212

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, NO. VOL.

AC-22,

2,APRIL 1977

Roger N . Schane (S66M70) was born i n Chicago, IL, on January 25, 1946. He received the B.S. degree in control systems engineering from the University of Illinois, Chicago, and the M.S.E.E. degree from Marquette University, Milwaukee, WI, in 1968 and 1970, respectively. From 1968 to 1971 he was a Project Engineer at General Motors Delco Electronics Division, Milwaukee, WI, and General Motors Technical Center,Warren, MI.AtDelco Electronics, he was involved in analysis of inertially stabilized fireantrol systems. At General Motors Technical Center, he par, .

ticipated in the design and mechanization of experimental automotive electronic systems. Since October 1971, he has been a member of the technical staff at T h e Analytic Sciences Corporation, Reading, Mass. where his activities have ranged from missile guidance error analysis, pilot modeling, and analysis of strapdowninertialtransfer alignment algorithms, to a broad spectrum of studies in the radio navigation field. These studies include: navigation algorithm analyses; airborne receiver performance specification;flight test planning, data reduction and analysis; and development of operational software used in computing published tables of OMEGA propagation corrections, and operational software used for synchronizing the OMEGA transmitter network.

Survey Paper VariableStructureSystems

with Sliding Modes

Abstmct-Variable stmctme systems consist of a set of continuous subsystems together with snitable switching logic. Advantageom properties r d t from changing struetnres according to this switching logic. Design and analysis for this clafs of systems are surveyed in this paper.

I. INTRODUCTION

HE basic philosophy of the variable structure apTproach is simply explained by contrasting it with the
linear state regulator design for the single-input system
i=Ax+bu.

In the linear state regulator design, the structure of the state feedback is fixed as

u=kTx
where the constant parameters are chosen according to various design procedures, such as eigenvalue placement or quadratic minimization. In variable structure systems the control is allowed to change its structure, that is, to switch at any instant from one to another member of a set of possible continuous functions of the state. The variable structure design problem is then to select the parameters of each of the structures and to define the switching logic.
Manuscript received April 7, 1976; revised December 21, 1976. Paper recommended by P. R. B e l k er, Chairman of the IEEE S-CS Optimal Systems Committee. This wort was supported by the Coordinated Sciof Electrical Engineering, Univerence Laboratory and the Department sitv of -~ Illinois. -The author iswith the Variable Structure Laboratory, Institute for Control Sciences, Moscow, USSR.
~~J
~~~

The idea of changing a structure is a natural one, and early utilization of this approach canbefound in the papers published about 20 years ago [18], [19], [26], [48], [50], [59], [61], [65], [66], [75], [88], [90]. A reward for introducing this additional complexity is the possibility to combine useful properties of each of the structures. Moreover, a variable structure system can possess new properties not present in any of the structures used For instance, an asymptotically stable system may consist of two structures neither of which is asymptotically stable. This possibility is illustrated by some early examples, which stimulated interest in variable structure systems ( V S S ) . Although very simple, two such examples are quoted here because they present the advantages of changing structures during a control phase. In the first example we consider a second-order system
x= - * x

having two structures defined by =a : and a$ where a:> a : . The phase portrait consists of families of ellipses [Fig. l(a), (b)] and hence, neither structure is asymptotically stable. However, asymptotic stability is achieved if the structure of the system is changed on the coordinate axes, that is, if the switching logic is
ifxi.0

*=

if x i <o.
The resulting phase portrait Fig. shown inis l(c).

SURVEY PAPER

213

stable Fig. 2. Asymptotically Fig. 1. Asymptotically stable VSS consisting of two stablestructures.

(4 VSS consisting of two unstable structures.

determines the behavior of the system in the sliding mode. It isuseful to note that this behavior depends on the parameter c. This invariance with respect to plant parame[>O x-&t+*x=o, ters and disturbances is of extreme importance when is considered, where the linear structure corresponds to controlling time-varying plants or treating disturbance negative and positive feedback when is equal to either rejection problems. a > 0 or to -a. Both structures are unstable [Fig. 2(a), In Section I1 the properties revealed in these simple (b)]. Note that the only motion converging to the origin is examples are utilized to design VSS in phase canonic - a. form. Methods developed to analyze VSS as differential along the stable eigenvector of the structure with If the switching occurs on this line and on x = 0 with the equations with discontinuous right-hand sides are switching law surveyed in Section 111. In Section IV, general design principles for multiinput variable structure systems are discussed. In the second example, the system

*=

11. VSS IN PHASECANONIC FORM In this section we consider the design of variable structhen the resulting VSS will be asymptotically stable. ture controllers for zeroing the output y = x, of the system In the above examples, new system properties are obtained by composing a desired trajectory from the parts of ii=xj+l, i=l;-.,n-l trajectories of different structures. An even more fundan mental aspect of VSS is the possibility to obtain trajectoin= a;x;+f(r)+u (1) ries not inherent in any of the structures. These trajectoi= 1 ries describe a new type of motion-the so-called sliding where u is control, f(t) is a disturbance, aiare constants or mode. To show how such motion occurs let us reconsider the time-varying parameters, f(t), a, may be unknown. Suppose that u as a function of the state vector x second example using 0 < c <A instead of c = A (Fig. 3). undergoes discontinuities on some plane s = 0, where The phase trajectories are directed towards the switching n line s = cx + i = 0 and hence once on this line the state s= cixi, q=const, c, = 1. (2) must remain on it. The motion along a line which is not a i= 1 trajectory of any of the structures is called the sliding mode. Then the velocity vector undergoes discontinuities in the The equation same plane. As in the second-order example of Section I, if the trajectories are directed towards the plane, an s = O i+cx=o

