Proceedings of the 39" IEEE
Conference on Decision and Control
Sydney, Australia December, 2000
An LMI Approach for Designing Sliding Mode Observers
Chee Pin Tan and Christopher Edwards
Department of Engineering, University of Leicester, Leicester LE1 7RH U.K.
e-mail : cpetl@le.ac.uk, ce@sun.engg.le.ac.uk
Abstract
This paper presents a method to design sliding mode
observers for a class of uncertain systems using Linear
Matrix Inequalities. The objective is to exploit the de-
grees of freedom available in the design which have hith-
erto been ignored because of the lack of a tractable so-
lution framework. The relationship between the linear
component of the sliding mode observer and a partic-
ular sub-optimal observer arising from classical Linear
Quadratic Gaussian theory is demonstrated. This helps
motivate ho& the design weighting matrices inherent in
the method may be chosen in practice.
1 Introduction
Sliding mode observers differ from linear Luenberger
observers in that there is a non-linear discontinuous
term injected into the observer depending on the out-
put estimation error. These observers are more robust
than Luenberger observers, as the discontinuous term
enables the observer to reject disturbances, and also
a class of mismatch between the system and observer.
The discontinuous term is designed to drive the trajec-
tories of the observer so that the state estimation error
vector is forced onto and subsequently remains on a
surface in the error space. This motion is referred to as
the sliding mode [lo]. In most cases, the sliding surface
is set to be the difference between the observer and sys-
tem output which is therefore forced to be zero. When
a sliding mode is achieved the system will experience
a reduced-order motion which is insensitive to a class
of system/plant mismatch. Utkin [lo] designed a sim-
ple observer, with only the discontinuous term being fed
back through an appropriate gain. Walcott and Zak [U]
designed an observer which also has the output error be-
ing fed back linearly and used a Lyapunov approach to
prove stability. The method in [ll]invariably requires
a symbolic manipulation package to solve the synthesis
problem which is formulated. Edwards and Spurgeon
[3, 51 proposed a canonical form for sliding mode ob-
server design subject to certain conditions relating to
the input and output distribution matrices, and also the
invariant zeros of the system. Their method described
in [3, 51 utilised both linear and discontinuous output
error injection. A method for computing the gain as-
O-7803-6638-7/00$10.000 2000 IEEE
sociated with the linear output error injection term is
presented. The solution is explicit, but does not exploit
all the degrees of freedom.
This paper builds on the work of Edwards and Spur-
geon [3, 51 and presents a method that will utilise Lin-
ear Matrix Inequalities (LMIs) [l]and seeks to exploit
this freedom. The relationship between the linear com-
ponent of the sliding mode observer and a particu-
lar sub-optimal observer arising from classical Linear
Quadratic Gaussian theory is demonstrated. This helps
motivate how the design weighting matrices inherent in
the method may be chosen in practice.
The notation used throughout the paper is standard;
in particular )I.)I will represent the Euclidean norm for
vectors and the induced spectral norm for matrices.
2 Preliminaries
Consider the dynamical system
k(t) = Az(t) + Bu(t)+ D[(t,z, U) (1)
!At> = W t ) (2)
where A E R n X n , BE RnXm,CE R P x n and D E
Rnxqwhere p 2 q. Assume that the matrices C and
D are full rank and the function 5 : R+ x Rn x R" 4
Rq is unknown but bounded so that
(3)
where r1 is a known scalar and CY : R+ x RP + R+ is
a known function.
Consider an observer of the form
k(t) = A2(t) + Bu(t) - Gle,(t) + Gnu (4)
$(t) = C q t ) (5).
where GI E R n X P and Gn E R n x p . The discontinuous
vector U is defined by
where ey = 6 - y and P2 E R p x p is symmetric positive
definite. The matrices P2 and V2 will be formally de-
fined in $3. The scalar function p : R+ x ' R P x Rm +
2587
R+ satisfies 0 The output distribution matrix has the form
At, Y,U ) L ~ 1 1 1 ~ 1 1 4+ t ,Y) + 70
(7) C=[O TI (15)
where 70 is a positive scalar. If the state estimation
error e := & - z,then it is straightforward to show where T E R P x P and is orthogonal.
from equations (1) and (4), (2) and (5) that
k ( t ) = Aoe(t) + Gnu- DC(t,2,U ) (8) 3 A Design Framework
where A0 = A-GlC. In 86.3 in [5], Edwards et al. show
that necessary and sufficient conditions for existence of Applying the linear change of co-ordinates TO to the
a stable sliding motion on S = {e E Rn : e y = 0) that observer system (4) and (5),
is independent of are <
&(t) = AP(t) + B u ( ~-) Gleg(t)+ Gnu (16)
1. rank(CD)=q $(t) = G ( t ) (17)
2. invariant zeros of ( A ,D, C ) lie in the open LHP
and define A0 = A - GlC. The gain matrix Gl is to be
determined but assume
Edwards et al.[5] suggest an explicit method for com-
puting the gains Gl and G,. However, the method
described in 151 does not exploit all degrees of freedom
available. The next section considers the design of the
matrices G1,G, and P2 so that a sliding motion takes where and L = [ L 0 ] with L E
E R(n-P)xP
place on S. This paper assumes that conditions 1 and 2 R(n-p)x(p-q). The orthogonal matrix T is part of the
are satisfied and sets up a new design framework which output distribution matrix 6 from (15).
seeks to exploit all of the available design freedom.
_Proposition 1 - If there exists a positive definite Lya-
Edwards et al.prove in Lemma 6.1 of [5] that if ( C D ) +
punov m a t r i x P , that satisfies PA0 A r P < 0 , with
is full rank, then there exists a linear change of coordi- the structure
nates TOso that the system can be written as :-
q t ) = Az(t)+ B U ( t ) + &(t, 2, U ) (9) = [ R
.ETPI P2 +pF2P 1 L ] >o (19)

