach has recently been reported by Hui & Zak (1993) and
126 SLIDING MODE CONTROL
Zak & Hui (1993), who propase an algorithm for output depcndent hyperplano
design, which is based upon eigenvector mcthods. The conceptual approach of
iak & Hui (1993) is similar to that of El-Khazali & DcCarlo in tha.t it relies on
establishing an appropriate eigenstructure for the rcduced-order sliding rnotion.
White (1990) considers thc ca.se when, in the notation of this chaptcr, FC B is rank
deficient. In this situation a.n ideal sliding mot,ion as given in Definition 3.1 ca.nnot
be guaranteed.
T he compensator-ba.sed a.pproach considered towards the end of the chapter is
similar to t he structure considered in El-Khazali & DeCarlo (1993). Thc framing
of t he problcm as a static output feedback pole placement problcm is described in
rktail in Bag et al. (1997) whern a so-called normal rnatrix approach is employed to
obtain the feedback gain. No single universally acknowledged algorithm for output
feedback pole placement has emerged, although a useful recent survey of the state
of the art is givcn in Syrrnos et al. (1997).
The quadratic stability approach used in the early sections relies on constructing
relative degree 1 mínimum phase systems. It. is well known that such systems are
st.abilisable via a linear static output feedback cont roller; see Gu (1990).
Work adopting a modcl-reference approach using only input/output information
has appcared within the context of adaptive control where the variable structnre
i<leas are utilised in thc paramcter adaptatíon mechanism. A survey of thcse meth-
ods appears in Hsu et al. (1994).
This chapter has not considerad a class of controllcrs found in the literature which
employ numerical methods to synthesise a linear static output feedback gain to
ensure that a reachabilily condition similar to t hat of equation (3.122) is satisficd.
For ::urther details see Heck et al. (1995) and Bag et al. (1997) .
The linear system in Example 1 of Section 5.4.3 is from Hui & Zak (1993).
The data for the invorted pendulum in Tablo 5.1 is from an experimental rig in the
Departmcnt of Electronic Engíneering at thc Univcrsity of Hull.
The lateral axis model of the L-1011 aircraft at cruise flight conditions in Section
5.7.1 appears in Sobe! & Shapiro (1986) . In Section 5.7.1, for control la.w design
purposcs, the states associated with the actuator dynamics ha.ve bcen removed.
Chapt er 6
6 .1 INTRODUCTION
Many of the theoretical developments in the area of sliding mode control systems
assume that the system state vector is available for use by the control scheme. In
order to exploit these control strategies, a suitable estímate of the state vector may
be constructed for use in the original control law. The idea of using a dynamical
system to generate estimates of the system states - an observer - can be traced
to Luenberger (1971), who proposed a method for linear systems which now bears
his name and which was described in Section 2.3.5. In this approach, t he observer
system is driven by the control input and by the difference between the output of
the observer and the output of the plant. The latter should ideally become zero.
This naturally suggests the exploration of ideas to generate a sliding mode on the
subspace for which the output error is zero. In particular, it is ofinterest to explore
the possibilities of using sliding mode observers for robust state reconstruction.
where A E ]Rnxn , BE lRnxm, CE lRpxn and p 2". m . Assume that the matrices
B and C are of full rank and the pair (A, C) is observable.
Te -- [ N¡]
e (6.2)
128 SLIDING MODE CONTROL
where the columns of Ne E lRnx(n-p) span the null space of C. This transformation
is nonsingular, and with respect to this new coordinate system, the new output
distribution matrix is
crc-l = [ o Ip ]
and
T = [ In- p L ] (6.10)
O Ip
and let é1 = e 1 + Ly. The error system with respect to the new coordinates can
be written as
¿1(t) Á11é1(t) + Á12ey(t) (6.11)
éy(t) = A21é1(t) + Á22ey(t) - v (6.12)
where Án = A11 + LA21, Á12 = A12 + LA22 - A11L and Á22 = A22 - A21L. It
follows from (6. 12) that in the domain
(6.13)
1 This is the dual result to Proposition 3.3.
SLIDING MODE OBSERVERS 129
is satisfied. Consequently, an ideal sliding motion will take place on the surface
(6.15)
lt follows that after sorne finite time ts, for all subsequent time, ey = Oand ey = O.
Equation (6.11) then reduces to
(6.16)
6.2.2 Example 1
where
C=[ l 1]
which represents a simple harmonic oscillator. For simplicity assume u = O and
consider the problem of designing a sliding mode observer. Define a nonsingular
matrix
Te= [ ~ ~] (6.19)
and change coordinates according to x f--+ Tcx to give the system triple
TA y-i
e e
=[ - 1
-3
1]
1
The system is now in the form of equations (6.3) and (6.4). An appropriate choice
of gain in the observer given in (6.5) is L = 0.5 which results in an error system
governed by A11 = - 2.5.
In the simulations which follow, the scaling constant M in the discontinuous compo-
nent in equation (6.7) has been set to unity. Figure 6.1 shows the state estimation
errors e 1 (t) and ey(t) resulting from the initial conditions e1 = - 1 and ey = O.
It can be seen that although the error system starts on the sliding surface S an 0 ,
ideal sliding motion cannot be maintained; only after approximately 1.2 seconds
is sliding established. At this point in the time interval 1.2 to 3.0 seconds, e1 (t)
exhibits a first-order exponentia.1 decay to the origin. After 3.0 seconds almost
perfect replication of the states takes place.
130 SLIDI NG MODE CONTROL
~
o
Jj 0.5
C
o
'i o
.§
"'
U:.l
O) -0.S
~
cií
Time, sec
In the original coordinates it ·c an be seen from Figure 6.2 that perfect tracking
occurs after approximately 3 seconds. T he dotted lines represent the true states
and the solid line the estimates from the observer.
g
·g
o> o
1-ll
~
""
cií -l
Time, sec
Finally Figure 6.3 shows the value of II with respect to time and shows switching
taking place from 1.2 seconds onwards.
E
"eo
o.,
E 0.5
o
u
"'
;;;)
o
o
::,
.§
g -0.5
ú
8 -J
o 2 3 4 5 6 7 8 9 10
Time, sec
Consider the effect of adding a negative output error feedback term to equations
(6.5) and (6.6) of the Utkin observer. This results in a new error system
By selecting G 1 = A12 and G2 = A22 -Ah, where A~2 is any stable design matrix
of appropriate dimension, t hen
=
In this form , the error system is asymptotically stable for v O because the poles
of the combined system are given by .\(An) U .\(A~2 ) and so lie in the open left
half complex plane. In the original Ut kin observer, the switching action v was
potentia.lly required to make the error system stable. The addition of a Luenberger
type gain matrix, feeding back the output error, yields the potential to provide
robustness against certain classes of uncertainty by virtue of the discontinuous
component v.
R emark
It should be noted t hat thus far the only restriction imposed on the nominal linear
system is that the pair (A, C) is observable.
where the function (: lR+ x IR" xRm-+ ]Rm is unknown, but bounded, so that
where
if FCe =/c O
(6.30)
otherwise
and the scalar function p(-) is any function satisfying
for some positive scalar r¡. To prove that this observer guarantees quadratic stability
of the error system, consider V(e) = eT Pe as a candidate Lyapunov function. It
follows from (6.24) and (6.29) that the error system satisfies
and therefore
Using an argument similar to that in Section 5.3 it can be shown that there exists
a domain in which sliding motion is induced on the surface in t he state error space
given by
Swz = {e E IRn : FCe = O} (6.35)
This is a very appealing approach but relies on establishing whether there exists
a gain matrix G such that, for the resulting closed-loop matrix A 0 , there exists a
Lyapunov matrix P for A 0 satisfying (6.28) for sorne FE 1Rmxp _ This represents
a nontrivial problem. Walcott & Zak (1988) propose an algorithm for the design
of P which can be summarised as follows:
(3) Compute Q(Pp) symbolically in terms of the entries of Pp using the expres-
sion Q(PF) = -(PpAo + AJ Pp ).
From a computational point of view, the algebra associated with the manipulation
and solution of t he inequalities resulting from step 4 is impractical without the
use of a symbolic manipulation package for systems of order 4 and above. Also no
indication is given of the class of systems for which the algorithm will produce a
successful design.
Remarks
where A E lRnxn' B E lRnxm) e E ]RP Xn and D E lRnxq where p 2'. q. Assume that
the matrices B, C and D are full rank and the function ( : lR+ x JRn x JRm -+ IRq
is unknown but bounded so that
(6.38)
Suppose that there exists a linear change of coordinates T 0 so that the system can
be written as
i1 (t) = A11x1(t)+A12y(t) +B1u(t)
(6.39)
y(t) A21X1 (t) + A 22y(t) + B2u(t) + 'D2f.
where x 1 E JR,(n - p), y E JRP and the matrix A 11 has stable eigenvalues. Consider
an observer of the form
The lower block triangular structure has been shown to occur quite naturally as a
result of the state-space representation chosen and the output error feedback gains
employed.
Proposition 6.1 There exist a family o/ symmetric positive definite matrices A
such that the uncertain dynamical error system above is quadratically stable.
Proof
Let Q 1 E JR,(n - p) x (n- p) and Q2 E JRP XP be symmetric positive definite design
matrices and define A E JR.P XP to be the unique symmetric positive definite solution
to the Lyapunov equation
(6.45)
(6.46)
(6.48)
Substituting the identity (6.50) into equation (6.49) and writing for notational
convenience (ey - Q21 PzA21e1) as ey then
S0 = {e E lRn : Ce = O} (6.51)
(6.52)
and hence in t he domain O = {(e1, ey) : IIA21eill < IIV2ll'Yo -17} where 1J is a
small positive scalar, it follows that
(6.53)
From Proposition 6.1 the output error ey enters nin finite t ime and remains there.
Integrating (6.53) and arguing as in Section 3.6.1 it follows that an ideal sliding
motion takes place on S0 in finite time.
•
If x represents the state estimate for x and e = x - x then the robust observer can
conveniently be written as
and
if Ce =/:- O
(6.57)
otherwise
Even in the special case when D = B the observer formulation (6.54) to (6.57) is
different from that of Walcott & iak (1987) for the case when p > m since their
results guarantee sliding will take place on the surface in the error space given
by {e E lR" : FCe = O}. This does not imply Ce = O since the null space of
F is nonempty and t herefore the observer <loes not necessarily track the system
outputs perfectly. In the above formulation this is guaranteed. The usefulness of
P roposition 6.1 will depend on being able to identify the class of systems which
can be placed in the canonical form of equation (6.39). This will be addressed in
the following section.
Let (A, D , C) represent the linear part of the uncertain system given in (6.37) which
represents the propagation of the uncertainty ( through to the output. Consider
the problem of constructing an observer for the uncertain systern of t he forrn
The presentation below is almost identical to that in Lemma 5.3 except a different
partition of the system matrix has been ernployed commensurate with the partition
of the outputs.
Lemma 6.1 Let the triple (A,D,C) representa linear system with p > q and
su.ppose rank(CD) = q. Then a change of coordinates exists so that the triple with
respect to the new coordinates (A, D, C) has the following structure:
A [ Au
(6.59)
and (6.60)
where Ar1 E IRrxr and A~l E IR,(p-q)x(n-p-r) far sorne r ~ Ü and the pair
(A~ 2 , A~1 ) is completely observable. Furthermore, the eigenvalues of AJ I are
the invariant zeros of (A,D,C) .
Jj = (6.61)
C = [O T ] (6.62)
Proof
Identical to Lemmas 5.3 and 5.5.
■
and Jj (6.63)
where D2 is defined as
(6.64)
The main result of this section (and the chapter) will now be proved.
138 SLID!NG MODE CONTROL
Proposition 6.2 A sliding mode observer of the form ( 6. 58) which rejects the
uncertainty class in ( 6. 37) exists if and only if the n ominal linear system satisfies
• rank(CD) =q
• any invariant zeros of (A , D, C) must lie in(;_ .
Proof
(proof of necessity)
Let G1 and Gn be appropriate gain matrices so that A 0 = A - G1C is stable, and
assume an ideal sliding mode insensitive to the uncertainty exists on the hyperplane
in t he error space given by S0 • The error system satisfies
(6.66)
To be insensitive to the uncertainty it follows that
or equivalently
(6.67)
Since by assumption rank(D) = q, it follows immediately from equation (6.67) that
rank(CD) = q. Therefore it can be assumed without loss of generality that the
system (A , D , C) is in the canonical form given in Lemma 6.1. If t he nonlinear gain
matrix is partitioned so that
Gn =[ g: ] ;;-p (6.68)
then CGn = TG2 and so det(G2) f:. O. From equation (6.66) and using arguments
similar to those presented in Section 3.4, it follows t hat the poles of the (linear)
reduced-order motion are given by
(6.69)
where (Ao)11 and (Ao)z1 represent t he top left and bottom left sub-blocks of the
closed-loop matrix Ao partitioned in a compatible way to the canonical form. By
definition the matrix Ao = A - G1C, so
where (G1C) 11 represents the top left sub-block ofthe square matrix G1C. However,
it is easy to check that (G1C)u = O for all G 1 E 1R.nxp and so (A0 ) 11 = A 11 .
Similarly it can be shown that (A0 )21 = A21 and consequently
(6.70)
3 T his
expression is borrowed from control terminology and is defined as the value of v required
to ma.intain an ideal sliding motion .
SLIDING MODE OBSERVERS 139
where GE rn.(n-p)x(p- q) and therefore from the definition of A21 it follows that
{zeros of (A, D , C)} = >.. (At1 ) C >.. (An - GA211 ) for ali GE rn.(n-p) x(p-q)
and therefore for a stable sliding motion any invariant zeros must lie in ((;_ .
(proof of sufficiency)
Conversely, let (A, D, C ) represent the system and suppose rank(CD ) = q and
any invariant zeros lie in {!_. Without loss of generality it can be assumed that
the system is already in the canonical form of Lemma 6.1 where the matrix Af 1
is stable. As a consequence there exists a matrix L E rn.(n-p)x(p- q) such that
Au + LA211 is stable. Define a nonsingular transformation as
T
L
= [ In-O p TL ] (6. 71)
where
L= [L Ü(n-p)xq ]
After changing coordinates with respect to 1'¡,, the new output distribution matrix
becomes
e = cr¡ 1 = [ o Ip ]
1
Finally, if A = T¡,A Ti , it can be shown by direct evaluation that
Au = Au +LA211
which is stable by choice of L. The system triple (A, V , C) is now in the canonical
form (6.39) and from Proposit ion 6.1 a robust observer exists. ■
This canonical form of Lemma 6.1 can be obtained by using the mfile commands
described below.
140 SLIDING MODE CONTROL
'l. Canonical form fo r the obser ver in section 6.3 .1 . If the dimension
'l. of the observable subspace is non zero, plisa vector of dimension
'l. nn-pp-r which specifies the desired stable poles of the subsystem
'l. where r is the number of invariant zeros of (A,D ,C).
A1 1=Aa(1:nn- pp,1:nn-pp) ;
A211=Aa(nn-pp+1:nn-qq,1:nn-pp);
'l. The aim is to put (A11,A21 1) in the obser vability canonical form
[Ab,Bb ,Cb,Tobs,k]=obsvf(A11,zeros(nn-pp,1),A211,1000*eps);
Af=Ta*Aa*inv(Ta);
Df=Ta*Da;
Cf=Ca*inv(Ta);
A22o=Af(r+1:nn- pp , r+1:nn-pp);
A21o=Af(nn- pp+1:nn- qq,r+1:nn-pp);
Ltemp=place ( A22o', A21o',p1)';
L=-inv(Tobs ) • [zeros(pp-qq,1:r) ,Ltemp];
else
L=zeros(nn- pp,pp- qq);
end
Lbar=[L zeros (nn-pp ,qq) ] ;
Build the final tranformation needed to obt ain the canonical f orro
TL=[eye(nn-pp) Lbar z eros(pp,nn- pp) T];
Remarks
T he gain matrices associated with the observer are generated by the following
commands.
142 SLIDING MODE CONTROL
1/. The cornmands below calcul ate the gain matrices which define the
1/. observer. The vector p2 is of dimension pp and specifies the
1/. desired stable poles of the error system.
