Proceedingsof the American Control Conference

San Diego, California June 1999

Sliding Mode : Control Engineering in Practice

K. David Young and Umit Ozgiiner

Abstmct- This paper presents sliding mode as viewed control design solutions for real life engineering appli-
from control engineering practice. It catalogs imple- cations;
mentable sliding mode control design solutions, and pro-
vides a tutorial for control engineers who are interested in initiate a dialog with practicing control engineers on
applying sliding mode. sliding mode control by threading the many analyti-
Kegword- Discrete time systems, Multivariable systems, cal underpinnings of sliding mode analysis through a
Nonlinear systems, Robustness, Sampled data systems, Sin- series of design exercises on a simple, yet illustrative
gularly Perturbed systems, Uncertain systems, Variable control problem; and
structure systems.
establish a frame of reference for future sliding mode
control research.
I. INTRODUCTION The flow of the presentation in this paper follows the
During t.he last two decades since the publication of the chronological order in the development of VSS and SMC:
survey paper in the IEEE llansactions of Automatic Con- First we introduce issues within Continuous Time Sliding
trol in 1977 [l],significant,interest on Variable Structure Mode in Section 11, then in Sectiou III, we progress to
Systems (VSS) and Sliding Mode Control (SMC) has been Discrete Time Sliding Mode, followed with Sampled Data
generated in the control research community wrldwide. SMC Design in Section IV.
One of the most intriguing aspects of sliding mode is the
discontinuous nature of the control action whose primary 11. CONTINUOUS TIME SLIDING MODE
function of each of the feedback channels is to switch be-
Sliding mode is originally conceived as system motion for
tween two distinctively different system structures (or com-
dynamic systems whose essential open loop behavior can be
ponents) such that a new type of system motion, called
Sliding Mode, exists in a manifold. This peculiar system modeled adequately wit,h ordinary differential equations.
characteristic is claimed to result in superb system perfor- The discontinuous control action, which is often referred to
mance which includes insensitivity to parameter variations, as Variable Structure Control (VSC), is also defined in the
and complete rejection of disturbances. The reportedly su- continuous time domain. The resulting feedback system,
the so-called VSS, is also defined in the continuous time
perb system behacvior of VSS and SMC naturally invites
domain, and it is governed by ordinary differential equa-
criticism and scepticism from within the research commu-
tions with discontinuous right hand sides. The manifold of
nity, and from practicing control engineers alike [2]. The
the state space of the system on which sliding mode occurs
sliding mode control research community has risen to re-
is the Sliding Mode Manifold, or simply, Sliding Manifold.
spond to some of these critical challenges, while at the same
For control engineers, the simplest, but vividly percepti-
time, contributed to the confusions about the robustness of
ble example is a double integrator plant, subject t,o time
SMC by offering incomplete analyses, and design fixes for
optimal control action. Due to imperfections in the iniple-
the so-called chattering phenomenon [3]. Many analyti-
mentations of the switching curve, which is derived from
cal design methods were proposed to reduce the effects of
the Pontryagin Maximum Principle, sliding mode may oc-
chattering [4], [5], [6], [7], [8] - for it remains to be the
cur. Sliding mode was studied in conjunction with relay
only obstacle for sliding mode to become one of the most
control for double integrator plants, a problem motivated
significantdiscoveries in modern control theory; and its po-
by the design of attitude control systems of missiles with
tential seemingly limited by the imaginations of the control
jet thrusters in the 1950's [16].
researchers [9], [lo], [ll].
In contrast to the published uorks since the 1977 arti- The chattering phenomenon is generally perceived as mo-
tion which oscillates about. the sliding manifold. There
cle, which serve as a status overview [12], a tutorial [13]
of design methods, or another more recent state of the art are two possible mechanisms which produce such a mo-
tion. First, in the absence of switching nonidealities such
assessment [14], or yet another survey of sliding mode re-
search [15], the purpose of this paper is to provide a com- as delays, i.e., the switching device is switching ideally at
prehensive guide to Sliding Mode Control for control engi- an infinite frequency, the presence of parasitic dynamics
neers. It is our goal to accomplish these objectives: in series with the plant causes a small amplitude high fre-
quency oscillation to appear in the neighborhood of the
offer a catalog of implementable robust sliding mode
sliding ma.nifold. These parasitic dynamics represent the
K. D. Young is with Applied Technology Center, Houg Kong Uni- fast actuator and sensor dynamics which, according to con-
versity of Science and Technology, Clear Water Bay, Kowloon, Hong trol engineering practice, are often neglected in the open
Kong, E-mail: kkdyoungQust.hk loop model used for control design if the associated poles
U . Ozguner is with t h e Department of Electrical Engineering, Ohio
State University, 2015 Neil Ave., Columbus, Ohio43210 USA, E-mail: are well damped, and outside the desired bandwidth of the
ozguner. lQosu.edu feedback control system. Generally, the motion of the real