214

IEEE TRAVSACTIONS

ON AUTOMATIC CONTROL, APRIL

1977

A necessary and sufficient condition for a sliding plane to exist [9], [37], [40] is
a;> cj- I - a, - tic,-,

+ c,a,,
(6)

~ i ~ c i ~ l - a i - c , c n ~ l + c i ai n = ,l ; . . , k , c , = O

Hence, coefficients c; needed for the design of a desired sliding mode (4) cannot be chosen freely. The inequalities in (6) may be satisfied by a proper choice of a; and pi but the equalities represent (n - k - 1 ) constraints for ( n - 1 ) coefficients c;. These constraints vanish only for k = n - 1 . For k < n - 1 a class of linear plants has been found [5], [38], [51], [40] for which a sliding plane with stable motion Fig. 3. Slidingmode i n a second-order VSS. exists. These plants are characterized by the following theorems. sliding mode will appear in this plane. The pair of inequalTheorem 1: Let X,; . ,& be eigenvalues of the system ities (1) with lim S > O and lim S < O (3)

Qjxj, Qi = cj- - a, - C,C,+ c,a,,. (7) are a sufficient condition for sliding mode to exist [6]. To i= I prove the invariancy of the sliding mode with respect to in a sliding plane is asymptotically the plant parameters a; and the disturbance f(r), we solve The sliding mode stable in systems ( l ) , ( 5 ) if and only if Re& < 0, i = the equation s = 0 for the variable x,, and substitute into 1, , n 1 ; one of the eigenvalues is equal to cn- I - c, (1). The resulting equations of the sliding mode and may be arbitrary. iI=xj+,, i=l;-.,n-2 Theorem 2: For the sliding mode in a sliding plane to n- 1 be asymptotically stable it is sufficient that for the ( n - k x,,-l= c;x; (4) 1)th-order VSS
u=

s-t-0

s++o

i= 1

a sliding plane s' = 0 with asymptotically stable sliding mode exists or, equivalently, that the condition of Thecrem 1 for the truncated system (8) is fulfilled. Theorems 1 and 2 show that the motion in a sliding Time-Invariant Plants plane may be stable even when none of the 2 , employed structures is stable. This is particularly clear for k = n - 2. Let the parameters of the plant a, be constant and In this case, (8) is a third-order VSS, whose control is a f ( t ) =O. The problem is to force the state to zero. Analopiece-wise linear function of only one variable xn-*. There gous to what was done with the second-order VSS, the exists such that the characteristic equation of this control u is chosen as a piecewise linear function of x with truncated system with \kL-,= Q n P 2 is discontinuous coefficients h3+a,A2+a,,~,X+(a,,_2+Q,~,)=0 k u=\Eixj-6,sgns, l<k<n-1 (5) which has only one eigenvalue with nonnegative real part. i=l It follows from Theorem 1 that the motion in a sliding plane for (8) is asymptotically stable. On the other hand, if s > O ai, if x,s>O sgns= Theorem 2 insures the asymptotic stability of the sliding ifs<O \E= mode in a sliding plane for the original system (l), (5) Pi, if xis <0 a;,p,, 6 , const, when k = n - 2. If either a,,- or a, is negative none of the 6, is small positive scalar. available structures is stable.

depend only on parameters c, [28]. Implications and uses of this result have been the subject of many early works summarized in [25], [40]. A design procedure based on the invariancy property can be outlined as follows. First, the desired sliding mode is formed by a choice of the parameters ci. Second, a discontinuous control is found which guarantees the existence of sliding modes at every point of the plane s =O. Such a plane will be referred to as a sliding plane. Third, the control must steer the state to the sliding plane. This approach is now applied to several control problems.

x. ',= x.
"

i=k;..,n-l
ajxi- +;x, - GAsgns'
i=k

in =-

i=

{ T:l

SURVEY PAPER

215

Asymptotic stability as determined by Theorems 1 and 2 isusually not the only requirement to be met by the desired sliding mode. In [52] the response time is reduced by placing the eigenvalues tofar as the left as possible providing a sliding plane exists. Under the same existence parameters minimizing a constraint sliding plane quadratic performance index are found [141. Time-optimal VSS of the second order were considered in [73]. Finally, it is necessary to guarantee that the designed sliding plane is reached from all initial states. Various reaching conditions have been proposed throughout the development of the vss theory. following necessary conditions were formulated in [37]. Theorem 3: For the state to reach s = O defined by (2) with ci >0, it is necessary that all real eigenvalues of the systems (I), (5) ,pi = ai ( i = 1,. . . ,k) be nonnegative. ci >0 is needed since the sliding mode Notethat governed by (4) must be asymptotically stable. Note also due to that (I), (5) with ,pi = ai is permitted to be complex eigenvalues with positive real parts. It was shown in [ 121 and [43] that the condition of Theorem 3 is also sufficient for second- and third-order VSS, respectively. TWO theorems concerning sufficient conditions are given below. Theorem 4 [6]: For the state to reach a sliding plane it is sufficient In that

(6) and asymptotic stability of the system (4) are met if

CI = & a1

> -CICZ~ & < -c]cz,

c,>o,

c,>o.

ne

The reaching condition of Theorem 6 is fulfilled if a I>O. Note that none of the controls which are linear functions of x, (even x1 and XJ reduces the state to zero. The above design principle constrains choice of parameter Ci (6) and hence it restricts the variety of sliding modes. VSS without sliding planes are difficult to analyze and their properties are established only for a few special cases [6], [lo]. They are also studied with the help of approximation methods 17417 [921When the system equation depends not Only On but also on its time-derivatives, control discontinuities result in discontinuous trajectories. Design principles have been modified for this Of linear plants [311,[331, [701. Additional possibilities appear if the coefficients ai,&,ci in (5) can be made large, since then the Order Of a mode equation may be reduced [61-[81. VSS have been applied to linear plants with pure delay [4], [27], [89], nonlinear plants with multiple stable and unstable equilibrium states [30], and plants with state variable constraints [3], [40], [60]. VSS discussed here the control is a function of the output and its derivatives. In practice pure derivatives are not available. VSS with estimators have been studied [35], [77], and variable structure estimators (filters with piecewise constant parameters) have been developed [24], [36], 1391,[781.
~~

c,-,

- a, Q 0.

