Variable Structure Control of Nonlinear Multivariable Systems: A Tutorial

This paper presents, in a tutorial manner, the design of variable structure control WSC) systems for a class of multivariable nonlinear time varying systems. By the use of the Utkin-DraienoviC method of equivalent control and generalized Lyapunov stability concepts, VSC design is described in a unified manner. Complications that arise due to multiple inputs are then described and several approaches useful in overcoming these complications are then developed. After this, the paper investigates recent developments and the kinship of VSC and the deterministic approach to the control of uncertain systems. All points are illustrated by numerical examples. In addition, the recent VSC applications literature is surveyed.

I.

INTRODUCTION

Variable Structure Control (VSC) i s a viable high-speed switching feedback control (for example, the gains in each feedback path switch between two values according to some rule). This variable structure control law provides an effective and robust means of controlling nonlinear plants. It has its roots in relay and bang-bang control theory. The advances in computer technology and high-speed switching circuitry, have made the practical implementation of VSC a reality and of increasing interest to control engineers (see References). Essentially, VSC utilizes a high-speed switching control law to drive the nonlinear plants state trajectory onto a specified and user-chosen surface in the state space (called the sliding or switching surface), and to maintain the plants state trajectory on this surface for all subsequent time. This surface is called the switching surface because if the state trajectory of the plant i s above the surface a control path has one gain and a different gain if the trajectory drops below the surface. The plant dynamics restricted to this surface represent the controlled systems behavior. By proper design of the sliding surface, VSC attains the con-

Manuscript received June 3,1987; revised November 16, 1987. R. A. DeCarlo and S. H. Zak arewith the School of Electrical Engineering, Purdue University, West Lafayette, IN 47907, USA. G . P. Matthews is with the Electrical and Electronics Engineering Department, General Motors Research Laboratories, Warren, MI 48090, USA. I E E E Log Number 8719082.

ventional goals of control such as stabilization, tracking, regulation, etc. The purpose of this paper is to furnish quick readable access to key design techniques in VSC (scatteredthroughout the literature) for a class of nonlinear time-varying systems. Because of the papers tutorial nature, the presentation includes only several of the basic forms of the many VSC design methods for multivariable, nonlinear, timevarying systems. These basic forms often need tweaking before application. To minimize confusion and to maintain a unified exposition, our discussion concentrates on systems linear in the control input. Such systems are amenable to Utkins methods [I], [2], [4]. Also for simplicity, the paper deals most of the time with ideal VSC-i.e., switching in the control law can occur infinitely fast. The ideal case is much easier to analyzeand provides a baselineagainstwhich onecan measure more realistic designs. Comments on the nonideal case are included for completeness at the end of the paper. Section II introduces the reader to the flexibility offered by the variable structure control strategy via the medium of a simple example. Section I l l crafts the setting in which the tutorial development is to proceed. It sets forth the basic definitions such as the system model, the switching surface, the associated notion of a sliding mode, and an overview of the two-phase VSC design process. Section Vexamines phaseoneof theVSC design process, that of designing a sliding surface so that the plant restricted to the sliding surface has a desired system response. This means that the statevariablesof the plant dynamics are constrained tosatisfyanother set of equationswhich definethe so-calledswitching surface. An example illustratesthe ideas and the relevant literature i s cited. Section VI discusses the construction of the switched feedback gains necessary to drive the plants state trajectory to the sliding surface. These constructions build on the generalized Lyapunov stability theory. The remainder of the paper deals with applications, the problem of nonideal switching, and the use of the boundary layer concept to alleviate the problem of chattering induced by the high-speed switching. Relationships to the theory of uncertain systems are also pointed out and dis-

0018-9219/88/0300-0212$01.00 0 1988 IEEE

212

PROCEEDINGS OF THE IEEE, VOL. 76, NO. 3, MARCH 1988

cussed along with a brief review of the recent VSC applications literature.

II.

BACKGROUND

U-

-1

U-

I1

The term "variable structure control" arises because the "controller structure" around the plant i s intentionally changed by some external influence to obtain a desired plant behavior or response. For example, consider a plant with two accessible states and one control input as described by the following state equations

s1 "small",

U = U I

A block diagram representation of (2.1) appears in Fig. 1. Let the so-called switching surface be u ( x l , x2) = S I X l + x2 = 0

(a)
x2

I
Relay

1
U-

t1

Fig. 1. Block diagram of a second-order system described by (2.1).

and the control law be given by (2.2) where

1
sgn (a) =
-1

u > o
U <

Fig. 3. Phase-planediagramsoftheclosed-loop system (2.1), (2.2) for different values of s,.

0.
our system on u = 0 is dependent only on the slope s1 of the switching line. This means the system i s insensitive to any variation-or perturbation of the plant parameters contained in the bottom row of the A matrix of (2.1), i.e., perturbations in the image of the " B matrix"[O I]'. This i s one dominate motivation for investigating variable structure systems. The motion of Fig. 3(b) is more complex. Here the state trajectory switches to a new parabolic motion every time it intercepts the switching line u = 0. Nevertheless, the parabolic motions "spiral" into the origin. As a second example consider a plant with two accessible states and one control input of the form U = k l ( x l , x 2 ) x 1+ k 2 ( x 1 ,x z ) x 2 where the gains k , ( x l , x,) take on two possible values, say a,or p,. To specifically illustrate the idea, consider the state model

A block diagram of the closed-loop system is depicted in Fig. 2. Let us now investigate the behavior of the system for

Fig. 2. Block diagram of the closed-loop system with the control strategy (2.2).

different values of the parameter sl,i.e., for different switching surfaces u.The phase-plane plotsof the system (2.1) with control law (2.2) are given in Fig. 3. Fig. 3(a) shows phaseplane trajectories for small s1 > 0 while Fig. 3(b) illustrates those for large s1 > 0. Upward motion in the trajectories is associated with U = + I and downward motion for U = -1. Suppose the relay element in the block diagram of Fig. 2 has a small delay when switching between the gains "1" and "-1". Consider the resulting system behavior as this delay tends to zero and s1 is small. The behavior of this second-order system on the switching line u = s l x l x2 = 0 (Fig.3(a))isdescribed equation x1 s l x l = 0 . It is important to note, that the behavior of

under the variable structure control law

u(t) = k ( x , ) xAt)

where k ( x l ) can be "-2" or "3". This system illustrated i n Fig. 4 has two linear structures, one each for k ( x l ) = 2 and k ( x l ) = -3. With k ( x l ) = -3, the system has complex eigenvalues and with k ( x l ) = 2, the system has real eigenvalues. With the switch in the upper position, the feedback pro-

DeCARLO er al.: C O N T R O L OF NONLINEAR MULTIVARIABLE SYSTEMS

213

Switching of course i s not random. It occurs with respect to a sliding or switching surface, generically denoted as U = 0. To illustrate this notion, consider the surface defined as U = ul(xl, = slxl x, = 0 with s1 > 1. If the feedback x,) is switched according to

k(x1) =

-3,

if ul(xl, x,)xl if ul(xl, xz)xl

2,

>0 <0

a behavior illustrated in the phase plane plot of Fig. 6 results.

Fig. 4. Block diagram of a second-order system with variable structure control.

duces an unstable free motion satisfying

p] [ '1 p]
=

-2O 2

x,

>1

as shown in Fig. 5(a). With the switch in the lower position, the feedback becomes positive and the system's free motion satisfies

XP

+ s1 x 1 = 0

E:] [" '1 p].

=

3 2

x2

Fig. 6. Path (dotted line) of state trajectory of system with control when perturbed slightly below the asymptote.

The "unstable" equilibrium point (0,O) i s now a saddle point with asymptotes x2 = 3x1 and x, = -xl, as shown in Fig. 5(b).

"'I
1

Observe from the dotted line trajectory that if the state vector i s perturbed below the surface, ul(xl, = slxl x, = x,) 0, at time to,it circles to the point tl before intercepting the surface again. On the other hand, if the switching surface is u2(xl,x,) = slxl x, = 0 with s1 < 1, then a perturbation off the surface i s always immediatelyforced back to the surface since the phase-plane velocity vectors always point towards the surface. Fig. 7 illustrates the phenomena.

x1 2)=x2+s1x1= SI<

x1

+ XI=

Fig. 7. Phase plane plot of system with control when a sliding mode exists on the switching surface.

(b) Fig. 5. Phase plane plots for feedback structures (a) kl(xl) = -3, and (b) k,(xl) = 2.

As suggested by Figs. 3, 6, and 7, different choices of switching surfaces produce radically different system responses.The richnessof variablestructurecontrol comes from this ability to choose various controller structures at different points in time. The above example also illustrates an important notion

214

PROCEEDINGS OF THE IEEE, VOL. 76, NO. 3, MARCH 1988

in VSC. For the switching surface, u2(xl, x = slxl + x2 = , ) 0 of Fig. 7, once the state trajectory intercepts the surface it remains on the surface for all subsequenttime. This property of remainingon the switching surfaceonce intercepted i s called a sliding mode. A sliding mode will exist for a system i f in the vicinity o f the switching surface, the state velocity vector (the derivative o f the state vector) is directed towards the surface. The lack of a sliding mode for the "a = u1 scenario" described in the second example disappears when using a full state feedback control law: u(x) = kl(xl, x2)x1+ k,(xl, x2)x2. With appropriate gain choice, the original system can always be forced to have a sliding mode on any surface U = slxl x2 = 0. It was the choice of partial state feedback ( U = kl(xl, x2)xl) which prevented the existence of a sliding mode on the surface ul = slxl x2 = 0 with s, > 1. Insuring the existence of a sliding mode on the switching surface i s a key necessity in VSC design. Designing the proper surface i s the complementary key problem. Thus VSC design breaks down into two major phases. The first i s theconstruction of the switching surface so that the original system or plant restricted to the surface responds in a desired manner. The second phase entails the development of a switching control law (i.e., appropriate switched feedback gains) which satisfies a set of "sufficient conditions" for the existence and reachability of a sliding mode [I], 1 1 1 1 [ I l l , [W, [151, [191-[211, [24l, [291, [301, 1 7 . 2, 8, 71

[14], [82]. Moreover, for a large class of systems, design of linear switching surfaces proves amenable to classical linear controller techniques. Thus for clarity, convenience, and simplicityof exposition, this tutorial will focuson linear switching surfaces of the form 4x1 = Sx(t) = 0 where S i s an m x n matrix. (3.4)

Sliding Modes
After switching surface design, the next important aspect of VSC i s guaranteeing the existence of a sliding mode. A sliding mode exists, i f in the vicinity o f the switching surface, a(~)= 0 the tangent or velocity vectors o f the state , trajectory always point toward the switching surface. Consequently, if the state trajectory intersects the sliding surface, the value of the state trajectory or "representative point" remains within an E neighborhood of {xla(x) = 0). As a point of information, i f a sliding mode exists on a(x) = 0 then a(x) is termed a sliding surface. As seen in Fig. 8, a ,

Ill. DEFINITIONS PRELIMINARIES AND

System Model This paper considers a class of systems having a state model nonlinear in the state vector x(-) and linear in the control vector U ( * ) of the form X(t) = f(t,
XI U )

= f(t, x)

+ B(t, x) u(t)

(3.1)

where the state vector x(t) E R", the control vector u(t) E R m , f (t, x) E R", and B(t, x) E R" '"'; further, each entry in f (t, x) and B(t, x) i s assumed to be continuous with continuous bounded derivative with respect to x. Each entry ui(t)of the switched control u(t) E R m has the form u;(t, x) = u:(t, u;(t, x) with q(x) x) with uj(x)

>0

-= 0

i= 1 ,

,m (3.2)
Fig. 8. A situation in which sliding exists only on the intersection of the two sliding surfaces.