0-7803-4990-6199$10.00 0 1999 AACC 150

system is close to that of an ideal system in which the par- where h > 0 is the observer feedback gain, and IC,is the
asitic dynamics are neglected, and the difference between output of the parasitic sensor dynamics. The SMC and the
the ideal and the real motion, which is on the order of the associated sliding manifold defined on the observer state
neglected time constants, decays rapidly. The mathemati- space is,
cal basis for the analysis of dynamic systems with fast and U = -sgn(*) . (2)
slow motion is the theory of singularly perturbed differen- The behavior of the closed loop system can be deduced
tial equations [17], and itsextensions to control theory have from the following fourth order system,
been developed and applied in practice [18]. However, the
theory is not applicable for VSS since they are governed by + A
k = -sgn(z - e) d ( t ) , e = z - 5, (3)
differential equations with discontinuous right hand sides.
The interactions between the parasitic dynamics and VSC
+
e = -he+ h(z - xs) d ( t ) , (4)
generate a non-decaying oscillatory component of finite am- 2xs+ 27-sxs + x s = 2. (5)
plitude and frequency, and this is generically referred to as First we consider the case when d ( t ) = 0. Using an infinite
chattering. gain linear function g(z - e ) to approximate the switching
Second, the switching nonidealities alone can cause such function sgn(z - e ) , and since T~ is finite, the above sys-
high frequency oscillations. We shall focus only on the de- tem is a singularly perturbed system with g-' being the
lay type of switching nonidealities since it is most relevant parasitic parameter. The slow dynamics which are of third
to any electronic implementation of the switching device, order can be extracted by formally setting 9-l = 0, and
including both analog and digital circuits, and micropro- x = e,
cessor code executions. Since the cause of the resulting
chattering phenomenon is due to time delays, discrete time e = --hzs, (6)
control design techniques, such as the design of an extrap +
$xS 27sxs +zs= e . (7)
olator can be applied to mitigate the switching delays [19]. It is possible to further apply a singular perturbation anal-
These design approaches are perhaps more familiar to con- ysis to insure that given r,, there exists h > 0 such that
trol engineers. the asymptotic observer dynamics are of first order, and
Unfortunately, in practice, both the parasitic dynamics its eigenvalue is approximately -h. Clearly, the adverse ef-
and switching time delays exist. Since it is necessary to fects of the parasitic sensor dynamics are neutralized with
compensate for the switching delays by using a discrete an observer based SMC design. If a switching function is
time control design approach, a practical SMC design may realized in the SMC design, the only remaining concern
have to be unmoidably approached in discrete time. will be switching time delays, and if the obsemr is to be
A . Observer based Sliding Mode Control implemented in discrete time, the entire feedback design
including the compensation of switching time delays may
Recognizing the essential triggering mechanism for chat- be best carried out in the discrete time domain. Figure 1
tering is due to the interactions of the switching action with is a block diagram of this design. Note that the switching
the parasitic dynamics, an approach which utilizes asymp- element is inside a feedback loop which passes through only
totic observers to construct a high frequency by pass loop the observer, bypassing both blocks of the plant dynamics.
has been proposed[4]. This design exploits a localization This is the so-called high frequency bypass effects of the
of the high frequency phenomenon in the feedback loop by observer based SMC [4], [26].
introducing a discontinuous feedback control loop which is When d ( t ) = 0, its effects on the convergence of asymp-
closed through an asymptotic observer of the plant [25]. totic observess are well known. If d ( t ) is an unknown con-
Since the model imperfections of t,he observer are suppos- stant disturbance, a multivariable servomechanism formu-
edly smaller than those in the plant, and the control is lation can be adopted to estimate both the state and ex-
discontinuous only with respect to the observer wiables, ogenous disturbance in a composite asymptotic obserwr.
chattering is localized inside a high frequency loop which The resulting feedback system is a Variable Structure (VS)
bypasses the plant. However, this approach assumes that Servomechanism [25], [27]. In general, d ( t ) can be the out-
an asymptotic observer can indeed be designed such that put of a linear time-invariant system whose system matrix
the observation error converges to zero asymptotically We is known, but the initial conditions are unknown.
shall discuss the various options available in observer based For bounded but unknown disturbances with bounded
sliding mode control in the following design example. time derivatives, the only known approach to ensure the
robustness of the asymptotic observer is to introduce a
A.l Design example of observer based SMC high gain loop around the obserwr itself to reject the un-
For the relay control example, we examine the util- known disturbance, i.e., by increasing the gain h in the
ity of the obserwr based SMC in localizing the high fre- observer such that the effects of d ( t ) are adequately atten-
quency phenomenon. For the nominal plant, the follow- uated. However, the requirements for disturbance at,ten-
ing asymptotic observer results from applying conventional uation and closed loop stability must be balanced in the
state space linear control design, design, and if sliding mode is to be preserved in the mani-
fold 2 = 0 , g must be sufficiently larger than h. A switch-
+
j: = h(xs - i ) U ! ing function implementation of the SMC muld seem t,o

151
ensure the necessary time scale separations, however, t,he issues are transferred to the SM disturbance estimator and
condition g << l / r s should also be imposed to avoid ad- its associated sliding mode. While there are many engi-
verse interactions with the parasitic dynamics. Note that neering issues to be dealt with in this approach, simulation
if the high gain loop in the asymptotic observer is imple- studies and experiment results [31] show that desired ob-
mented with a switching function, it is referred to as a jectives are indeed achievable.
Sliding Mode Observer [28], [29], [30]. Since two sliding
manifolds are employed in the feedback loops, the closed B.l An SM disturbance estimator
loop system robustness must be carefully examined when Once again we return to the simple relay example with
less than infinite switching frequencies are to be expected. parasitic sensor dynamics for our design of a disturbance
In such robustness analysis, the relative time scales of t,he estimator. The plant model is
various motions in the system can be managed with singu-
lar perturbation methods, similar to that applied to high i =u+d(t), (8)
gain observers. T,’$, + 2TsXs + 2, = 2 . (9)
The performance of the observer based SMC can be eval-
uated by simulation. We let the senmr dynamic time con- We shall design a disturbance estimator with sliding mode
stant be T, = 0.01, and assume the same unity reference as follows:
command and constant disturbance as in the boundary i=u+sgn(z,-i). (10)
layer control example, d ( t ) = 0.5. A linear feedback gain Suppose sliding mode occurs on e = 0. Since u(t) is con-
approximation with a boundary layer of 0.002 is used in tinuous and differentiable, from the error dynamics,
place of the switching control in the observer based SMC.
The closed loop eigenvalues are at {-5lO(due to bound- A
e = -sgn(z. - 2 ) + d ( t ) + i , - i , e=x,-2. (11)
ary layer), -12.92(from observer), -59.44, -127.62(shifted
sensor poles)}. Figure 2 shows the error response between The “equivalent control” is the control which keeps the
the reference and the obsened state. The steady state er- trajectories of the system on e = 0. It can be solved from
ror of -0.001 reflects the attenuation of the disturbance e= 0,
by the high gain of 500. Note that sliding mode in the [ S d x s - 211eq = d(t) + o(~s
1. (12)
observer state space can be implemented with high gain
with no adverse interactions with the parasitic dynamics. Note that from Eqn.(9), 12, -XI = O(T,). Thus, within
Figure 3 shows the observer state’s tracking of the unity this estimator, there exists a signal which, under the sliding
reference command despite the constant disturbance. The mode condition, is O(T,) close to the unknown disturbance
superb rejection of the disturbance by sliding mode in the d ( t ) . This forms the basis of a feedback control design
observer state space is expected since a large gain value which utilizes this signal to compensate the disturbance to
can be chosen freely when the constraints imposed by the O(r,). The resulting control law has a convent,ional lin-
parasitic dynamics are no longer present. Howver, also ear feedback component, and a disturbance compensating
shown in this figure, the plant state response has a steady component, and for this system
state error of 0.05 which is due to the observation error
caused by the relatively low feedback gain of the obsener,
U= -kf - [sgn(z, - *)le,. (13)
h = 10. This error can be reduced by increasing the value The extraction of the equivalent control from the sliding
of gain h, provided that the time scales and stability of the model control signal is by low pass filtering. While theo-
system are preserved. retically there exists a low pass filter such that the equiv-
alent control can be found, in practice, the bandwidth of
B. Disturbance Compensation the desired closed loop system, the spectrum of the distur-
In SMC, the main purpose of sliding mode is to reject bance, are all important considerations in the selection of
disturbances and to desensitize against unknown paramet- the cutoff frequency of this filter. A recent analysis of the
ric perturbations. Building on the observer based SMC, equivalent value filtering in disturbance estimation prob-
a sliding mode disturbance estimator which uses sliding lem [41] provides an initial attempt to formalize the prac-
mode to estimate the unknown disturbances and paramet- tical implementation concerns in a theoretical domain.
ric uncertainties has also been introduced [SI. In this ap- For a closer examination of the behavior of this distur-
proach, the control law consists of a conventional continu- bance estimator, we let the sensor time constant be once
ous feedback control component, and a component derived again T. = 0.01, and simulate the system’s responses with
from the SM disturbance estimator for disturbance com- the same unity reference command, and constant distur-
pensation. If the disturbance is sufficiently compensated, bance d(t) = 0.5 as before. After canceling the distur-
there is no longer the need to evoke a discontinuous feed- bance, w design a closed loop system with a time con-
back control to achieve sliding mode, thus, the remaining stant of one seconds which can be attained with k =
control design follows the conventional wisdom, and issues 2. A boundary layer of 5 x replaces the switch-
regarding unmodeled dynamics are no longer critical. Also ing function in the estimator. The closed loop eigenval-
chattering becomes a non-issue since a conventional feed- ues are {-2000.(from the boundary layer), -l,(the domi-
back control instead of SMC is applied. The critical design nant closed loop pole), -96.753~~101.83(theshifted sensor