y ( t ) = &(t) (10)
where PI E R ( n - p ) x ( n - p ) and p2 E R P x P , t h e n the
where (A,D , C) has the following structure - error s y s t e m an equation (8) is quadratically stable.

The system matrix can be written as Proof : Consider the quadratic form given by

V ( e )= B T P E (20)
as a candidate Lyapunov function where E := T o e . No-
tice that if Pl,P2 > 0 then P > 0 from the Schur
expansion. From (8) the derivative along the system
trajectory :-

v = eT(ATP+ PA0)e + 2eTPGnv - 2eTPDt (21)

(12)
where 4l E R ' X ' and A"2L E R(P-k')X(n-P-') for From the definitions in (15), (18) and (19):
some r 2 0 and the pair ( A ; 2 , A $ 1is ) completely
observable. Furthermore, the eigenvalues of 4,
are the invariant zeros of ( A ,D , C ) . PG, = [ pZTT ] = C T P ~
0
(22)
0 The disturbance distribution matrix has the form where PZ := TP2TT.

D=[ i2] Using the special structures of L and D2, ED2 = 0 and
where 0 2 E R p x q has the structure
PD = [ P2i2] = FP2V2
D2 = [ j 2] where Vz := T D z . Note this implies /1Vz11= 1
10~
11.
and D2 E RqxQis nonsingular.

2588
 The remainder of this paper focuses on design methods
to synthesise the gain G l and the Lyapunov matrix P
for A0 which has the structure given in (19). The prob-
lems will be posed in such a way that Linear Matrix
Inequalities (LMIs) can be used to numerically synthe-
Using the bounds for E in equations (3) and (7) sise the required matrices.

4 Synthesis Procedure for Matrices P and Gl

In this section, P and Gl will be chosen so that the

matrix inequality
Since (AFP + FAO)< 0 it follows that V < 0 for all A,TP + PA0 < -PWP - PGiVGFP (27)
af0.
is satisfied, where the design weighting matrices W and
Corollary 1 - An ideal sliding motion takes place on V are assumed t o be symmetric positive definite, and P
S in finite time. firthermore the sliding dynamics are has the structure in (19). The rationale for the matrix .
given by the system matrix 1111 LA211.+ inequality (27) will be given in the sequel. Substituting
for &, the inequality (27) can be written as :-
Proof: Using Proposition 1, a modification to Corol-
lary 6.1 in [4]shows that sliding takes place on S in ATP+PA-(pC)T-pC+PWP+pVpT < 0 (28)
finite time. Using the concept of equivalent output er-
ror injection, the sliding motion is governed by where := PG1. Using standard matrix manipula-
tions, inequality (28) is identical to
1111 + LA211 A12 + in,,]
(I - Gn(CGn)-lC)Ao= i o 0 P A + ATP + (FT--V-1C)TV(FT-V-1C)
Hence the reduced order sliding motion is governed by -CTV-'G + PWP < 4x29)
+
A11 LA211 as claimed.
Using inequality (29), the necessary and sufficient con-
Remarks: If a further linear change of co-ordinates dition for (28) to hold is that P satisfies