[ I 3 ] : For the state to reach the plane s=o with ci > 0 it is sufficient that

k = n - l , a i > - a i , & < -ai,

(i=l;.-,n-l)

Time-Varying Plants

and that the system (l), (5) with qi= ai (i = 1, - . . , n - 1) has no nonnegative eigenvalues. The condition of Theorem 4 is easily verified but it is valid only for positive a, since cn- >0. The most effective reachability result is the following theorem. Theorem 6 [ 1 5 ] ,[16]: The necessary condition of Theorem 3 is also sufficient if the plane s =0 is a sliding plane and the sliding mode is asymptotically stable. An estimate of the time needed to reach the plane s =0 is given in [23]. To summarize, the design procedure beginswith the determination of a continuous control u = - 2%lG?ixi placing (n - 1) eigenvalues in desired locations. Then the control (5) assures the existence of the sliding plane with the sliding mode determined by these (n - 1) eigenvalues [82]. Reaching conditions are met by increasing ai. This will eliminate nonnegative eigenvalues when \ki = q and thus satisfy Theorem 6. For instance in the third-order system
&=x2, i2=x3,

The variable structure approach to plants with varying parameters q ( t ) is based on the invariancy of the sliding mode with respect to parameter variations [l I], [29], [34], [42], [44]. Parameters q ( t ) are assumed to be inaccessible for measurement and are arbitrarily varying in some known ranges:
a , , Q ai(?) Q aimax,

i = 1;

- - ,n.

The problem is to steer to zero the state of the system without disturbance (f( t ) =0). Again the control (5) is to guarantee both the existence of a sliding plane with asymptotically stable sliding mode and the reaching conditions. The equality constraints for ci in (6) cannot be satisfied when parameters ai are unknown. It has already been mentioned that these constraints vanish if k = n - 1. Then the ( n - 1) inequalities in (6) can always be satisfied for any ci and bounded parameters ai(t) with
q~sup[Ci-]-ai(t)-cic,-l+cia,~~(t)]
I

i3= u

with u= -CilxI, 9, < O two of three eigenvalues have negative real parts. This means that variable structure control u= - * , x , assure the existence of a sliding plane with asymptotically stable motion. Indeed, the condition

~ i ~ i ~ [ c i - l - a i ( t ) - c i c , - l - c i a , _ l ( t ) i] = , l,---,n-l.

216

IEEE TRANSAC~ONS ON AUTOMATIC CONTROL, APRIL

1977

When the control (5)is a piecewise linear function of (n - 1) variables the inequalities (9) are necessary and sufficient conditions for a sliding plane to exist [42]. Then there are no constraints on ci and any desirable sliding mode (4) can be designed. The reaching conditions for time-varying VSS have been obtained as a result of generalization of the theorems I for time-invariant systems. The reaching condition for Fig. 4. Disturbancerejection in VSS. time-varying second-order systems [12] is that the reaching condition for time-invariant systems be satisfied for every fixed set of parameters from the range of their then in the desired state xi= 0 the control is of the same variation. Theorem 4 holds if c,-, - u,,(t) < 0 holds for order of magnitude as f ( t ) and it is able to maintain the any u,,(t) [84]. The sufficient conditions for time-varying sliding mode. Here *,,6,s are the same functions as in (5), VSS of arbitrary order given in [ 131 are and m is the order of the differential equation for SM. It should be noted that the output of SM is usually available C i > 0 , q > - q ( t ) , pi< -a,(t), u,(t)>O. for measurement. An implication is that the coefficients u,(t) is either The conditions of the existence of a sliding plane, the greater than cn-l (which is alwayspositive) or positive. reaching conditions for time-invariant and time-varying For the VSS with an arbitrary a,,(t) reaching conditions VSS and different types of disturbances have been estabfor third- and higher order systems are formulated in [ 101 lished in [ a ] , [67], and [79H81]. In these references it is and [40] and [41], respectively. According to these results shown that the control (1 1) is able to reject the disthere exists a positive scalar a. such that the state reaches turbances and solve a tracking problem with zero error for the switching plane s = O if a1 a,,and PI < - ao. Thus, disturbances from the class for VSS with control (5) and k = n - 1 the existence of a sliding plane which is reached can always be guaranteed. When the plant parameters u,(t) are not equal to their extreme values, the coefficients c, may be varied without violation of the conditions (9). As the coefficients ci de- which encompasses exponential, harmonic functions, termine the behavior of the system in sliding mode these polynomials, etc. The implementation of the control law degrees of freedom maybeutilized to improve the dy- (10) does not require knowledge of the disturbance f ( t ) the parameter B and namic characteristics of the motion in sliding mode. The nor the plant parameters. Only ranges of plant parameter variations are needed. To design methods for adaptive VSS based on this idea are clarify the nature of this type of VSS suppose that the developed in [40] and [69]. transfer function of SM has m poles at the origin. Then its behavior for xi = 0 is governed by Rejection o f Disturbances Let us now consider the case when f ( t ) # O in (1) and is not accessible for measurement. Let xI be the difference between some reference input and the system output whichisto be steered to zero together withits n - 1 derivatives. The idea is to introduce a sliding mode for disturbance rejection. This should bepossiblesince the sliding mode is described by the homogeneous differential equation (4) whose solution tends to zero, if all the ci satisfy Hurwitz conditions. It is easily seen that the control (5) is not suitable for this task, since at x = 0 it is close to zero, u = 6. To introduce an additional control term counteracting the disturbance f ( t ) we note from the system diagram in Fig. that 4 the output y of the servomechanism (SM) should coincide with -f ( t ) . (Without loss of generality it is assumed that the reference input g ( t ) = 0.) If u is a piecewise linear function not only of x but also of the state variables of SM, that is, if
u=
i=O

x
k
i= I

m- 1
\kkiX;-6i=O

x +?p

When a sliding mode occurs all the coefficients oscillate at high frequency and to find the solution of ( 1 ; ) these coefficients shouldPe replaced by average values ' k i of \k{ such that - < \ k , ( t ) < If a : = B and =B then the output of SM belongs to the class (11) and therefore SM is able to generate the function coinciding with f(t). This will occur automatically if sliding mode exists. Similar designprinciples assuming that the parameters of the disturbance model are known using continuous control can be found in [21] and [52]. When disturbances can be measured, the control is a piecewise linear function of the state and disturbances [22], [45], [46], [71].An important property of these "comto the bined" VSS is that they are muchlesssensitive inaccuracy of the measurement of disturbances than disturbances rejection schemes based on a compensation principle [58].
consideration i s aualitative one and the physical meaning of an average value of contro? in a sliding mode will be &cussed i n the next section.

+ :

T 'h i s

SURVEY PAPER

217

Rejection of disturbances has muchin common with the control of interacting systems. To provide decoupling, interactions are considered as disturbances and rejected by the above methods. Moreover, a desirable interaction between subsystem outputs maybe established [68] by letting the equation of a sliding plane for each subsystem depend on the state variables of the others. The design methods described in Section I11 have been worked out only for single-input systems. Another limitation for these methods results from the necessity to have high-order derivatives of theoutput. Sinceallrealdifferentiators have denominators in transfer functions, the system representation (1) may be an inadmissible idealization.