Consider the uncertain linear system (6.24) given earlier in Section 6.2.4. It will be
shown that if a Walcott- Zak observer represented by t he pair (G, P) exists for a
given system and, in particular if the Lyapunov matrix (6.28) is known, then this
observer can be analysed within the framework developed in P roposition 6.1. It is
necessary to introduce first a lemma which considers t he effect of a change of basis
on the structura.1 constraint.
Lemma 6 .2 Let the system ( A 0 , B, C) be given with A 0 stable, and let (Á0 , .B, C)
be related to (A 0 , B, C) by a nonsingular similarity transformation T. lf P is a
=
Lyapunov matrix for A 0 which satisfies the structural constraint cT pT P B then
the matrix P = (r - 1 ? pr- 1 is a L yapunov matrix for Á 0 which satisfies the
structural constraint 6T pT = PB.
Proof
Since P is a Lyapunov matrix for Ao
(6.73)
Since Í' and (T- 1 )'r Qr- 1 are symmetric positive definite, equation (6.73) implies
P is a Lyapunov matrix for Á 0 . It can be verified by direct evaluation that the
structural constraint 6T pT = PB is satisfied.
■
Proposition 6.3 Let (A, B, C) be a system for which there exists a Walcott-iak
observer characterised by the matrix pair (P, F). Then there exists a nonsing·n-
lar similarity transformation so that the triple with respect to the new coordinates
(A, E, C) and the observer pair (P, F) exhibits the following properties:
- _ [ A11
A- A-
21
Proof
Let (P, F) represent the nonlinear components of the observer and let G be an
appropriate (Luenberger) gain matrix so that A 0 = A - GG is stable. If the
output distribution matrix is not in the form C = [O Ip], apply transformation
(6.2) in Section 6.2 and use Lemma 6.2 to calculate the observer pair in these
new coordinates. For convenience assume this is already the case. Partition the
Lyapunov matrix as
p =
where P11 E lR(n-p)x(n- r), Pi2 E 1R(n - p) x p and P22 E JRPxr _ Because Pis sym-
metric positive definite, the sub-block P 11 is also symmetric positive definite and in
particular is nonsingular. Change coordinates with respect to the transformation
Ín - p
T =
o
In the new coordinate system
144 SLIDING MODE CONTROL
p = O ] [ P11
Ip P(,¿ ] [ In-po
= (Pn - P&P111 A2) ]
for sorne Q which is symmetric positive definite. If (.40 ) 11 and Q11 are the top left
sub-blocks of A0 and Q respectively then from equation (6.74) it follows that
(6.75)
where both P 11 and Qu are symmetric positive definite. Equation (6. 75) represents
a Lyapunov equation for (.A0 ) 11 which is therefore stable. However, beca.use of
the structure of C, if A is partitioned as in the proposition statement, it follows
immediately that Au = (Ao) 11 and the first property has been demonstrated.
■
The above proposition has shown that given a system (A, B, C) for which there
exists a Walcott-Zak observer then there exists a nonsingular similarity transfor-
mation so that in these coordinates the system triple (A, E, C) can be written
as
= A11x1 (t) + A12v(t) (6.76)
= A21X1 (t) + iby(t) + B2 u(t) + B2~
where the matrix A11 is stable. This is identical to the uncertain system given
in equation (6.39) and is thus appropriate for the design and analysis method
presented in Proposition 6.1. This suggests an explicit method for designing a
Walcott- Zak observer:
SLIDING MODE OBSERVERS 145
(2) Establish the canonical form of Section 6.3. 1 using the coordinate transfor-
mation (6.71) in the proof of Proposition 6.3.
(3) The linear gain G can be used from the observer design in Section 6.3.1 and
the matrix
(6.77)
lt can then be verified that the block diagonal Lyapunov matrix associated
with the quadratic form in equation (6.48) satisfies the constraint (6.28).
MATLAB commands to carry out the observer design process are given below under
the assumption that D = B.
mfile: computes the gains for a Walcott- Zak observer
'l. The commands below designa Walcott-Zak observer for the triple
'l. (A,B,C) in the canonical form of Lemma 5.6. The vector p2 is of
'l. dimension pp and specifies stable design poles for the error system.
'l. Compute F and the linear gain matrix Gin the original coordinates
P2=lyap(a22s',eye(pp));
G=inv (Te) *inv(Tb) *inv(TL) * [a12; (a22-a22s)] ;
F=D2'*P2;
±1 X2
±2 = - sin(x 1 ) (6.78)
Therefore
A = [~ ~ ] and B = [~ ]
and the matched 'uncertain' bounded function ((t, x 1 , x 2 ) = - sin(x 1 ). By design
the output distribution matrix C = [ 1 1 ] .
Consider the problem of finding a robust observer. Change coordinates with respect
146 SLIDING MODE C ONTROL
to
Te [ ~ ~]
so that the output is a state variable. The system triple becomes
Á = TAT-
e e
1
=[-l - 1
1
1 ]
and
and C
which is in the canonical form (6.39). A robust observer exists for this system
because Á.11 = - 1, which is stable. Letting the design matrix A~2 = - 1 results in
>-(A0 ) = {-1, -1}. Defining Q2 = 2 and solving the Lyapunov equation for A 22
and Q2 gives P2 = l. In the original coordinates the gain matrices become
Figure 6.4 demonstrates the nonlinear observer tracking the output from the pen-
dulum when the initial conditions of the true states and observer states are delib-
erately set to different values
2 4 6 8 lO 12 14 16 18 20
Time, sec
Figure 6.4: Comparison between outputs from the observer and system
It can be seen from Figure 6.4 that after 0.5 second a 'sliding motion' takes place.
A comparison of the true and estirnat ed states is given in Figure 6.5. After ap-
proximately 4 seconds, visually perfect replication of the states is taking place. If
the nonlinear component is removed by setting p to zero, the resulting Luenberger
Observer behaves as shown in Figure 6.6. There appears to be a distinct phase
discrepancy between the outputs of the system and the outputs of the observer;
this is due to the presence of the nonlinear sine term.
SLIDING MODE OBSERVERS 147
23
;:,
Ul
~
Vl
·=
[J..
-1
2 4 6 8 10 12 14 16 18 20
Time, sec
B
'"
(/)
"O
e o
o
¡5
Ul
-1
2 4 6 8 10 12 14 16 18 20
Time, sec
2 '
/
=Í
o.
=Í o . ·/ ··
o
....,r.-;-,.
-2
-4
o 2 4 6 8 10 12 14 16 18 20
Time, sec
Fig ure 6.6: Comparison between outputs from t he observer and the system
The fundamental purpose of a fault det ection and isolation (FDI) scheme is to
gen er ate an alarm whcn an abnormal condition, such as a component malfunction
or variation in operating condition, develops in the process being monitored and
to identify the source or location. An extensively studied FDI methodology is the
observer-based approach (Patton et al., 1989) which analyses residuals formed from
the difference between the actual outputs and the outputs from an observer. In
the absence of faults, the residuals a re designed to be small: once a fault occurs,
148 SLIDING MODE CONTROL
residuals are intended to react by becoming larger (or greater than sorne predefined
threshold), thus signifying the presence of a fault. These schemes usually employ
full-order Luenberger observers in which the gain matrix is designed either to malee
certain residuals sensitive to certain faults and not others, or else to make t he
residual vector lie in a specific direction in response to a particular fault, thus
enabling the source to be identified. This section considers the use of the sliding
mode observer defined in Section 6.3 to reconstruct the fault rather than analyse
residuals.
Consider a nominal linear system subject to certain faults described by
where A E lRnxn, E E lRnxm, CE lR¡,xn, D E lR.n xq with q ::; p < n and the
matrices B, C and D are of full rank. The functions f;(t) and f 0 (t) are deemed to
represent actuator and sensor faults respectively. It is assumed that the states of
the system are unknown and only the signa.Is u(t) and y(t) are available.
The objective is to synthesise an observer to generate a state estímate x(t) such
that a sliding mode is established in which the output error
G n == y o- 1 [ Ip
O ] (6.83)
and
if ey :f. O
(6.84)
otherwise
Notice how this represents a slight modification to (6.56) and (6.57). It will be
shown that, provided a sliding motion can be attained, estimates of fi(t) and f 0 (t)
can be computed from approximating the equivalent control.
(6.85)
SLIDING MODE OBSERVERS 149
where V eq is the equivalent control. From (6.43) it follows that e 1(t) -+ O and
therefore
(6.86)
One way to recover the equivalent control is by t he use of a low pass filter as de-
scribed in Section 1.2. Here an alternative approach will be employed: suppose that
the discontinuous component in (6.41) is replaced by t he continuous approximation
vs --
2
P e"
- p JJD2 11 II P2e,ll+ó (6.87)
(6.88)
The key point is that the signal on t he right-hand side of t he equation above can
be computed on-line and depends only on the output estimation error ey, thus the
fault Íi can be approximated to any degree of accuracy.
Now consider the case when Íi = Oand consider the effect of f 0 (t).
In t his situation since y(t) = Cx(t) + f (t) it follows that
0
After sorne manipulation, it follows that in the coordinates of the canonical form
iu SecUon ü.3 .1 the state estimation error is given by
Note that f 0 (t) and j 0 (t ) appear as output disturbances and thus p in equation
(6.84) must be chosen to be sufficiently large to maintain sliding in the presence of
these disturbances. Arguing as before, provided a sliding motion can be attained,
(6.92)
Thus, for slowly varying faults, provided the dynamics of the sliding motion are
sufficiently fast,
(6.93)
As in the previous section, the equivalent control Veq can be calculated from (6.87)
and consequent ly, if (A 22 - A21A1/ A12) is nonsingular, t he fault signa! can be
obt ained from equation (6.93).
Note that from the Schur expansion
Remar ks
Note that even if (A22 - A21A1} A 12) is singular, it may still be possible to re-
construct sorne of the sensor faults f 0, depending on the structure of the rank
deficiency. An example of this will be considered in the next section.
The fault detection approach adopted here is based on the equivalent control con-
cept and thus <loes not interact in any way with the use of the sliding mode observer
as a state estimator. A set of states compatible with the current plant output will
always be produced; this is not the case with those fault detection methods which
rely on breaking the sliding motion in order to detect faults.
Consider the inverted pendulurn introduced in Section 5.6.3. Recall that a lin-
earisation of the system about the equilibrium point at the origin generated the
triple
o
[i
1
o o
A = - 1.9333 -1.9872
1
0.0091 o1 B
[ -1.0095
03t5 1
36.9771 6.2589 -0.1738
o o
~l
1
e =
[ o 1 o
o o 1
(6.95)
In the analysis which follows, the control law developed in Section 5.6.3 will be
used to control the system.
In this particular situation any actuator faults will occur in the input channel,
hence in the notation of Section 6.3 the fault distribution matrix D = B. It can
readily be established that the conditions of Proposition 6.2 are satisfied and thus
all the preceding analysis is valid.
In the canonical form of (6.39) the system described in (6.95) becomes
¡-10.ogoo o
o
-67.6603
O
31.4960
1.0000
1
=
A
1.0000
0.0091
o
o
9.8548 - 3.1496
- 1.8437 - 2.0158
D
[OL1
e =
[io 1
o o
1
n
In this particular case Ah= diag(-11,-12,-13)
which furnishes the linear component of the observer with poles approximately
(6.96)
three times faster than the closed-loop poles of the controlled plant. The symmetric
positive definite matrix P2 has been selected as the unique solution to the Lyapunov
equation
(6.97)
In this particular design the scalar function p = 75 and the observer design is
complete.
SLIDING MóDE OBSERVERS 151
It can be verified that the eigenvalues of A are {O, 5.8702, -6.3965, - 1.6347}, so
from the argument in Section 6.5.2 the steady-state gain from f 0 to Veq is singular.
In fact
which is clearly rank deficient. However, if Veq,i and fo,i represent the ith com-
ponents of the vectors Veq and f 0, from equation (6.93) and using the particular
A22 - A21A1/ A12 from (6.98) it is apparent that
Veq,1 ~ fo,3 (6.99)
Veq,2 ~ 3.0888fo,2 (6.100)
It is also clear that any fault in the first output channel has no direct long-term
effect on Veq. Furthermore, because of the structure of D 2 in (6.96) it can be verified
that
0.l ~ - -º~e_te_ct_or_S~i~n~al_l_ _~
0.8
O.OS
0.6
0.4
0.2
o~--~-- ~- - ~
o 5 10 15 5 10 15
Time Time
0.l ~ - ~ Dc..c¡eFte~ct~or~S=i=n,al~2~- - ,
0.05
01--- - -- - -- ----j
-0.05
-0.1 ~ - - ~ -- ~ - - ~
o 5 10 15 15
Time Time
Figure 6.7: Response of the detection signals to a fault in the input channel
152 SLIDING MODE CONTROL
0.08
0.05
0.06
o
0.04
-0.05
0.02
o -0.l
o 5 10 15 o 15
Time Time
0.1
0.05
o
-0.05 ·
' -0 .1
() 15 o 5 JO 15
Time Time
Figur e 6.8: Response of detection signals to a fault io thc first output channel
0.1
Detector Si na! 1
0.08
0.05
0.06 -
o
0 .04
-0.05
0.02
0.05
0.5
() t-------<·
-0.05
-0 .l
o 5 10 15 o 5 !O 15
Time Time
Figure 6.9: Response of detectioo signals to a fault in the second output channel
SLIDING MODE OBSERVERS 153
0.1 ~----=r='-'--'-'=+----~
0.08
0.02
5 IO 15 5 10 15
Time Time
Detector Si nal 3
0.1
0.05
0.5
o
-0.05
o
-0.1
o 5 10 15 o 5 10 15
Time Time
Figure 6.10: Response of detection signals to a fault in the third output channel
The following nonlinear simulation results show the response of these three detector
signals to different fault conditions. Figure 6. 7 shows the effect of a ramp in the
input channel. As predicted by the theory, the third detector signa! reproduces the
fault signa! whilst not affecting the other two signals. Figure 6.8 shows the effect
of a ramp in the first output channel. As predicted by the theory, the detector
signals do not reproduce this fault signa! (although the second detector signa!
approximates the gradient of the fault signa!). Figures 6.9 and 6.10 show that the
appropriate detector signa! reproduces the ramp fault signals in output channels 2
and 3. In both cases the detector signa! 3 is inftuenced as well.
Remark
This example highlights the difference between the Walcott- Zak approach and the
observer described in Section 6.3. In the Walcott-Zak approach, sliding takes place
on the surface in the error space given by {e E IRn : FCe = O} , so in t his particular
example the equivalent control would have only one component, making it difficult
to distinguish between faults in different channels.
6.6 SUMMARY
Different slidin g mode strategies have been employed for the problcm of state re-
construction using only input and output information. A framework has been in-
troduced to incorporate the effects of bounded, matched or indeed other structured
system uncertainty. This effectively considers the problem first posed by Walcott
and iak. Here, however, the emphasis is placed upon the numerical tractability
of the solution method. A different nonlinear structure is used which guarantees
154 SLIDING MODE CONTROL
sliding on the surface in the error space described by the null space of the input
distribution matrix. Also the precise class of systems to which these results apply
has explicitly been identified. In the case of matched uncertainty, the class of ap-
plicable systems is identical to those for which the direct output feedback designs
described in Chapter 5 were applicable. In practice such an observer may forro part
of a control strategy where the estimated states would be used in a state feedback
control law. The next chapter examines the overall closed-loop performance of the
observer when used in conjunction with a robust output-tracking state feedback
control law.
Despite fruitful research a.nd development activity in the area of variable structure
control theory, relatively few authors have considered the application of the un-
derlying principies to the problem of observer designs. The earliest work in this
field, which was the approach described in Section 6.2.1, appears originally in Utkin
(1992).
The approach described in Section 6.2.3 is conceptually similar to that proposed
by Slotine et al. (1987) in that the output errors are fed back in both a linear
and discontinuous way. The problem for mulation, however, is quite different since
Slotine et al. propase an observer for the system x = f (t, x) and seek only to extend
the region in which sliding takes place, the so-called sliding patch, by incorporating
a linear term.