152
poles)}. For low pass filtering, a third order buttenvorth tivity to parameter variations in sliding mode [32] which
filter with a 3dB corner frequency of 50 rad./sec. is used to are satisfied in the nominal plant model are violated. This
filter the equivalent control. Figure 4 shows the error be- results from having dominant dynamics inserted between
tween the reference command and the plant state which the physical input to the plant, such as force, and the con-
exhibits the desired one second time constant transient troller output, usually an electrical signal. The design of
behavior, with the exception of initial minor distortions SMC which incorporates the actuator dynamics as a pre-
which are due to the convergence of the disturbance esti- filter for the VSC was proposed in [28]. This design uti-
mate shown in Figure 5. Despite the constant disturbance, lizes an expansion of the original state space by including
the steady state error is zero. While standard PID con- state derivatives, and formulates an SMC design such that
trollers can achieve the same zero steady state error in the the matching condition is indeed satisfied in the extended
presence of unknown constant disturbance, the tracking er- space. Another alternative approach is to utilize sliding
ror is regulated to zero even when &(t)is time varying [8]. mode to estimate the disturbance for compensation as dis-
cussed earlier. Since sliding mode is not introduced priniar-
C. Actuator Bandwidth Constraints ily to reject disturbances, the matching conditions axe of no
Despit,eits desirable properties, VSC is mostly restricted significance in this design. Provided that a suitable sliding
to control engineering problems where the control input of mode exists such that the disturbance can be estimated
the plant is, by the nature of the control actuator, nec- from the corresponding equivalent control, this approach
essarily discontinuous. Such problems include control of resolves the limitations imposed by actuator bandwidth
electric drives where pulse-width-modulation is not the ex- constraints on the design of sliding mode based controllers.
ception, but the rule of the game. Space vehicle attitude C.l An SMC Design with Pre-Filter
control is another example where reaction jets operated in
an on-off mode are commonly used. The third example, We shall use the example with a nominal integrator
which is closely related to the first one, is power converter plant, and actuator dynamics,
and inverter feedback control design. Fbr these classes of
applications, the chattering phenomenon still needs to be 2 = xa +d(t) , (14)
addressed. However, the arguments against using sliding a2&+2aka+& = U , (15)
mode in the feedback design are weakened. The issue in
this case is whether VSC should be utilized directly to to illustrate this design. The actuator bandwidth limita-
improve system performance while at the same time pro- tion isexpressed in the time constant a. Given adiscont,in-
duces the required PWM control signal, or a standard PID uous input u ( t ) ,the rate of change of the actuator output
X a ( t ) is limited by the finite magnitude of a. However, in
type controller should first be designed, and then the ac-
tual PWM control signal is to be generated by applying order for the disturbance d ( t ) to be rejected, x,(t) must
standard PWM techniques to approximate the continuous be an SMC. Also if 2, can be designed as a control input,
linear control signal. If VSC is to be used, by adopting an then the matching condition is clearly satisfied. But since
U is the actual input, the matching condition does not hold
observer based SMC, the high frequency components of the
discontinuous control can be bypassed, and consequently, for finite a. The design begins with an assumption that
adverse interactions with the unmodeled dynamics which d ( t ) has continuous first and second derivatives, and with
cause chattering can be woided. the introduction of new state variables,
In plants where control actuators have limited band-
width, e.g., hydraulic actuators, there are two possibili- x1=x, x2=x, x3=x, (16)
ties: First, the actuator bandwidth is outside the required the control U is designed as an VSC with respect to the
closed loop bandwidth. Thus the actuator dynamics be- sliding manifold,
come unmodeled dynamics, and our discussions in the pre-
vious sections are applicable. While it is possible to ignore s(x1,52,x3)= c121 + (2324- 2 3 =0 . (17)
the actuator dynamics in linear control design, doing so in
VSC requires extreme care. By ignoring actuator dynam- With the equivalent control ueq computed from
ics in a classical SMC design, chattering is likely to occur
since the switching frequency is limited by the actuator
dynamics even in the absence of other parasitic dynam-
ics. Strictly speaking, sliding mode cannot occur, since the the resulting sliding mode dynamics are found to be com-
control input to the plant is continuous. posed of two subsystems in series:
Second, the desired closed loop bandwidth is beyond the
actuator bandwidth. In this case, regardless of whether !I?+Q*+clx= 0 , (19)
g C I ) +
SMC or other control designs are to be used, the actuator Fa = ( ~2 - -
1 ~ 2 d2.
~ ~ (20)
dynamics are lumped together with the plant, and the con-
trol design model encompasses the actuator-plant in series. This design shows that. although the embedded pre-filter in
With the actuator dynamics no longer negligible, often the the plant model destroys the matching condition, an SMC
matching conditions for disturbance rejection and insensi- can still be designed to reject the unknown disturbance.