Fig. 5. Sliding mode i n the intersection of &continuity surfaces.

Differential Equations of Sliding Mode

The differential equations (13) and (14) do not formally satisfy the classical theorems on the existence and uniqueness of the solutions, since they have discontinuous righthand sides. Moreover, the right-hand sides usually are not defined on the discontinuity surfaces. A series of works 111. SLIDING MODES IN DISCONTINUOUS DYNAMIC simply postulate the equations of sliding mode [64], [72], SYSTEMS W I , [~51+71. These mathematical difficulties arise because nonidealiWe now consider VSS of a general type described by ties of real systems are neglected in their models. To treat x= f (X,t,U) (13) the sliding mode carefully, various small nonidealities of time-delay,hysteresis, and other types should be taken where the x E R", u E R", f E R". Each component of into consideration. These nonidealities determine the real control vector may be equal to one of the set of continu- sliding mode, that is the motion in the vicinity of the ous functions of x and t . Analogous to what was done for discontinuity surfaces. Loosely speaking, as the nonidealiVSS with scalar control each component of control is ties tend to zero this motion tendstothe ideal sliding assumed to undergo discontinuity on an appropriate mode. If different types of nonidealities and different surface in the state space, limiting processes result in the same sliding mode equations, then itis reasonable to consider them as correct if s,-(x)> O descriptions of the ideal sliding mode. In the opposite ui+ (x, t ) , i = 1;. ,m. (14) ui= case, we have to admit that (13) and (14) describing the if s j ( x ) < O ui-(x,l), system outside the discontinuity surfaces do not allow an The design problem consists of choosing the continuous unambiguous formulation of the equations for ideal slidfunctions ui+,uj- and m-dimensional vector s (s E R") ing modes on these surfaces. Such limiting processes will always be used as the criterion of the validity of sliding with the functions si(x) as components. Suppose that the state of the system (13), (14) has mode equations? A formal technique named the equicalent control method reached the intersection of the discontinuity surfaces. In [83] will be used for finding equations of ideal sliding certain cases the state would inevitably return to this modes. In this technique a time derivative of the vector manifold should any small deviation occur. As is seen in s(x) along the system trajectory of (13)is set equal to Fig. 5 such a situation can arise evenin the casewhen sliding mode does not exist on each of the surfaces taken zero, and the resulting algebraic system is solved for the separately. Having reached the intersection the state will control vector. This "equivalent control" (if itexists)is stay in this manifold. Such a motion along the intersection substituted into the original system. The resulting equaof a number of discontinuity surfaces will also be referred tions are the equations of ideal sliding mode. From a to as sliding mode. A sliding mode of this type is a new geometric point ofview the above method means finding a continuous control which directs the velocity vector phenomenon and an interesting subject for investigation. From a practical point of view we are interested in utiliza- along the intersection of the discontinuity surfaces. This formal technique is now substantiated for systems tion of this motion for the design of control systems. The linear with respect to control, methods for the study of sliding modes in discontinuous systems and the design of systems with discontinuous f =f (x, t ) B (x, t)u (15) controls will be considered in the remaining part of the (14). The equivalent control where u is determined by paper. The two problems treated in this section are the deriva- method results in the sliding mode equations for the tion of the differential equations of sliding mode and system (14), (15), conditions for the existence of sliding mode. (For m = 1 i=f-B(GB)-'Gf, (16) the inequalities (3) provide existence conditions and the equations (4)describe the sliding mode when the discon%is approachwasappliedtothesystem with scalar control and tinuity surface is chosen in the space of phase variables.) nonidealitm of hysteresis or time-delay type in [I], [2], [17], and [63].

218

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, APRIL

1977

where m X n-dimensional matrix G = as/ ax and det (GB) is assumed to be different from zero. Introduction of nonidealities into an ideal model would result in the appearance of realslidingmodes. For examplein the scalar control case nonidealities of time-delay or hysteresis type lead to oscillations in some vicinity of a switching surface [ 11, [Z], [ 171, [63]. The amplitude and frequency of these oscillations depend on the value of the time-delay and hysteresis loop width. In the vector case when an ideal sliding mode occurs in an (n - m)-dimensional manifold s = 0, these nonidealities allow the trajectories in state space to occur in some A-vicinity of this manifold. Let the behavior of the system with nonidealities be described by the equation i = = f ( x , t ) +B(x,t)u. (17)

frequency component, it is reasonable to expect that the original control without a high-frequency component is to be close to the equivalent control. In [83] this suggestion has been verified using first-order low-pass filters whose time-constants tend to zero slower than the nonidealities. The proof consists in showing that if the filters input is u, their output tends to ueq. Existence o f a Sliding Mode The condition (3) for the existence of a sliding mode in discontinuous systemswith scalar control was obtained from the evident geometric consideration: the velocity vectors should be directed towards the discontinuity surface in its small vicinity. As is seen in Fig. 5, the sliding mode may exist in the vector case ( m > 1) even if it does not occur in each of the discontinuity surfaces. This example also shows that the conditions for sliding modes to exist are closelylinkedwith the convergence of the state to the manifold s = 0 or to the origin in the ( n - m)dimensional subspace (si.- ,sm). Therefore, in the vector case itis reasonable to use stability theory to formulate the concept sliding mode and the conditions for its existence. For the systems of the type (14), (15) this approach leads to the following results which are systematically considered in [84]. Definition: A domain S in the manifold s = 0 is a sliding mode domain if for each & >0, a IS > O exists such that any motion starting in the n-dimensional &vicinity of S may leave the n-dimensional & -vicinity of S only through the n-dimensional & -vicinity of the boundaries of S (Fig. 6)Theorem 8: For the (n - m)-dimensional domain S to be the domain of sliding mode, it is sufficient that in some there exists a continun-dimensional domain 52, S cQ, ously differentiable function c(x, s, t ) , satisfying the following conditions. 1) u ispositive definite with respect to s and for any X E Sand t

Suppose that all the nonidealities are taken into account by the control i. Their nature is not specified, i is only known to guarantee the motion in a domain llsll < A and the existence of the solution of (17). For this real sliding mode the following theorem holds. Theorem 7 [ 8 3 ] : If for any finite interval of time the solution x(t) of (17) is such that llsll < A , then for this interval, lim,,G(t)=x*(t), where x*(t) is the solution of (1 5) resulting from the equivalent control m e t h ~ d . ~ In [83] a class of the systemsisgiven for whichthis theorem is true for an infinite interval of time. Theorem 7 servesas a substantiation of the validity of equivalent control method for the systems which are linear with respect to discontinuous control. For singular caseswhen det (GB)= 0 the equivalent control is either not unique or does not exist. The first situation may result in both unique equations of sliding mode or in a variety of these equations depending on the types of nonidealities and limiting processes. When equivalent control does not exist sliding modes cannot appear, that is, the state leaves the intersection of discontinuity surfaces. For the systems (13) which are nonlinear with respect to control, even if an equivalent control exists and is unique, in general, the differential equations for sliding modes are not unique and depend on the types of nonidealities and limiting processes. All these facts are established in [83], where also the reasons for ambiguity are shown and the systems of the type (13) with unique equations for sliding modes are delineated. Slidingmodes in discontinuous systems are analyzed with the help of an auxiliary continuous equivalent control. It is of interest to show the physical meaning of this function. In sliding modes each component of control may be considered asa function consisting of a low frequency, or average, component, and a high frequency component. The behavior of the system primarily depends on the average rather than on the high frequency component. Since the equivalent control does not have a high3The assumptions about functions f, B, s, and initial conditions x ( t J and x*(@ are given i n [83].