The work of WaJcott & Zak (1987) seeks global error convergence for a class of
uncertain systems. This strategy, although intuitively appealing, necessitates the
use of algebraic manipulation tools to effectively solve an associated constrained
Lyapunov problem for systems of reasonable order. In addition to the original
papers (Walcott & Zak, 1987,1988) the approach is discussed in detail in Zak &
Walcott (1990) . This collection also describes a hyperstability approach to observer
design by BaJestrino & Innocenti (1990), based on the concept of positive realness.
The pendulum in Section 6.4.1 is the example used by Walcott et al. (1987) to
demonstrate that t heir sliding mode observer compares favourably with direct ap-
proaches to nonlinear observer designs when the nonlinearities present in the sys-
tems are assumed to be perfectly known.
Sreedhar et al. (1993) considera sliding mode observer approach to fault detection,
although in their design procedure it is assumed that the states of the system are
available. A different approach is adopted by Hermans & Zarrop (1996), who
attempt to design an observer of the form used in Section 6.3 in such a way that
the sliding motion is destroyed in the presence of a fault .
Chapter 7
7.1 INTRODUCTION
where x E JRn, 'U E IRm and y E ffi.P with m :S: p < n . Assume that thc nominal
linear system (A, B, C) is known and t h at the matrices B and C are both o:· full
rank. The function ( : IR+ x nr x ]Rm ---+ 1an is unknown and represents any
nonlinearities plus any model uncertaintics in the system and is assumed to satisfy
for sorne known positive scalar k1 < 1 and known function a: : lR+ x JRP ---+ lR+-
In addition, it is assumed that the nominal linear system satisfies
The assumption on the input/ output dimensions is required since the sliding mode
observer formulations of Chapter 6 require there to be at least as many outputs as
inputs.
It can be assumed without loss of generality that the system is already in the
regular form, so that
A= [ (7.3)
where Au E IR(n-m) x(n-m), B2 E IRmxm and the matrix C2 E IRpxm. The square
matrix B 2 is nonsingular because t he input distribution matrix is assumed to be
of full rank.
The sliding mode observer that will be considered is essentially that of Walcott &
Zak (1988) which was introduced in Section 6.2.4. If z(t) represents an estímate
for the true states and
e(t) = z(t) - x(t) (7.4)
is the state estimation error, then the proposed observer has the form
The output error feedback gain matrix G is chosen so that the closed-loop matrix
A0 = A - GC is stable and has a Lyapunov matrix P satisfying
for sorne positive definite design matrix Q and the structural constraint
(7.7)
if FCe f. O
(7.8)
otherwise
w here p0 (u¡, y) is a sea.lar function depending on the linear component of the control
law which, loosely speaking, is an upper bound on the norm of the uncertainty.
As shown in Section 6.4, assumptions A2 and A3 are necessary and sufficient con-
ditions for the existence of such an observer which induces a sliding motion on
Assume for simplicity that the system is square, i.e. p = m. This additional
assumption, although not essential, is convenient since the observer formulation
requires there to be at least as many outputs as inputs; conversely, the proposed
OBSERVER-BASED OUTPUT TRACKING CONTROLLERS 157
control law requires at least as many inputs as outputs - a square system is there-
fore essential. 1 In this situation the square matrix C2 is also nonsingular because
C2 B 2 = CB which is non.singular by assumption.
The assumption of squareness results in a further convenience - namely an explicit
solution for t he choice of gain matrices F and G can be obtained in the original co-
ordinates without the apparent need for the coordinate transformation introduced
in Section 6.3. This will be demonstrated in the following subsection.
The approach employed here uses the ideas developed in Chapter 6. For non-
square systems, the observer structure in equations (7.5) to (7.8) is different from
that described in Section 6.3. However, for square systems they are identical except
for a nonsingular scaling of the hyperplane. In the special case of a square plant
satisfying A2 and A3, an analytic expression will be obtained for the components
G and F. Assume the linear part of the uncertain system (7.1) is in the regular
form of (7.3). Since the sub-block C2 is nonsingular, the linear t ransformation
(7.10)
(7.11)
with A H = All - A1 2C2 1 C1 and whcrc thc input and output distribution matrices
are given by
C=[O (7 .12)
From Lemma 5.1 it follows that the eigenvalues of A11 are the invariant zeros of
(A, B, C) and consequently by assumption the matrix An is stable. The system
in (7.11) and (7.12) is thus in t he canonical form for sliding mode observer design
described in Section 6.3. Using the synthesis procedure described there, for a
given stable matrix Ah, the Lyapunov matrices A and A can be obtained from
equations (6.47) and (6.45). Because the transformation to obtain the canonical
form can be explicitly determined, it follows that the Lyapunov matrix P and the
gain matrix G can be determined explicitly. In the coordinates of the regular form,
via t he mapping in Lemma 6.2, the gain matrix
(7.13)
t he scaling matrix
(7.14)
1 Technically, for a nonsquare system, a subset of m outputs could be identified as the 'con-
trolled outputs'. In this way an alternative output distribution matrix Cm E !Rmx n can be
isolated to be used for the development of the control law; recall Section 5. 7. l.
158 SLIDING MODE CONTROL
and
p = r
[ C1I C2º r PiO 22 ) [ ~l
o
C2 ]
= [ Pi C[P2C1
+C¡PiC1 C¡P2C2
C!P2C2 ] (7.15)
Remark
It should be noted that from a design point of view, only the Lyapunov matrix A
is required for the observer synthesis.
A suitable mfile to accomplish an observer design is given below.
mfile: designs an observer explicitly for a square system
'l. Commands to produce an observer for the special case of square systems
'l. It is assumed t hat (A,B,C) is in regular form and the condit ions for
'l. the existence of a robust observer are satisfied.
'l. The pth order vector po is assumed to contain stable poles
'l. Q2 is a symmetric positive definite matrix of order p
C1=C(:,1:nn-pp);
C2=C(:,nn-pp+1:nn);
T=[eye(nn-pp) zeros(nn-pp,pp); C1 C2];
'l. Partition the system matrix to use in the observer gain matrix
A11=Astar( 1:nn-pp,1:nn-pp);
A21=Astar(nn-pp+1:nn,1:nn-pp);
A12=A(1:nn-pp,nn-pp+1:nn);
A22=A(nn- pp+1:nn,nn-pp+1:nn);
B2=Bstar(nn-pp+1:nn,:);
Consider initially the development of a tracking control law for the nominal linear
system
:i:(t) = Ax(t) + Bu(t) (7.16)
where the matrix pair (A, B) is assumed to be in regular form as in (7.3) . The
control law that will be described here is similar to that of Section 4.4.2 except
for a modification which makes allowance for the fact that an observer is used to
generate state estimates. As in Section 4.4.2 consider the introduction of additional
states Xr E lRT satisfying
Xr(t) = r(t) - y(t) (7.17)
where the differentiable signa! r(t) satisfies
r(t) =f (r(t) - R) (7.18)
with r E JR,PXP a stable design matrix and R a constant demand vector. Augment
the states with the integral action states to obtain
(7.19)
(7.20)
where S E IRm x(n+p) and Sr E IR"xp _ Assume, as in Section 4.4.2, that the
hyperplane system matrix has the form
(7.22)
where M E IRmxn and the scaling matrix S 2 = AB:21 . The linear component of
the control law proposed in Section 4.4.2 is given by
(7.23)
where
L -A- 1 (SA.- <PS) (7.24)
1
Lr = -A- (<PSr+S2MB,.) (7.25)
L;- = A- 1 sr (7.26)
<I> is a stable design matrix and Br is defined in equation (4.103). The discontinuous
vector is given by
(7.28)
for sorne positive definite design matrix Q2 • The reason for the use of t he modified
Lyapunov equation (7.28) will be made clear in Sect.ion 7.3.3 in the proofofLemma
7. l. The positive scalar function which multiplies the unit vector component of the
controller is given by
(7.29)
,e
where 'Yo and are positive design scalars and A E lRm xm is a nonsingular diagonal
design matrix satisfying
(7.30)
A corresponding scalar function to premultiply the unit vector component in the
observer from equation (7.8) is
(7.32)
In this section the effect of using the state estimates in the control law wiU be ex-
plored. In particular, stability of the combined closed-loop system will be demon-
strated.
From equations (7.1), (7.5) and (7.17) the system representing the uncertain error
and observer dynamics is given by
where ey = Ce and u is the control action obtained from using the state estimates
z(t) instead of x(t) in equations (7.23) and (7.27). It is convenient to repartition
equations (7.34) and (7.35) to obtain coordinates
(7.36)
(7.37)
It can be verified that, in terms of (z1 ,z2 ), equations (7.33) to (7.35) become
i1(t) Á11z1(t) + Á12 (z2(t) - Srr(t)) + (Br + A12Sr) r(t) - G1 ey(t) (7.38)
.z2(t) 1> (z2(t) - Srr(t)) + A.í)c + Srf(t) - (S1G1 + S2G2)ey(t) + A110 (7.39)
(7.40)
and therefore from equation (7.18) it follows that the evolution of er(t) satisfies
(7.42)
(7.43)
is introduced, then
where
G2 ~ S1G1 + S2G2 (7.46)
Equations (7.44) and (7.45) may then be written more conveniently as
for appropriate choices of gain matrices G, Gr and X. Thus the closed-loop system
is determined by (7.47) together with the observer error system
which is stable, with eigenvalues given by >.(A 11 ) U >.(<I>). For the closed-loop
analysis which follows, a block diagonal Lyapunov matrix for Ac will be sought of
the form
P- -- [ Ao (7.49)
where the lower right matrix sub-block A is t he Lyapunov matrix for <I> used in
the unit vector component of the control law. Using the Lyapunov matrix P for
A 0 given in (7.6) define
~ [ po pQ ]
p e;- (7.50)
where the matrix f> will be chosen so that P0 is a Lyapunov mat rix for the overall
closed-loop system matrix
(7.51)
-CT(_;Tf> ]
(7.52)
f>Ac +A;f>
where Q is defined in t he Lyapunov equation (7.6). The expression on the right-
hand side of equation (7.52) is negative definite if and only if
(7.53)
where Q2 is any symmetric positive definite matrix such that Q2 > G2 Q 22 Gf; and
A is the unique symmetric positive definite solution to
(7.56)
Proof
Solving the Riccati equation given in (7.54) is equivalent2 to finding a block diag-
onal solution to the Lyapunov equation
(7.58)
where
(7.59)
From the definition of Ac it follows that
where the matrices A and A solve the pair of Lyapunov equations (7.55) and (7.56)
where Q1 and Qz are arbitrary symmetric positive definite matrices of appropriate
dimension. If
(7.60)
then it follows that
which by careful choice of Q1 and Q2 can be made positive definite. This statement
will be justified as follows. A necessary condition for Q to be positive definite is
that
(7.62)
Let Q2 be any positive definite matrix satisfying the matrix inequality above. Solv-
ing the Lyapunov equation (7.55) provides the matrix A. Consequently, each
element in the matrix Q given in equation (7.61) is specified except for Q1 . A nec-
essary and sufficient condition from (7.61) for Q to be positive definite is that Q1
satisfies (7.57). A can then be obtained from (7.56) anda block diagonal solution
to the Riccati equation (7.54) is given by
- [A o]
P= o A
2
•
Since the intention is to find a positive definite solution, the matrix P will be invertible.
164 SLJDJNG MODE CONTROL
Utilising the structural constraint (7.7) and the uncertainty structure given in equa-
tion (7.2) it follows t hat
Note the second equality uses the fact that Pe = IIAIIPo + 'Ye from (7.32). From
inequalities (7.65) and (7.66) it follows that
(7.67)
Also
since IIAIIPo = Pe - 'Ye · Using inequalities (7.67) and (7.68) in equation (7.64) it
follows that
Proof
It is sufficient to demonstrate that a sliding motion occurs on the surface in the
combined state space given by
2
S e= {(e,(1,(2) E IR. n+p : C e= O and (2 = O}
To this end, let r¡ be a small positive scalar such that r¡ < í'c and r¡ < 10 , and define
a domain
n,, ~{(e,(1,(2) E IR.2 n+p : ll(CB)-1CAoell < í'o - 7/ and IIG2Cell < 'Ye -17}
Frorn P roposition 7.1 it follows that in finite time the states (e,() enter n,, and
remain there. It will be shown that a sliding motion takes place in this domain.
Consider the quadratic form
Ve = eT CT P2Ce + T -
( 2 Pz(2 (7.70)
This is clearly positive definite with respect to the surface S e defined above. From
(7.48) it follows that
Ce(t) = CAoe(t) + C2B2 (v 0 (t) - ((t,x,u))
and from the definition of Fin equation (7.14) it follows that Pz = pT (CB)- 1 .
Therefore
PzCe(t) = pT (CB)- 1 CAoe(t) + pT (vo(t) - ((t, x, u))
and using inequalities (7.67) it can be verified that
eTcT P2Ce '.S IIFCellll (CB)- 1 CAoell - í'ollFCell
Taking derivatives of the quadratic form (7.70) along the trajectories, and using
equation (7.45) in conjunction with inequality (7.68) it follows that
respectively. This represents a sufficient condition for sliding on the surface Se, ■
Corollary 7.2 The output of the uncertain system y(t) tracks the constant demand
vector R asymptotically.
Proof
From Corollary 7.1 a sliding motion is achieved, which implies from equation (7.44)
that
where
er(t) = f er(t)
and both ..4 11 and r are stable. For this linear system, stability of ( 1 implies ( 1 ➔ O
which implies i 1 ➔ O. From the definition of z1 , the first p st ates are the integral
action states. Consequently, Xr ➔ O and it follows that y ➔ r asymptotically. ■
Remark
An intuitive argument supporting this analysis runs as follows. Consider equation
(7.33); although it is related to the uncertainty and hence to the estimated st.ates,
the arguments presented in Section 6.3.1 indicate that a sliding motion is attained
on S0 in finite time. It follows therefore that in finite time equation (7.35) becomes
i(t) = Az(t) + B u(t) - Bveq (7.71)
where lleq is the 'equivalent control' necessary to maintain the motion on S0 and is
a function of the uncertainty. The control law u can be thought of as 'controlling'
the dynamical system (7.71) given above. The term Bveq can be considered as
matched uncertainty and, by an appropriate choice of sliding mode control law, it
can therefore be rejected. The output y = Cz therefore behaves in an ideal way.
However, since sliding occurs on the hyperplane S0 , it follows that the output of the
uncertain system y is identically equal to y. Therefore the output of the uncertain
system behaves in an ideal nominal way.
'!'he observer described in Section 7.3.1 has only the stable matrix A22 and its
associated Lyapunov matrix A in the way of design freedom. As a result, this
section concentrates on the design of the nonlinear control law. The hyperplane
defined in Section 7.3.2 is given by
(7.72)
OBSERVER-BASED OUTPUT TRACK!NG CONTROLLERS 167
where r represents the reference signal, Sr E IRmxm is a free design parameter and
where M E IRmxn_ The square design matrix S2 has no effect on the dynamics
of the reduced-order motion and serves only as a scaling of t he hyperplane. The
sliding motion is given by
(7.73)
where the pair (Á11 , Áiz) is defined in (4.102) and Br is defined in equation (4.103).
It was shown in Lemma 4.1 that assumptions Al and A3 guarantee that (All, A. 12 )
is completely controllable and consequently an appropriate matrix M may be cho-
sen by any of the design techniques described in Chapter 4.
The choice of Sr may be viewed as affecting the values of the integral states since,
from equation (7.73), at steady state
One possibility is to choose a value of Sr so that, for the nominal system at steady-
state, the integral action states are zero. Equation (7.74) can be rewritten as
(7.75)
because A11 is stable by design and therefore invertible. The steady-state values
of the integral action states are given by
(7.76)
because of the special form of Br in (4.103). Define K s to be the static gain of the
square system (.An, Á1 2, BJ'), then provided Ks is nonsingular, choosing
(7.77)
det [ Áu
BTr
168 SLIOING MODE CONTROL
Substituting for Au and Á12 from equation (4.102) and for Br from (4.103) it
follows that
det [ Au
B; det
det
This is guaranteed by assumption A3 in Section 7.2 and so the gain 1{9 is nonsin-
gular.
■
'l.
Commands to make a unit vector controller incorporating integral action.
