Proceedings of the 34th

Conference on Decision 81Control

FM04 1:38 -
New Orleans, LA December 1995

SLIDING MODE OBSERVERS. TUTORIAL.

Sergey Drakwnov Vadim Utkin
Department of Electrical Department of Electrical
Engineering Engineering
Tulane University The Ohio State University
New Orleans, LA 70118 Columbus, OH 43210

ABSTRACT given by Krener and Respondek [6]. These

Tutorial discusses the problem of designing conditions are considerably more restrictive
observers for state estimation using sliding than the observability condition of the sys-
modes. The theory and design principles are tem 1, 2. On the other hand the problem of
presented for linear and nonlinear systems. integration of a corresponding partial differ-
For linear systems the observers are devel- ential equation for obtaining this transforma-
oped using a block-observable form which is tion makes sometimes this method difficult to
similar to a lower triangular matrix form. Com- implement in practical situations.
paring with known approaches such observers Observers based on sliding mode approach
have better robustness properties. In case of [l,21 first were developed for linear case [3,
nonlinear system an equivalent control con- 5 , 4, 9, 81.
cept makes possible to develop finitetime ob- x = Ax Bu, + (3)
servers for a wide class of systems.
y = ex. (4)
Summary By structure they are very similar to the stan-
dard full order observers with replacement of
The problem of state estimation for the non- the linear innovative term by a discontinuous
linear system function:

x = f(x) + B(s)u, (1) 1 = A? + Bu + Lsign(y - C?), (5)

where x E R", by output vector measure- where sign is understood componentwise for
ment s vector argument z = cok(z1,. . . ,zn) :
y = h(x) E R" (2) sgn(z) = c d ( s g n ( z l ) ,. . . ,s g n ( z n ) ) . (6)
has been considered by many authors (see,
for example, survey [7]). For exactly known Under a suitable choice of a gain matrix L
smooth nonlinearities f and h and if B = sliding occurs in the system 5 on the manifold
0, the method of global linearization 191 by y - C4 = 0, and it becomes equivalent to the
using nonlinear state transformation permits reduced order observer.
to obtain a system in the new variables z E The observer design can be based on se-
Iz" which is linear with respect to z and has quential application of t he st ate transforma-
in the right hand side nonlinear function only tions reducing the system to the block form
of the measurements y. Then the standard [3, 51-
linear pole-placement technique can be used Let us illustrate this method for the sys-
for designing observer. tem 3, 4. By using linear transformation of
For successful application of this method state variables it can be written in the form:
the nonlinear transformation must be deter-
mined. The necessary and sufficient condi-
tions of existence for this transformation were

0-7803-2685-7/95 $4.00 O 1995 1EEE 3376

The corresponding sliding mode observer for
the first subsystem is
$2 = A4i~i + A4222 + B4u. (17)
The observer for 16 is
S = Ail6 + A12& + B i u + Lisgn(y - 6) (9) Dl = A3161 + A32& + +
~ 3 u ~Bsign(y1- $I),
The system for the error tj = y - 6 obtained (18)
by subtracting 7 from 9, has the form which requires the measurements of y1.
Since according to 13, 15
5 = Aiiy + Ai2Z1- Lisgn(y) (10)
YI - $1 = (sign(y))ep,
By appropriate choice of the observer gain
matrix L1 sliding occurs in 10 along the man- this equivalent value can be used in the right
ifold y = 0. hand side of the observer 18:
For the second subsystem 8 we can use the
observer equation of the form
= ~31c1+~3z~i.2+~3~+~3sig72((sign(y))e~ 1.
(19)
$1 = A219 + A2221 + Bzu - LzLlsign(y - c). As it is shown in [I] for real-life system
(sign(@),qby its physical meaning is an av-
(11) erage value of the chattering with high fre-
which gives the following equation for the er-
ror quency variable sign(y) . (sign(y))ep can be
$1 = A 2 2 i i 1 - L2Li~ign(tj). (12) obtained by using low pass filter whose time
constant is “sufficiently small to pass the slow
According to the equivalent control method, component of the motion (which is an equiva-
the system in sliding mode behaves as if Llsign(y) lent control) but is large enough to eliminate
is replaced by its equivalent value (Llsign(y))eq, the high frequency component” caused by the
which can be calculated from the subsystem presence of nonidealities. Repeating sequen-
10 assuming p = 0, ?j = 0. Hence tially the described procedure we obtain the
observer which converges in finite time.
(&sign( y))eq = A 1 2 ~ 1 - (13) The idea described above also can be im-
Substituting 13 into 14 we obtain plemented for a nonlinear case [4].As in lin-
ear case the main idea of the design is to use
i 1 = (A22 - L2A12)Zi. (14) the equivalent control method to obtain ad-
ditional information from the system.
It can be shown that if the system 3, 4 (the Let us consider the system 1, 2 with m =
pair (A, C)) is observable then the pair (A22, A21) l , B = 0 and introduce the following nota-
is also observable and, therefore, by appro- tions H ( z ) = col(hl(z), . . .,h,(z)) where h l ( z ) =
priate choice of Lz the desired rate of conver- h ( z )and hi+l (z) is ith Lie (directional) deriva-
gence 51 -+ 0 can be obtained. tive of the function h ( z )along the trajectories
Under special conditions sliding mode ob- of the system 1
server similar to 5 can be used for estimating
state of the uncertain system (see [9, 81). hi@) = L>-lh(z).
The other way of designing observer is to
continue transforming the system like it was or
done on the first step instead of using 11. In-
troducing a variable
Let H ( z ) = col(hl(z), .. .,h,(z)), is nonde-
y1 =L312z1 (15) generating in X i.e.:
the second subsystem 8 can be decomposed
to the form

$1 = A31Y1 + A32z2 + B3u (16) for some S > 0 and every x E X .