where ai(x) = 0 is the ith switching (also called discontinuity) surface associated with the (n - m)-dimensional switching surface a(x) = [aq(x), The Switching Surface The switching surface a(x) = 0 i s a (n - mbdimensional manifold in R" determined by the intersection of m (n 1)-dimensional switching surfaces aj(x) = 0. The switching surfaces are designed such that the system response restricted to u(x) = 0 has a desired behavior such as stability or tracking. Switching surface design i s taken up in a later section. Although general nonlinear switching surfaces (3.3) are possible, linear ones are more prevalent in design [I], [6],

* *

, um(x)lT 0. =

(3.3)

sliding mode may not exist on u,(x) = 0 separately, but only on the intersection. An ideal sliding mode exists only when the state trajectory x(t) of the controlled plant satisfies cr[x(t)] = 0 at every t 2 tofor some to. This requires infinitely fast switching. In actual systems, all facilities responsible for the switching control function have imperfections such as delay, hysteresis, etc., which force switching to occur at a finite frequency. The representative point then oscillates within a neighborhood of the switching surface. This oscillation i s called chattering. If the frequency of the switching is very high compared with the dynamic response of the system, the imperfections and the finite switching frequencies are often but not always negligible. Hence our subsequent development considers primarily ideal sliding modes. The

DeCARLO et al.: C O N T R O L OF NONLINEAR MULTIVARIABLE SYSTEMS

215

problem of chattering and techniques for circumventing the problem are discussed in a later section of the paper. Conditions for the Existence o f a Sliding Mode Existence of a sliding mode [I]-[3], [5] requires stability of the state trajectory to the sliding surface a(x) = 0 at least in a neighborhood of {x la(x) = 0}-Le., the representative point must approach the surface at least asymptotically.The largest such neighborhood is called the region ofattraction. Geometrically, the tangent vector or time derivative of the state vector must point toward the sliding surface in the region of attraction [I], [8]. For a rigorous mathematical discussion of the existence of sliding modes for such systems see [ ] [8], [Ill, I[2], , [45], [&I. These types of systems are referred to as discontinuous systems in the literature. The existence problem resembles a generalized stability problem, hence the second method of Lyapunov provides a natural setting for analysis. Specifically, stability to the switching surface requires selecting a generalized Lyapunov function V(t, x) which i s positive definite and has a negative time derivative in the region of attraction. Formally stated: Definition I [2]: A domain D in the manifold U = 0 i s a sliding mode domain if for each E > 0, there i s 6 > 0, such that any motion starting within a n-dimensional &vicinity of D may leavethe n-dimensional e-vicinity of Donlythrough the n-dimensional e-vicinity of the boundary of D. See Fig. 9.

ii) sup V(t, x, a) = Hw

llull = P

H, > 0

(3.6)

hold, where h and H, depend on p (hp # 0 if p # 0). , 2) The total time derivative of V(t, x, a) for the system (3.1) has a negative supremum for all x ECI except for x on the switching surface where the control inputs are undefined, and hence the derivative of V(t, x, U)does not exist. I . Proof: The proof i s given in [ ] A sliding mode i s globally reachable if the domain of attraction i s the entire state space. Otherwise the domain of attraction is a subset of the state space. Thestructureof thefunction V(t, x,a)determines theease with which one computes the actual feedback gains implementing a VSC design. For poorly chosen Lyapunov functions, the feedback gain computations can be untenable. For all single input systems a suitable Lyapunov function is V(t, x) = .5a2(x)which clearly is globally positive definite. In VSC, uwill depend on the control and hence if switched feedback gains can be chosen so that
.s-=u-<o

du2
dt

da dt

(3.7)

in the domain of attraction, then the state trajectory converges to the surface and is restricted to the surface for all subsequent time. The feedback gains which would implement an associated VSC design are straightforward to compute in this case [I], 1 1 [81, [Ill, 2, [MI, 1291. Illustrative Design Examples To illustrate the single input VSC design procedure consider the single pendulum system of [22], [23] having nonlinear state model x(t) = A(x) x(t) Bu(t) where x = [xl, x2IT

- vicinity of boundary point of D

6

roi
L
X I

_I
if a(x)xi if o(x)xi

boundary p o i E o f D
X,

with control u(t) = kl(x)x, k;(x) =

+ k2(x)x2where
>0 <
0

ai(x), pi(x),

and a(x) = [sl s2]x.The feedbacks aj(x)and &(x) are chosen so that ~ ( xir(x) < 0. Some simple substitutions show that )

Fig. 9. Two-dimensional illustration of domain of sliding mode. Since the region D lies on the surface ~ ( x= 0, dimension ) [D] = n - m. Hence: Theorem 1 For the n - m-dimensional domain D to be : the domain of a sliding mode, it i s sufficient that in some n-dimensional domain Q 3 D, there exists a function V(t, x, a) continuously differentiable with respect to all of its arguments, satisfying the following conditions: 1) V(t, x, U) is positive definite with respect to U , i.e., V(t, x, U ) > 0 with U # 0 and arbitrary t, x, and V(t, x, 0) = 0; and on the sphere 1u1 = p for all x E Q and any t the relations 11

if
al(x) = a1 < min
x1

LsinxlxI

-=

-1

x, [inxxl] pl(x) = p1 > max - = 1

i) inf V(t, x, a)
IIUII = p

h ,

h >0 ,

(3.5)

and a2 < -(s1/s2) and p2 > -(s1/s2). Hence, computation of the feedback gains i s straightforward. For multivariable systems, useful Lyapunov functions prove difficult to find, except in special cases [I], [21. Specifically: 1) Suppose there exists a positive definite, symmetric transformation W(t, x) such that R(t, x) = - W(t, x) SNt, x)

216

PROCEEDINGS OF THE IEEE, VOL. 76, NO. 3, MARCH 1988

where R = [rij] has the property rii > EY=l,j+i (rijl for i = 1, , m, i.e., R(t, x) is diagonally dominant. If so, the recommended formof V(t,x, a) isaquadratic in a(t,x)withcoefficients depending on t and x. 2) Suppose SB(t, x) i s symmetric. The recommended form here i s V(t, x, U ) = a'Ru where R i s symmetric and diagonalizes SB. 3) Suppose SB(r,x) is diagonallydominant. Here the recommended form of V(t, x, a) is a simple quadratic V(t, x, a) = a(x)' W(t, x) a(x) where W(t, x) i s a nonsingular diagonal matrix. 4) Finally suppose SB(t, x) i s diagonal. The recommended form of v(t, x, a) i s V(t, x, a) = u'u. For these cases, the necessary feedback gains needed to implement a VSC design are straightforward to compute. For other cases, one usually executes some type of transformation to obtain such a form. Section VI provides a detailed discussion of case 4) above. To illustrate that a poor choice of a Lyapunov function for a particular sliding surface combined with a naive choice of a control law may lead to extreme difficulties in solving for the necessary control gains, consider the hypothetical multivariable system

...

Case 3: xlal Case 4: xlul

<0 <0

and xlq and xlq

>0 < 0.

After some calculations, one concludes that

i) 9.2:C O

ii)

*i > 0

iii) 1

> 4; >

+ 24;.

Given i) and ii), it i s not possible to satisfy iii). Because of this and the obvious difficulty of solving (3.12) directly, the use of the Lyapunov function V = .5aro and the control law (3.10) make this problem more difficult than necessary. By using a different Lyapunov function or another control law it is relatively straightforward to compute the control gains. For example, let

v = .5aJu
and
u(x) =

-(Sf?)-' SAX - (SB)-' 2 I #' ?

x(t) = AX@) t Bu(t)

(3.8a)

where
0
A = [

This control forces o r b = -1 < 0; thus a sliding mode exists on the sliding surface (3.9) and is reachable for all x E R 3 . Also, a different sliding surface will work, for example if
2 E

1 0

0 1
-3-

-1

0 -1

s-[:

-1

-2

with the sliding surface

where E i s a small positive constant, it is possible to solve for the gains *$ since SB i s diagonally dominant. Lastly, we point out that using a multi-input diagonalization method, described in a later part of the paper, controller design for this problem i s easily accomplished. Design Procedure Overview

and the control law u(t) = *x(t) where \k =

[$ij]

and

f o r i = 1 , 2 and j = 1 , 2 , 3. Observe that

a r b = (1 -

*lq

+ 2*',,)x,a,

+ (-1 + 9 1 1 - *2l)XlUZ
422)XZU2

+ (3 - 4 1 2 + 2 4 2 2 ) X 2 f f l + (*12 + (5 - 9 1 3 + 2423)xg~l+ ( - 4

From the above discussion it becomes clear that VSC design breaks down into two phases. Phase 1 entails constructing switching surfaces so that the system restricted to the switching surfaces produces a desired behavior. Phase2 entails constructing switched feedback gainswhich drive the plant state trajectory to the sliding surface and maintain it there. The actual details of this procedure are developed in Sections IV through VII. IV. EXISTENCE SYSTEMS
AND

+ *I3

- Q23)xj~p (3.11)

UNIQUENESSSOLUTIONS OF

TO

VSC

A sliding mode exists if (3.11) i s negative in the domain of attraction. In the single-input case this i s usually accomplished by making each term in the sum negative. By considering just the first two terms in (3.11), it i s necessary to simultaneously satisfy (1 (-1
911

+ 2*21)XlI71

< <

0 0.

+ 4 1 1 - *'21)Xlff2
and xlq

(3.12)

There are four cases to consider. Case 1: xlal Case 2: xlal

>0
>

>0
<0

0 and xla2

VSC produces system dynamicswith discontinuous righthand sides due to the switching action of the controller. Thus they fail to satisfy conventional existence and uniqueness results of differential equation theory. Nevertheless an important aspect of VSC design i s the presumption that the plant behaves in a unique way when restricted to a(x) = 0. Therefore the problem of existence and uniqueness of differential equations with discontinuous right-hand sides is of fundamental importance. Various types of existence and uniqueness theorems can be found in [I], [8], [35]. Supplementary material also exists in [52], [53]. However, one of the earliest and conceptually straightforward approaches i s the method of Filippov [46]. We will briefly review this method as background to the

DeCARLO et al.: C O N T R O L OF NONLINEAR MULTIVARIABLE SYSTEMS

217

above referenced results and as an aid in understanding variable structure system behavior on the switching surface. Consider the following nth order, single input system
X(t) =

The Method of Equivalent Control The method of equivalent control i s a means for determining the system motion restricted to the switching surface a ( ~= 0. Suppose at to,the state trajectory of the plant ) intercepts the switching surface and a sliding mode exists for t 2 to. The existence of a sliding mode implies 1)u(x(t)) = 0, and 2) a(x(t)) = 0 for all t 2 to. From the chain rule [au/ax]x = 0. Substituting for x yields

f (t, x, U)

(4.1)

with the following general control strategy

U

+ ( t ,x),

if u(x)

>0
(4.2)

u-(t, x),

if u(x) c 0.

It can be shown from Filippov's work in [46] that the state trajectories of (4.1) with control (4.2)on ~ ( x= Oare the solu) tions of the equation (see Fig. IO)
+

[E]x =

[E]

[ f ( t ,x )

+ B(t, X ) U e q ] = 0

x u

= af

+ (1 - a)f /

= fO,

0 5 a 51

(4.3)

where ueqis the so-called equivalent control which solves this equation. After substituting this ueq into (3.1), the motion of (3.1) describes the behavior of the system restricted to the switching surface provided the initial condition x(t,J satisfies u(x(to))= 0. To compute ueq,let us assume that the matrix product [aalax] B(t, x ) i s nonsingular for all t and x. Then

Therefore, given u(x(to))= 0, the dynamics of the system on the switching surface for t 2 to i s given by
x = [I

- B(t, x )

[E

-1

B(t, x)]

E]

f ( t , x).