'l.
It assumes the tripl e (A , B,C) is already in regul ar forro
'l.
For simplicity robust pol e placement is used for the hyperplane design
'l.
(with suitable modification any of the methods in Chap 4 could be used)
% the nth order vector p1 specifies the reduced order motion peles
% the mth order vector p2 specifies the range space dynamics
[nn,mm]=size(B);
[pp,nn]=size(C);
'l. Computes the gains for the feedback and feed-forward components
L= -inv (S•bigB)*S*bigA + inv(S*bigB)•Phi*S;
Lr= - inv(S*bigB)•(Phi *Sr+S(:,1:pp) );
Lrdot=inv(S•bigB)•Sr;
From the design perspective, apart from the scalar functions which premultiply the
unit vector components, the only dependence between the controller and observer
is through the choice of the controller Lyapunov matrix A. This matrix defines
the unit vector component of the control law
if Sx =/= Srr
(7.78)
otherwise
and satisfies
(7. 79)
where <I> is the stable design matrix defining the linear range space dynamics. From
Lemma 7.1 the matrix Q2 is a symmetric positive definite design matrix which must
satisfy equation (7.62), namely
For a given observer design, the right-hand side of the above equation is completely
deterrnined and therefore appears to restrict the choice of A. However, this is not
so. Let Ch be any symmetric positive definite matrix and !et A be the unique
solution to the Lyapunov equation (7. 79). If A is any positive scalar then the
rna.trix pair (f>.¡, Q~) ~(½A, AQ2) satisfies
(7.80)
(7.81)
and consequently for any given (h, for a large enough value of >., the rnatrix P.¡ is an
appropriate choice for the unit vector control component ve, However, frorn (7.78)
it can be seen that 1,c is independent of the value of >.. As a result, the controller and
observer can be designed independently in the sense that any observer designed as
in Section 7.3.l and any control law of the form given in Section 7.3.2 will provide
closed-loop stability for the uncertain system, provided the (uncertainty dependent)
scalar functions multiplying the unit vector cornponents are chosen so that equation
(7.32) holds.
170 SLIDING MODE CONTROL
A == O
1
-0.0082
º1
- 0.1029
l
e == [ 0.0001 0.0022 0.0053 ) (7.82)
This linear model has been obtained by system identitication from a nonlinear
furnace simulation based on the zone method (Hottel & Sarofim, 1967). In this
approach the surfaces and volumes that comprise the furnace are divided into sub-
surfaces and subvolumes which are small enough to be considered isothermal. The
integro-differential equations governing radiation exchange become algebraic and
finite difference equations which can be solved numerically. The solution of these
equations is numerically intense and such models do not lend themselves easily to
the problem of designing control laws. As such the linear model given in (7.82)
will be used for this purpose; the inherent nonlinearities present will be treated
as bounded uncertainty and will be incorporated into the uncertainties associated
with the valve dynamics and any externa! disturbances. The states of the linear
system representation have no physical meaning and are therefore not available to
the control law; an observer-based approach is therefore necessary. The control law
will then be tested on a nonlinear simulation model.
Conveniently the realisation given in (7.82) is already in regular form. The invariant
zeros of this linear syRtP.m arn givP.n by {-0.3749, -0.0358} anc\ thP.reforP. ;iJl the
preceding theory is valid. The rest of the section looks at synthesising and analysing
sliding mode temperature controllers based on this nominal linear model.
Before discussing the design procedure adopted, it should be noted that because
of the structure of the realisation, any variation in the elements in the last row
of the system matrix (which are the coeflicients of the characteristic polynomial)
occurs in 'R(B), and so can be considered as matched uncertainty. T he design will
therefore be insensitive to changes in the poles of the system.
1
o
-0.0082
and
If the stable design matrix A~2 = -0.2, it follows immediately frorn equation (7.13)
t hat
G = [ A C1}
22 2
~f\ A•
2
2 22
] = [ 188.~498
18.3328
l
Because the system is single-input single-output there is no need to compute the
sliding surface matrix F. Ali that remains is to compute the nonlinear scalar gain
function given in (7.31). In this case, the approach that has been adopted is to !et
(7.83)
where uL(·) represents the linear component of the control action and the positive
scalars r 1 , r-2 and 'Yo are to be chosen empirically. This will be discussed in more
detail in a later subsection.
Because the system is alrea.dy in regular form, the augmented system frorn equation
( 4.102) can easily be identified as
- 0.0001 -0.0022 -0.00531
o 1 O
o O 1
- 0.0001 -0.0082 -0.1029
As predicted from the theory, the pair (A.11 , A. 12 ) is completely controllable. The
poles of the ideal sliding motion have been chosen to be {- 0.025, - 0.03 ± 0.025j},
w hich represent dynarnics marginally faster t han t he dominant pole of the open-
loop plant. The unique M such that A(Á11 - Á 12 M) = {-0.025, -0.03 ± 0.025j}
is given by
M [ -0.5372 0.0019 0.0822 ]
For single-input single-output systems, no additional design freedom is provided
by the scalar A. For simplicity it has been chosen so that S2 = l. The remaining
pole, associated with the range space dynamics, has been assigned the value -0.1
by selecting the design rna trix 1> = -O. l. Again, because of the single-input single-
output nature of the system, there is no need to calculate the Lyapunov matrix
A associated with the unit vector controller. The nonlinear component of the
controller is given by
(7.84)
where the gain function Pe(·) =p 0 (·) + 'Ye for sorne positive scalar 'Ye i and the
hyperplane matrix is given by
S = [ - 0.5372 0.0019 0.0822 1.0000 ]
As argued in Section 7.3.4, because of the minimum phase relative degree 1 as-
sumption, it is always possible to choose Sr so that X r = O for the nominal sys-
t em at steady st ate. For the system under consideration, equation (7.77) gives
172 SLJDING MODE CONTROL
0.8
~
o>
:, 0.6
.&
::,
o
0.4
0.2
Tjme, sec
The stable matrix r from equation (7.18) affects the closed-loop performance. Here
it has been chosen to tailor the step response of the nominal closed-loop system.
Figure 7.1 shows the step response of t he nominal closed-loop system for different
values of r in the interval [ - 0.04, - 0.015 ). In practice, overshoot in a thermal
process is very undesirable. The heating and cooling charact eristics are obviously
very different, and overshoot will essentially produce unnecessary time delays as
the process is left to cool down. As a consequence, the value of -0.025 has been
chosen as a compromise between the conflicting objectives of overshoot and rise
time. Ali that remains is to calculate the scalars that comprise the gain functions
in equations (7.83). T his is discussed in detail in the following subsection.
allowable inputs. For the system under consideration, the input is restricted to the
interval [O, 4] volts. Consequently, it is reasonable to require that the inequality
r- 1 IYI < 4 is satisfied. In practice, as a result of the chosen temperature profile to be
tracked, it was found that y is of the order l. In this way, a sensible upper bound
on r- 1 can be obtained.
uo
....
o
t::
t.Il
bJ)
o
e::
;Q
ü
....
""
r,-
-1
-2
-3
o 20 40 60 80 100 120
Time, mins
Figure 7.2 represents nonlinear simulation tests, using t he observer designed in the
previous subsection, with the nonlinear gain
for r 1 E {O,½, 1, J, 2}
This demonstrates the increase in performance obtained as a result of increasing r 1
and hence the nonlinear component of the control action. The final value of r-1 was
chosen as a trade-off between the tracking error and the control effort required. In
t he simulations shown in the next section, a scaling function of the form shown in
(7.83) has been employed.
Under typical operating conditions, the controller is required to take the furnace
from one operating temperature to another , along a specified trajectory. Figure
7.3 shows a typical temperature demand signa!, which will subsequently be used
for the nonlinear simulations. The demand comprises a period of low fire, a ramp
up to a high temperature, a period of soak then a return to low fire.
174 SLIOJNG MODE CONTROL
850
(.)
o 800
i::,
;¡¡
.,... 750
o.
.,e 700
E-<
650
o 20 40 60 80 100 120
Time, min
which effectively replaces (7.18) in Section 7.3.2. It is assumed that the intervals
which define the piecewise linear components of the reference are large enough
compared to the time constant of r so that 'steady state' occurs. It can be seen that
on each interval R(t) = O and therefore R(t) = a for sorne scalar. If er = r(t)-R(t)
then from equation (7.87) it follows immediately that
Figure 7.4 shows the closed-loop response of the final controller design on t he
nonlinear simulation. The tracking error is very good even on those parts of t he
demand profile that comprise transients between steady-state operating points and
for which asymptotic tr acking is not guaranteed t heoretically.
OBSERVER·BASED OUTPUT TRACI<ING CONTROLLERS 175
(7.80)
where w E IRn is the state vector of the model, r E IR" is an input vector and Am
and Bm are compatibly dimensioned matrices. lt is assumed that t he ideal model
is stable. Define an error state
as the difference between the plant and model state responses. This error is required
to tend asymptotically to zero. Substituting from equations (7 .88) and (7.89), the
dynamics of t he model-following error system can be determined as
As in Section 4.4.1, if the conditions for perfect model-following hold, there exist
compatibly dimensioned matrices L ,, and L ,. such that
(7.92)
(7.93)
In this case, ali forcing ter ms to t he error system (7.91) occur in the range space
of the plant input distribution matrix. Equation (7.91) can then be written as
The required system performance is usually specified via the model, and the con-
troller is synthesised to minimise the error between the model states and the con-
trolled plant. If the states x (and hence the error e,,) are available, then a control
law of the form described in Section 4.4.1 can be used to ensure e,, -+ O. Here,
however, it will be assumed that only the output from (7.1) is available and hence
an alternative approach is necessary.
To this end, consider the observer structure from Section 7.2 and define
i.e. the difference between t he ideal model and observer states. Let the control law
be of the form
u(t) = u¡(t) + u(t) (7.96)
176 SLIDING MODE CONTROL
where u¡(t) = Lxz(t) + Lrr(t) and the component ü(t), which includes a discon-
tinuous element, will be described later. Arguing as before, and substituting from
(7.5) and (7.89), it can easily be shown that
(7.97)
where e(t) represents the state observation error. The objective is to establish a
control law ü to induce a sliding motion on the surface
(7.98)
where the switching function s(ez) = Sez and the matrix SE JRmxn . For conve-
nience, partition the hyperplane matrix so that
(7.99)
with S1 E JRm x (n- m) and S2 E JRm x m . By design, assume that det(S2) i- O and
(7.100)
where A is a nonsingular diagonal matrix. Since it has been assumed that (A, B)
is in regular form, it follows that (Am, B) is also in regular forrn. Also, since
Am =A+ BLx it follows that the sliding motion on S is governed by
(7.101)
where M = S2 1 S 1 . A mínimum requirement is that M is chosen to ensure that
A. 11 is stable, and again any of the rnethods described in Chapter 4 may be utilised.
For a given A the matrix S2 is specified, hence the switching function matrix is
fully determined by the choice of M and A.
Remark
The notation used in this section replicates that of Section 7.3 in the sense that,
for example, A.11 was used to denote the system rnatrix governing sliding in (7.73),
although the orders of the matrices are different. This abuse of notation is delib-
erate and will allow sorne results to be conveniently quoted since the disparity in
order makes no difference to the fundamental results.
The control law component ü(t) comprises a linear element to stabilise the nom-
inal linear system, and a discontinuous term to induce an ideal sliding motion.
Specifically
ü(t) = U¡(t) + Ve (7.102)
where the linear component is given by
(7.103)
where <I> E IRmxm is a stable design matrix. The nonlinear component is defined
to be
for s(t) i- O (7.104)
(7.106)
,4c = [ A¿ 1
¾ 2
]
where A11 is defined in (7.101) and A 12 = A1 2S2 1 . Further, for the analysis which
follows, partition t he gain matrix so that
G= [ Q1 ] tn-m (7.108)
G2 tm
The closed-loop configuration is determined by these equations:
w(t) = Amw(t) + Bmr(t) (7.109)
é(t) = .4.oe(t) + B110 - B((t, x, u) (7.110)
ez(t) .4 cez(t) - GCe(t) + B(vc + 1.10 ) (7.111)
It should first be noted that equation (7.109) is known, stable and independent of
e and ez. It follows that it is sufficient to prove the stability of subsystems (7.110)
a nd (7.111).
The stability analysis which follows parallels the development in Section 7.3.3. A
block diagonal Lyapunov matrix for Ac of the form
- [A
P- (7.112)
- o
will be shown to exist where the lower right matrix sub-block A is the Lyapunov
matrix satisfying
--1 T - -1 "
P2 cp + cp P2 = - Q2 (7.113)
where Q2 is any symmetric positive definite matrix satisfying
6, 2 > a2 cQ- 1cTaT2 (7.114)
Using the Lyapunov matrix P for A 0 given in (7.6), define
pG ~[ pQ ºp ] (7.115)
where the matrix P will be chosen so that P0 is a Lyapunov matrix for the overall
closed-loop system matrix
A ~
o-
[ -~o
-GC (7.116)
178 SLIDING MODE CONTROL
(7.120)
is a Lyapunov function for the system given in equations ( 7.11 O) and ( 7.111 ) and
an ideal model-matching takes place.
Proof
Taking derivatives along the trajectories implies
V = - eT Qe + 2eT PBv0 - 2eT PE~+ e; (P A.e + A; P)ez
- 2e; Pace+ 2e;'PfJ(vc + Vo) (7.121)
Substituting from equation (7.119), the expression
1/ = -e;PQPez - (e+ Q - 1cT aT Pez)TQ(e + Q- 1 cT(;TPez)
+2eTPBv0 - 2eTPB~ + 2e;FB(vc + v 0 ) (7.122)
can be established. It follows from the definition of p0 (u,y) in (7.31) and manipu-
lations as in the proof of Proposition 7.1 that
2eTPBr10 - 2eTPB~ :S -2jjFCeil'Yo (7.123)
By an argument similar to the proof of Proposition 7.1, it follows that
T - - -
2ez PB(vc + vo) :S - 2-ycllPzsll (7.124)
Using inequalities (7.123) and (7. 124) in conjunction with (7.122), the derivative
V satisfies
v :::;- e; PQPez- (e+Q - 1 cT(;T Pez?Q(e+Q·-lcT(;T P ez)- 2IIAsll'Yc - 2IIFCelb o
and therefore V < O for e, ez -::/ O and quadratic stability is proved.
Furthermore because e., = ez - e, since ez ➔ O and e ➔ O, it follows that e., ➔ O
and perfcct modcl-following is attained. ■
Corollary 7.3 An ideal sliding motion takes place on S 0 and S in finite time.
Proof
Similar to Corollary 7.1.
■
OBSERVER-BASED OUTPUT TRACKING CONTROLLERS 179
Consider the lateral axis model of an L-1011 at cruise flight conditions which was
discussed in Section 5.7.1. The system triple (A ,B,C) is given by
o 1.0000 o o
-0.1540 - 0.0042 1.5400 o
A 0.2490 - 1.0000 -5.2000 o
[00~86 - 0.9960 -0.0003 -0. 1170
0.5000 o o
o
- 0.5000
B
[-07~~ -0.0320
o
0.3370 - 1.1200
0.0200
o
o
o
e ~ l o 1 o o
o o 1 o
o o o 1
1 o o o
-¡ l
The ideal model that will be used here is the model of Section 5.7.1. (Recall that
a gain matrix Lx was designed using eigenstructure assignment to yield a stable
matrix Am =A+ BLx with >.(Am) = {-0.05, -2 ± l.5j, -1.5 ± l.5j}) . T he feed-
forward gain matrix Lr was chosen so that the steady-state gain of (Am, BLr, C,,i)
was unity where the distribution matrix Cm defined in (5. 120) identified the bank
angle </> and sideslip angle /3 as the 'controlled outputs'.
In the design which follows, the switching function matrix S from (7.99) has been
obtained using the quadratic minimisation approach described in Section 4.2.2.