153
However, it is necessary to restrict the classof disturbances D. 1 A frequency shaped SMC design
to those which have bounded derivatives. Furt,hermore, For the nominal integrator plant with parasitic sensor
derivatives of the state, i,f are required in the feedback dynamics, we introduce a frequency shaping post-filter,
control implenientat ion.
xp + 2w$p + w;xp= 2, , (23)
C.2 A Disturbance Estimation Solution 9 p = PI% + P25p. (24)
For the nominal integrator plant with limited bandwidth The sliding manifold is defined as a linear operator, which
actuator dynamics given by Eqns.(14,15), we introduce the can be expressed as a linear transfer function,
same set of sensor dynamics as in Eqn.(S) and use a distur-
bance estimator similar to Eqn.( lo), only with IC, replacing 2 +1s +p2
2wps + w;
) IC,(S) = 0 . (25)
U,
Given an estimate of the lowr bound of the bandwidth
& = xu + sgn(xs - 2 ) . of parasitic dynamics, the post-filter parameter wp can be
(21)
chosen to impose a frequency dependent weighting function
With slidingmode occurson x-i = 0, the disturbance d ( t ) in a linear quadratic optimal design whose solution provides
is estimated with the equivalent control given by Eqn.(l2) an optimal sliding manifold. The optimal feedback gains
to O(r,). With the disturbance compensated, the remain- are implemented as pl ,p2 in Eqn.(25), and they ensure
ing task is to design a linear feedback control to achieve that the sliding mode dynamic response has adequate roll
the desired transient performance. The resulting feedback off in the specified fi-equencyband.
control law is given by
111. DISCRETETIMESLIDING MODE
While it is an accepted practice for control engineers to
consider the design of feedback systems in the continuous
With rs = 0.01, and a = 0.2, the feedback gains ICl = time domain - a practice which is based on the notion that,
31.25, and = 6.25 place the poles of third order sys- with sufficiently fast sampling rate, the discrete time imple-
tem dynamics, which consists of the actuator dynamics mentation of the feedback loops is merely a matter of con-
and the integrator plant, at (-2.5, -2.5, -5}. Again, we venience due to the increasingly affordable microprocessor
use the same third order butterworth low pass filter with hardware. The essential conceptual framework of the feed-
a 50 rad./sec. bandwidth as before to filter the equivalent back design remains to be in the continuous time domain.
control signal. Figure 6 shows the effects of the constant For VSS and SMC, the notion of sliding mode subsumes a
disturbance d ( t ) = 0.5 are neutralized since the error be- continuous time plant, and a continuous time feedback con-
tween the reference command and the plant state is reduced trol, albeit its discontinuous, or switching characteristics.
to zero in steady state. The disturbance estimate is shown However, Sliding Mode, with its conceptually continuous
in Figure 7 to reach its expected value in steady state. time characteristics, is more difficult to quantify when a
discrete time implementation is adopted. When control en-
D. Frequencv Shaping gineers approach sampled data control, the choice of sam-
pling rate is an immediate, and extremely critical design
An approach which has been advocated for attenuating decision. Unfortunately, in continuous time Sliding Mode,
the effectsof unmodeled parasitic dynamics in sliding mode desired closed loop bandwidth does not provide any useful
involves the introduction of frequency shaping in the design guidelines for the selectionof sampling rate. In the previous
of the sliding manifold [5]. Instead of treating the sliding section, w indicate that asymptotic observers or sliding
manifold as the intersection of hyperplanes defined in the mode observers can be constructed to eliminate chatter-
state space of the plant, sliding manifolds which aredefined ing. Observers are most likely constructed in discrete time
as linear operators are introduced to suppress frequency for any real life control implementations. However, in or-
components of the sliding mode response in a designated der for these observer-based design to work, the sampling
frequency band. For unmodeled high frequency dynamics, rate has to be relatively high since the notion of continuous
this approach implants a low pass filter either as a pre-filter, time sliding mode is still applied.
similar to introducing artificial actuator dynamics, or as a For Sliding Mode, the continuous time definition and its
post-filter, functioning like sensor dynamics. The premise associated design approaches for sampled data control im-
of this Frequency Shaped Sliding Mode design, which was plementation have been re-defined to cope with the finite
motivated by flexible robotic manipulator control applica- time update limitations of sampled data controllers. Dis-
tions [33],is that the effects of parasitic dynamics remain crete time Sliding Mode (DSM) was introduced [34] for
to be critical on the sliding manifold. However, robustness discrete time plants. The most striking contrast between
to chattering was only implicitly addressed in this design. SM and DSM is that DSM may occur in discrete time sys-
By combining frequency shaping sliding mode and the SMC tems with continuous right hand sides, thus discontinuous
designs introduced earlier, the effects of parasitic dynamics control and Sliding Mode, are finally separable. In discrete
on switching induced oscillations, as well as their interac- time, the notion of VSS is no longer a necessity in dealing
tions with sliding mode dynamics can be dealt with. with motion on a Sliding Manifold.

154
 IV. SAMPLED DATASLIDING MODECONTROLDESIGN where F, G and D result from integrating the solution of
We shall limit our discussions to plant dynamics which Eqn.(26) over the time interval t E [kT,( k + 1)T] wit,h
can be adequately modeled by finite dimensional ordinary
u(t)= U ( k T ) , d ( t ) = d(lcT), (31)
differential equations, and assume that an apriori band-
width of the closed loop system has been defined. The
feedback controller is assumed to be implemented in dis- E = exp(tlT), G = r B , (32)
crete time form. The desired closed loop behavior includes
insensitivity to significant parameter uncertainties and re-
F = F + I ' A A , r = s,'exp(Ar)d.r , (33)
jection of exogenous disturbances. Without such a demand
G=G+I'AB, D = I ' E . (34)
on the closed loop performance, it is not worthwhile to
evoke DSM in the design. Using conventional design rule This discrete time model is an O(T2)approximation of the
of thumb for sampled data control systems, it is reason- exact model which is described by the same F and G ma-
able to assume that for the discretization of the contin- trices, but because the exogenous disturbance is a continu-
uous time plant, we include only the dominant modes of ous time function, the sample and hold process yields a D
the plant whose corresponding corner frequencies are well matrix which renders the matching condition for the con-
within the sampling frequency. This is always achievable in tinuous time plant to be only a necessary, but not sufficient
practice by anti-aliasing filters which attenuate the plant condition for the exact discrete time model [37]. However,
outputs at frequencies beyond the sampling frequency be- by adopting the above O(T2)approximated model, it fol-
fore they are sampled. Actuator dynamics are assumed to lows from (34) that, if the continuous time matching condi-
be of higher frequencies than the sampling frequency. Oth- tion (28) is satisfied, the following matching condition for
erwise, actuator dynamics will have to be handled as part this model holds :
of the donfinant plant dynamics. Thus, all the undesirable
parasitic dynamics manifest only in the between sampling rank[rB i FAA! I'AB irE]= rank(rB). (35)
plant behavior, which is essentially the open loop behavior
of the plant since sampled data feedback control is applied. LFi-om an engineering design perspective, the O(T2) models
Clearly, this removes any remote possibilities of chattering are adequate since the between sampling behavior of the
due to the interactions of sliding mode control with the continuous time plant is also O(T2) close to the values at
parasitic dynamics. the sampling instants. Let the sliding manifold be defined
We begin to summarize sampled data sliding mode con- b
trol designs with the well understood sample and hold pro- sk = c x k = o , k =0,1, .. . , s(m)= s k (36)
cess. This may seem to be elementary at first glance, it is
however worthwhile since the matching conditions for the Two different definitions of discrete time sliding mode have
continuous time plant are only satisfied in an approxima- been proposed for discrete time systems. While these def-
tion sense in the discretized models. We shall restrict our initions share the common base of using the concept of
discussions to linear time-invariant plants with uncertain- equivalent control, the one proposed in [34] uses a defini-
ties and exogenous disturbances, tion of discrete time equivalent control ugq = U(&') which
is the solution of
k = A ~ z + B u + E d , z E l F t n , U E IR", d E Et.' (26)
~ & +=l 0 , k = 0 , 1 , . .. (37)
where A, B , E are constant matrices, and d ( t ) is the exoge-
nous disturbance. For the plant (26), we assume that the On the other hand, u E ~is defined in [35] as the solution of
system matrices are decomposed into nominal and uncer-
tain components, A: = s k + l - s k = O , k = O , I , . . . (38)
Note that (37) implies (38), however, the converse is not
A=A+AA, B=B+AB, (27) true. Herein, the first,definition given by Eqn.(37) shall be
where A , B denote the nominal components. Let the used.
admissible parametric uncertainties satisfy the following A . DSM Control Design for Nom,inal Plants
model matching condition [32]
Given the nominal plant with no external disturbance,
the DSM design becomes intuitively clear. In DSM, by
rank([Bi A A ; AB ; E ] )= rankB. (28) definition,
The discrete time model is obtained by applying a sample Sk+] = cZ&+]
= C ( F 5 k -f- G W ) = 0 , (39)
and hold process to the continuous time plant with sam-
pling period T, which to O ( T 2 ) ,is given by: and provided that, CG is invertible, the DSM control which
is also the equivalent control, is gixn by the linear contin-
g + ]= F X k -tG'Uk + D d k , 20 = Z(to), (29) uous feedback control,
X ( w ) =Xk , U ( k T ) = U k , d ( k T )= d k , (30) Uk,, = - [ c G - ' c F X k . (40)