3377
 To estimate system 1 state variables by Generalization when control enters the right
measurements 2 one can use an observer ([4]) hand side of the equation was also obtained
of the form: but in contrast with linear case in this situa-
tion the observation problem depends on the
control.
In conclusion let us note that a wide vari-
(23) ety of parameter estimation problems can be
where V ( t )= c o l ( q ( t ) ,. . .,v n ( t ) ) solved by means of sliding mode observers.
Assuming that these parameters are some of
v d t ) = Y(t), (24) the unknown state variables one can use an
v;+1(t) = (m;(+ign(v;(t) - k ( q t ) ) ) ) e q , observer for nonlinear case described above.
(25) The other way is to use Lyapunov approach
i = 1, ...,n - 1 and matrix M is n x n di- to derive a sliding mode observer specifically
agonal matrix with positive elements m ; ( ( ~ ) , for parameter estimation or simultaneous pa-
i = l,...,n rameter and state estimation in linear system

M(P) = d i ~ g ( m l ( ? ).,. .,m,(P)). (26)

References
By suitable choice of M , observer 23 con-
verges in any prescribed finite time interval. Utkin V.I., Sliding Modes and Their Application In
Variable Structure Systems. Moscow: MIR, 1978.
From 22 it follows that the map H : R” w
R” is an injection therefore it is sufficient to DeCarlo R. A., S. H. Zak, and G.P. Matthews, Vari-
show that the modified observation error e ( t > = able Structure Control for Nonlinear Multivariable
Systems: A Tutorial, IEEE Trans. Aut. Contr., Vol.
H ( s ( t ) )- H(ii(t)) converges to zero in finite 76, NO. 3, pp. 212-232. 1988.
time.
Utkin V.I. A Method of Separation of Motion for
From 1, 2 and 23 it follows that
Designing a State Identijier, in: Problems of Con-
trol of multivariable Systems, Nauka, MOSCOW, 1983.
i ( t ) = dH(z(t)) - M(?)sign(V(t)- H(?(t))).(27)
dt Drakunov, S.V. Sliding-Mode Observers Based on
Equivalent Control Method, Proceedings of the 31st
Therefore from 21, 27 IEEE Conference on Decision and Control (CDC),
Tucson, Arizona, December 16-18, 1992, pp. 2368-
6; ( t )= hi+l(Z(t )) -mi(?)sign(vi(t )-hj( ?(t ) ) ) 2369.

(28) Draknnov, S.V., Izosimov D.B., Luk’yanov A.G.,

f o r i = l , ..., n. Utkin V.A. and Utkin V.I., The Block Control Prin-
If lhZ(x(t))l 5 ml(?(t))then in the first ciple, Part II,Automation and Remote Control, V.
51, no. 6 , 1 9 9 0 , ~ 737-746.
~.
equation 28 sliding mode occurs since 01 ( t )=
h l ( z ( t ) )and hence v l ( t ) - hl(G(t))= el@). Krener, A.J. and Respondek, W., Nonlinear ob-
servers with Linearizable Error Dynamics, SIAM J.
During sliding, e l ( t ) =’O. According to the Control and Optimization, V.23,no.2,March 1985.
equivalent control method ((ml(2)sign(vl(t)-
hl(2(t))))eq = hZ(x(t)),or, as it follows from Misawa, E.A. and Hedrick,J.K., Nonlinear Ob-
servers A State-of-the-art Survey, ASME J. Dy-
25, vz(t) = h,(z(t)). Again if the condition namic Syst. Measurement Contr.,v.lll,No.3,1989.
lh3(z(t))l5 ma(ii(t)) is satisfied sliding mode
Walcott, B.L., and Zak, S.H., Observation of
occurs in the second equation and we will Dynamical Systems in the Presence of Bounded
have eZ(t) = 0 and so on. Nonlineariries- Uncertaities, Proceedings of 25th
When the sliding mode occurs in the last CDC, pp. 961-966, 1986.
equation 28 we have e ( t ) = 0. The conditions Walcott, B.L., Codess, M.J. and Zak, S.H., Com-
for sliding mode to exist in every equation 28 parative Study of Nonlinear State Observation Tech-
are niques, Int,J.Control, V.45, no. 6 , pp. 2109-2132,
1987.
Ih;+1(.(t))l 5 mi(qt)). (29)

3378