Here, as in Section 5.7.1, the symmetric positive definite matrix which defines the
cost function has been chosen as
Q = diag(5, 1, 1,5,5)
which gives sliding rnode error poles at
Ali t hat remains is to design the sliding mode observer. The canonical forrn of
Section 6.3 can be obtained via the output feedback canonical form in equation
(5.121) by suitable repartitioning. From (5.121) it follows that
and thus pair (A 11 , A211 ) is trivially observable. Furthermore, since the poles of
An + LA211 can be placed arbitrarily, there is no loss of design freedom in the
180 SLIDING MODE CONTROL
- 0.9167 -0.0026
-0.1637
2.0019 -0.0000
0.2292 -0.9994 -5.9468 -0.0000
2.8702 0.0386
l
+ LA21 1 = -3.
(7.125)
O O 1 O o
l
and the input and output distribution matrices are
o o
~ ººº
l
- 0.7440 - 0.0320
13 = 0.3370 - 1.1200 e -- [ o O
1 1 O O
O O 1 O
(7.126)
0.0200 O
o o o O O O 1
F=[ -0.0930 ~]
0.0381 0.0022
- 0.0040 -0.1266 o
and
o o
G=
3.5399
0.2292
1
0.0055 - 11.4944
3.4256 - 5.9468
-0.9167 -0.0026 7.3702 0.0386
5T]
0.1816 0.0101 - 13.4963 . o
Figure 5.12 shows the response of the closed-loop system to a step change on bank
angle. The dotted line represents the ideal model response and the solid line the
syst em response. As in Section 5.7.1, in arder to make the error system apparent,
the initial state of the observer has been deliberately chosen injudiciously. As can
be seen from Figure 7.5, the mismatch between the pla.nt model and the ideal model
disappears after a few seconds and the responses are indistinguishable.
0.3
0.2
"'
"5
B-
::,
0.1
o
o
-0.1
o 20 40 60 80 100 120 140
Time, sec
Figure 7.6 shows the first 10 seconds of the simulation so that the plant and observer
discrepancies can be seen. Figure 7.7 shows that a sliding motion takes place on S
after approximately 1 second.
OBSERVER-BASED OUTPUT TRACKING CONTROLLERS 181
O. l
V)
g'" O.OS
';,"o.'
;,
o
-0.05
o 2 3 4 5 6 7 8 9 10
Time, sec
Figure 7.6: Evolution of the errors between plant and observer outputs
0.1
e:
_g
si; O.OS
¡.r..
01)
e:
:.a o
B
·;
V)
-0.05
o 2 3 4 5 6 7 8 9 10
Time, sec
7.6 SUMMARY
P ractical nonlinear control strategies have been presented which provide robust
output tracking despite the presence of uncertainty. The strategies use only mea-
sured output information and construct estimates of the internal states, required for
the control law, using a nonlinear observer. It has been demonstrated for the class
of controllers and observers considered that a form of separation principie holds. A
rigorous analysis of the closed-loop performance of the controller/observer pair has
been undertaken; it indicates that in finite time a sliding motion can be attained
which rejects any bounded matched uncertainty.
The problem of analysing the properties of sliding mode controllers used in con-
junction with sliding modc obscrvcrs has not bccn studicd cxtcnsivcly. Work
analysing the closed-Ioop stability when using a sliding mode controller and an
asymptotic observer appears in Utkin (1992), Young & Kwatny (1982) and Breinl
& Leitmann (1987). In the latter the well-known 'separation principie' for lin-
ear systems is shown to be valid. Walcott & iak (1988) consider a sliding mode
controller/observer pair; they appeal to the results of Breinl & Leitmann (1987)
182 SLIDING MODE CONTROL
and claim that the separation principie is valid in this situation, but they do not
attempt any formal analysis.
The nonlinear furnace simulation in Section 7.4 is based on work by Palmer (1989)
which has subsequently been adapted for use as a test bed for controller design.
Chapter 8
8.1 INTRODUCTION
The actuation system under consideration is severcly nonlinear and conta.ins signif-
icant unmatched as well as rnatchcd uncertainty. A critica] component is thc large
stiction torque. Thc stiction contribution to the unmatchcd uncertainty poses a
particular problem for control system dcsign. Stiction is very difficult to model
184 SLIDING MODE CONTROL
Disturbance
Td(s)
Power +
!--->{}--
Sta.ge Vmot +
and/or measure and worst case estimates of stiction effects are often the only in-
formation that is available to the designer. If tight performance specifications are
required, as is the case for this automotive application area, robust performance
across the operation range can be difficult to achieve.
The basic behaviour of the actuator is best discussecl in terms of the simplified
schematic diagram given in Figure 8.1. The actuator comprises a DC motor which
directly drives a spring-loaded inertia . The model is developed from the standard
linear representation of a DC motor. The winding resistance Rm is assumed linear
and constant. This is a considerable simplification as the resistance will change
significantly with temperature. However, it will be shown that the effect of Rm
is wholly matched, so the proposed control configuration can be made completely
insensitive to its effects without further modelling investigations. The inductance
Lm varies with both the position output 0 and the current flowing in the motor
coi!. The motor gain Kmot also shows significant dependence upon the motor
current. The disturbance Td(s) incorporates stiction effects, the offset torque of
the spring, the cogging torque a.nd endstop effects. The magnitude of the stiction
effects depends upon output position and the direction depends on angular velocity.
The two torques are functions of output position. Endstop effects are significant
when the output position nears its limits of movement. In Figure 8.1, Vemf denotes
the back emf. The controller must be tolerant of changes in rotor inertia due to
differences in the type of motor which can be used to build the system. The
automotive environment is harsh due to large temperature variations and large
levels of vibration and shock. It is expected that many of the system parameters
will vary significantly during the life of the actuator. The dynamic behaviour of
t he actuator is determined by the following states:
The design specification requires both large and small step responses to be accom-
plished within a settling time of 0.1 second, with a 0.1° steady-state error and
ideally no overshoot. In addition, the controller must be able to cope with a slow
ramp input; practica! experience indicates that it is particularly hard to overcome
stiction effects swamping the performance under this condition. For reasons of
safety, the controller must be able to tolerate a sudden drop to zero spring offset
torque corresponding to the spring breaking during continuous operation of the
actuator.
Having selected one nominal operating condition, three state variables are defined
in terms of the deviations (.)d from this nominal condition; x 1 = 0 - 0d, x 2 = w,
x3 = i 0 - (i0 )d, u= Vmot - (Vmot)d . Note that the power stage is simply modelled
by an ideal voltage source. The control requirement is to provide close tracking of
a reference demand. A further state variable is therefore introduced:
(8.1)
Here Bref is the difference between the actual position demand and 0d, and T; is a
design parameter. If, as in Section 4.4.2, an augmented state vector
(8.2)
- [o
A= (8.4)
o
where eT = [l O O] and the matrix (A, B) is the state-space model of a DC motor
given in equation (3.112). As observed in Section 3.6.4, the nonlinearities associ-
ated with the electrical dynamics and motor gain occur as matched uncertainty, i.e.
in the channels implicit in the input distribution matrix, and therefore belong to
the class of uncertainty to which any appropriately designed sliding mode control
strategy is inherently insensitive. However, the disturbance Td(s), which is domi-
nated by the friction torque, appears as unmatched uncertainty a.nd its effect must
be minimised by an appropriate choice of switching surface. Here a time invariant
switching function of the form
s(t) = Sx(t) (8.5)
will be employed where
(8.6)
and the vector m E IR-3. The performance specifications needed for the actuator
translate to a requirement for a fast well-damped sliding mode dynamic behaviour
which must be achieved for the range of possible torque disturbance effects. As
shown in Lemma 4.1 , since the triple (A, B, e'f) <loes not have any invariant zeros at
the origin, the augmented pair (A, ÍJ) is completely controllable. Here a design was
186 SLIDJNG MODE CONTROL
perforrned using the robust pole placernent technique described in Section 4.2.1,
although any preferred robust design philosophy could be used.
A control strategy to ensure that this dynarnic performance is attained and rnain-
tained rnust ensure satisfaction of a reachability condition which ensures the sliding
function is at least locally attractive. Subject to this constraint, many forms of
control are possible. One appropriate choice is given by
- - 1 -
U = -(SE) SBw0ref + kT lxl sgn(s) (8.7)
with
(8.8)
where (.); indicates the ith element of a vector and !xi represents the vector con-
taining the moduli of the state vector components Xi . In practice, it was found
that good performance could be obtained using a sea.lar upper bound for kTlx[,
which is very attractive from the implementation point of view. The possibly dis-
continuous sign function is replaced with a smooth boundary !ayer approximation
as detailed in Section 3.7. It should be noted that the sliding rnode control strat-
egy (8. 7) requires full-state feedback. However, it is not a practica! proposition to
measure the angular velocity of the rotor. Sorne form of estimation or observation
is required for practica! implementation of the controller. Investigation showed
that straightforward numerical differentiation of measurements of output position
produced adequate results.
A prelimina.ry performance analysis was first carried out on the controller using
a nonlinear simulation model of the actuator. This investigation considered the
robustness of the controller to changes in the assumed stiction model and also
the effects of noisy mea.surement devices on the performance of the controller. In
addition, power stage nonlinearities were incorporated in the tests. The simulation
results ma.intained the required specifica.tion a nd the controller proceeded to bench
test assessment. Practica! implementa.tion was ca.rried out on a microprocessor
using integer arithmetic.
80
OJl 60
",::
-o
40
:~
o
"'-
o
o 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Time, sec
F igure 8.2: Response of the sliding mode controller to a large step demand
Figure 8.2 shows the closed-loop response of the sliding mode controller to a large
step dema.nd. For comparison, the response of a third-order linear controller with
a n integral element is shown in Figure 8.3.
AUTOMOTIVE CASE STUDIES 187
80
.,, 60
"¿
"O
o 40
:~
o
. ... ,. ....... --~' ---··!·•
"" 20 .
- - -- -- ._
.
-_____ ·------
.
-----
.
.
This linear controller was designed using a large magnitude frequency response and
consequently the requirement of ideally zero overshoot is achieved. Consider now
the response of both schemes to a small step demand (Figures 8.4 and 8.5).
48
46
.,.,,
"O
¿ 44
o
:~ 42
o
""
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Time, sec
Figure 8.4: Response of the sliding mode controller to a small step demand
48
,- -- . -- -----+ ·--- - -- :---- --- :--- ---- :----,
' . . ~ .·/~ .....
46
.,.,,
"O
¿ 44
o
·.::
·cñ 42
o
o..
40
38
o 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Time, sec
The performance of the linear controller has deteriorated significantly and is well
outside the desired performance specifications. The sliding mode scheme which
188 SLIDING MODE CONTROL
was not tuned after implementation to any particular mode of operation provides
superior performance that is very close to the required specifications. Figure 8.6
shows a typical sliding mode response to a slow ramp input. Stick-slip behaviour
is evident, but stiction effects have not swamped t.he behaviour of the actuator.
100
80
.,
ºº
-o
¿ 60
.g
·;;; 40
o
/l.,
20
Figure 8.6: Sliding mode controller response to a slow ramp output trajectory
55
~
-o
¿
.g
·;;;
o
/l.,
45
40L-- --J._ __.___ __,__ _. L . - _ - - l_ __.___ _ _¡__ _.L.-_ __,__ _ _,
Figure 8.7 shows a sliding mode response corresponding to the spring being dis-
engaged at approximately t = 0.14 second. After a small transient, the controller
regains control with no loss in performance. This is particularly important since
robustness to spring failure is an essential safety requirement. It should be noted
that the linear controller shows sorne stick-slip behaviour on the slow ramp and
tales approximately three times as long to recover from a broken spring.
When an automotive engine is idling andan accessory such as headlights, heater , air
conditioner, etc, is t urned on, an additional torsional load is placed on the engine.
If there were no compensation arrangements, the engine speed would fall drastically
and the engine would shut down. This effect is not observed when the vehicle is
moving since the driver compensates for the additional load by using the throttle.
Traditionally, in the automotive industry, this problem of idle speed regulation has
AUTOMOTIVE CASE STUDIES 189
been solved by increasing the airflow into the cylinder. This increase is different for
different loads or accessories, so this feed-forward control is accomplished vía look-
up tables. These demand extensive time and effort to formulate. The alternative is
to treat the on and off switching of the accessories as a disturbance to be rejected
vía feedback control.
. _ AIR INTAKE
In spark ignition engines the speed is regulated by controlling the amount of air
consumed (Figure 8.8). This is done by throttling the air mass flow rate entering
the cylinder through the inlet manifold. An additional valve is placed in parallel
to the throttle to act as a bypass path for the incoming a.ir; it is called the air
bypass valve (ABV). The air intake is controlled using this valve during idling,
since the throttle is not used. The input voltage is sent to the valve as a pulse
width modulated signa! (PWM) and the designed controller varíes the duty cycle
of the PWM signa!.
In an engine cylinder the air- fuel mixture is ignited by a spark plug just before
190 SLIDING MODE CONTROL
the piston reaches the top end of the cylinder. The position of the piston with
respect to the top end of the cylinder affects the output torque and in turn the
engine speed. Thus, the timing of the spark ignition or spark advance (SPA) can
be used to regulate the idle speed. However , this input has only a transient effect.
It is worth noting that SPA should not be used during steady-state conditions as it
generates emissions. The limits on this input are largely dependent on the engine
speed, but a rough measure on the limit is between 5 and 25°. The unit of this
input is the angle of the crankshaft with respect to the vertical when the spark
ignition takes place.
Obviously the output of the system being considered is the engine speed in rotations
per minute or rpm. The speed of the engine is sensed twice during each revolution,
which means that the sampling time is dependent on speed; the higher the speed,
the shorter the time constant and vice versa. Thus the effective sampling time
of the implemented controller is variable, resulting in time delays which alter the
phase of the system. The controller should be robust enough to cope with these
time delays.
The controller/observer scheme presented in Section 7.3 will be employed. A Ford
1.6 litre Zitec engine is used for the design study; it is a four-cylinder, four-stroke
engine. ldentified linear models from the two inputs (the ABV and the SPA) to
the output (engine speed) were used for controller design. These linear identified
models are mínimum phase and the product CE is also nonsingular.
The design issues which need to be considered can be summarised in the following
steps:
l. The dynamics of the demand r(t) are specified by the spectrum of r. The
eigenvalues of r used in equation (7.18) should nót be much faster than the
open-loop plant spectrurn to avoid unrealistically large controller effort.
2. The sliding surface dynamics are specified by the design matrix M from
equation (7.22), which acts as a state feedback matrix for the null space
dynamics. Previous knowledge of the plant can be used to define the required
closed-loop characteristics; this information can be translated into a set of
required eigenvalues for the sliding mode system and the eigenvalues can be
assigned via M. This translation can be done using LQR or mínimum entropy
methods.
3. The scalar constants used in the nonlinear part of the controller and the
observer are determined from the error bounds of the identified model. Later
on, these constants can be tuned during the nonlinear simulation to improve
performance.
Having discussed the general concepts of the controller design, particular attention
will now be focused on the controller design issues pertaining to the automotive
engine control problem.
AUTOMOTIVE CASE STUDJES 191
1. The engine inputs ABV and SPA used for speed control are severely restricted
in magnitude. In addition, the engine process is a discrete time process with
variable time interval, so it is more prone to input saturation if the controller
dynamics are fast. Hence the choice of r was made realistically, so as not to
saturate the inputs.
2. As discussed earlier, the hyperplane design procedure can be accomplished vía
traditional linear methods such as LQR or mínimum entropy design. Both
these methods rely on appropriate weighting of the plant states and plant
inputs. Usually these methods cannot be used efficiently with an identified
linear model, as the linear model states do not have any physical meaning.
However , in the present sliding mode controller formulation, the first state
of the augmented plant is the integral error state and it should be as small
as possible. Hence this state is heavily weighted. In order to avoid input
saturation the inputs are also grossly penalised. (This is particularly true for
the SPA input because it has much less authority than the ABV input.) The
hyperplane was designed using the mínimum entropy method as it gives good
disturbance rejection properties.
3. The linear part of the observer was designed by specifying the matrix A~2
in equation (7.13). This observer specification should be as fast as possi-
ble. However, as the sample rate for implementation had to be the same as
the sample rate of the controller, there was an upper limit on the speed of
observer.
4. The nonlinear parts of the controller and observer are fully determined by
specifying the scalars k1, 'Yo in equation (7.31), and 'Ye in equation (7.29).