IIsII = R

inf u = h,,

llsll = R
HR

sup u = H,, depend only on R ).

h, # O if R # 0, (h,,

2) Time derivative of u for (15) has negative supremum on small enough spheres llsll= R with removed points on the discontinuity surfaces where this derivative does not exist. The equation o f the motion projected on subspace (spa * * J , ) , i = Gf GBu, (18)

should be considered in using Theorem 8, which is an analog of a Lyapunov stability theorem. The domain S is s , ) is the set of x for which the origin in subspace (si; . . , an asymptotically stable equilibrium point for the dynamic system (14), (18). Unfortunately there are no standard methods to find the functlon u (as there areno methods to generate Lyapunov functions for arbitrary

SURVEY PAPER

219

problem of desirable characteristics of the sliding mode and the problem of the existence of a sliding manifold may be decoupled, as the following theorem shows. Theorem 9 [84]: The equation of the sliding mode is invariant with respect to nonsingular transformations s*=H,(x,t)s, u*=H,(x,t)u, detH,#O, detH,#O. (19) This theorem means that the sliding mode is governed by the same equation (16) if the components of the control vector undergo discontinuities on new surfaces nonlinear systems). Therefore, special cases have been si* = O or the components of the new control vector u* classified when the domain of the sliding mode may be undergo discontinuities on already chosen surfaces si= 0. derived. These cases are as follows. At the same time, the equation of the motion projected 1) There exist matrices W ( x , t ) and L ( x , t ) such that on subspaces (s:, * s , ) : in the first case and on L = - WGB, WT= W , W>O, L=llZ,ll, Zji>Zm,=llZul, i = (s,, s , ) , in the second case do depend on H, or H,, j# 1 1 ; . ,m. respectively. In general, for any matrix G the matrices H, 2) Matrix GB in (18) is symmetric. or H, can be chosen such that this mth-order equation is reduced to one of the cases (listed in Section 111) when the 3) Matrix GB is diagonal. 4) Matrix GB has a dominating diagonal (for each row existence conditions can be obtained. the absolute value of the diagonal element is greater than The methods for study of the existence conditions may the sum of absolute values of the rest of the elements). be applied to analyze reaching conditions. These two 5) The systems with so-called hierarchy of controls, problems are closely related because the first one means that is when luil >Ej=j+,lduu,l,d,some functions of x and asymptotic stability in small, and the second in large. Wenow consider the properties of multiinputs VSS t. For the first two cases the domain of the sliding mode resulting from the above design procedure [84].Let the may be found with the help of Lyapunov functions which control system be represented by are quadratic forms of s with coefficients depending on x X=AAx+Df(t)+Bu (20) and t . Cases 3 and 4 may be decoupled into m scalar cases. For case 5, hierarchy of controls implies that one of where x E R n, u E R , A , D , B-constant or time-varying the components of control assures the existence of sliding matrices, and f ( t ) E R is the disturbance. The compomode on the first discontinuity surface, the second on the nents of u undergo discontinuities on m planes si = 0, that intersection of the first two surfaces, etc., up to the inter- is section of all m discontinuity surfaces. The analysis of the s = C x , s E R , C = const. (21) cases 3, 4, and 5 is based on the inequalities (3) which are the conditions for the existence of a sliding mode for the For time-invariant VSS when f = O the control is a systems with scalar control. piecewise linear function of some of the state variables
Fig. 6. Definition of a domain of sliding mode.

IV. DESIGN OF MULTIINPUT VSS In this section we treat control systems linear with respect to control (14), (15). The design idea is similar to the basic approach in the single-input case. First, a sliding mode isdesigned to have some prescribed properties. Second, it is guaranteed to exist at any point of the intersection s = 0 of the discontinuity surfaces, which is then referred to as a sliding manifold. Third, it is guaranteed that the state reaches a sliding manifold. As it follows from (16) the motion in a sliding mode depends on the m X n elements of matrix G. That means that desirable properties for this motion may be obtained by a proper choice of discontinuity surfaces. It should be noted that the system order is thus reduced from n to n - m due to the m-dimensional equation s=O, allowing the elimination of m state variables. Although the existence conditions discussed in Section 111 are unknown for an arbitrary matrix GB in (18), the

u = -*xk-CY,(xk)

= ( x , ; . .,Xk),*,
,m ( Y , , , E $ - const * ,k

{ Po,
au,

if six,>O, i = 1; if sixj<O, j = 1 ;

(22)

8 = (CYl, . . ,C Y , ) , = 8oi sgnsi, aOi are small positive scalars. To assure the existence of a sliding manifold with asymptotically stable sliding motion itisnecessary and sufficient that the system (20)-(22) with = Qe, where some constant values between a , and & has n - m eigenvalues with negative real parts. The design procedure consists of finding R = const such that u = R x places n - m eigenvalues at desired locations. Then the parameters of control in (22) are chosen to result in the sliding mode determined by these (n - m) eigenvalues. Different types of reaching conditions depending on the matrices H, and H, in (19) are given in [84].