(5.2)

In the special case of a linear switching surface u(x) = Sx = 0, a d a x = S, (5.2) reduces to

x = [l B(t, x ) [SB(t, x)l-'Sl f ( t , x). rr-0

(5.3)

Fig. 10. illustration of the Filippov method of determining thedesired velocityvector f'for motion in the sliding mode.

where f = f (t, x , U +), f - = f (t, x, U -), and f o is the resulting velocityvector of the state trajectory in a sliding mode. Solving the equation (du, f ' ) = 0 for a yields
+

a=

(do, f - ) (da, ( f - - f +))

+

provided 1) (du, ( f - - f ')) > 0, and 2) (du, f ) 5 0 and (du, f - ) 0 where the notation (a, b ) denotes the inner product of a and b also written as "a * b", and du = grad

44.
Therefore one may conclude that on the average, the solution to (4.1) with control (4.2) exists and is uniquely defined on a(x) = 0. Notice also that this technique can be used to determine the behavior of the plant in a sliding mode.

This structure can be advantageously exploited in switching surface design. Observe that (5.2) in conjunction with the constraint a(x) = 0 determines the system motion on the switching surface. As such, the motion on the switching surface will be governed by a reduced order set of equations. This order reduction comes about because of the set of state variable constraints, u(x) = 0. The remaining parts of the section will describe 1) how one determines a reduced order set of dynamical equations governing the system motion on the switching surface, and 2) howtochoose surface parameterssfor a linear switching surface u(x) = Sx = 0, so that the system in a sliding mode exhibits the desired behavior. Before closing this subsection, the reader should note that some control applications require a time-varying switching surface u(t, x ) = 0. In this case, U(t, x ) = (adat) (au/ax)x and the equivalent control takes the form

V. SLIDING SURFACE DESIGN

Filippov's method i s one possible technique for determining the system motion in a sliding mode as outlined in the previous section. In particular, computation of f o represented the "average" velocity ( x ) of the state trajectory restricted to the switching surface. A more straightforward technique easily applicable to multi-input systems i s the method of equivalent control, as proposed by Utkin in [ I ] , [4], and Draienovit in [77.

ueq =

-[E

-1

B(t, x ) ]

[E + "1
f(t, x)

at

(5.4)

For simplicity of exposition, we have avoided the added complexity of the time-varying surface throughout most of the paper. Generalizations incorporating a time-varying component of a(t, x ) are straightforward to construct. The time-varying surface structure will appear briefly when discussing diagonalization methods in Section VI and more commonly when discussing uncertain systems and VSC in Section VII.

218

PROCEEDINGS OF THE IEEE, VOL. 76, NO. 3, MARCH 1988

Reduction o f Order For sake of clarity, we concentrate on the case of linear switching surface, u(x) = Sx = 0. As mentioned above, in a sliding mode, the equivalent system must satisfy not only the n-dimensional state dynamics (5.2), but also the "m" algebraic equations, a(x) = 0. The use of both constraints reduces the system dynamics from an nth-order model to an (n - m)th-order model. Specifically, suppose the nonlinear system of (3.1) i s restricted to the switching surface of (3.4), i.e., a(x) = Sx = 0, with the system dynamics given by (5.3). Then it i s possible to solve for m of the state variables, in terms of the remaining n - m state variables, if the rank [SI = m. The condition that rank [ S I = m holds under the earlier assumption that [aalax] B(t, x ) is nonsingular for all t and x. To obtain the solution, solve form of the state variables (e.g., x,-,+~, , x,,) i n terms of the n - m remaining state variables. Substitute these relations into the remaining n - m equations of (5.3) and the equations corresponding to the m state variables. The resultant (n - m)th-order system fully describes the equivalent system given an initial condition satisfying d x ) = 0. Example 5.4:To clarify the above procedure and to pave the way for later examples consider the system x ( t ) = A(t, x ) x(t) BuW, where

(SB)-' =
This produces
X(t) = [I

(5.9)

- S(SS)-lS] A([, X ) X ( t )
1 0 0 0

(5.10)

subject to a(x) = 0, in which case

[' '1 1";1

1 1 xg

:I[-

s22 s12

s24 '14]

111.
r - l

(5.11)

x4

Solving for x3 and x5 yields

- -

The reduced order equivalent linear time-invariant system is

(5.5)

where i1= x l , i2= xZl 2, = xq. To see how control design might be accomplished, suppose a design constraint requires the spectrum of the equivalent system be { -1, -2, -3}; the desired characteristic polynomial is

Assume the third and fifth rows of A(t, x ) have nonlinear time-varying entries which are bounded: ayin 5 aij(t,x ) 5 a y for all x E R" and t E [tola . ) The method of equivalent control produces the following equivalent system (as per (5.3))
X(t) = [I - S(SS)-'S] A(t, X ) x(t)

aA(h)= h3

+ 6X2 + 11X + 6.

The characteristic polynomial of the equivalent system given in (5.13) is

a,4(X) =

(5.6)

x3 + (SI2 - s22 + 2s24 - s14)X2

provided a(x(ro)) = 0 for some to. If the linear switching surface parameters are
s11 s21 s12
s22

+ (s12s24 - s14s22 + s11 - s2l)X + (s11s24 - s14s21).

Equating coefficients of like powers of X produces the set of equations
(5.7)

513

s14 s24

'151
s25

s23

then

s12

SB =

1: l:].
r
-7

(5.8)

~ 2 4 -sZ

-1
SZ

To simplify the example let us choose ~ 1 3 . ~ 2- s15s23 = 1. 5 Specifically, choose ~ 1 = 2, s15 = ~ 2 = sZ5 = 1. Then 3 3

DeCARLO

et al.: C O N T R O L OF NONLINEAR MULTIVARIABLE SYSTEMS

219

One solution accomplishing the control design objective is S=[

face consider
a(x) = a,(x,)

1 1.8333 2 - 6 1 1.8333 1

0 1

1-

+ S2x, = 0

(5.19)

In conclusion, the reduced order equivalent system with the desired eigenvalues i s = An, where

which i s linear in x2 and possibly nonlinear in x,. For this case, the reduced order dynamics in a sliding mode will have the form x, = f,(t, XI, -S;lu,(x,)).
(5.20)

A=

io
0 0
L-1
-1.83333 -61

This example worked out so cleanly because the original system dynamics were given in the Luenberger canonical form. Systems not in this form often require a transformation to a more general form called the regular form [28]. Regular Form and the Reduced Order Dynamics The regular form of the plant dynamics (3.1), is
x, = fl(t, x)
Xi(z

An example of designing a nonlinear switching surface will be given later. The next important question i s how one transforms the given system dynamics (3.1)to the regular form of (5.14). We first consider the case of a linear switching surface of (5.15) and a nonsingular linear time invariant transformation z = Tx. Taking the time derivative of z yields z = Tx = T f ( t , X ) If it i s true that
(5.22)

+ TB(t, X ) U .

(5.21)

= f&t,

X)

+ B2(tr X ) U

(5.14)

then in the new coordinates the dynamics of the plant (3.1) become
Z, = f,(t,
2,

where x1 E R"-m and x, E Rm. A system in this form has simply computed reduced order equivalent dynamics, also referred to as the system equations of slow motion. The computation of this form assumes BZ(t,x ) i s an m x m nonsingular mapping. This assumption i s necessaryforthe existence of the equivalent control. To compute the reduced order dynamics, assume a linear switching surface (this will be generalized later) of the form
a(x) = [S, S,]

z)

= f2(t, z)

+ B,ct, z)u.

(5.23)

Hence in a sliding mode the equivalent reduced order dynamics are given by (5.17) modulo the coordinate change, i.e., z, =

p]

il(t, Z,, -S;'S,Z1)

(5.24)

= 0.

(5.15)

where = [SI & I T - ' . If there i s no linear transformation such that (5.22) i s satisfied, then one must resort to nonlinear transformations of the form z = T(t, x ) =

[s, s2]

Without loss of generality assume S2 i s nonsingular. Thus in a sliding mode

x, = - s ; ' s 1 x ,

(5.16)
e ) e):

[;:
e):

I:
+

(5.25)

and
x, =
fl(t,

x) = f,(t, XI, -S;'S,x1)

(5.17)

which i s the reduced order dynamics. Observe that if f, has the very desirable linear structure
XI

where 1) T ( . , i s a diffeomorphic transformation, 2) T,(., R x R" + R"-"', and T 2 ( * , R X R" Rm. Diffeomorphic [66], [67l means that there exists a continuous differentiable inverse mapping T(r, z) = x satisfying T(t, 0) = 0 for all t. Differentiating z in (5.25) with respect to time produces
z = - (t, x)X aT
ax

= fi(t,

X)

= AiIx,

+ A12~2
(5.18)
Z =

+ - (t, X I . aT
at

(5.26)

then the reduced order dynamics becomes

XI

= [A,,

- A12S;1S1]~1

Substituting (3.1) into (5.26) yields

-f(t, X)
aT ax aT aT + - B(t, X ) ~ ( t+) -. ax at

which has the feedback structure "All + A12F" with F = -S;'S1 and A,, playing the role of the input matrix. If the pair (A,,, A,,) is controllable, then it is possible to effectively use classical feedback control design techniques to compute an F such that A l l AI2F has desired characteristics. Having found F, one can compute [S, Sd such that F = -S;'S1, thus completing the switching surfacedesign. Note that one can use pole placement techniques, linear optimal control techniques, etc., to design F. For more details of the linear case see Young et a/. [25] and El-Ghezawi et a/. [14]. For the more general case of a nonlinear switching sur-

(5.27)

If the transformation has the property that

ax

B(t, x ) =

B(t, X ) =

(5.28)

LaxJ

then in the new coordinates, the equations describing our system or plant are

220

PROCEEDINGS OF THE IEEE, VOL. 76, NO. 3, MARCH 1988

Thus if the entries

U*+

and

U*-

are chosen to satisfy

qi(t,X)

U;+

< -vu;(x) f ( t , X)
=

-igls j j f i ( t , x)
-Vu;(t) f ( t ,x)

when ui(x) > 0

(6.5a)

i2 ?,U, z) + s^,(t,z)u.

(5.29)

9;(t,x)

U;-

>
=

The problem of converting a nonlinear system to a canonical form, in particular the regular form, was explored in 1281, 1661, 1671 among others.
VI.

-jFl s i i f i ( t , x)

when ui(x)

<

(6.5b)

CONTROLLER DESIGN

Controller design is the second phase of the VSC design procedure mentioned earlier. Here the goal i s to determine switched feedbackgains which will drive the plant state trajectory to the switching surface and maintain a sliding mode condition. The presumption is that the sliding surface has been designed. I n general, the control is an m-vector u(t) each of whose entries have the structure of the form

then sufficient conditions for the existence and reachability are satisfied where sji equals thej-entry of Vu,(x) which is the ith row of (adax). In particular, the conditions of (6.5) force each term in the summation of arbto be negative definite. As mentioned, the control actually implemented i s

u(t) =

c:

- B(t, X)

)-I

Q(t,X) u*(t).