155
The only other complication is that since 1 1 ~ 1 1 = O(T), B. DSM Control with Delayed Disturbance Compensation
the required magnitude Of 'his may be large. If The earlier DSM control design for nominal plants can be
the bounds ii On are taken into account1 the following modified to compensate for unknown disturbances in the
feedback has been shown [I9] to force the system system [36], [37]. LFrom the discrete model in Eqn.(29),
into DSM: the one step delayed unknown disturbance

An alternative control law which replaces the control mag- can be computed, given the measureme_nts-xk,xk-1 and
nitude constraint with an adjustable reaching phase behav- Uk-1, and the nominal system matrices F, G. Let
ior has also been reported [39]:

where M is a diagonal positive definite matrix whose en-

tries are in the sampling time order and B is a user-defined U;*$, if Il4J <
boundary layer expression which has the following form: -ij "kea if 11~*ke,I12 .ii . (47)
l14cqlil
= {XklSk = C X k O(T)}. (43)
The effectiveness of this controller is demonstrated by ex-
A.l DSM Control of the integrator plant amining the behavior of s k when the control signal is not
For the nominal integrator plant with parasitic sensor saturated,
dynamics, Eqns.(9-9), we design an DSM controller based s k + l = cD(dk - dk-1). (48)
on Eqn.(41). Let the sensor time constant T. = 0.02, and If the disturbance has bounded first derivatives, Le., 181 5
the control magnitude ii = 1. The desired closed loop d < CO, dk -dk-l is of O(T), and from the definition given
bandwidth is given to be one Hz. A good choice of the in Eqn.(34), = O(T),hence l s k l = O(T2), implying
sampling frequency muld be 10 Hz (T=O.l) since the sen- that the motion of the system remains within an O ( T 2 )
sor dynamics are of 50 Hz, and therefore can be neglected neighborhood of the sliding manifold. This controller has
initially in the design. The DSM control takes the form of also been shown [19] to force the system into DSM if the
control signal is initially saturated.
On the sliding manifold, the system dynamics are, to
O ( T 2 ) invariant
, with respect to the unknown disturbance.
where = z,(kT) is the sampled value of the sensor out- Since similar matching conditions exist for the O ( T 2 )dis-
put z,(t). Note that due to the control bounds, a linear crete time models we have adopted, it follows from contin-
feedback control law is applied inside a boundary layer of uous time sliding mode [28] that by using a change of state
thickness 2T about the sliding manifold x i = 0. Without variables, the discrete model can be transformed into
sensor dynamics, the behavior inside the boundary layer
is that of a deadbeat controller. The sensor dynamics im- (49)
pose a third order discrete time system inside this bound-
ary layer, and its eigenvalues are inside the unit circle at
{ -0.002 ,O. 1k30.436). For reference, the discrete model of with the sliding manifold given by
the open loop nominal plant and the sensor dynamics has a
pair of double real pole almost at the origin (4.54 x Sk = C I X k +
c24 = 0 , (51)
which result fromsampling at a frequency much lower than
the sensor's corner frequency, and a pole at unity which is and C2G is nonsingular. By eliminating x:, the reduced
due to the integrator plant. The third order system re- order sliding mode dynamics are O(T2) appraimated by
sponse can be seen in Figure 8 where the sample values of
the error between the constant unity reference command x;+1= (F11- F12c,-1cl)x:. (52)
and the sensor output is plotted. Note that only the be-
havior inside the boundary layer is shown, and it agrees B.l Discrete time disturbance compensation for the inte-
grator plant
well with the predicted third order behavior. The steady
state error magnitude of 0.05 is due to the constant distur- We continue with the DSM control design using the same
bance d(t) = 0.5 as applied to this plant as before, and the sampling frequency and system parameter values. The con-
effective loop gain being T-l = 10. Figure 9 displays the troller which takes into account the one step delayed dis-
continuous time error of the plant state and the discrete turbance estimates is given by
time error of the sensor output where the time lag due to
the sensor dynamics can be seen during the transient pe-
riod.