E ,

S 2 ,

220

WEE TRANSACTIONS ON AUTOMATIC CONTROL, APRlL

1977

In general, the motion in sliding mode depends not only on matrix C in (21), but also on the parameters of the system (20) and disturbances. If the influence of disturbances and parameter variations can be reduced enough by a proper choice of discontinuity surfaces (or matrix C ) , then it is reasonable to design VSS such that the slidingmodein the intersection s = O alwaysarises. This problem may be solved with the help of a hierarchy of controls when only the ranges of the parameter variations and the class of disturbances (11) are known. The sensitivity of sliding mode with respect to plant parameters is studied in [62]. The invariancy conditions of a sliding mode are studied in [20] for the system i=Ax+h(x,t)+Bu where h(x, t ) E R n, h(x, t ) depends on state variables, disturbances, and time-varying parameters. The equations of the sliding mode on manifold s = O do not depend on h(x,t) if rank ( B ,h ) = rank B. (23)

plant in a sliding mode, then the average values of discontinuous coefficients of the model depend on the plant parameters. These average values (which are equal to the equivalent control) can be measured employing first-order filters (see Section 111) and then used for calculation of the unknown parameters. In conclusion we list some results of practical application of VSS. Control methods for chemical processes are considered in [30]. A universal system of control devices for a wide range of processes in steel, power, nonferrous, chemical and food industries is described in [33]. Applications of VSS t o automation of a power station and flight are given in [47] and [76], respectively.

ACKNOWLEDGMENT
The author wishes to thank Prof. J. B. Cruz, Jr., Prof. P. V. Kokotovic, and Prof. W. R. Perkins for their invaluable assistance in preparing this survey as a technical paper in English. The author also wishes to thank the Coordinated Science Laboratory and the Department of Electrical Engineering,University of Illinois, Urbana, for support in preparation of this paper.

The conditions of invariancy with respect to plant parameter variations and disturbances, and the conditions of so called selective invariancy (when one set of state variables do not depend on another set of variables) follow from (23), [20]. Note that the equations in phase canonic form satisfy condition (23). The next set of problems considered in the VSS theory pertains to self-optimizing systems and nonlinear programming [54], [55]. In such problems we search for the extremum of some function y=f( yx E )R , ,
xER, ~ 9 .

REFERENCES
J. Andre and P. Seibert, Uber stuckweise linear differentialgleichungen, die bei regelungsproblem aufretten, Vols. I, I1,Arch der Math. vol. 7, no. 2, pp. 140-156, no. 3, pp. 157-164, 1956. A. A. Andronov, A. A. Vitt, and S. Khaikin, Theoly of Oscillations (in Russian). Moscow: Fizmatgiz, 1959. A. V. Bakakin, M. A. Bermant, and V. B. Exerov, Application of variable structure systems to stabilization of time varying plants in f thecontrolling thepresence of restrictions on themovement o device, Automat. Remote Contr., no. 7, pp. 1016-1021, 1964. A. A. Bakakin and V. I. Utkin, Variablestructuresystemswith pure delay in switching devices,in Variable Structures Systems and their Application to Flight Automation (in Russian). Moscow: Nauka, 1968, pp. 64-71. , Conditions of stability of sliding mode in variable structure systems, in Analysis and Design o f Control Systems ( i nRussian). Moscow: Nauka, 1968. E. A. Barbashin, Introduction to the 73eoly o f Stabilig ( i nRussian). Moscow: Nauka, 1967; also Wolters-Noordhoff hblishing, Groningen, The Netherlands, 1970 (in Enghsh). E. A. Barbashin and E. I. Gerashchenko, Forced sliding regimes in automatic control systems, Differential Equations, vol. I , no. 1, pp. 1620, 1965. , On a design principle for control systems, Automat. Remote Contr., no. 6. pp. 989-997, 1965. E. A. Barbashinand v. A. Tabueva, A method of stabilizing third-order controlsystems with largegains, Automat. Remote Conrr., vol. I, no. 10, pp. 1215-1222, vol. 11, no. 5, pp.562-567, 1963. E. A. Barbashin, A. Tabueva, and R. M. Eidinov, On the stability ofvariablestructure control systems when sliding conditions are violated, Automat. Remote Contr., no. 7, pp. 81-16, 1963. M. A. Bermant and S. V. Emelyanov, On a metho! of controlling a certain class of plants with varyingparameters, Automat.Remote Conrr., no. 5, pp. 568-579, 1963. - On stabihty of a class of time-varying control systems, Eng.Cybern., no. 6, pp. 147-152, 1964. M. A. Bermant and V. I. Utkin, On stability of time-varying variable structure control systems, in Variable Structure Systems and their Application to Flight Automation (in Russian). Moscow: Nauka, 1968, pp. 72-74. V. I. Buyakas,Optimalcontrol by systemswithvariablestructure, Automat. Remote Contr., no. 4, pp. 579-589, 1966. T. A. Benodinskaya and E. F. Sabaev, Study of state space features in variable-structurecontrolsystems, Automat. Remote Contr., no. 7, pp.1105-1108,1973. , "Stability conditions i n large for variable structure systems, Automat. Remote Contr., no. IO, pp. 15961599, 1974.