(6.6)

Other sufficient conditions for the existence of a sliding mode can also be used. The second method of diagonalization requires a nonsingular transformation of U rather than the control U . In particular, consider the new switching surface
U*&

where 4x1 = [ul(x),*

. , um(x)lJ = 0.

x) =

nct, x) u(x) = 0

(6.7)

Diagonalization Methods Our purpose hereistodescribetwodifferentapproaches to controller design labeled in the literature as diagonalization methods. The essential feature of these methods i s conversion of a multi-input design problem into m singleinput design problems. Method 1entails constructing a new control vector U* via a nonsingular transformation, Q-'(t,x) [aulax] B(t, x) of the original control U defined as

for an appropriate transformation Q(t, x). This method i s based on the fact that the equivalent system i s invariant to a npnsingular switching surface transformation as verified in the following theorem. Theorem 2 [?I: Suppose that the original system is given by (3.1), (3.2) with switching surfaces uj(t, x) = 0 (i = 1, , m), then the sliding motion (trajectory of the equivalent system) i s invariant to the transformation of the switching surface u*(t, x) = h(t, x) u(t,x) = 0 E R m , if llhll and llfl-'l/ are bounded for all t, x E S E R x R". Proof: First, by the method of equivalent control
a . .

where Q(t, x) i s an arbitrary m x m diagonal matrix with elements qi(t,x) (i = 1, . . , m) such that inf Iqi(t,x)) > 0 for all t 2 0 and all x. The actual conversion of the m-input design problem to "m" single input design problems is accomplished by the [adax] B(t, x)-term with the diagonal entries of Q-'(t, x) merely allowing flexibility in the design, for example by weighting the various control channels of
U*.

U*

au n(t, X) - (x) ( f (t, X) + Bu,,) ax

au + Cl - 00 = 0. at

(6.8)

Since O(t, x) i s a nonsingular m x m matrix

(6.9)
U*

Often Q(t,x) i s chosen as the identity. In terms of state dynamics become

-1

the which differs from U,, only by the term ((adax) B ) - ' Q - ' hu. However, on the switching surface, U = 0, thus

x(t) =

f ( t ,x)

+ B(f, x ) [ E (x) B(t, XI] Q(f,x) u*(t).

(6.3)

Although this new control structure looks more complicated, the structure of U(x) = 0 permits one to independently choose the m-entries of U* to satisfy the sufficient conditions for the existence and reachability of a sliding mode. Once U* i s known it can be unraveled by inverting thetransformation to yield the required u.To see this, recall that for existence and reachability of a sliding mode it i s enough to satisfy the condition uT(x)U(x) < 0. In terms of
U*

(6.10) Hencethe equivalent systems are identical and the motions in the sliding mode coincide. Loosely stated, Theorem 2 says that the motion in the sliding mode is independent of a nonsingular possibly timevarying transformation of the switching surfaces. Observe that any nonsingular transformation fi with bounded derivatives will produce the same "equivalent" system. In this second diagonalization procedure, we select

U(X) =

- (x) f (t, X) + Q(t,X) u*(t).

aa ax

(6.4)

DeCARLO er al.: CONTROL OF NONLINEAR MULTIVARIABLE SYSTEMS

221

Wt, x ) so that Wt, x ) @a/ax)(x) B(t, x ) i s a diagonal matrix, say

Q ( t , x ) = diag[q;(t,x)]whoseentriesare bounded awayfrom zero. Specifically select Q(t,x ) as
-1

design employs method 1 which transforms the control U as per (6.2)

u*(t) = Q-'(t, x ) SB(t, X ) u(t)

(6.17)

Nt, x)

= Q(t, x)[$

(x) B(t, x,]

(6.11)

where Q(t,x) i s a nonsingular diagonal matrix such that inf Iqi(t, x)l > 0. For simplicity choose

for appropriate Q(t,x). Again Q(t,x ) is often chosen as the identity matrix. In order to determine the existence and reachability conditions it i s necessary to compute U* as

[:

Q-' =

[i:d.

(6.18)

+ Q ( f , X ) U + o(t,X ) n - ' ( t , x)a* A = ao(t,X ) + Q(t, X ) U + a& x)

*

The choice was random. However, the diagonal entries of Q can be chosen to weight different control channels or to compensate somewhat for the "distortion" introduced by [SB(t, x)]-'. As per (6.3), the state dynamics driven by U* are
x(t) = ~

f(t, x)

( tX ), ~ ( t+ ~ ( t ), [ s B ( ~ x1i-l Q(t,x ) u*(t). (6.19) ) x ,

(6.12)

where again the control term U "enters" U* via the diagonal matrix Q ( t , x). Sufficient conditions for reachability/existence of sliding mode are met if for any point in the state space and for all t a to,$ ( t , x ) and $ ( t , x ) are of opposite sign. Specifically, this requires that

In computing the feedback gains to meet the existence conditions, (6.4) becomes
ir(t) = Sx(t) = SA(t, x ) x(t)

+ Q(t,x ) u*(t, x).

si51 A(t, X) x(t),
= 1, 2

Since Q(t,x ) i s diagonal, using (6.5), sufficient conditions for the existence of a sliding mode are

q;(t,X ) U < :

-[s;l,

* *

.
*

q,(t,x ) U +
qi(t, x ) U;

- u f ( t , x ) - ajn(t,x),

for for

U:

>o
0
(6.13)

if

U;
*

> 0, i

>

-af(t, x )

- a&,

x),

U:

9;(CX )

U;

>

-[sill

, si51 A(t, X ) x(t),

(6.20)

where a,& x) is the ith entry of

U&

if a; < 0, i = 1 2. ,
X)

X)

A = B(t, X)

n- l ( t ,

a*(t, x ) .

Example 6.74: To clarify the diagonalization methods, consider the system of Example 5.4 where

It follows tha for the first switching surface al(x) = ,s151x(t) = Slx(t)one has

...

[sll,

SIA(r, x ) x(f) = {(2all

+ aZ1)x1 + (1 + 2a12 + a22x2)x2

(6.21)

x(t) = ~ ( t , x(t) X)

+ Bu(t)

(6.15a)

+ (1.8333 + 2a13 + az3)x3 + (2a14+ a24)x4

+ (.al5 + aZ5 - 6 ) x 5 } .
Recall the assumption: arin 5 aji(t,x ) 5 a y , i = 1 , 2 , j = 1, , 5 . Under this assumption and control law U * = Kx where K = [ k p I, to satisfy the existence condition of (6.20), kli must satisfy the following:

and

kll

k2 1

i i

B =

1 0

(6.15b)
'13

= =

t>
[

k14

The surface a(x) = Sx = 0 was designed to have

k15

+ a y ) , if ulxl > 0 + a$n), if alxl < 0 if u l x 2 > 0 < -(I + 2al;"" + a r ) , > -(I + 2aEin + ag"), ifa1x2< o if (I1x3 > 0 4 . 8 3 3 3 + 2a;3"" + a F ) , if <0 -(1.8333 + 2 a p + agn), if 61x4 > 0 - ( 2 a r + a,","), > + a;"), if 01x4 < O if a l x s > 0 - ( 2 a y + a r - 6), if ~ 1 x 5 0. < (6.22) > -(2aZn + agn - 6),
-(2a,mlax

>

-(2a?;'"

. . . , sZ5]x(t) A Szx(f), one has =

S2A(t,x ) x(t) = {(all

With regard to the second switching surface, u ~ ( x= [ s ~ I , )

+ a24 x1 + (1 + a12 + aZ2x2)xz

(6.23)

The objective of this example is to illustrate phase 2 of the VSC controller design process using the first and second diagonalization methods described above. The first

+ (1.83333 + a13 + a23)x3 + (al4 + az4) x4 + (al5 + a Z 5 ) x 5 } .

222

PROCEEDINGS OF THE IEEE, VOL. 76, NO. 3, MARCH 1988

Let U: = K2x; to satisfy the existence conditions (6.201, k , must satisfy the following:
k1 2

The conditions for the existence of a sliding mode (6.13), for this example are

k2 2

I I
[

< >

-1/2(ar,a" -1/2(a;;'" -1/2(1

+ ay), + agin),

if q x l if q x l

>0 <o

qlu: <
q,u;

-4'(t,
-U%

XI,

i = 1,2

(6.29)

>

x),

k23

k24

k5 2

I I

+ a y + a y ) , if u2x2 > 0 -1/2(1 + a;;'" + ag"), if u2x2 < 0 -1/2(1.8333 + a y + a y ) , if ~ 2 x 3 0 > > -1/2(1.8333 + aT," + ag"), if <0 if ~ 2 x 4> 0 < -1/2(a;",ax + a?), > -1/2(a$" + a$"', if ~ 2 x 4 o < if ~ 2 x 5> 0 -1/2(ay + a y ) , > -1/2(a?'5'"+ ag"), if u2x5 < 0. (6.24)
< >
g2x3

where up(x, t) i s the ith component of uo(t, x) = Q(t, x) SA(t, x) x(t). . , 5. Since Let U = Kx where K = [k,], i = 1, 2, j = 1, ay'" a,(t, x) 5 a y , i = 1 , 2 , / = 1, . , 5 , then to satisfy (6.29), it is required that

kll =

k1 2

Summarizing, u*(t) = Kx(t) where the entries of K are specified by (6.22) and (6.24). Since u(t) = (SB)-'Qu*(t, x), the actual control i s u(t) = [U#), u2(t)] so that

k2 = 1

',

k2 2

I I
x1

r l

k13 =

k23

Lx5

k14

This completes the illustration of the first method of diagonalization. Attention now turns to the second method. Again assume that the switching surface design is complete. In the second diagonalization method, the objective is to decouple the controls by making a nonsingular transformation of the switching surface. The control components, the entries of U , now switch on u:(x) = 0, i = 1, 2, with U: (x) given by (6.7). To perform the diagonalization choose Q(t,x) according to (6.11) where (adax) = S. If Q i s chosen as per (6.18), then

k4 2

k15

k5 2

i i i i i i i i i i

< >

-ay, -a;",

if ulxl if ulxl if u2xl if u2xl

>0 <0 >0 <0

< -ay, > -ag", < >

c
-ay,

if u1x2 > 0

-a;", -(I -(I

if u1x2 < 0

> < > < >

+ay), + a;"),

if u2x2 > 0 if uzx2 < 0

-ay, -aT;",

if ~ 1 x 3 0 > if u1x3 < 0

-(1.8333 -(1.8333 -ay,

- p i n 14

+ay),

if ~ 2 x 3 0 > if ~ 2 x 3 0 <

+ az"),
if u1x4 > 0 if ~ 1 x 4 0 < if ~ 2 x 4 0 >

< >
< >

-ay, -a;",

if ~ 2 x 4 0 <

< > >

-(a;;"" - 6), -(a;" -(6 -(6 - 6),

if ulxs

>0 if u1x5 < 0

if ~ 2 x 5> 0 if ~ 2 x 5 0. <
(6.30)

+ay), + ag"),

nu, X) = Q(t, x)[SB(t,

=

XI]-'

[:

-1

2]

1 -1

= [-2

41-

(6*26)

To construct a controller meeting the existence conditions of a sliding mode consider the derivative of U* given by (6.12), noting that h = 0
u*(t, X) =

Q(t,x) SA@,x) x(t)

+ Q(t, x) u(t)

(6.27)

where

Q(t,x) SA(t, x) x(t)

+ a12x2+ a13x3 + a14x4 + (al5 - 6)x5 + 2(1 + a22)x 2 + 2(a23 + 1.8333)x3 + 2a2,x4 + 2(a25 + 6 ) ~ s
and Qu = [ul 2 ~ 2 1 ' .

Since Q(t, x) in (6.26) is constant, h = 0 and the above are also sufficient conditions for reaching the sliding surface. As a second rather important illustration of the first diagonalization method, recall the example developed in (3.8) through (3.11).This multi-input example demonstrated how a poor choice of Lyapunov function in conjunction with a naive choice of controller led to an inconsistent solution of the equations defining the switched feedback gains needed to drive the state trajectory of the plant (3.8)to the switching surface of (3.9). The use of a diagonalization method circumvents this difficulty by converting the problem to one where the intuitive choice of Lyapunovfunction actually will work. Using the first diagonalization method with Q = I and with B defined in (3.8b)and the switching surface S defined in (3.9), then as per (6.2), the new control has the form

(6.28)

DeCARLO et al.: C O N T R O L OF NONLINEAR MULTIVARIABLE SYSTEMS

223

where

kl,
ki,
SAX +

>0 if ujxi c 0.
if uixi
X1

(6.32)

To compute the values of these gains consider

a =
U*

= =

+ 3x2 + 5x3
- x , - 4x,

] .+I:[
(6.33)

This leads to
U101

(1

+ k l l ) ~ 1 x 1+ (3 + k12) ~ 1 x 2+ (5 + k 1 3 ) ~ 1 ~ 3
- 1) ~ 2 x 2+ k 2 2 ~ 2 ~+2 ( k 2 3 - 4 ) U 2 X 3 .