156
Note the PID controller structure of this controller when of the second order system governing the motion outside
the system is inside the boundary layer. Figure 10 shows the manifold is plotted in Figure 13 for -1 5 y 5 0.34.
the sampled error between the reference command and the For y = -1, there is a pair of double poles at unity, and
sensor output. The practically zero steady state error is for y = 1/3, one of the poles becomes -1. The case for y =
much better than our O ( T 2 )estimate due to the PID con- -0.5, corresponding to a pole of complex pairs 0.5 30.5, *
troller st,ructure. The one step delayed disturbance esti- is simulated wit#hthe same reference and disturbance as in
mate is given in Figure 11, showing convergence to the ex- the previous studies. Figure 14 shows the conxrgence of
pected value. Figure 12 displays the continuous time error the sampled error between the reference command and the
between the plant state and the reference, and its discrete sensor output to zero. Figure 15 displays the estimates of
time measurements. the exogenous disturbance and the residue control signal
due to the gain uncertainty. The continuous time error of
C. DSM Control with Pammeter Uncertainties and Dis- the plant state and the discrete time error of the sensor
turbances output are shown in Figure 16 for comparison.
With the presence of system parameter uncertainties, the
above approach which uses one step delayed disturbance es- V. CONCLUSIONS
timates can still be applied. However the one step delayed We have systematically examined SMC designs which are
signal contains both delayed state and control d u e s , firmly anchored in sliding mode for the continuous time do-
main. Most of these designs are focused on guaranteeing
fk-1 e AFxk-I-tAGuk-1SDdk-1 = X k - F x k - l - G u k - 1 , the robustness of sliding mode in the presence of practical
(54) engineering constraints and realities, such as finite switch-
where AF = r A A ,AG = r A B . The DSM control is of ing frequency, limited bandwidth actuators, and parasitic
the same form as Eqn.(47), with G F 1replaced by f k - 1 . dynamics. Introducing DSM, and restructuring the SMC
The behavior of s k is prescribed by design in a sampled data system franiemrk are appropri-
ate, and positive steps in sliding mode control research. It
+
S k + l = c ( f k - f k - 1 ) == cD(dk - d k - 1 )
directly addresses the pivotal microprocessor implements
+
+CAF(xk - X k - I ) CAG(Uk - U k - 1 ) . (55) tion issues; it moves the research in a direction which is
more sensitive to the concerns of practicing control engi-
Since If1 is bounded, x k - x k - 1 is of O ( T ) , and since neers who are faced with the dilemma of whether to ignore
IlAFll = O ( T ) ,we have this whole branch of advanced control methods for fear of
the reported implementation difficulties, or to embrace it
Sk+l =CAG(Uk - U k - 1 ) + o ( T 2 ) . (56) with caution in order to achieve system performance oth-
Due to the coupling between s k and uk, it has been erwise unattainable. However, as compared with the ideal
continuous time sliding mode, we should also be realistic
shown [37], [38] that the behavior outside the sliding man-
about the limitations of DSM control designs in reject-
ifold is governed by the following second order difference
ing disturbances, and in its ability to withstand parameter
equation,
variations. The real test for the sliding mode research com-
S k + l = -CAG(CG)-’[2sk - S k - I ] 4-o ( T 2 ), munity in the near future will be the willingnessof control
(57)
engineers to experiment with these SMC design approaches
which has poles inside the unit circle for sufficiently small in their professional practice.
11A B ] [ . The permissible control matrix uncertainties are
dictated by the above stability condition which determines REFERENCES
the convergence on the sliding manifold. Note that pro- V. I. Utkin, “ Variable structure systems with sliding modes’’,
vided the parameter uncertainties are in the system m& IEEE P a n s . Automat. Contr., Vol.AC-22, No.2, pp. 212-222,
1977.
trix, they do not impact the conwgence, nor they affect Friedland, B. Advanced Control System Design, pp. 148-155,
the motion on the manifold. Prentice Hall, Englemcd Cliffs, N.J., 1996.
Asada, H. and J-J. E. Slotine, Robot Analysis and Contro1,
C. 1 Compensation for gain uncertainties in integrat,or pp. 140-157, John Wiley and Sons, 1986.
Bondarev, A. G., S. A. Bondarev, N. E. Kostyleva, and V. I.
plant Utkin, “Sliding Modes in Systems with Asymptotic State Ob-
servers,” A utomataonand Remote Contml, pp. 679-684, 1985.
We shall introduce gain uncertainties in the integrator Young K. D. and U. Ozgiiner, “Requency Shaping Compensator
plant to examine their effects on the convergence of the Design for Sliding Mode,” Special Issue on Sliding Mode Control,
sliding manifold. The actual plant is given by International Journal of Control pp. 100.5-1019, 1993.
Young, K. D. and S. Drakunov, “Sliding mode control with chat-
2 = (1 + y)u + d ( t ) (58)
tering reduction,” in Proceedings of the 1992 American Contml
Conference, Chicago, Illinob, p p . 1291-1292, June 1992.
Su, W. C., S. V. Drakunov, U.Ozgiiner and K . D. Young, “Sliding
where y represents the gain uncertainty in the int,egrat,or. inode with chattering reduction in sampled data systems , Pro-
The DSM controller in Eqn.(53) can be used again because ceedings of the 32nd IEEE Conference on Decision and Control
San Antonio, Texas, pp. 2452-2457, December 1993.
the right hand side of the one step delayed signal is t,hesame Young, K. D., and S. V . Drakunov,“Discontinuous frequency
regardless of the parametric uncertainties. The root locus shaping compensat.ion for uncertain dynamic systems,” Proceed-