(24)

Forf(x) to be maximized, the variable y is compared with a monotonously increasing reference input g ( t ) . Discontinuous control reduces the difference between y and g to zero inslidingmode. As a result y ( t ) tracks g ( t ) and reaches the maximum point. If the constraints are active, an auxiliary discontinuous control guarantees the existence of sliding mode on the boundary of the admissible domain which leads to the search of maximum point in this domain. The above method does not need the measurement of the gradients of the function to be minimized nor those of the constraint functions. Application of the technique of Section I11 substantiates the convergence of the gradient procedures for piecewise differentiable penalty functions [%]. These functions with finite penalty factors are known to have maximum point coinciding with the solution of a nonlinear programming problem [91]. Identification methods based on the models with variable structure are described in [57], [84]. It is assumed that the order of the differential equation representing the plant is known and its right-hand side may be expanded into finite series with unknown coefficients. If the state of themodelwith variable structure tracks the state of the

SURVEY PAPER

22 1

P. V. Bromberg, Matrix Methods in the Theory of Switched Sampled-Data Systems (in Russian). Moscow: Nauka, 1967. J. Clegg, Timedependent nonlinear compensating network, Elec. Eng., no. 10, 1956. S. J. Garret, Linear switching conditions for a third order positive-negative feedback control, Appl. Znd., no. 54, 1961. B. Draxenovic, The invariance conhtions in variable structure systems, Automatica, vol. 5, no. 3, pp. 287-295, 1969. E. J. Davison, The output control of linear time-invariant multivariable systems with unmeasurable arbitrary disturbances, IEEE Trans. Automat. Contr., vol. AC-17, no. 5, pp. 621430, Oct. 1972. E. Dud$, Conditions of switchings in tracking variable structure systems, Eng.Cybern., no. 2,pp.106-114,1964. R. M.Eydinov,On an estimate of transient time in variable structure systems, Eng. Cybern., no. 5, pp. 129-135, 1965. V. B. Ezerov and V. I. Utkin, A variable structure filter as a phase-lead network, in Variable Structure Systems and their Application to Fiight Automation (in Russian). Moscow: Nauka, 1968, pp. 171-182. S . V. Emelyanov, Variable Structure Control System (in Russian). Moscow: Nauka, 1967; also Oldenburg Verlag Munchen-Wien (in German). , The use of nonlinear corrective devices of the key type to improve the behavior of second order control systems, Automat. Remote Contr., no. 7, pp. 844-859, 1959. , Control of first orderdelayed systemsby means of an astatic controller and nonlinear correction, Automat. Remote Contr., no. 8,pp. 983-991, 1959. , On high quality control of a class of nonlinear time-varying plants, Eng. Cybern., no. 4, 1962. S. V. Emelyanov and M. A. Bermant, On high-quality control of time-varying plants (in Russian), DoklaQ ANSSSR, vol. 145, no. 4, pp. 748-751, 1962. S . V. Emelyanov et ai., Mathematical models of industrial of industrial processes and design of variable structure control systems (in Russian), Metalurgia, 1964. S . V. Emelyanov and N. E. Kostyleva, On peculiarities of variable structure control systemswith discontinuous switching functions 153, no. 4, pp. 776778, (inRussian), Dokla& ANSSSR, vol.
1963

[32] , Design of variable structure controlsystemswithdiscontinuous switching functions, Eng. Cybern., no. 1, pp. 156160, 1964. [33] S. V. Emelyanov, B. P. Matic, and N. E. Kostyleva,Auniversal uniform control system with variable structure, vols. I, I1 (in Russian), I n s f . Contr. Syst., no. 12, pp. 8-16, 1973, no. I , pp. 5-10, 1974. [34] S. V. Emelyanov et al., Designprinciples for variable structure control systems, in Proc. 3rd IFAC Congr., vol.1,book3,pp. 40C.140C.6, 1966 [35] S. V. Emelyanov and V. A. Taran, Use of inertial elements in the design of a class of variable structure control systems, vols. I, 11, Automat. Remote Contr., no. 1, pp. 2942, no. 2, pp. 183-190, 1963. , Use of passive switched filters in variable structure control [36] systems, Eng. Cybern., no. 5, pp. 143-149,1963. , On design of variable structure control systems for control [37] of linear plants, Eng. Cybern., no. 2, pp. 69-78, 1963. , Supplement to the article by E. I. Gerashchenko, Eng. [38] Cybern., no. 4, p. 92, 1963. , Stabilization of variable structure control systems by means [39] of inertial elements with variable time constants, Auromaf. Remote Contr., no. 6, pp. 79&794, 1964. [ a ] S . V. Emelyanov et a/., Theory o f Variable Structure S y s f e m (in Russian). Moscow: Nauka, 1970. 1411 S. V. Emelyanov and V. I. Utkin, On stability of a class of variable structure systems, Eng. Cybern., no. 2, pp. 115-1 17, 1964. [42] , Variable structure systems for control of plants withlarge parametervariations (inRussian), Doklady ANSSSR, vol.152, no. 2, pp. 299-301,1963. [43] S. V. Emelyanov, V. I. Utkin, and V. A. Taran, On reaching conditions for variable structure systems of the third order, Eng. Cybern., no. 3, pp. 123-128,1965. [44] S. V. Emelyanov and V. I. Utkin, Invariant solutions of differential equations with discontinuous coefficients. Design principles of variable structurecontrol systems, in Mathematical Theory of Control. NewYork:Academic,1967. [45] S . V. Emelyanov and A. I. Fedotova, Design of astatic tracking systems with variable structure, Automat. Remote Contr., no. 10, pp. 1223-1235,1962. , Reproduction of a reference input by astatic tracking sys[46] Automat. Remote Confr., no. 1, pp. tems with variable structure, 67-12,1965. [47] J. Erschler, F. Roubellat,and J. P. Vernhes, Automation of a hydroelectro power station using variable structure systems, Automatica, vol. 10, pp. 31-36, Jan. 1974. Synthesis of nonlinearcontrol systems, Trans. [48] I. Fliigge-Lotz, IRE PGAC, vol.11, May 1956.