6 2 0 2 = (k21

Therefore, sufficient conditions for stability to the switching surface are

k $ c -1,
kll

if ulxl if ulxl

>0 <0

mode on u1 = 0. The third control u,drives the system along the intersection of the surfaces uI = 0 and u2 = 0 to the intersection of the first three switching surfaces. This hierarchy of controls i s continued until the last control U, drives the system to a sliding mode on the intersection of all the m switching surfaces. Design of the control entry u presupposes i) existence k of a sliding mode on ai = 0, j = 1, . ,k - 1for any possible value of the controls uk through U,, and ii) knowledge of the system structure in these sliding modes. Since all controls Uk, k c m, depend on the values taken on by the control U,,,,U, must precede the design of u,,,-~,u , - ~ , * * * , U. In addition, design of the control u2 presupposes the system structure obtained assuming a sliding mode exists on u1 = 0. This system structure results by replacing u1with the Utkin-Draienovik equivalent control uleq. Call the resulting system structure E. To determine E,consider (aul/ax) = 0, which, using (3.1), x implies that
9

k , > -1,

k12

k23

i i

k:2 < -3,

if u1x2 > 0 if ulx2 c 0

(6.35)

kG > -3,

k2: < 4,
k s > 4,

if if

~ 2 x 3 ~ 2 x 3

>0

< 0.

where bi is the j t h column of B(t, x). Note that this relationship requires (dullax)bl # 0. (In fact, it i s necessary to assume that (auj/ax) # 0 for all i.) bj Substituting (6.35) into (3.1) produces the equivalent system model

Thus in terms of the original plant controller x = f(t, x )

+ B(t,

x)

[I]
um

An implementation of the control can be accomplished using the structure of Fig. 11.

P x(O

Q
(sei
Fig. 11. Implementation of the controller of (6.34).

for an appropriately formed f (f, x ) and B(t, X I . Design of u3 presupposes the system structure obtained by supposing a sliding mode exists on U, = 0 for the system structure E. This implies a sliding mode exists on u1 = u2 = 0. Call the resulting structure E. Of course E 2 is specified by replacing u2 in E by uleq. I n general, uk+l is designed supposing a sliding mode exists on uk = 0 for the system structure C k - and hence on ai = 0, j = 1, . * , k . The new system structure is Ek and i s denoted by the dynamics

rUk+i
Before designing the first control U,,, it isclearly necessary to sequentiallydeterminethesetof equivalent systems {E, ~ 2 , . , ~ m - }. Given Em-, U, has gains chosen to satisfy 1 the reachabilityand existenceconditionsforasliding mode , on U = 0. After this, one presumes the system structure Em-2 and proceeds to find gains for U,,,- so that a sliding mode exists on U,- = 0 given the gains computed for U,. To see how the existence and reachability conditions are determined at the (k 1)-step, realize that a sliding mode exists and i s reachable on u + = 0 provided uk + is chosen k so that

Method of Control Hierarchy As an alternative to the diagonalization methods described earlier it i s often possible to define a hierarchy of controls. One then employs this control hierarchy in designing a controller. With this approach a hierarchy of control channels is established so that, for example, the first control u1 drives the system from an initial condition onto the surface u1 = 0. The second control then drives the system onto the intersection of u1 = 0 and 0, = 0, while u1 maintains a sliding

Jk+luk+l

(6.37)

224

PROCEEDINGS OF THE IEEE, VOL. 76, NO. 3, MARCH 1988

for ail values of form

uk+i

uk+2,

, U,.

Observe that uk+' has the

m-k

Let u = [Ul,u2, u J r = Sx = 0 where

1 2 1 1 0 1
= VUk+lfk

,=1

,z
+

Vuk+lbfUk+;

S = O 1 0 1 1 1 0 1 0 0

where b : i s the ith column of Bk(t, x ) and Vuk+' = (auk +,,/ax. To insure the existence of a sliding mode on uk+' = 0, (6.37) must hold for all U ; , i = k 2, * * , m. Specifically, this condition has the form
Vuk+lb$u;+l

Following the preceding algorithm, first find U, = S'Ax S'BU = 0 where S i s the first row of S. Hence '
S'A = [2, 5 , 2, 4, -1, -21,

d I:[
A

= s 2

S'B = [2, 1, I].

Solving for

U',

in terms of x, u2 and ugyields

<

min
Uk+2,"'rUm

Uleq =

-3[2, 5, 2, 4, -1, -21 x - i [ U Z , Ugl.

- i=k+2

Vuk+lb,k-kuj,

if uk+l > 0 (6.38a)

Inserting ul, into x = Ax where

+ Bu, yields x
0
-1

= A'x

+ B'u',

0
-2

0
112

0 -112

if uk+'

< 0.
(6.38 b)

0 0 - 1 - 2

'1

The maximal and minimal values in (6.38) indicate that uk+luk+l < 0 no matter which of the two values U : or U,: i s taken on by the components of the control U ; (i = k 2, ,m). Summarizing, we introduce a hierarchy of controls whereby u1 guarantees motion in a sliding mode along u1 = 0 for any value of u2, * . , U,. The second component u2guarantees motion along the intersection of u1 = 0 and u2 = 0 for any values of uj, * . * , U,,,, and so on. The significance of this method is that a sufficient condition for a sliding mode i s obtained from the sliding mode existence condition for the scalar case. Example 6.39:To clarify the method of control hierarchy, consider the state model

x = Ax

+ Bu

Next, solve u2 = S2A'x S2B1u' = 0 for u2,, in terms of x and u 3 , where S2A1 = [0, -1/2, 1, 1, -1/2, 0 , S2B1 = [1/2, 1 1/21, and u2, is given by
u2, = [O, 1, -2, -2, 1, O I X - u3.

where

Inserting u2, into x = A'x where

+ B'u'
0
0

0
A =

1 0

0 - 1 0 - -1
0 0
1 0
A 2 = i

1
1

0
1

0 0

0 -1

-- 2

Now that A', A', B', B 2 are known, the second half of the algorithm i s used to determine the controls. Starting with k = 2, (6.38) gives
S3b:u: S3b:u;

,1 :j
0
2
B2=

yields x = A2x

+ B2u2

< >

-S3A2x -S3A2x.

DeCARLO et al.: C O N T R O L OF NONLINEAR MULTIVARIABLE SYSTEMS

225

Since S3b: = 1 and S3A2= [0,1,0,1, -1, -21, to satisfy the , above inequalities it i s sufficient that
U; U;

< >

-S3A2x, -S3A2x,

if if

U, U,

> <

0 0.

sliding mode, i.e., the condition that ujuj < 0 when uj(x) # 0. Below are five possible discontinuous control structures for UN. I ) Relays with constant gains:

TO determine the second control we again apply (6.38) obtaining S2b:u; and S2b:u;

< >

min [-S2Ax
U3

- S2b:u3]

Observe that this controller will meet the sufficiency condition for the existence of a sliding mode since
uiuj = ajuj(x) sgn (ui(x)) < 0,

if u,(x) # 0.

max [-S2Ax - S2b:u3].

U3

2) Relays with state dependent gains:

UiN(X) =

From our previous calculations we have S2b: = 1/2, S2bi = 112, and S2A = [0, -112, 1 1 -112, 01. To satisfy (6.38) it is , , sufficient that

u2 =

y
U;

I
CY;

sgn (uj(x)),

uj(x) # 0,
U;(X)

mi(.) < 0

= 0,

<

-(2S2A1x -(2S2Ax
U,

>

+ U?), + U;),

if u2 > 0 if u2 < 0.

Again it i s straightforward to check that

u;u; = al(x) ul(x) sgn (u,(x)) < 0,

if u,(x)

z 0.

For an example of cq(x) consider q ( x ) = &(U;~(X)

+ 7 ; )with
>
0

Finally the first control Sblu: and Sblu;

must satisfy

Pi < 0, y i > 0, where k i s a natural number.

3) Linear feedback with switched gains:
uiN(x) = $x; $ =
[$;j], $,j

<

min [-SAx - Sb2u2 - Sb3u31

U2,U3

>

max [-SAx - Sb2u2 - Sb3u3]

U. U 3 ?,

with ai, < 0 and Pi,

ujuj =

>

0. Thus again

jj,

ujxj

Pij, uixl < 0

where B(t, x) = [b,, b2, b3]. Since Sbl = 2, Sb2 = 1 and SA , = [2, 5, 2, 4, -1, -21 to satisfy the above inequalities it i s sufficient that
U1

Uj($jX

+ $;2x2 +
=
CU;U;(X)

* *

. + $,nXn) < 0.

4)

Linear continuous feedback:

u;N(x)

=[
U:

< ul- >

-0.5(S1Ax -0.5(S1Ax

+ U; + U;), + U: + U ; ) ,

if u1 if u1

> <

0 0.

and a; < 0.

The condition for the existence of a sliding mode i s

ujuj =

In this example the control hierarchy i s u1 then u2 and then u3,but i s it the optimum order? Is there another order
in which thecontrol gains are smaller?Theanswer depends on the initial condition. Other Approaches In addition to the diagonalization methods and hierarchial control method mentioned above, other approaches are possible. In theory an infinite variety of control strategies of the form (6.1) are possible. An alternative structure for the control of (6.1) i s
uj = uj4

aju;(x)
=&.)I !

<

or more generally
uN(X)

where L E R m X mi s positive definite constant matrix. The condition for the existence of a sliding mode is easily checked uT(x)u(x) = -uT(x) Lu(x) < 0, if u(x)

+ 0.

5) Univector nonlinearity with scale factor:

+ UjN

(6.40)

The existence conditions are uT(x)u(x) = Ilu(x)II p

where uj4 i s the ith component of the equivalent control (which i s continuous) and where ujN the discontinuous is or switched part of (6.1). For controllers having the structure of (6.40), the following i s true:

< 0,

if u(x)

+ 0.

Given a nonlinear system, if a linear behavior is required in a sliding mode then one may need to use a nonlinear switching surface. A good control for this purpose i s the controller of (6.41). To illustrate the above point in the context of this control structure consider the following example. Example 6.42: Consider a simple robotic manipulator driven by a dc armature control dc motor [72] modeled by the following equations

Let us assume that (adax) B(t, x) = I the identity. Then ~ ( x ) , = U,.+ This condition allows an easy verification of the sufficiency conditions for the existence and reachability of a

[l:n(xl)

+ x3] +

x2 + x3

[!]

f(x)

+ Bu.

(6.43)

226

PROCEEDINGS OF THE IEEE, VOL. 76, NO. 3, MARCH 1988

For any switching surface u(xl, x x3) = 0, development , of the control requires [(adax) B] be nonsingular. The structure of B then implies that adax, # 0. Without loss of generality, we set adax, = 1. Hence a general structure for a nonlinear switching surface in this example (similarto (5.19)) is a(x) = x3

+ fJI(X,,

x) = 0. ,

Using a control structure of (6.40) and (6.41) yields

U

= ueq

01

sgn [a(x)],

CY

<

0.