These scalars were initially set to sorne small values and tuned on the nonlin-
ear simulation model of the engine. Later on, they were tuned on the actual
rig. The parameter 'Ye in particular provides a very effective knob for improv-
ing the disturbance rej ection properties. The function a(.) in equation (7.31)
was taken to be zero for the sake of simplicity.
5. In Chapter 7 it is assumed that the plant is square (i.e. the number of inputs
equals the number of outputs) . However, in this ISC problem there are two
inputs and one output. The square plant assumption is needed for observer
design only. The first input channel, ABV, has most of the authority as
compared to the SPA input channel, which is only available during transient
motion, so the observer was designed for the first input only.
The controller was implemented on t he test rig using dSPACE, 1 which is an efficient
too! for rapid prototyping. lt can convert SIMULINK block diagrams into optimised
executable C code and then download it onto a microprocessor running as controller
host . This greatly facilitates controller t uning during testing. The sliding mode
controller/observer pair was first implemented in SIMULINK and subsequently the
1 dSPACE is the registered trademark of dSPACE GmbH.
192 SLIDJNG MODE CONTROL
observer was translated into a MEX file2 written in C. The controller structure in
SIMULINK is shown in Figure 8.9.
demaoded
speed in 1pm Muxl3
Sum1 Demux2
Line•r
Controllor
lnput2
Mux12
As described previously, the controller should exhibit good tracking, robustness and
disturbance rejection properties. The set-point speed was chosen to be 880 rpm and
the tracking and disturbance rejection properties were verified by applying different
loads to the engine and observing the controller performance. Figure 8.10 shows
the closed-loop performance on application of the following sequence of loads: first
the electric drive fan and headlights were turned on (point A); after the cont roller
had returned the engine to the set-point, the heater and rear window defroster
were turned on (point B). The two dips (points A and B) and fiares (points C
and D) represent the application and release of these loads respectively. From
Figure 8.10 it is observed that the controller keeps the speed within the tolerable
limits. For the dips, the controller returns the engine to set-point speed at a rate
similar to that of the descent and the overshoots are less than 4%. In the case of
fiares, the magnitude of the speed <loes not exceed the 10% tolerable limit and the
maximum dip is around 17%. This dip can be reduced, but at the cost of increased
oscillations, high overshoots and lower robustness. The load presented in Figure
8.10 is a scaled version of measured load (obtained from a hot-wire anemometer
located in the air intake).
1000
~
a. 900
'-
.¿
<l)
<l)
o.. 800
V)
700
o 5 10 15 20 25 30 35 40 45
Time, sec
0.1
. . . .
;:; 0.05
"'
tt:
i. ~.
o
>
il'.l
-< o
-0.05
o 5 10 15 20 25 30 35 40 45
Time, sec
0.5
;:;
<f:
......o
o
-< -0.5
p..
V)
•1
5
Time, sec
3.5 . .
. .
3 . .-· -
. - .... ~- .. - . -. .' .. . -~ - .
.
-o
..... ,: ... .. .. ····:..
j"' 2.5
. .
1.5
o 5 JO 15 20 25 30 35 40 45
Time, sec
1000
8
o. 900
'-<
..¿
V
V
o. 800
V)
700
o 5 10 15 20 25
Time, sec
0.06
:¡, 0.04
:E
o 0.02
>
a:l
<(
o
-0.02
o 5 10 15 20 25
Time, sec
0.5
:¡,
o
:E
o -0.5
<(
P-
V)
-1 m-m1iW~
, LUMITTlITTl/[Thffifllf~
-1.5
o 5 10 15 20 25
Time, sec
2.5
-o
cd
.3
2
~~
1.5
o 5 10 15 20 25
Time, sec
1000
E
e-
-g
O)
o.
V)
10 20 30 40 50 60 70 80 90
Time, sec
0.1
'¡j
0.05
tl: ~ ··
o
>
~ ..... .• . .. . •··· ..•. . 't,r,,I
<!'.
-0.05
o 10 20 30 40 so 60 70 80 90
Time, sec
o
"
~
.....
o -l
·~
. . .... . ..... ... ..
····
<!'.
o..
V)
-2
10 20 30 40 50 60 70 80 90
Time, sec
3.5
3
"O
"'o 2.5
...:¡
2
~~~~\~
1.5
o 10 20 30 40 50 60 70 80 90
Time, sec
1200
1000
E
e-
.¿ 800
<l)
<l)
p.
V)
600
400
o 10 20 30 40 50 60 70 80 90
Time, sec
0.15
-0.05
o 10 20 30 40 50 60 70 80 90
Time, sec
0
~ o
o
t
V)
-1
-2
-3
o 10 20 30 40 50 60 70 80 90
Time, sec
3.5
. .
3 - . .. . . . ' .
'O
"'o
,....¡
2.5 . .. . . ·• ·
1.5
o 10 20 30 40 50 60 70 80 90
Time, sec
The next test involves the simultaneous application and release of the air condi-
tioning fan and rear window defroster (Figure 8.11). The dip (point A) corresponds
to t he application of the accessory load, and the fiare (point B) is the consequence
of turning t he accessory off. The fiare has a magnitude of 8% and the dip is
around 15%. It should be noted that the controller is rejecting these fiares and
disturbances within t he maximum time limit of 3 seconds. Hence the controller is
exhibiting good tracking and disturbance rejection properties. In both these tests,
the controller effort is quite low, guaranteeing that the input valve saturations are
avoided. The second input channel, SPA, seems to exhibit sorne chattering, but
this is the consequence of noise generated during the combustion process.
The robustness of the controller is demonstrated by the fact that the controller
design was performed ata set-point speed of 720 rpm whereas the tests were per-
formed at a speed of 880 rpm.
A further disturbance rejection test was performed with different gains and time
delays included. The controller easily stabilised these load tests with a delay of
five sample intervals and the gain increased by a factor of 1.5. Figure 8.12 shows a
performance test for various loads a.nd accessories, applied and withdrawn at the
marked points A to H. The same test was performed with a delay of 5 sample
periods on the measured output (Figure 8. 13). By comparing Figures 8.12 and
8.13 it is seen that, despite the heavy disturbance dueto the dela.y, the controller
is still maintaining rea.sonable performance.
8.4 SUMMARY
This chapter has considered the application of sliding mode control techniques to
automotive problems. The first case study has demonstrated that sliding mode
control is an appropriate design technique for nonlinear systems which include
significant unmatched uncertainty as well as the usual matched components. The
design approach is systematic, involving classification of matched and unmatched
uncertainty and switching surface selection to minimise unmatched effects. The
methodology provides good performance a.cross a wider range of operation t.han
may be possible with a classical design. A classical design can be tuned well
at one operating point but the performance can degrade severely away from this
condition. Bench test results support these assertions. The second ca5e study
considers the application of a sliding mode controller/ observer pair to the problem
of idle speed control. The controller shows good performance and robustness,
hence demonstrating that the previously described theoretical developments are
suitable for practica! application. Since the controller/observer pair is designed in
a multivariable framework, and assumes that only a subset of state information is
measurable, it is applicable to other automotive control problems such as air-to-fuel
ratio control.
Further details pertaining to t he idle speed control problem and the controller de-
sign are given in the thesis of Bhatti (1998). This problem has been addressed
198 SLIDING MOOE CONTROL
in the past by various other controllcr design techniques. Puskorius & Feldkamp
(1993,1994) used rccurrent neural networks to design a controllcr for thc ISC prob-
lem. They used throttle dcmand and spark advance as control inputs, demonstrat-
ing good simu!ation results. Shim et al. (1995) derivcd a nonlinear autornotive
engine modcl and used it for rernoving engine speed ripple fluctuations. Thc li
optimisation is used in Butts et al. (1995) for feedback a.nd fecd-forward controllcr
dcsign. Fuzzy logic (Vachtsevanos et al., 1992) and 1{00 (Williams et al., 1989)
ha.ve also been used for this purposc.
The details of the linear identification methods used in the second ca.se study can
be found in Astréim & Wittenmark (1989) and Soderstrom & Stoica (1989).
Boyd & R arrntt (1991) givc a detailed treatm<~nt of the LQR. and minimmn en-
tropy methods which were used to determine the desired sliding mode transient
performance.
Chapter 9
FLUE
EXHAUST
AIR
SUPPLY T HERMOCOUPLES
EMBEODE□ IN
WALLS
BURNER
NOZZLE
Until recently control of temperature was the primary concern. However, in view
of fut ure environmental legislation with regard to harmful emissions resulting from
combustion, attention is now being focused on the problem of combustion efficiency
and its effect on the formation of noxious by-products (Goodhart, 1994). One way
of monitoring combustion efficiency is to measure the concentration of excess oxy-
gen present in t he flue gases. Oxygen analysers, adapted from automotive tech-
nology, situated in the exhaust flue are able to provide on-line measurements of
oxygen concentration for use by the control system. Currently, control of both tem-
perat ure and excess oxygen is achieved by independent single-input single-output
control loops. This is clearly not necessarily t he optima! solution. This chapter
describes the development of a nonlinear multivariable control scheme from de-
sign t hrough to implementation for a test facility at the Gas Research Centre at
Loughborough, UK.
From a control systems perspective, the outputs of the 'system' are the furnace tem-
perature (as measured by t he thermocouple) and the percentage of oxygen present
in the combustion products (as measured by t he oxygen analyser in the flue). In an
industrial situation, the interna! furnace temperature would be required to exhibit
a specific time/temperature profile comprising, say, a period of low fire, a ramp to
a higher temperature, a period of soak and finally a return to ambient tempera-
ture. During normal furnace operation, efficient fuel combustion is desirable. For
a given mass of fue!, a theoretical mass of oxygen is required to completely oxidise
the hydrocarbons - so-called stoichiometric combustion. An inadequate air supply
wi ll result in incomplete combustion with a corresponding loss in thermal energy
release. Conversely, excess air will guarantee complete combustion but will give
rise to unnecessary enthalpy losses through the flue due to the increased flue flow
ra.te. Therefore, efficient combustion is ensured by minimising the amount of excess
oxygen present in the combustion products in the flue. In addition to the potential
energy savings, it is argued by Disdell et al. (1994) that accurate control of both
temperature and excess oxygen is of paramount importance from t he viewpoint of
reducing pollutant emissions.
For safety reasons, such furnaces usually operate at a fixed fuel/air ratio as a result
of an electronic ratio cont roller (ERC). The fuel and air flows are measured using
heated t hermistors placed in bypass lines around orífice plates in the flow paths of
the air and gas. The relative flows are compared and fedback to the ERC, which
makes appropriate adjust ments to the air valve position. The device also provides
an additional safety feature in t he form of a 'shutdown' alarm which isolates the
fuel supply when the furnace persistently operates 'off ratio', i.e. away from the
required fuel/air ratio set-point. This framework is usually described as 'gas led'
since t he airflow is modulated as a result of changes in the fue! flow to maintain
the appropriate fuel/air ratio necessary for efficient combustion. An additional
input to the ERC is the 'trim signa!'. This allows the fuel/air ratio set-point to
be dynamically adj usted. In this way, as far as this chapter is concerned, the
control inp11ts t.o the system are the fue! valve posit.ioner signa! and the ERC trim
signal. These inputs will be manipulated to ensure the temperature and excess
oxygen levels follow specified reference profi.les. This multivariable approach is
distinct from current practices which control the excess oxygen and temperature
vía independent single-input single-output control loops. The proposed control
scheme is illustrated in Figure 9.2.
FURNACE CONTROL CASE STUDY 201
Oxygen Signal
Ajr Valve
Oxygen
Positioo
Relerence 1
Sliding : Air Flow
Mode
Flow Measurement
Controller :Fuel Flow
1------------'
Temperature Gas Val·1e Position
Reference
Thermocouple Temperature
P hysical modelling of such a system, by the zone method 1 for instance, is c!early
nontrivial and <loes not lead to a systern of equations conducive to controller syn-
thesis; instead, an identification approach has been adopted. In order to obtain
a linear model around which to base the controller designs, identification of the
real plant was chosen as the most viable option. A pseudo-random binary signa!
(PRBS), taking the discrete values O and 1.5 volts, was applied to the valve po-
sitioner signa! in the fuel line. Unfortunately, such a signa! could not be used to
excite the oxygen trim channel. For reasons of safety, t he burner is only permitted
to operate 'off ratio' for a short period of time, after which a relay 'alarm' shuts
down the flow of fuel to the burner. Using a PRBS signa! in the oxygen trim channel
with its large, sudden amplitude changes would undoubtedly provoke a shutdown,
hence it is not suitable for identification purposes. Instead, an excitation signa!
comprising the sum of severa! sinusoids of different frequency and amplitude was
used. Such a signa! is not ideal for identification purposes but representecl the best
that could be done under the circumstances.
A low-order, relative degree 1 mínimum phase model, which provides good agree-
ment with the input/output data, is given in regular forro by
l
- 0.0186 - 0.0065 0.0190 0.01291
0.0026 - 0.1354 0.0310 0.0040
A
- 0.0972 0.0695 -0.1273 0.0530
-0.0193 -0.0155 - 0.1121 -0.4934
e = [ 0.6101 -0.1085 - 0.0286 0.0086 ] (9.1)
-0.2750 -0.1933 - 0.2175 0.0060
This system is stable and has invariant zeros at {-0.1633, - 1.2059}. The first
input channel represents the fue! valve position signa! and the second represents
the oxygen trim signa!. The first output is the furnace temperature as measured
by the thermocouple; and the second output is the signa! from the zirconia probe
1 See Section 7.1 for further details and references.
202 SLIDING MODE CONTROL
Since the system is square, the reduced-order dynamics of the estimation error are
completely determined by the invariant zeros of the system. Following the design
procedure in Section 7.3.1, by choosing a priori the stable matrix A~2 from (7.13)
to be diagonal, any diagonal positive definite matrix Pz will always be a Lyapunov
matrix for Ah. The diagonal matrix P2 can then be considered independen:tly as
a design matrix. Here it has been chosen in an effort to ensure the diagonal entries
of the matrix
from equation (7.14) are of compatible order. It was reasoned that the diagonal
elements of F act as weighting parameters which govern the distribution of the
nonlinear action between the individual input channels. By choosing
F = [ 0.0426 0.0060 ]
0.0313 0.0423
The diagonal elements of A~ 2 were chosen to make the condition number of the
observer closed-loop matrix A 0 = A - GC small. By inspection, it was found that
choosing
A• _ [ -0.5 o
22 - Q -0.5
l. 7201 -0.31321
G = 0.6269 -0.2251
[
8.1321 -2.7820
0.4512 0.4562
As in the single-input single-output design in Section 7.4, the nonlinear gain func-
tion will need to be determined empirically.
FURNACE CONTROL CASE STUDY 203
Forming the augmented plant from (4.102) it follows that in the notation of Section
4.4.2
o o - 0.6707 0.1085 0.0286 -0.0086
o o 0.2750 0.1933 0.2175 -0.0060
o o
[ ~ 11
A21
A12
A22 ] o o
- 0.0186 - 0.0065
0.0026 -0.1354
0.0190
0.0310
0.0129
0.0040
o o -0.0972 0.0695 -0. 1273 0.0530
o o -0.0193 -0.0155 -0.1121 -0.4934
Early attempts to select the poles of the reduced-order sliding motion by inspection
resulted in controllers which exhibited high levels of control activity. To circumvent
this, a preliminary linear LQR design was made for the augmented system, with
the cost function biased towards penalising the use of control effort. The four
slowest poles of the resulting closed-loop matrix were used as initial values for the
poles of the reduced-order motion. These were subsequently manually adjusted in
an effort to improve the conditioning of the closed-loop matrix. Ultimately the
sliding mode eigenvalues were taken to be { - 0.0849, - 0.0118, -0.0357 ± 0.0220j}.