I. Fliigge-Lotz and W. S. Wunch, Mechanical reproduction of an arbitrary function of time, TR-N19, Office of Naval Res. Div. Eng. Mech., Stanford Univ., Stanford, CA, Mar. 1952. - On a nonlinear transfer system, J . Appl. Phys., vol. 26, no. 4,ld55. E. I. Gerashchenko, On stability in a sliding plane of a class of variable structure systems, Eng. Cybern., no. 4, pp. 92-98, 1963. , On adegree of stability of nonlinear systemsinsliding mode, Eng. Cybern., no. 2, p. 91, 1964. C. D. Johnson, Further study of the linear regulator with disturbances-The case of vector disturbances satisfying a linear differential equation, IEEE Trans. Automat. Contr., vol.AC-15, pp. 222-228, Apr. 1970. 1541 S. K. Korovin and V. I. Utkin, Application of sliding mode to static optimization, Automat. Remote Contr., no. 4, pp. 57tL519, 1972. 1551 S. K. Korovin and V. I. Utkin, Using sliding modes in static optimization and nonlinear programming, Automafica, vol. 10, pp. 525-531,1974. , On the method of piece-wise differentiable penalty functions, Automat. Remote Contr., no. 4, 1976. N. E. Kostyleva, Recovery of differential equations of a plant by method of variablestructure systems, Automat. Remote Contr., no. 10, pp.1588-1595,1971. A. I. Kuhtenko, Inuariang Problem in Automation (inRussian), Gos. izd. tech. lit. Kiev. 1963. A. M. Letov, Conditionally stable control systems (on a class of optimal control systems), Automat. Remote Contr., no. I , pp. 649464. 1957. ~ - -I. M: Makarov and V. Z . Rakhmankulov, Improvement of variablestructure systemstimeresponseby combination of sliding hyperplanes, Automat. Remote Contr., no. 6, pp. 896905, 1969. V. A.Maslennikov,High-quality control of neutralplants by means of an astatic controller, Automat.RemoteContr., no.2, 1956. B. Matic, Design of asensitivitymodel for variable structure systems with sliding modes (in Serbian), Energoincest, no. 2, 1971. Yu. I. Neimark, On sliding modes in relay control systems, Automat. Remote Confr.,no. 1, pp. 29-35,1957. - Aremark for the report byA. F. Filippov,in Proc. Ist IFAC Congr., 1960, pp. 926927. G. M. Ostrovsky, Application of nonlinear correction to control no. 11, systems of the second order, Automat.RemoteContr., 1956. , Control system with discontinuous parameters, in Proc. lsr IFAC Congr.. 1960, pp. 529-534. design B. N. Petrov, S . V. Emelyanov, and V. I. Utkin, A principle forinvariant variable structure systems(inRussian), DokIa& ANSSSR, vol. 154, no. 6, pp. 1294-1296, 1964. B. N. Petrov, S . V. Emelyanov, and M. B. Gritsenko, Autonomity in interacting control systems with variable structure (in Russian), DoklaQ ANSSSR, vol. 169, no. 1, pp. 52-54,1966. B. N. Petrov, S. V. Emelyanov, E.N. Dubrovsky, and A. V. Kortnev, Ondesignprinciple for adaptive systemwithvariable pp. structure (in Russian), DoklaQANSSSR, vol. 170, no. 6, 1279-1282,1966. B. N. Petrov, S . V. Emelyanov, and N. E. Kostyleva, On control of linear time-varying plants (in Russian), Doklady ANSSSR, vol. 155, no. 1, pp. 61-64, 1964. B. N.Petrov, G. M. Ulanov, and S. V. Emelyanov,Optimality and invariancyin control systemswithfixed and variablestructure, in Proc. 2nd IFAC Congr., vol. 1, 1963, pp. 421429. A. M. Popovsky,Linearization of sliding mode of acontroller with constantoutput velocity, Automat. Remote Contr., no. 3, 1950. B. G. Pittel, Some problem of optimal control systems, Auromat. Remote Contr., no. 9, pp. 1078-1091,1963. N. A. Smetanin,Harmonic linearization of a component with variable structure, Automat. Remote Confr., no. 6, pp. 998-1001. 1965. ~751 G. M. Stout, Nonlinearity in control systems, Contr. Eng., Apr. 1956. application of a new nonlinear [761 W. Sobotta, Realization and attitude control andstabilization system. in Proc. 4th IFAC Symp. Control in Space, 1974. V. A. Taran, Control of alinear plant with an astatic variable structure controller without pure derivatives in control law, vols. I, 11, Automat. Remote Contr., no. 10, pp. 1278-1288, no. 11, pp. 1406-1412,1964. , Design of control systems applying switched filters, Eng. Cybern., no. 4, pp. 172-180,1965. V. I. Utkin,Oncompensation of theforcedcomponent inthe motion of control systems with variable structure, Eng. Cybern., no. 4, pp. 167-171,1965. , Quasi-invariant control of linear plants with disturbances, Eng. Cybern., no. 5, pp. 149-155,1966. , Stability conditions for forced motion in variable structure
I

222
systems (in Russian), in Variable Structure @stems and their Application to Flight Automation. Moscow: Nauka, 1968, pp.

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, APRIL

1977

75-80.

, Sliding modes in variable structure systems (in Russian), Automatika, no. 3, 1970. , Equations of sliding mode in discontinuous system, vols. I, 11, Automat.RemoteContr., no. 12, pp. 1897-1907, no. 2, pp.
211-219, 1972.

[92] K. K. Zhiltsov, A method of describing functionsfor systems with logic control, Automat. Remote Contr., no. 2, pp. 286-289, 1966.

their Application to VariableStructure System (in Russian). Moscow: Nauka, 1974. A. F. Filippov, Differential equations with discontinuous rightVadim I. Utkin was born in Moscow, USSR, in hand sides (in Russian), Mat. Sb., vol. 5, no. 1, pp. 99-128, 1960. 1937. He received the Dipl. Ing. from the , Application of the theory of the differential equations with Moscow Power Institute, Moscow,USSR, and discontinuous right-hand sides to nonlinear control problems, in the CandidateandDoctor of Science degrees Proc. Ist IFAC Congr., pp. 923-925, 1960. from the Institute for Control Sciences, in 1960, Ya. Z. Tsypkin, Theoly of RelayControlSystems ( i nRussian). 1964,and 1972,respectively. Moscow: Gostechizdat, 1955. E. K. Shigin, On improvement of transient processeswith the Since 1960 he has been with the Institute for help of correction networks with varying parameters, Automat. Control Sciences and is the Director of the VariRemote Contr., no. 4,pp. 299-304, 1958. able Structure System Laboratory. He is Pro, Controller with,witched parameters for the plants with fessor of Control Theory at the Moscow Metalpure-delay, vols I, 11, Automat. Remote Contr., no. 10, pp. lurgical Institute. From 1975-1976 he was a 1612-1618, 1965, no. 6, pp. 999-1007, 1966. W. S. Wunch, Reproduction of an arbitrary function of time by Visiting Professor at the Coordinated Science Laboratory and the Dediscontinuous control, Ph.D. dissertation, Stanford Univ., Stan- partment of Electrical Engineering, University of Illinois, UrbanaChampaign. His research interests are variable structure systems, theory ford, CA, May 1953. [911 W. Zanpdl, Nonlinearprogramming via penalty functions, of discontinuous dynamic systems, and optimization. Management Sci., no. 13, 1967. Dr. Utkin is an Associate Editor of IFAC Journal Automatica.

, SlidingModesand -