Note that the above control somewhat resembles a control strategy resulting from the quadratic form-based design of Utkin [ ] I . It i s easy to check that

air =

O1

-< 0

U2

1 4

since sgn (U) = u/(a(.Thus as expected we are assured the existence of a sliding mode. The next step i s to determine the structure of the switching surface and the equivalent system. Suppose we desire the system to exhibit a linear behavior in a sliding mode described by the equation

fication of real-world systems introduces parameter errors. Hence models contain uncertain parameters which are often known to lie within upper and lower bounds. Awhole body of literature has arisen in recent years concerned with the deterministic stabilization of systems having uncertain parameters lying within known bounds. Such control strategies are based on the second method of Lyapunov. On the other hand, VSC controllers are based on the Generalized Lyapunov Second Method. Hence, one expects some fundamental links in the two theories. Using the control structures described in Section VI under the heading "Other Approaches" we will establish these links. First, a description of the uncertain plant and a brief review of the basic definitions of deterministic control of uncertain systems will be given. Following this we will outIinetheVSCapproach to uncertain system control.Thiswill then lead us to the expected fundamental connections. Lastly, we will focus on improving controller performance by reducing or eliminating chattering through the introduction of the so-called boundary layer controllers. This, in fact, i s a natural outgrowth of the theory of uncertain systems although it was developed independently in the VSC context [75], [76], [30], [7l. Deterministic Control o f Uncertain Systems To represent uncertainties in the plant from parameter uncertainties consider the following state dynamics

for appropriate a, and a2, both positive. Computing ueq according to (5.1) results in x2

x(t) = [ f (t, x(t))

+ A f (t, x(t)),
(7.1)

+ [B(t, x(tN + A M , x(t), r(t))l u(t)

Substituting the above into the plant dynamics implies that on a(x) = 0

which follows according to (5.20). Comparing (6.44) and (6.45) requires that al(x,, x) = sin(x,) ,

where r(t) is a vector function (Lebesgue measurable) of uncertain parameters whose values belong to some closed and bounded set. The formulation presumes no statistical information on the uncertainties. The plant uncertainties A fand A B (arising from r(t))are required to lie in the image of B(t, x) for all values of t and x. This requirement is the socalled matching condition [47l, [50], [51] [61]. Assuming the satisfaction of the matchingconditions, it i s possibleto lump the total plant uncertainty into a single vector e(t, x(t), r(t), u(t))and represent the uncertain plant as x = f(t, x) X(t0) = xg.

+ B(t, x) U + B(t, x) e(t, x, r, U )

(7.2)

+ alxl + a2x2

a nonlinear switching surface, i.e., thedesired linear behavior with the controller structure of (6.40), (6.41) requires the use of a nonlinear switching surface. For other examples of using a nonlinear switching surface see for example [73]. As a final point, the type of controller discussed in this subsection i s further developed in the next section to account for the problem of uncertain parameters in the plant model. VII. OVERVIEW UNCERTAIN OF SYSTEM THEORY, VSC, CHATTERING Introduction The purposes of this section are the exposition of VSC for uncertain systems, unification of the theories of VSC and deterministic methods of controlling uncertain systems, and a discussion of chattering. The motivation for exploring uncertain systems i s the fact that model identiAND

With regard to a stabilization analysis of the above model (7.2) the following definitions are pertinent: Definition 2: Let x(-): [to, w] -, R" be a solution of (7.2). x(*) i s uniformly bounded if for each xo there i s a positive finite constant, d(xo), (0 < d(xo) < 00) such that IIx(S112 < d(xo) all t E [ t o , a ] where 11 1, is the usual Euclidean vector for norm [47l, [a]. Definition 3: Solutions to (7.2) are uniformly ultimately boundedwith respect to some closed bounded set S C R" if for each xothere is a non-negative constant T(xo, S ) < 03 such that x(t) E S for all t > to T(xo, S . I The problem i s to find a state feedback u(t, x): R x R" -P Rm such that for any initial condition xo and for all uncertaintiesr(t)asolutionx(.):[t,,, OD) -+ R"of (7.2)existsandevery such solution is uniformly bounded. The literaturecontainstwo main approachesforthe solution of the above stabilization problem, the so-called minmax controller discussed by Gutman and Palmor [50] and

DeCARLO et al.: CONTROL O NONLINEAR MULTIVARIABLE SYSTEMS F

227

the Corless-Leitmann approach [47J. These approaches begin with a nominal system defined by x = f (t, x) X(t0) = xg (7.3)

If B r ( t , x) V , V ( t , x) i s zero then take

U

{U U

E Rm

and

I( U 11 5

pft, x)}

(7.8)

The reader should note that the set

assuming that x = 0 is an equilibrium point, i.e., f(t, 0) = 0 for all t. Both approaches require this nominal system to be uniformly asymptotically stable, i.e.,
1) for any E > 0, there is a 6(e) > 0 such that a trajectory starting within a &)-neighborhood of x = 0 remains for all subsequent time within the -neighborhood of

"O",
2) there i s a 6, such that a trajectory originating within a &-neighborhood of x = 0 tends to zero as t W.
+

It turns out that if there exists a (Lyapunov) function U . 1 : R x R" R + with a continuous derivative, and there exist continuous strictly increasing functions n(-): R, R i= , 1 2, 3, with the properties ,
+

y;(O) = 0,

i = 1, 2, 3

such that for all (t,x) E R x R"

can be thought of as a switching surface. One of the main goals of this section i s to show that the controller of (7.7) can be made to behave as a VSC controller with the switching surface (7.9). A close inspection of (7.7) reveals that this control i s discontinuous in the state since, for example, in the single input case it reduces to U = -sgn ( B T ( t , x) V , V ( t , x)) p(t, x). Since the above control is discontinuous it may excite unmodeled high-frequency dynamics of the plant. To avoid this problem, it i s necessary to modify this controller by introducing the so-called boundary layer controller which continuously approximates the discontinuous action of (7.7) in a neighborhood of the switching surface, (7.9). Let p(t, x) be any continuous function such that p(t, x) = -[ulllu)llp when llall = E. Then the structure of the boundary layer controller is

U =

u(t, x) =

and

'-I

Ibll
p(t, x),

if llull
if

(7.10)

llull < E.

where V,Visacolumn vector, then the nominal system (7.3, i s uniformly asymptotically stable. The objective i s to use this nominal Lyapunov function V ( * ) and bounds on the uncertainty e(t, x, r, U ) to develop conditions on the state feedback control U = u(t, x) guaranteeing uniform boundedness of the closed loop state trajectory of the plant of (7.2). This background sets up a discussion of the two methodsof stabilization of uncertain systems.The min-maxcontrol method comes first. I n the min-max approach one assumes a stable nominal system (7.3) with Lyapunov function V(t, x). A Lyapunov function candidate for the closed loop plant, (7.2), with U = u(t, x), i s again V(t, XI. objective i s to choose u(t, x) The to make the derivative of V(t, x) negative on the trajectories of the closed loop system, i.e., choose U = u(t, x) such that
V(t, x) =

Unfortunately, this controller does not guarantee asymptotic stability but rather uniform ultimate boundedness as per Definition 3. For a proof of this fact i n a slightly modified context see Corless and Leitmann [47l. This completes our brief review of the deterministic control of uncertain systems. The VSC Approach to the Control of Uncertain Systems Here again consider the uncertain plant as described in (7.2). In theVSCapproach it i s not necessaryforthe nominal system (7.3) to be stable. However, the equivalent system, i.e., the restriction of (7.3) to the switching surface u(t, x) = 0, must be asymptotically stable. The VSC control structure for plant (7.3) will be
U

uq

+ UN

(7.11)

- + (VLV)x = - + at + (v,Tv) ~ ( + e) c 0. u

av

[a,':

(V,TV)f

1
(7.5) (7.6)

Since (7.4b) holds, (7.5) holds if that min max

u e

= u(t, x) is chosen such

(v,Tv)

B(u

+ e) Io

where ueq i s the equivalent control for (7.3) assuming all uncertainties e(t, x, r, U ) are zero and uN i s to be designed to account for nonzero uncertainties. Recall from Section VI under the heading "Other Approaches" that the structure of (7.11) makes determination of the reachability and existence conditions for a sliding mode more straightforward to compute. Proceeding in the usual fashion, with the switching surface u(t, x) = 0, one may compute

for all ( t ,x) E R x R" and all admissible controls and admissible uncertainties. Assuming B T ( t ,x) V , V ( t , x) i s nonzero, the control

Ueq =

-[E

-1

B]

;[ + 5 f ]

(7.12)

where p(t, x) i s a scalar function satisfying p(t, x) 2 IIe(t, x, r, u)l12can be shown by direct substitution to satisfy (7.6).

assuming as usual that [(adax) B] is nonsingular and that r, U ) = 0. It i s now necessary to account for uncertainties and develop an expression for U,.,. To develop this thread assume as in the previous subsection that
e(t, x,

IIe(t, x, r, u)l12

s d t , x)

(7.13)

228

PROCEEDINGS OF THE IEEE, VOL. 76, NO. 3, MARCH 1988

where p(t, x) i s a non-negative scalar valued function. Also introduce the scalar valued function a t , x) = a

+ dt,

x)

(7.14)

where a > 0. This particular structure simplifies some of the derivations. Before specifying the control structure, we choose the most simple generalized Lyapunov function V(t, x) = .5aJ(t, x) a(t, x). (7.15)

As usual, in order to insure the existence of a sliding mode and attractiveness to the surface, it i s sufficient [ I ] tochoose a variable structure controller so that

dV - (t, x)
dt

V = arb

(7.16)

whenever a(t, x) # 0 where b(t, x) =

aa aa - + - x.
at ax

(7.17)

Thecontroller form given in (6.41) in conjunction with the controller of (7.7) suggests the VSC form
(b) Fig. 12. Typical closed-loopsystem trajectoriesfor the control (a) (73,and (b) ( . 8 . 71)

(7.18) when a(t, x) # 0 and where (7.19) where V, V(t, x) i s the gradient of the generalized Lyapunov function (7.15). This is different from the V(t, x) used to develop (7.7). If a(t, x) = 0, then set u(t, x) = ueq(t,x). In ordertoverifythevalidityof this controller notice that (suppressing t and x arguments)
V=aJ-+aJ-(f+Bu+
at ax

aa

aa

Be).