The matrix M, from equation (4.106), which acts as a feedback matrix for the
pair (.4 11 , Á 12), was obtained using the robust eigenstructure assignment method
described in Section 4.2.1. The resulting matrix which defines the hyperplane is
given by
A _ [ 0.1 O ]
- O 0.1
which completes the hyperplane design. If the stable design matrix <I>, which assigns
the poles of the range space dynamics, is block diagonal, a diagonal solution for
A can be attained from the rnodified Lyapunov equation (7.28) by a suitable
choice of Qz. A diagonal structure for A is preferable since this 'decouples' the
nonlinear components of the control action. The eigenvalues for the range space
dynamics were take to be t he unused poles from the initial LQR design, namely
{- 0.1800 - 0.2509}. The diagonal elements of <I> were arranged so that the slowest
pole was associated with the oxygen trim channel. Consciously, every effort was
made not to apply an aggressive control signa! into this channel to guard against
violating the fuel/air ratio error shutdown alarm. The linear components of the
control law from equations (7.24) to (7.26) can be shown to be
0.0745 0.0107 -3.6948 - 0.5455 -0.3455 0.2117 ]
L [ 0.0729 0.0777 - 3.8278 -0.4594 - 0.6343 0.2124
6.9795 - 0.7794 ]
( 5.9178 2.2003
26.6345 -3.2763 ]
L.;- = [ 30.6273 9.8265
204 SLIDING MODE CONTROL
The sliding surface design parameter Sr was chosen so that, in the nominal case, the
steady-state values of the integrator states are zero using t he procedure described
in Section 7.3.4. In t his particular case study
The design matrix r from equation (7.18) has been chosen to tailor the step re-
sponse of the closed-loop system in th e nominal case. The stable design mat rix has
been chosen to be diagonal, since it makes no sense to introduce coupling between
the reference signals. The multivaria ble equation (7.18) can be represented as the
pair of scalar equations
(9.3)
was used and the scalars were chosen experimentally duríng the trials a t the Gas
Research Centre, Loughborough.
850
u
e 800
i
3
.,,...es: 750
o..
..,
E
700
¡.....
650
o 20 40 60 80 100 120
Time, min
850
u
o 800
i
a,...es: 750
o
o..
..,E
¡..... 700
650
o 20 40 60 80 100 120
Time, min
e::
o
12
00
>,
o
>< 10
V,
"'
o
ü
8
><
µ.¡
..,
o{:¡
6
1:::
5 4
2
o
c.
20 40 60 80 100 120
Time, min
was considered to be undesirable in the oxygen t rim channel b ecause of the safety
shutdown mechanism which exists in this input channel.
Figures 9.4 and 9.5 present results obtained from using t he scalar gains
and
(9.5)
Tracking accuracy of approximately 2° was obtained, equivalent to the precision of
the measuring device.
9.5 SUMMARY
This chapter has considered the problem of controlling both the temperature and
the oxygen content in the combustion products. The oxygen trim signa! to t he
ERC, which alters the fuel/air ratio set-point, has been used as an additional
input. The theoretical results of Chapter 7 were once again successfully used to
design a controller/ o bserver pair. Because a model was not available to test the
controller befare implementation, the design scalars that define the nonlinear gain
functions were selected in situ during the trials.
The electronic ratio controller (ERC) and its operation are discussed in Hammond
(1992). A comparison of different control methodologies which have been applied
to this type of high temperature furnace is given in Goodhart (1994).
Appendix A
Mathematical Preliminaries
In general, the definition of a vector space involves the notion of an arbitrary field
whose elements are called scalars.
Definition A.1 Let K be a given field and let V be a nonempty set with rules
of addition and scalar multiplication which assigns to any u, v E V a sum
u + v E V and to any u E V, k E K a product ku E V. Then V is called a
vector space over K, and the elements of V are called vectors, if the following
axioms hold:
of the form
(A.2)
where the ai E K is called a linear combination of v1 , v2 , • .. , Vm. The following
theorem applies.
T h eorem A .1 Let S be a nonempty subset of V. The set of ali linear combinations
of the vectors in S , denoted by I,(S), is a subset of V containing S. Furthermore,
if W is any other subspace of V containing S, then I,(S) e W.
The above theorem states that C(S) is the smallest subspace of V containing S; it
is called the subspace spanned by S.
D efinition A.4 A vector space V is said to be of finite dimension n or to be
n -dimensional, written dim V = n, if there exist linearly independent vectors
e 1 , e2 , ..• , en which span V. The set {e1, ... , en} is then called a basis f or V.
Let A and B be arbitrary sets. Suppose to each a E A there is assigned a unique
element of B; the collection f of such assignments is called a function or mapping
from A into B and is written
(A.3)
(A.4)
ª m2
where the a;j are scalars in K defines a matrix. Every such m x n matrix A
determines the mapping T : Kn -+ K= defined by
v-+ Av (A.5)
where the product of the matrix and vector on the right-hand side is defined com-
ponentwise as
n
(A v); =¿ a ;jVj (A.6)
j=l
P= (A.9)
is the transformation matrix from the original bo1:is { e1 , . . . , en} to the new basis
{!1 , . .. , fn}- As the vectors Ji, ... , fn are linearly independent, the matrix P is
invertible. It follows that the inverse p - 1 is the tr ansformation matrix from the
basis {/;} back to the basis {ei}.
For matrices A E 1Rnxm and BE 1Rnxm, the sum of the two matrices is given by
the matrix C E 1Rnxm where each element
for i = 1, .. . ,n, j = l , ... ,m
The product of the two matrices A E 1Rnxr and BE 1Rr x m is given by the m atrix
C E 1Rn xm where
r
c;J =¿ a ; kbkJ for i = 1, ... , n, j = l , ... ,m
k=l
where A11 E 1Rn1xm1, A12 E 1Rn 1Xm2' A21 E 1Rn2xm1 and A22 E 1Rn2xm2 with
n 1 + n 2 = n and m 1 + m2 = m. In this way the matrix A may be thought of as a
two-dimensional square matrix whose elements are t hemselves matrices. Let
B = [ Bu (A.11)
B21
MATHEMATICAL P RELIMINARIES 211
where B11 E lRn i xm,, B12 E 1Rnixm 2 , B21 E IR.n 2 xm, and B22 E 1Rn 2 xm2, then
the corresponding matrix sum may be written as
Let t he matrix A represent a linear transformation from the vector space V into t he
vector space W. Two important vector subspaces of V and W can be associated
with the transformation A . The null space of A is defined by the set
N(A) = {v E V : Av = O} (A.14)
To every square matrix A over a field K t here can be assigned a specific sea.lar
called the determinant of A and denoted det{A) . Consider the matrix of dimension
2 given by
A= [ a 11 (A.17)
a21
The determinant of A is given by
(A.18)
212 SLIDING MODE CONTROL
a-,2
= a11 det [ -
a32
The form of the above equation shows how determinants can be defined in terms of
determinants of lower order. To develop this idea, the minor Mrs of any element
ars of a general matrix A of dimension n is the determinant of the array obt ained
by omitting the row and column coutaining ars · The cofactor uf ars, denoted by
A rs , is given by
Note that the determinant of a single-element matrix is equal to the element itself.
Sorne useful properties of determinants will now be listed:
8. The matrix A has rank equal to its dimension if it has a nonzero determinant .
R emark
The matrix manipulations described in items 3, 4 and 5 are called elementary row
( column) operations.
MATHEMATICAL PRELIMINARIES 213
Av= .Xv
A- [ A11
- o
where the top left and bottom right sub-blocks are square.
Proposition A.1 The determinant and the eigenvalues of A satisfy:
A useful decomposition of the nonsquare matrix A defined over the field of complex
numbers <C is given by the so-called singular value decomposition. Before this
decomposit.ion can be defined it is necessary to consider the notion of a unitary
matrix.
Definition A.6 The matrix U E <Cnxn is unitary if
utt = u- 1 (A.20)
k = min(n,m)
214 SLIDING MODE CONTROL
where
In the special case when A is a real matrix then the factors U and V may be cho sen
as orthogorw.l matrices rather thrm 1mitary matrices.
It is useful to note that the singular values may be obtained by computing the
eigenvalues of AH A or AA1·1 and then computing the square root of the result.
It is straightforward to verify that when n :::; m the singular value decomposition
provides a useful tool for determining a set of vectors which span the null space of
the matrix A. For instance, suppose A E 1Rnxm where n:::; m and A has full rank,
i.e. rank(A) = m. Let A be decomposed as in equation (A.21) and let
V = [Vi Vi]
where Vi E 1Rmxm-n and notice that because V is real VTV = lm, Using this fact
and consequently
Since the m - n columns of Vz are linearly independent and A Vz = Oit follows that
they span the null space of A.
A.2.5 QR Decomposition
1. llxll ~ o
2. llxll = O if and only if x = O
3. llkxll = lkl llxll where k is a scalar and lkl is the absolute value of k
l. The 1-norm
(A.23)
2. The Euclidean norm (or 2-norm)
3. The oo norm
llxlloo = max lxJ;, (i= l , ... ,n) (A.25)
The Euclidean norm is perhaps t he most intuitive and corresponds exactly to the
notion of distance, i.e. the length of the straight line between two points. Ali of
the above are merely special cases of the p-norm, which is defined by
(A.26)
where p ~ l.
The vector p-norm in equation (A.26) induces a corresponding matrix norm which
is given by
IIAIIP = max JJAxJIP
- - - (A.27)
x,to 11 x 11 p
Such matrix norms satisfy the properties of a vector norm. In addition, the follow-
ing condition, relating to the norm of a product of matrices, is satisfied:
l. The 1-norm
(A.29)
An inner product is essent ially a rule which assigns to any pair of vectors in a
vector space an associated scalar quantity.
D efinition A.8 Let V be a real or complex vector space defined over the field
K, which is either the real field lR ar the complex field C. Suppose that to each
pair of vectors u, v E V there is assigned a scalar (u, v) E K . The mapping is
called an inner product in V if it satisfies the following axioms:
(A.32)
(It can be readily shown from the axioms of the inner product t hat t he four norm
properties are satisfied.) The inequality in (A.31) is often called the Cauchy-
Schwarz inequality.
In the case when V = 1Rn an inner product is given by
(A.33)
An important fact is t hat the norm induced by t his inner product is the Euclidean
norm.
In general, if two vectors u and v satisfy (u, v) = O, they are deemed to be orthogonal
(or perpendicular).
The orthogonal complement of a vector subspace W ~ V is the vector subspace
The concept of a projection of a vector into a given subspace will now be intro-
duced. Assume that the columns of the matrix L are independent and span a given
subspace .C. It is necessary that a vector Va resides in this prescribed subspace, i.e.
Va= Lx (A.34)
for sorne appropriately dimensioned x. Assume that a desired form Vd is known for
the vector Va- In general, this desired form may not lie in the required subspace.
In this case, it is useful to select Va so that it is as clase as possible to the desired
form Vd - To this end, an appropriate vector x in equation (A.34) is a vector chosen
to minimise
(A.35)
which is a measure of the distance between the desired vector Vd and the achievable
vector Va- It can be shown that
(A.36)
for ali x
In particv.lar if
x:¡i:O
p = [ Pi..¡ P12 ]
P 12 P22
where the matrix sub-blocks P 11 and P 22 are square.
Proposition A.3 The symmetric matrix P satisfies
P > O <⇒ P n > O and P22 > Pl;Pi11Pi2
<⇒ A2 > O and P11 > Pi2P221P'[i
As a corollary
P>O =;, P11 > O and P22 > O
Another property that will be exploited is that, given a symmetríc matrix P and
a nonsingular matrix T of the same dimension, then
p >o TTPT> O
A symmetric matrix Pis said to be negative definite if - Pis positive definite.
Assorted mfiles
This appendix provides a listing of sorne mfiles which would have caused a distrac-
tion from the main argument had they been placed in the body of the text.
The function below has been used in severa! places, notably Chapters 4 and 7, as
a robust pole placement method. It is fundament ally the same as the MATLAB
command place except for a preliminary transformation to the system and input
distribution matrix. This precaution has been added because invariably t he com-
mand is used on sorne subsystem - usually (A11 , A12 ) ora variation thereof - which
although guaranteed to be controllable does not necessarily have the property that
A12 is of ful] column rank. QR reduction is performed on Af2 so that
A12 = [ Á12 Ü ] QT
where Q is orthogonal and A12 has full column rank. The pole placement is then
performed on the pair (A11 ,A 1 2).
mfile: a modified v ersion of the 'place' command
function K=vplace(A,B,p)
[nn,mm]=size(B);
rnk=rank(B);
if rnk==mm,
K=place(A,B,p);
else
[Q,Btilde]=qr(B') ;
Btilde=Btilde' ;
Btilde=Btilde(:,1 :rnk);
Ktilde=place(A,Btilde,p);
K=Q*[Ktilde ;zeros(mm-rnk, nn)];
end
return
220 SLIDJNG MODE CO NTROL
and
(B.2)
Because the eigenvalue and eigenvector must satisfy
(B .3)
(B.6)
and
(B.7)
Define
(B.8)
2
and let K(>.;) E IR.( n+m)xm be a matrix whose columns span the null space of
Ge(>.;). If
K(.\;) =[ N(A;)
P(.\;)
M(>.;)
l tn
tn
tm
(B.9)
it follows t hat
[.. l- v,m
- FVre
E span {
N(.\;)
P(>.;)
M(>.;) } and
[
Vím
Vre
- F vim l E span {
N(>.; )
P(.\;)
M(.\;) }
which implies
[i:; ] N(.\;)
E span { - P(.\;) }n span {
P (.\;)
N(>.;) } (B.10)
ASSORTED MFILES 221
To calculate a basis for this vector space numerically, the key observation is that,
given any matrix H , if KH is a matrix which spans the null space HT then
which implies that the R(H) = N(KJ). With this fact in mind, define
and (3 - [ P(.\;) ] (B.11)
- N(.\;)
and Jet Ka and l<fJ be matrices whose columns span the null spaces of the matrices
aT and (3T respectively. Next define
(B.12)
and !et K 'Y be a matrix whose columns span the null space of K'Y" Notice t hat
KJ K'Y = O and K J' K'Y = O and consequently, from the observation above, the
columns of K'Y span the intersection of R(a) n R((3) . This construction is easily
implemented in MATLAB using the singular value decomposition.
Comparing this situation with that in Section 4.3, the matrix K 'Y plays the role of
the matrix N(.\;); consequently, the rows of K'Y need to be reordered to isolate an
equation of the form
(B.13)
where vf is a vector comprising the real and imaginary components which have
been specified and ó; represents a vector which will determine the eigenvector.
In the mfile fragment below, the variable nocomp represents the number of complex
conjugate pairs specified in lambda. The first loop calculates the required ó;'s which
determine complex eigenvalues using the method described above. The secon d loop
is identical to the mfile fragment in Section 4.3 and calculates the eigenvectors
associated with the remaining real eigenvalues.
Note that the matrix V contains only real values. For complex eigenvalues, V
contains the real and imaginary components as two columns of real numbers. Since
the eigenvectors must occur in complex conjugate pairs, the real and irnaginary
components for only one eigenvector of the pair need be stored; in this way, as in
the real case, V E lRnx(n-m).
mfile: hyperplane design via eigenstructure assignme nt
% For the systern pair (A ,B), nocornp defines the nurnber of pairs of desired
% cornplex conjugate poles. The vector lambda defines the n-rn desired
% s liding rnode poles;the first 2i-1 (i=1, nocornp) entries contain the
% real parts of the complex conjugate poles and the entries 2i (i=1,nocornp)
¡, contain the irnaginary parts . In t he matrices described below, the first
1, 2i-1 (i=1 , nocornp) colurnns contain inforrnation about the desired real
1, parts of the eigenvectors and the first 2i (i=1,nocomp) contain the
1, irnaginary parts. The columns of the nx(n- rn) rnatrix specpos are used
% to distinguish vector entries which are arbitrary (denoted by O in the
1/. relevant matrix entry) from entries which are to be specified (denoted
222 SLIDING MODB CONTROL
'l. by 1 in the relevant matrix entry) . The columns of the nx(n-m) matrix
'l. specent specify any desired eigenvector entries.
despos=find(specpos(:,2*i-1));
numspecr=length(despos);
n1= []; desent= [] ;
for j =1:numspecr,
nl(j,:)=(kgamma(despos(j),:));
desent(j)=specent(despos(j),2*i-1) ;
end
despos=find(specpos(:,2*i));
numspecc=length(despos);
for j=1 :numspecc,
n1(j+numspecr,:)=kgamma(despos(j)+nn,:);
desent(j+numspecr) =specent(despos(j),2*i);
end
delta=n1\desent' ;
vector=kgamma*delta;
V(1:nn,2*i-1)=vector(1:nn);
V(1:nn,2*i)=vector(nn+1:2*nn);
end
desent2=[]; n2=[];
for j=1:numspec,
n2(j,:)=nlambda(despos( j) ,:);
desent2(j)=specent(despos(j),i);
end
o o - 1 o - 1 o o o o
o o -1 o o o o o o
o o o o 1.74 0.08 0.59 o o
A= o o o -1.99 - 13.41 - 18.95 -3.60 B= o o
o o o 1.00 - 1.74 -0.08 -0.59 o o
o o o o o -20 o 20 o
o o o o o o -20 o 20
In Section 4.5.2 the peles of the ideal sliding mode dynamics were chosen to be
and the desired eigenstructure was indicated in equation (4.144). In terms of the
matrices used in the mfiles, this information translates into
o o o o o * * * * *
o o o o o
1 1 1 1 1 o* o* * * *
1 o 1
specpos = 1 o 1 1 1 specent = 1 o 1 o (B.15)
*
o 1 1 o 1 * 1 o * o
o o o o o * * * * *
o o o o o * * * * *
224 SLIOING MOOE CONTROL
It can be seen that in t he vectors in which only two components have been specified,
the desired values have been obtained exactly. When three components have been
specified, an approximation has been returned.
http://www.le . ac.uk/eg/ce14/vscbook/mfiles.html
Astrom, K.J. and Wittenmark, B., 1989, Adaptive Control, Addison-Wesley, Read-
ing MA.