157
 ings 12th IFAC World Congress, Sydney, Australia, pp. 39-42. [34] Drakunov S. V. and V. I. Utkin, “Sliding mode in dynamic sys-
1993. tems;’ International Journal of Control Vo1.55, pp. 1029-1037,
[9] Young, K. D. (editor), Variable Structure Control for Robotics 1990.
and Aerospace Applications, Elsevier Science Publishers, 1993. [35] Furuta, K., “Slidingmode control of a discrete system,” Systems
[lo] Zinober, A. S. (editor), Variable Structure and Lyapunov Con- and Control Letters,,.Vol.l4, pp. 145-152, 1990.
trol Springer Verlag, London, 1993. I361 R. G. Morgan and U. Ozgiiner, “A Decentralized Variable Struc-
[11] F. Garofalo and L. Glielmo (editors), Robust Control via Vari- ture Control Algorithm for Robotic Manipulators, IEEE Journal
able Structure and Lyapunov Techniqueq Lecture Notes in Con- of Robotics and Automation, 1, 1, pp. 57-65, 1985.
trol and Information Sciences Series, Vol. 217, pp. 87-106, [37] W-C. Su, S. V. Drakunov and U. Ozgiiner, Sliding Mode Control
Springer-Verlag, Berlin, Heidelberg, New York, 1996. in Discrete Time Linear Systems, Preprints of IFAC 12th World
[12] Utkin, V. I . , “Variable Structure Systems: Present and Future,” Congress, Sydney, Australia, 1993.
Avtomatika i Telemechanilca, No. 9, pp. 5-25, 1983 (in Russian), [38] Su, W. C., S. V. Drakunov, U. Ozguner, “Implementation
English ’Ikanslation, pp. 110.51119. of Variable Structure Control for Sampled-Data Systems,” Ro-
1131 DeCarlo, R. A, S. H. Zak, and G. P. Matthews, “Variablestruc- bust Control via Variable Structure and Lyapunov Techniqueq
ture control of nonlinear multivariable systems: A tutorial,” Proc. F.Garofalo and L.Glielmo (editors), Lecture Notes in Control
O f IEEE, Vol. 76, NO. 3, pp. 212-232, 1988. and Information Sciences Series, Vol. 217, pp. 87-106, Springer-
[14] Utkin, V. I., “Variable Structure Systems and Sliding Mode - Verlag, Berlin, Heidelberg, New York, 1996.
State of the Art Assessment,”, Young, K. D. (editor), Variable [39] I. H a s b r a , U. Ozgiiner andV. I. Utkin 1997, “Variable Structure
Structure Control for Robotics and Aerospace Applioations, pp. 9- Control for Uncertain Sampled Data Systems,” Proceedings of the
32, Elsevier Science Publishers, 1993. 36st IEEE Conference on Decision and Control, San Diego, CA,
1151 Hung, 3. Y., W . B. Gao, and J . C. Hung, “Variable structure U.S.A., 3226323.1.
control: A survey,” IEEE Trans. Ind. Electron., Vol. 40, No. 1, [40] I. Haskara, U. Ozgiiner and V. I. Utkin, “On Sliding Mode Ob-
pp. 2-22, 1993. servers via Equivalent Control Approach,” International Journal
[16] Flugge-Lutz, I., Discontinuous Automatic Control Princeton of Control, vol 71, no 6, pp. 1051-1067, 1998.
Uniwrsity Press, 1953. (411 I. Haskara and U. Ozgiiner “Equivalent Value Filters in Dis-
[17] Tikhonov, A. N., “Systemsof differential equations with a small turbance Estimation and State Observation,” Pmceedzngs of
parameter multiplying derivations,” Mathematicheslcii Sbornilc, the 5th International Workshop on Variable Structure Systems
Vol. 73, No. 31, pp. 575-586, 1952 (in Russian). (VSS’gS), pp. 77-83, Longboat Key, FL, 1998.
[18] Kokotovic, P. V., H. K. Khalil, and J. O ’ b i l e y , Singular pertur-
bation methods in control :analysis and design, Academic Press,
1986.
1191 Utkin, V. I. , “Sliding Mode Control in Discrete-Time and
Difference Systems,” Variable Structure a n d Lyapunov Control,
A.S.Zinober (editor), Springer Verlag, London, pp.83-103, 1993.
I201 Utkin, V. I., Sliding Modes and their applimtaons in Variable
Structure Systems, Moscow:MIR, 1978 (translated from Russian).
1211 Young, K-K. D., P. V. Kokotovic, and V. I. Utkin, “A Singular
Perturbation Analysis of High Gain Feedback Systems,”, IEEE
‘Pmns. Auto. Contr., Vol. AC-22, No. 6, pp. 931-938, 1977.
[22] Slotine J-J. and S. S. Sastry, “Tracking Control of Nonlinear
Systems using Sliding Surfaces with Application to Robot Ma-
nipulator,” Znt. J. Control, Vol. 38, No. 2, pp. 465-492,1983.
(231 Burton, J. A., A. S. I. Zinober, “Continuous approximation of
variable structure control,” Int. J. System Sci., Vol. 17, No. 6,
pp. 875-885, 1986.
(241 Young, K-K. D. and P. V. Kokotovic, “Analysis of Feedback
Loop Interaction with Parasitic Actuators and Sensors,” Auto-
matiat, Vol 18 , September 1982, pp. 577-582.
(251 Kwatny, H. G. , K. D. Young, “The Variable Structure Ser-
vomechanism,”, Systems and Control Letterg Vol. 1, No. 3,
pp. 184-191, 1981.
(261 Young, K. D., and V. I. Utkin, “Sliding Mode in Systems with
Parallel Unmodeled High Frequency Oscillations,” Proceedings of
the Third IFAC Symposium on Nonlinear Control Systems De-
sign, Tahoe City, California, J u n e 25-28, 1995.
1271 Young, K-K. D. and H. G . Kwatny, ‘Variable Structure Ser-
vomechanism Design and its Application to Overspeed Protection
Control,” Automatzm, vo1.18,No. 4, pp. 385400,1982.
[28] Utkin, V. I., Sliding Modes in Control Optamimtion, Springer-
Verlag 1992.
(291 Slotine, J-J. E., J. K. Hedricks, and E. A. Misawa, “On Sliding
Observers for Nonlinear Systems,” ASME J. Dynamic Systems,
Measurement and Control, Vol. 109, pp. 245-252,1987.
[30] Xu, J-X., H. Hashimoto and F. Harashima, “On t h e design of a
VSS O b s e m r for Nonlinear Systems,” Tmns. of the Society of Ira-
strument and Control Engineers (SICE), Vol. 25, No. 2, pp. 211-
217, 1989.
[31] Korondi, P., H. Hashimoto, K. D. Young, “Discrete-time Sliding
Mode Based Feedbadc Compensation for Motion Control,” Pro-
ceedings of Povter Electronics and Motion Control (PEMC’96),
Budapest, Hungary, Sept. 2-4, 1996, V01.2, pp. 21244-21248,1996.
I321 Drazenovic, B., “The invariance conditions in variable struct,ure
systems,” Automati?, vo1.5, No. 3, pp. 287-295, 1969.
[33] Young K. D., U. Ozgiiner, and 3-X. Xu, “Variable Structure
Control of Flexible Manipulators,” Young, K. D. (editor), Vari-
able Structure Control for Robotics and Aerospace Applications,
pp. 247-277, Elsevier Science Publishers, 1993.

158
 Figure 1: Block diagram of Observer based Sliding Mode
Control.

0 01 , Observer based snc aaunaary layer

I @
o 002.dtstumance o s
1
0008 ......... ..: . . . . . . . . . . . . . . . . j ......... .:... .......< .............i ..........j ............ ~
Figure 4: SM Disturbance estimator Control: error be-
0.006- . . .......................... tween reference command and plant state.
0.004 ........... : ........ .(. ....... j .......................... ;............. ;.............:............ _
0.002 ........... .;. ..... ....; . . . . ....j. . . . ......; ............;............. i.. ......... .i. .........

g 0
_............1............ .............i ...........................
~ i.............;.........................._
-0.002 _..........: ... .". I .... .:.. ........ .: ...........:.............:.. ..........:. ..........
...........j ....... ...:.............j ............ i ............. ............. i . . ..........
~

s
~ 0.4
.o.ow .................... ..I. . . . ...;......... ...;.......................... ;.............f ............ ~

;. .: .: i .....;............i i
1 .I
-0.006 _..
.......................................... :. ........................ ..; .......................... 80.2 . . . .. . . . . ..

.O.OOB .......... .;.............;.. .........: .........i ............j ............. ;. ....... ..i,,

.........
1 0 .......... ............. ....... ............. .......................... i...........;. .....

-0.2 ...........(..............:.........i.. ...........;. .............;._.