(7.20)

some cases for sufficiently large p(t, x), the response will behave as a VSC response in a vicinity of the origin. I n fact, if p(t, x) i s sufficiently large the controller exhibits the standard VSC response once the trajectory intercepts the surface B 'V,V = 0 [50]. In Fig. 12(b)we have the usual VSC responsewith respect to the user-chosen switching surface a(t, x) = 0. In addition to stabilization of the closed loop system one may obtain tracking properties as implicitly built into the user-designed switching surface. Chattering

Substituting (7.18) into (7.20) and manipulating produces

i / = a Jaa + + J - f f g J - f - a J aa aa

aa
at

at

ax

ax

(7.21) verifying the negative definiteness of V. This establishes attractiveness to the switching surface.
A Comparison of the VSC and Deterministic Controllers

The VSC controllers developed in Section VI and the uncertain system controllers of this section assure the desired behavior of the closed loop system. These controllers, however, requirean infinitely(in the ideal case) fast switching mechanism. The phenomenon of nonideal but fast switching was labeled as chattering (actually the word stems from the noise generated by the switching element). The high frequency components of the chattering are undesirable because they may excite unmodeled high-frequency plant dynamics which could result in unforseen instabilities. The boundary layer controller of (7.10) helps to eliminate the effects of chattering. Let us now refine this notion of boundary layer and boundary layer controller. Define the set
{XI Ila(x)II

The similarity of the structure of (7.7) and (7.18) i s clear in view of the uNterm. The controllers, however, generate different responses with respect to different switching surfaces. This is best seen by viewing Fig. 12. Fig. 12(a) illustrates a typical trajectory of the closed loop system (7.2) driven by the controller (7.7). Notice that the control switches with respect to the surface BJV,V = 0 where V i s the Lyapunov function of the nominally stable system. In

E,

>

0)

as the so-called boundary layer of thickness 2 ~ Consider . the control law (suppressing t and x arguments)

CUeq

+ Pf

if )1u11 <

DeCARLO er al.: C O N T R O L OF NONLINEAR MULTIVARIABLE SYSTEMS

229

where ueqi s given by (7.12) and wherep = p(t, x) is any continuous function such that

whenever ) ) U ) ) = E and \)p)l 6 . This control guarantees attractiveness to the boundary layer and inside the boundary layer, (7.22) offers a continuous approximation to the discontinuous control action of (7.18). As shown in Corless and Leitmann [47], one i s not guaranteed asymptotic stability but ultimate boundedness of trajectories to within a neighborhood of the origin depending on E. The reader might peruse [30], [51] which offer an alternate class of controllers than those above. Specifically [50], [60], [61] offer a discussion of such controllers for linear plants. Finally, the reader might note that in the control of dc motors chattering i s of minimal concern since switching can occur in the high kilohertz range if not megahertz range due to advances in power electronics. This, of course, i s well beyond the structural frequencies of mechanical systems involved. VIII. APPLICATIONS I n [26] Young developed an adaptive VSC for an aircraft control. Calise and Krammer [I31 also investigated VSC for aircraft control. An alternative approach to aircraft control using uncertain system controllers was proposed by Petersen [12]. Other applications to spacecraft control can be found in Sira-Ramirez and Dwyer in [63], [64]. I n the area of robotic control, Young [27] developed an algorithm based on hierarchical VSC, later refined by Morgan and Ozguner [36]. Also Slotine and Sastry [7l used VSC for tracking control of robot manipulators. A model following VSC scheme was developed by Ambrosin0 eta/. [20] and applied to a simple model of a robot manipulator. A similar application can be found in Bailey and Arapostathis [70]. The use of the Filippov method applied to robotic control can be found in Bartolini and Zolezzi [MI, Paden and Sastry [74], and [70]. A combination of VSC and deterministic approach to the control of uncertain systems was proposed by Spong and Sira-Ramirez [54]. There are various applications of VSC to power systems. Among these are Young and Kwatny [6] (who developed an overspeed protection control for an electric power generating plant), Hawley and DeCarlo [22], Lefebvre et a/. [23] and Richter et al. [24]. This was later extended by Matthews and DeCarlo in [78], [81]. Also Sivaramakrishnan et a/. [34] applied the VSC to the design of a variable structure loadfrequencycontrollerfor a singlearea power system. In Bengiamin and Kauffmann [32] an application of VSC to the dc motor position control i s described. While Utkin and Orlov [I91 treat the problem of distributed control using VSC. In contrast to the above VSC applications other methods applicable to real-world systems can be found in Hunt et a/. [66], [67l and Sain and Peczkowski [65]. A nice review paper on recent trends in nonlinear system feedback control is Kokotovic [58].

of the control law which forces the systems trajectory to and maintains it on the sliding surface. In discussingthe multi-input case,wesawthatthedesign process was complicated by the coupling of the controls through the switching surface. Several different methods were then developed (the diagonalization and hierarchical methods) toeffectivelydecouple thecontrolsand thus simplify the design process. Essentially, these techniques reduced the overall control problem to a series of single input problems. After this, important connections to the theory and application ofthedeterministiccontrol of uncertain systemwere introduced and developed. The two theories were seen to have a close alignment. To illustrate that VSC theory i s sufficiently advanced to allow the design of sophisticated control systems, a brief survey of VSC applications as found in the literature was given. Thewarnings of the previous sections combined with the application survey indicate that when VSC can be applied to a particular problem a very high-quality control system results. However, significant research i s needed to successfullyapply VSC to even more general classes of nonlinear systems, for example, systems nonlinear in the control U = u(t, x) and systems in which the matching conditions are not satisfied. Another problem with VSC i s the need for complete state information. Development of switching surfaces and controllers based on measurable output signals represents an open problem and an areaof important research.Thedeve1opment of nonlinearobservers usingVSCconceptsisastep in this direction. See, for example, Walcott and Zak [69]. Another area filled with fertile research soil i s the relationship of VSC with the recently fabricated Lie Algebraic approach to the control of nonlinear systems. Preliminary resuIts in this direction have been published by Marino [57]. Other exciting open problems fall in the categories of tracking and output regulation [59], [80], discrete variable structure control [62] and large scale systems. In the large scale systems area a number of promising results (especially in the decentralized control framework) have been obtained by DeCarlo and co-workers [22]-[24], 1311, [781-[821. A very interesting application of the VSC theory is in the design of Pulse-Width-Modulated (PWM)control strategies in nonlinear systems. In particular, Sira-Ramirez [83] established an equivalence betweenthe sliding modes, resulting fromaVSCstrategy,and the responseresuItingfromaPWM control law in nonlinear analytic systems. Specifically, under certain conditions the PWM controlled response represents an ideal sliding motion on an invariant manifold associated with an ideal average system. Details of these results and an application to the control design of switchmode dc-to-dc power converter circuits can be found in [831* Finally let us close with the point that this paper attempts to present to the reader possible solutions to some questions raised in the report [71].
ACKNOWLEDGMENT The authors gratefully acknowledge constructive remarks of the reviewers.
REFERENCES
[I] V. I. Utkin, Sliding Modes and Their Application in Variable Structure Systems. Moscow, Soviet Union: MIR Publishers,
1978.

IX. CONCLUDING REMARKS

This paper has developed and surveyed the essential concepts of VSC. Recall that the design of a VSC has two steps: 1) design of the switching surface to assure the desired behavior of the plant in a sliding mode, and 2) development

230

PROCEEDINGS OF THE IEEE,

VOL. 76, NO. 3,

MARCH 1988

[2] -, Variable structure systems with sliding modes, /E Trans. Automat. Contr., vol. AC-22, no. 2, pp. 212-222, 1977. structure systems-present and future,Auto[3] -,Variable mat. Remote Contr., vol. 44, no. 9, pp. 1105-1120,1983. [4] -, Equations of the sliding regime in the discontinuous systems 1,Automat. RemoteContr., vol. 32, no. 12, pp. 18971907,1971. [SI V. I. Utkin and K. D. Yang, Methods for construction of discontinuity planes in multidimensional variable structure systems, Automat. RemoteContr., vol. 39, no. IO, pp. 1466-1470, 1978. [6] K.-K. D. Young and H. G. Kwatny, Variable structure servomechanism design and applications to overspeed protection control, Automatica, vol. 18, no. 4, pp. 385-400, 1982. [;7 J . J. Slotine and S. S. Sastry, Tracking control of non-linear systems using sliding surfaces, with application to robot manipulators, lnt. 1. Contr., vol. 38, no. 2, pp. 465-492,1983. [8] U. Itkis, Control Systems of Variable Structure. New York, NY Wiley, 1976. [9] - Dynamic switching of type-l/type-ll structures in track, ing servosystems, / Trans. Automat. Contr., vol. AC-28, no. 4, pp. 531-534, 1983. [IO] B. A. White, Reduced order switching functions in variable structure control theory, I froc. D. Control Theory & Appl., VOI. 130, pp. 33-40,1983. [Ill B. A. White and P. M. Silson, Reachability in variable structure control systems, IProc., vol. 131, pt. D., no. 3, May 1984. [I21 I. R. Petersen, A procedure for simultaneouslystabilizing a collection of single input linear systems usingnon-linearstate feedback control,Automatica, vol. 23, no. 1, pp. 33-40,1987. [I31 A. J. Cake and F. Kramer, A variable structure approach to robust control of VTOL aircraft, in froc. of the American Control Conf., pp. 1046-1052, June1982. [I41 0.M. E. El-Chezawi, A. S.Zinober, and S. A. Billings, Analysis and design of variable structure systems using a geometric approach, lnt. /. Contr., vol. 38, no. 3, pp. 657-671, 1983. [I51 E. P. Ryan, A variable structure approach to feedback regulation of uncertain dynamical systems, Int. /. Contr., vol. 38, no. 6, pp. 1121-1134,1983. [I61 I. Kadushin and A. M. Steinberg, Stabilization of nonlinear systems with a dither control, 1. Math. Anal. Appl., vol. 43, no. 1, pp. 273-284, July 1971. [17] C. Zames and N. A. Shneydor, Dither in nonlinear systems, I Trans. Automat. Contr., vol. AC-21, no. 5, Oct. 1976. [I81 W. M. Wonhamand C. D. Johnson,Optimal bang-bangcontrol with quadratic performance index, in 4th joint AutomaticControlConf., Minneapolis,MN, pp. 101-112, June1963. [I91 Y. V. Orlov and V. I. Utkin, Use of sliding modes in distributed system control problems, Automat. RemoteContr., vol. 43, no. 9, pp. 1143-1148,1982. [20] G. Ambrosino, C. Celentano, and F. Garofalo, Variable structure model reference adaptive control system, lnt. 1. Contr., vol. 39, no. 6, pp. 1339-1349, 1984. [21] W. Hejmo, On the sensitivity of atime-optimal position control, / Trans. Automat. Contr., vol. AC-28, no. 5, pp. 618621,1983. [22] P. Hawley and R. DeCarlo, Variable structure control of interconnected systems with applications to power systems, Purdue University, School of Electrical Engineering Tech. Rep. TR-EE-83-9, 1983. [23] S. Lefebvre, S. Richter, and R. DeCarlo, Decentralized variable structure control design for a two-pendulum system, I Trans. Automat. Contr., vol. AC-28, no. 12, pp. 112114, Dec. 1983. [24] S. Richter, S. Lefebvre, and R. DeCarlo, Control of a class of nonlinear systems by decentralized control, I Trans. Automat. Contr., vol. AC-27, no. 2, Apr. 1982. [25] K.-K. D. Young, P. V. Kokotovic, and V. I. Utkin, A singular perturbation analysis of high-gain feedback systems, I Trans. Automat. Contr., vol. AC-22, no. 6, pp. 931-938, 1977. [26] K.-K. D.Young,Designofvariablestructure modelfollowing systems, /E Trans. Automat. Contr., vol. AC-23, no. 6, pp. 1079-1085,1978. [27] -, Controller design for a manipulator using the theory of variable structure systems, / E Trans. Syst. Man Cybern., vol. 8, no. 2, pp. 101-109, 1978. [28] A. C. Lukyanov and V. I. Utkin, Methods of reducing equa-

[29]

[30] [31]

[32] [33] [34]

1351 [36] [371 [38] [39] [40]

[41] [42] [43] [44] [45] [46] [47]