Dag, S.K., Spurgeon, S.K. and Edwards, C., 1997, Output feedback sliding modc
design for linear uncertain systems, Proceedings of IEE, I'art D 144, 209- 216.
J3alcstrino, A. and Innoccnti, M., 1990, The hypcrstability approach to VSCS de-
sign, in Zinobcr, A.S.I. (ed.), Deterministic Control o/ Uncertain Systems,
Peter Pcregrinus, Stevenage, UK, pp. 170- 199.
Dalestrino, A., DeMaria, G. and Zinober, A.S.I., 1984, Nonlincar adaptive model-
following control, Av.tomatica 20, 559- 568.
Bélanger, P .R., 1995, Control Engineering: A Modern Approach, Harcourt Brace,
London.
Dhatti, A.I., 1998, Advanced Sliding Mode Controllers for Industrial Applications,
PhD thesis, University of Leicester.
Iloyd, S.P. and Barratt, C.H., 1991, Linear Controller Desi_gn, Limits of Perfor-
mance, Prcntice Hall, Englewood Cliffs NJ.
Drcinl, W. and Leitmann, G., 1987, State feedback for uncertain dynamical sys-
tems, Applied Mathematics and Computation 22, 65- 87.
Ilrogan, W.L., 1991, Modern Control TheonJ, Prentice Hall, Englewood Cliffs NJ.
Ilroussard, J.R. and O'Brien, M ..J., 1980, Fecdforward control to track the out.put
of a forced model, IEEE 1'ransactions on Automatic Control AC-25, 851- 853.
13urton, J.A. and Zinober, A.S.L, 1986, Continuous approximation ofvariable struc-
ture control, Jnternational Journal of Systems Science 17, 876-885.
Butchart, R.L. and Shakcloth, B., 1966, Synthesis of model-reference adaptive
con Lrol sysLems by Lyapunov's second meLhod, Prnceediny:; uf lhe 2nd IFAC
Symposium on the 1'heory of Self-adaptive Control Systems, pp. 145- 152.
Ilutts, K. , Siva.shankar, N. and Sun, J., 1995, Feedforward and feedback design
for engine idlc spced control using l 1 optimi:1,ation, Proceedings of the 1995
Amer·ican Control Conference, Seattlc WA., pp. 2587- 2590.
226 SLIDJNG MODE CONTROL
Chan, Y.T., 1973, Perfect model-following with a real model, Proceedings of the
Joint Automatic Control Conference, pp. 287-293.
Chen, C.T., 1984, Linear System Theory and Design, Holt, Rinehart & Winston,
New York.
Corless, M., 1994, Robust stability analysis and controller design with quadratic
Lyapunov functions, in Zinober, A.S.I. (ed.) , Variable Strucfore and Lyapunov
Control, Springer-Verlag, Berlín, pp. 181-203.
Corless, M., Goodall, D.P., Leitmann, G. and Ryan, E.P., 1985, Model-following
controls for a class of uncertain dynamical systems, Proceedings of the 7th
IFAC/IFORS Symposium on Identification and System Parameter Estimation,
pp. 1895-1899.
Davies, R. and Spurgeon, S.K., 1993, Robust tracking of uncertain systems via a
quadratic stability condition, Proceedings of the 12th IFAC World Congress,
pp. 10:43-46.
DeCarlo, R.A., iak, S.H. and Matthews, G.P., 1988, Variable structure control of
nonlinear multivariable systems: a tutoría!, Proceedings of IEEE 76, 212- 232.
deJager, B., 1992, Comparison of methods to eliminate chattering and avoid steady
state errors in sliding mode digital control, Proceedings of the IEEE VSC and
Lyapunov Workshop, Sheffield, pp. 37- 42.
Disdell, K.J ., Burnham, K.J. and James, D.J.G., 1994, Developments in fur-
nace control for improved efficiency and reduced emissions, IEE Digest No.
1994/018, Colloquim on improvements in furnace control: current develop-
ments and new technologies.
Dorling, C.M. and Zinober, A.S.I., 1986, Two approaches to hyperplane design
in multivariable variable structure control systems, International Journal of
Control 44, 65-82.
Drazenovié, B., 1969, The invariance conditions in variable structure systems, Au-
tomatíca 5, 287- 295.
EI-Khazali, R. and DeCarlo, R.A., 1991 , Variable structure output feedback con-
trol: switching surface design, Proceedings of the 29th Allerton Conference on
Communications, Control and Computing, pp. 430- 439.
El-Khazali, R. and DeCarlo, R.A., 1992, Variable structure output feedback con-
trol, Proceedings of the American Control Conference, pp. 871- 875.
REFERENCES 227
El-Khazali, R. and DeCarlo, R.A., 1993, Output feedback variable structure con-
t rol design using dynamic compensation for linear systems, Proceedings of the
American Control Conference, pp. 954- 958.
Gu, G., 1990, Stabilizability conditions of multivariable uncertain systems vía out-
put feedback control, IEEE Transactions on Automatic Control 35, 925- 927.
Harnmond, P.S., 1992, Advanced kiln and furnace control, Proceedings of the In-
ternational Gas Research Conference, Florida.
I-Iarvey, C.A. and Stein, G., 1978, Quadratic weights for asymptotic regulator prop-
erties, IEEE Transactions on Automatic Control AC-23, 378- 387.
Heck, B.S., Yallapragada, S.V. and Fan, M.K.I-1., 1995, Numerical methods to
design the reaching phase of output feedback variable structure control, Au-
tomatica 31, 275- 279.
Hermans, F.J.J. and Zarrop, M.B., 1996, Sliding rnode observers for robust sensor
monitoring, Proceedings of the 13th IFA C World Congress, pp. 211- 216.
Hottel, H.C. and Sarofim, A.F., 1967, Radiative Transfer, McGraw-Hill, New York.
Hsu, L., Araújo, A.D. De and Costa, R.R., 1994, Analysis and design of
input/output based variable structure adaptive control, IEEE Transactions
on Automatic Control 39, 4-21.
Hui, S. and Zak, S.H., 1993, Robust output feedback stabilisation of uncertain
dynamic systems with bounded controllers, International Journal of Robust
and Nonlinear Control 3, 115- 132.
Hung, J.Y, Gao, W. and Hung, J .C., 1993, Variable structure control: a survey,
IEEE Transactions on Industrial Electronics 40 , 2- 22.
Itkis, U., 1976, Control Systems of Variable Structure, Wiley, New York.
Kailath, T., 1980, Linear Systems, Prentice Hall, Englewood Cliffs NJ.
Kautsky, J. and Nichols, N.K., 1983, Robust eigenst ructure assignment in state
feedback controls, Numerical Analysis R eport NA/2/83, Department of Math-
ematics, University of Reading, UK.
228 SLJDING MODE CONTROL
Kautsky, J. , Nichols, N.K. and Dooren, P. Van, 1985, Robust pole assignment in
linear state feedback, International Journal of Control 41, 1129-1155.
Khalil, H.K., 1992, Nonlinear Systems, Prentice Hall, Englewood Cliffs NJ.
Kwakernaak, H. and Sivan, R., 1972, Linear Optima[ Control Systems, Wiley, New
York.
Luo, N.S. and Feng, C.B., 1989, A new method for suppressing chattering in vari-
able structure feedback control systems, Nonlinear Control Systems Design:
Science Papers of the IFAC Symposium, Pergamon, Oxford, pp. 279-284.
Monopoli, R.V., Gilbert, J .W. and Thayer, W.D. , 1968, Model-reference adap-
tive control based on Lyapunov-like techniques, Proceedings of the 2nd IFAC
Symposium on System Sensitivity and Adaptivity, pp. F24- 36.
Mudge, S.K. and Patton, R.J., 1988, Variable structure control laws for aircraft
manoeuvres, Proceedings of the IEE Conference Control '88, pp. 564- 568.
Nichols, N.K., 1987, Robustness in partial pole placement, IEEE Transactions on
Automatic Control AC-23, 728- 732.
O'Brien, M.J. and Broussard, J.R., 1978, Feedforward gains to track the output
of a forced model, Proceedings of the 17th IEEE Conference on Decision and
Control.
Ogata, K., 1997, Modern Control Engineering, 3rd edn, Prentice Hall, Englewood
Cliffs NJ.
Palmer, M.R., 1989, A practica! computer package for the thermal design of high
temperature industrial plant, Proceedings of the International Gas Research
Conferen ce, Tokyo.
Patel, R.V. and Toda, M., 1980, Quantitative measures ofrobustness for multivari-
able systems, Proceedings of the American Control Conference, pp. TP8-A.
Patton, R.J., Frank, P.M. and Clark, R.N. , 1989, Fault Diagnosis in Dynamic
Systems: Theory and Application, Prentice Hall, New York.
REFERENCES 229
Puskorius, G. and Feldkamp, L., 1993, Autornotive engine idle speed control with
recurrent neural networks, Proceedings o/ the 1993 American Control Confer-
ence, San Francisco CA.
Ridgely, D.B., Silverthorn, J.T. and Banda, S.S., 1982, Design and analysis of a
multivariable control system for a CCV-type fighter aircraft, AIAA Technical
report 82-1350.
Ryan, E.P., 1988, Adaptive stabilisation of a class of uncerta.in nonlinear systems:
a differential inclusions approach, Systems and Control Letters 10, 95- 101.
Ryan, E.P. and Corless, M., 1984, Ultimate boundedness and asymptotic stability
of a class of uncertain dynamical systems via continuous and discontinuous
control, IMA Journal o/ Mathematical Control and Information 1, 223- 242.
Shim, D. , Park, J. , Khargonekar, P. and Ribbens, W., 1995, Engine idle speed
control, Proceedings of the 1995 American Control Conference, Seattle WA.,
pp. 2582- 2586.
Slotine, J.J.E., 1984, Sliding controller design for nonlinear systems, International
Journal o/ Control 40, 421-434.
Slotine, J.J.E. and Li, W., 1991, Applied Nonlinear Control, Prentice Hall, Engle-
wood Cliffs NJ.
Slotine, J.J.E., Hedrick, J.K. and Misawa, E.A., 1987, On sliding observers for
nonlinear systems, Transactions o/ the ASME: Journal o/ Dynamic Systems,
Measurement and Control 109, 245-252.
Smith, K.L. and Anderson, D.C., 1982, Optima! control applications to AFTI/F-16
and advanced flight control system design, AIAA Technical report 82-1572.
Sobe!, K.M. and Shapiro, E.Y., 1986, Flight control synthesis via eigenstructure
assignment- t he discrete version, Proceedings o/ the the 28th Israel Conference
on Aviation and Astronautics, pp. 1- 6.
Soderstrom, T. and Stoica, P., 1989, System Identification, Prentice Hall, Heme!
Hempstead, UK.
Spurgeon, S.K., 1991, Choice of discontinuous control components for robust sliding
rnode performance, Intemational Journal o/ Control 53, 163- 179.
Spurgeon, S.K. and Davies, R., 1993, A nonlinea.r control strategy for robust sliding
rnode performance in the presence of unmatched uncertainty, International
Joumal o/ Control 57, 1107- 1123.
230 SLIOING MODE CONTROL
Sreedhar, R., Fernández, B. and Masada, G.Y. , 1993, Robust fault detection in
nonlinear systems using sliding mode observers, Proceedings of the IEEE Con-
ference on Control Applications, pp. 715-721.
Strang, G. , 1988, Linear Algebra and its Applications, Harcourt Brace Jovanovich,
London.
Syrmos, V.L., Abdallah, C.T., Dorato, P. and Grigoriadis, K., 1997, Static output
feedback - a survey, Automatica 33, 125- 137.
Utkin, V.I., 1977, Variable structure systems with sliding modes, IEEE Transac-
tions on A utomatic Control 22, 212- 222.
Utkin, V.I. , 1981, Principies of identification using sliding regimes, Soviet Physics
Doklady 26, 271-272.
Utkin, V.I. and Young, K.-K.D., 1978, Methods for constructing discontinuity
planes in multidimensional variable structure systems, Automation and R e-
mote Control 39, 1466 - 1470.
Vachtsevanos, G., Farinwata, S. and Kang, H., 1992, A systematic design method
for fuzzy logic control with application to automotive idle speed control, Pro-
ceedings of 31st IEEE Conference on Decision and Control, Tuscon AZ.
Walcott, B.L. and Zak, S.H., 1987, State observation of nonlinear uncertain dy-
namical systems, IEEE Transaction on Automatic Control 32, 166- 170.
Walcott, B.L. and Zak, S.H., 1988, Combined observer- controller synthesis for un-
certain dynamical systems with applications, IEEE Transactions on Systems,
Man and Cybernetics 18, 88-104.
Walcott, B.L., Corless, M.J. and Zak, S.H., 1987, Comparative study of nonlinear
state observation techniques, International Journal of Control 45, 2109- 2132.
Wilkinson, J.H., 1967, The Algebraic Eigenvalue Problem, Oxford University Press,
Oxford.
Williams, S., Hrovat, D. , Davey, C., Maclay, D., Crevel, J. and Chen , L., 1989,
ldle speed control design using an H 00 approach, Proceedings of the American
Control Conference, P ittsburgh PA.
Young, K-K.D., 1977, Asymptotic stability of model referencc systcms with variable
structure control, IEEE Transactions on Automatic Control AC-22, 279-281.
Young, K.-K.D. and Kwatny, H.G., 1982, Variable structure servorncchanism design
and applications to ovcrspccd protcction control, A utomatica 18, 385- 400.
Young, K.-K.D., Kokotovié, P.V. and Utkin, V.I., 1977, A singular pcrturb,"tion
analysis of high-gain feedback systern, IEEE Transact·ions on A utomatic Con-
trol 22, 931 -937.
Zak, S.H. and Hui, S., 1993, Output feedback in variable structurc controllcrs and
state estimators for uncertc1in/nonlinear dynamic systems, Proceedin_qs of the
IEE, Part D 140, 41-50.
%ak, S.H. and Walcott, B.L., 1990, State observation of nonlinear control systems
via the mcthod of Lyapunov, in Zinober, A.S.I. (ed.), Deterministic Control
of Unccrtain Systems, Petcr Pcregrinus, Stcvcnage UK, pp. 333- 350.
Zhou, F. and Fisher, D.G., 1992, Continuous sliding mode control, Intcrno.l,ional
Journal of Control 55, 313 327.
Zinobcr, A.S.I. and Dorling, C.M., 1990, Hyperplane dcsign and CAD of variable
structure control systcms, in Zinober, A.S.I. (ed.), Deterministic Control of
Uncertain Systems, Pet.er Pr.regrinus, Stcvr,nage UK, pp. 52--79.