.......... (............:.............

Figure 2: Observer based SMC: error betwen reference

command and o b m r state.
Figure 5: SM Disturbance estimator Control: disturbance
estimate.
Limited bandwidth actuator w i t h Sti disturbance estimator.
f i l t e r t i m e constant 0 02. disturbance 0.5

. . . .~~ i.............;... ........ :............. ; .................... .-

.................... .;...............................................
1.w- ...........:............5 ............;.. .........;............. ;.............;.............;. .......... ~

31.02 ~.
: . . ..........................
........ ...... ......... .... ........ ............. ........ .........
-................ Fr;.:.............................................
B 6-i ;
1 :..................... 1
m
$0.9@ - ............i.. ................... i .........................
F
!............. :. ........ I ' \
g0.96 - ...................I. ....... : ................... ........................ ...... :.

0.w -...... . ..i......... :. ....... .: ......... i ........ ;..................... ...I.. .... .~

092 - . . . . . . . . . . .

Figure 6: Limited bandwidth actuator with SM

Figure 3: Observer based SMC: plant state (upper curve) Disturbance estimator: error between reference command
and its estimate (lower curve). and plant, state.

159
 06/
Llmlted bandwldth actuator w l t h SM disturbance estlmator
, rii~ericme~onstant~02,dis!urbsnce~5 , , o,,

U U5

006-
~ ,
Discrete time SMC with disturbance Bstlmatlon
sarnpling,period 0,l, distrt?ance 0 5 , , ,
OM-

0 OB

O10 5 10 15 &2Zw 25 30 35 40

Figure 7: Limited bandwidth actuator with SM Distur- Figure 10: DSM Control with disturbance compensation:
bance est,imator: disturbance estimate. error between reference command and sensor output.
Dlscrete tlme Snc with dlsiurbance estlmailon:
ampling period 0 1, disturbance 0 5
08

#03
~~

02

01

OO 5 I

Figure 8: DSM Control for nominal plant: error between Figure 11: DSM Control with disturbance compensation:
reference command and sensor output. one step delayed disturbance estimate.
Dlscrete tlme 5MC sampling perlod 0.1. disturbonce 0.5

Figure 9: DSM Control for nominal plant: continuous Figure 12: DSM Control with disturbance compensation:
time and discrete time error responses. continuous time and discrete time error responses.

160
 Discrete t i m e SMC w i t h disturbance estimation
Roat locus f o r gain uncertainty between - I en4 0.34 sampling period 0.1, gain uncertainty - 0 5 . disturbance 0.5
1, , 1 12

1- -
m
U

E
30s-

-0a

1
1 05 0 05 1
Real part

Figure 15: DSM Control with disturbance compensation:

one step delayed parameter and disturbance estimate.
Dlscrete t l m e SMC W l t h d1StUrtiOnCe estimation
samplig period 0.1, gain uncertainly -0 5 , disturbance 0.5
1- s r p l t n g P ~ l o d ~,,gern
o uncfrtacnty ;OS. dlstTUi3mce15

02,
0.8 ....j . . . . . . . . . . . . . . . . . . . . . . . . ; ...........;............. ;. . . . . . . ;.. . . . .

0.6
.....-..:.. .................. : ..........;............;........................ i .......... -
.... ..........i .......... j . ...........i... ..........i............. i... .........;.. ........

0.4- .......T..; i ..............i............. j.............i ............_

.............i......... . i ..............

-
0 2 - ....................................... :................................................ .~

I I
-0.20
5 10 15 20 25 30 35 40

Figure 14: DSM Control with control parameter

variations: error between reference command and sensor Figure 16: DSM control with control parameter
output. variations: continuous time and discrete time error
responses.

A . Other DSM Contrvl Designs

On going research in DSM control designs is expected
to fill in the missing links corresponding to many of the
exist,ing continuous time SMC designs. In particular, the
design of discrete time sliding mode observers and the re-
lated SM disturbance estimators can serve as the robust
discrete time counterparts of the existing continuous time
SMC techniques without requiring a discountinuous control
action.

A.l A Discrete Time Sliding Mode Observer

We first present a recently developed discrete time sliding
mode observer [40] which can be considered as a direct
discrete time counterpart of the sliding mode realization of
a reduced order asymptotic observer based on the discrete
time equivalent control definition. Consideration is given

161
 Pi represents the ith entry of DI, and q is the one step
time forward shift operator, the following control law has
been shown [39] to force the system into DSM by providing
complete disturbance rejection:

Uk = -(cG)-'[CFxk - Rk]. (69)

where I& E IR" is inserted to the control law to filter out
the disturbance which is given by

Note that, even if the disturbance dynamics are not

known, the time history of Pk can be deducted from the
equation

p= [ :] for any T > 0 since all the terms at the right hand side
are known. The previous values of Rk provides us with an
where R E IR(n-p)xn is arbitrary provided that P is in- intuition about the structure of the disturbance and this
vert ible intuition can be used to fit a disturbance model. The accu-
racy of the disturbance model increases while the number
of past values used for model fitting increases at the ex-
pense of making the controller more complicated. Further
details of this approach can be found in [39].
with Fi1 E R p x pF12
, E RPX(n-p)l F21 E IR(n-p)xpand
From a theoretical point of view, the disturbance com-
F22 E R("-p)x(n-p); z k is the state of the following error
pensation by a coniplementary disturbance cancelation
driven system
Z k + l = Fzk + e. term in the control input can eliminate only the distur-
(64) bance terms which enter the system through the same chan-
with = (F22 +LF12) , G = (F21+ LFll)ey,k-l - Le,,k; nel with the control which is often referred to as the match-
L is the observer gain matrix which places the eigenvalues ing condition in the literature. If the disturbance does not
+
of (F22 LF12)at the origin. satisfy the matching condition the motion on the Mani-
This obserw decouples the overall motion into two inde- fold will also be affected by the disturbance even though
pendent motions which are controlled by the same parame- an ideal sliding motion is formed. The usage of SM dis-
ter L . This decoupling property is expected to provide the turbance estimation techniques for a larger class of distur-
designer with additional flexibility to robustify the nominal bances without requiring the matching condition is cur-
design. rently being investigated.
A.2 Robust Sampled Data Sliding Mode Control Design
In this section, we present an alternative sampled data
sliding mode control design which exploits the one step
delay disturbance estimation technique. The system con-
sidered is given by

xk+l= FXk f GUk 4- Ddk . (65)

where the system matrices F and G are assumed to be
known. If the discrete time equivalent control law was ap-
plied, one would get
A
S k + i = CDdk = a. (66)
and a linear transformed version of the disturbance would
appear in s. If the disturbance dynamics are known, i.e;

E(q-1)PL = 0 (67)
where
E(4-I) = eo +elq-l+ ... + erq--'. (68)

162