[48] [49] [SO] [SI] [52] [53] [541

tions of dynamic systems to regular form, Automat. Remote Contr., vol. 42, no. 4, pp. 413-422, 1981. A. S. Zinober, Controller design usingthe theory of variable structure systems, in Self-Tuning a n d Adaptive Control. London, England: Peter Peregrinus Ltd., 1981, ch. 9, pp. 206229. J. J. Slotine, Sliding controller design for non-linear systems, lnt /. Contr., vol. 40, no. 2, pp. 421-434, 1984. C. P. Matthews, S. H. Zak, and R. A. DeCarlo, Decentralized control for aclass of nonlinear systems usingavariable structure approach, in Proc. ofthe22Allerton Conf., pp. 192-200, Oct. 1984. N. N. Bengiaminand B. Kauffmann, Variable structure position control, Control Syst. Mag., pp. 3-8, Aug. 1984. A. R. Bazelow and J. Raamot, On the microprocessor solution of ordinary differential equations using integer arithmetic, /E Trans. Comput., vol. C-32, no. 2, pp. 204-207,1983. A. Y. Sivaramakrishnan,M. V. Hariharan, and M. C. Srisailam, Design of a variable-structure load-frequency controller using pole assignment technique, lnt. 1. Contr., vol. 40,no. 3, pp. 487-498,1984. 0. Hajek, Discontinuous differential equations I, 11,). Differential Equations, vol. 32, no. 2, 1979. R. C. Morgan and U. Ozgunner, A decentralized variable structure control algorithm for robotic manipulators, /E 1. Robotics Automat., vol. RA-I, no. 1, pp. 57-65, Mar. 1985. T. A. Bezvodinskayaand E. F. Sabaev, Stability conditions in the large for variable structure system, Automat. Remote Contr., vol. 35, no. IO, pp. 1596-1599,1974. G. I . Lozgachev, The construction of a Lyapunov function for variablestructure systems, Automat. RemoteContr., vol. 33, no. 8, pp. 1391-1393,1972. A. J. Calise and K. V. Raman, A servo compensator design approach for variable structure systems, presented at the 19thAllerton Conf., Universityof Illinois, Sept. 30-Oct. 2,1981. Y. F. ltkis and A. V. Leibovitch, Application of systems variable structure for centralized control of objects with variable parameters, Automat. RemoteContr., no. 8, pp. 6770, 1970. 0. M. E. El-Ghezawi, A. S. I.Zinober, D. H. Owens, and S. A. Billings, Computation of the zeros and zero directions of linear multivariable systems,lnt.I. Contr., p p 833-843,1972. M. A. Aizerman and E. S. Pyatniskii, Foundations of a theory of discontinuous systems 1,Automat. RemoteContr.,vol. 35, no. 7, pp. 1066-1080,1974. -, Foundations of a theory of discontinuous systems 11, Automat. Remote Contr., vol. 35, no. 8, pp. 1242-1263, 1974. C. Bartolini and T. Zolezzi, Variable structure nonlinear in the control law, I Trans. Automat. Contr., vol. AC-30, pp. 681-684, Jul. 1985. V. I. Utkin, Equations of the sliding regime in discontinuous systems 11, Automat. Remote Contr., vol. 33, pp. 211-218, 1972. A. F. Filippov, Differential equations with discontinuous right hand sides, Am. Math Soc. Transl., vol. 42, pp. 199-231, 1964. M. J. Corless and G. Leitmann, Continuous state feedback guaranteeinguniform ultimate boundedness for uncertain dynamic systems, I Trans. Automat. Contr., vol. AC-26, no. 5, pp. 1139-1144,1981. C. Ambrosino, G. Celentano, and F. Carofalo, Robust model tracking control for a class of non-linear plants, /E Trans. Automat. Contr., vol. AC-30, no. 3, pp. 275-279,1985. Y.-Y. Hsu and W.-C. Chan, Optimal variable-structurecontroller for DC motor speed control, froc. Inst. lec. ng., vol. 131, no. 6, part D, pp. 233-237,1984. S. Cutman and Z. Palmor, Properties of min-maxcontrollers in uncertain dynamical systems, SIAM 1. Contr. Optimization, vol. 20, no. 6, pp. 850-861,- 1982. B. R. Barmish, M. Corless, and G. Leitmann, A new class of stabilizing controllers for uncertain dynamical systems, SAM/. Contr. Optimization, vol. 21, no. 2, pp. 246-255,1983. F. H. Clarke, Optimization a n d Nonsmooth Analysis, Canadian Math. Soc. Ser. in Math. New York, NY: Wiley-lnterscience, 1983. R. T. Rockafellar, Convex Analysis, Princeton Mathematics Ser., vol. 28. Princeton, NJ: Princeton Press, 1970. M. W. Spong and H. Sira-Ramirez, Robust control design

DeCARLO et al.: C O N T R O L OF NONLINEAR MULTIVARIABLE SYSTEMS

231

techniques for a class of nonlinear systems, in Proc. 7986 American Contr. Conf. (Seattle, WA), pp. 1515-1522, June1986. M. W. Spong, K. Khorasani, and P. V. Kokotovic, An integral manifold approach to the feedback control of flexible joint robots, / E 1. RoboticsAutomat., vol. RA-3, no. 4, pp. 291300,1987. M. W. Spong, Robust stabilization for a class of nonlinear systems, presented at the 7th Int. Symp. on the MTNS, Stockholm, June1985. R. Marino, High-gain feedback in nonlinear control systems, Int. J. Contr., vol. 42, no. 6, pp. 1369-1385, 1985. P. V. Kokotovic, Recent trends in feedback design: An overview, Automatica, vol. 21, no. 3, pp. 123-132, 1968. A. S. Vostrikov, V. I. Utkin, and C. A. Frantsuzova, System with state vector derivative in the contro1,Automat. Remote Contr., vol. 43, no. 3, pp. 283-286, 1982. L. C. Chouinard, J. P. Dauer, and C. Leitmann, Properties of matrices used in uncertain linear control systems, SIAM I. Conk Optimization, vol. 23, no. 3, pp. 381-389, 1985. I. R. Petersen, Structural stabilization of uncertain systems: Necessity of the matching condition, SIAMI. Contr. Optimization, vol. 23, no. 2, pp. 286-296, 1985. C. Milosavljevic, General conditions for the existence of a quasisliding mode on the switching hyperplane in discrete variable structure systems,Automat. RemoteContr., vol. 46, no. 3, pt. 1 pp. 307-314,1985. , H. Sira-Ramirezand T. A. Dwyer, Ill, Variable structure controller design for spacecraft nutation damping, / Trans. Automat. Contr., vol. AC-32, no. 5, pp. 435-438, May 1987. T. A. W. Dwyer, Ill, and H. Sira-Ramirez, Variable structure control of spacecraft attitude manuevers, to be published J. Guidance, Dynamics and Control, 1987. M. K. Sain and J. L. Peczkowski, Nonlinear control by coordinated feedback synthesis with gas turbine applications, in Proc. American Contr. Conf. (Boston, MA), pp. 1121-1128, June1985. L. R. Hunt, R. Su, and C. Meyer, Global transformation of nonlinear systems, /E Trans. Automat. Contr., vol. AC-28, no. 1, pp. 24-31, 1983. -, Design for multi-input nonlinear systems, in Differential Geometric Control Theory, R. W. Brockett et al., Eds. Cambridge, MA: Birkhauser Boston, 1983, pp. 268-298. S. M. Madani-Esfahani, R. A. DeCarlo, M. J. Corless, and S. H. Zak, On deterministic control of uncertain nonlinear systems, in Proc. 7986AmericanContr. Conf. (Seattle, WA), pp. 1523-1528, June1986. B. L. Walcott and S . H. Zak, State observation of nonlinear uncertain dynamical systems, I Trans. Automat. Contr., vol. AC-32, no. 2, pp. 166-170,1987. E. Bailey and A. Arapostathis, Simple sliding mode control scheme applied to robot manipulators, lnt. I. Contr., vol. 45, no. 4, pp. 1197-1209,1987. Challenges to control: A collective view, Report of the workshop held at the University of Santa Clara on September 18-19,1986, /Trans. Automat, Contr., vol. AC-32, no. 4, pp. 275-285, Apr. 1987. S. H. Zak and C. A. MacCarley, State-feedback control of non-linearsystems,lnt. J. Contr., vol. 43, no. 5, pp. 1497-1514, 1986. H. Sira-Ramirez, Harmonic response of variable-structurecontrolled Van der Pol oscillators, /E Trans. Circuits Syst., vol. CAS-34, no. 1, pp. 103-106, Jan. 1987. B. E. Paden and S. S. Sastry, A calculus for computing Filippovs differential inclusion with application to the variable structure control of robot manipulators, /E Trans. Circuits Syst., vol. CAS-34, no. 1, pp. 73-82, Jan. 1987. A. S. Vostrikov, Control of Dynamical Systems (in Russian). Novosibirsk, Soviet Union: Novosibirsk Institute of Electrical Eng., 1979. -, Optimal a n d Adaptive Systems (in Russian). Novosibirsk, Soviet Union: Novosibirsk Institute of Electrical Eng., 1977. B. Draienovit, The invariance conditions in variable structure systems, Automatica, vol. 5, no. 3, pp. 287-295, 1969. C. Matthews,R. DeCarlo, P. Hawlev, and S. Lefebvre,Toward afeasiblevariablestructurecontroldesign for asynchronous

[79]

[80]

[SI]

[82] [83]

machine connected to an infinite bus, /E Trans. Automat. Contr., vol. AC-31, no. 11, Dec. 1986. C. Matthews and R. DeCarlo, Decentralized variable structure control of interconnected multi-input multi-output nonlinear systems, Circuits, Systems, a n d Signal Processing, vol. 6, no. 3, 1987. -, A variable structure control approachto the nonlinear decentralized tracking problem, presented at the Proc. of the 24th Annual Allerton Conf. on Communication, Control, and Computers, Monticello, IL, Oct. 1986. C. Matthews, C. DeMarco, and R. DeCarlo, A decentralized VSC design for two interconnected synchronous power machines, presentedat the Proc. of the 1987AmericanControl Conf., Minneapolis, MN, June 1987. C. Matthews and R. DeCarlo, Decentralized tracking for a class of interconnected nonlinear systems using variable structure control, Automatica, Mar. 1988. H. Sira-Ramirez, A geometric approachto pulse-width-modulated control design, presentedat the Proc. 26th IEEE Conf. Decision and Control, Los Angeles, CA, Dec. 1987.

RaymondA. DeCarlo (SeniorMember, IEEE) was born in Philadelphia in 1950. He receivedthe B.S. and M.S. degrees in electrical engineering from the University of Notre Dame in 1972 and 1974, respectively. In1976, he receivedthe Ph.D. degreeunder the direction of Dr. Richard Saeks from Texas Tech University. His doctoral research centered on Nyquist stability theory with applications to multidimensional digital filters. After graduating, he became a Lecturer at Texas Tech for one year, and then became an Assistant Professor of Electrical Engineering at Purdue University in the Fall of 1977, and an Associate Professor in 1982. He has worked at the GeneralMotors Research Laboratories during the summers of 1985 and 1986. His research interests include interconnected dynamical systems, analog and analog-digital fault diagnosis, decentralized control of large-scale systems, decentralized identification and estimation of interconnected systems, and computer-aided circuit design. Dr. DeCarlois past Associate Editor for Technical Notes and Correspondence and past Associate Editor for Survey and Tutorial papers, both for the IEEETRANSACTIONSON AUTOMATIC CONTROL. Presently, he i s the Secretary-Administrator of the Control Systems Society and also a member of the Boardof Governors of the Society. Stanislaw H. Zak (Member, IEEE) received the Ph.D. degree from the Technical University of Warsaw, Poland, in 1977. From1977 to 1980, hewas a FacultyMember in the Institute of Control and Industrial Electronics, Technical University of Warsaw. From 1980 until 1983, hewas a Visiting Assistant Professor in the Department of ElectricalEngineering, University of Minnesota. InAugust 1983, hejoined theschool of ElectricalEngineeringat Purdue University. His primary areas of research aresystemstheory, time-delay systems, nonlinear systems control, and computer algebra. Gregory P. Matthews (Member, IEEE) was born in Pontiac, MI, on December 3,1956. He received the B.S. degree in electrical engineering from the University of Michigan in 1979, and the M.S. and Ph.D. degrees in electrical engineering from Purdue University, West Lafayette, IN, in 1982 and 1985, respectively. Since 1986 he has been a member of the General Motors Research Laboratories Electrical and Electronics Engineering Department,Warren, MI. His interests include the control of large scale and nonlinear systems.

232

PROCEEDINGS OF THE IEEE, VOL. 76, NO. 3, MARCH 1988