P A + A T P - GTV-lC + PWP < 0 (30)

since choosing
is applied to the triple (A,D , C) and its Lyapunov ma- PT = v-1C
(31)
trix p, the system matrix, disturbance distribution ma-
trix and the output distribution matrix will be in the eliminates the third term in (29). The problem consid-
form ered here is one of minimising truce(P-') subject to
satisfying inequality (30). The observer gain G1 can
then be directly calculated as G l = P-'CTV-l which
follows from equation (31) and the definition of I'.The
(25) matrix inequality in (30) is equivalenk to
where dl1 = All + LA211 and 272 = TD2. In the
new co-ordinate system, the Lyapunov matrix will be
P = ( T L - ~ ) ~ P ( T L and
- ~ ) hence
by using the Schur complement. If x
E R n x nis sym-
metric positive definite then (again using the Schur
complement) the LMI
As argued in [4], the fact that P is a block diagonal
Lyapunov matrix for A0 = A - GlC implies that A11 is
stable and hence the sliding motion is stable. Further- [ -;
4 < 0 (33)
more
is equivalent to X > P-l. Thus minimising truce(P-l)
subject to (30) is equivalent to minimising truce(X)
and the canonical observer form from [5] is recovered. subject to the LMIs (32) and (33). Writing from
(19) as
Finally note that equation (23) is (effectively) the
'structural constraint' employed by Walcott & Zak [ll]. (34)

2589
where 5 1 E R(n-p)x(n-p), P22 E R P x P and P12 := This section considers the sliding motion design prob-
[ P121 0 ] with 5 2 1 E R ( n - p ) x ( p - q ) , it follows there lem and shows how additional LMI constraints can be
is a one-to-one correspondence between the variables augmented with those in $4 to tune the sliding mode
(SI, S z 1 , P22) and (PI,E , 4 )since performance. Using the new co-ordinates obtained af-
ter applying the transformation TL in equation (24),
PI1 = PI (35) the matrix inequality (30) becomes
E = P;;'P12 (36)
4 = P22 - P;rzPG1P I 2 (37) Pd + ATP - CTV-lC + PWP < 0 (40)

It follows that the constrained minimisation problem is where W := TLWTLT.The top left block of (40) is
convex with regard to 9 1 , P 1 2 1 , P 2 2 and X. Standard
LMI software, such as [SI, can be employed to synthesise PIA11 + + PiWiP1 <0 (41)
numerically P and X.
where W1 E % L ( n - P ) X ( n - p ) > 0 is the top left sub-block
4.1 Connection with the ARE
+
of W and d l 1 = A 1 1 LA211. If the weighting matrix
W is partitioned as
The motivation for the choice of the inequality posed
in (27), and the minimisation of trace(P-') subject to
(30) and (33), is that in the absence of uncertainty and
as 'yo + 0 the observer tends to a sub-optimal Linear
Quadratic Gaussian formulation. This will be demon- where W11 E R ( n - p ) x ( n - p ) and W12 is the null space
strated using an argument similar to that on page 115 of ET for all L , then (41) is
of [l].Defining 0 = P - l , then pre and post multiply-
ing inequality (30) by Q, the following inequality can PiAii + ATlP1 + PiW11Pi + PiETW2zLPi < 0 (43)
be obtained :
This is identical in structure to inequality (27) and
AQ + &AT - QCTV-'CQ +W < 0 (38) hence Wll and W22 may be interpreted as playing the
and G l = QCTV-l. The objective is thus to minimise roles of performance and noise attenuation matrices in a
trace(Q) subject to (38). Linear Quadratic Gaussian sense for the observer prob-
lem associated with the pair ( A 1 1 , A 2 1 1 ) . Thus the
The Linear Quadratic Gaussian 191 optimal observer de- choice of W11 and W22 can be used to tune the slid-
sign method uses the stabilising solution QaTe to the ing motion. However since
Algebraic Ricatti Equation (ARE)
PIA11 = piiA,i + Pizi&ii (44)
AQare+ QareAT - Q a , e ~ ~ V - l c Q a+r ew = 0 (39)
and this is linear with respect to the LMI optimisation
to calculate the optimal observer gain as variables Pll and P121, additional LMIs can be em-
Glare := QareC'Tv-l ployed together with (32) and (33) to tune the sliding
mode performance. One approach is to use Root Clus-
The associated optimal cost is given by trace(Qare). tering [7] methods to achieve pole placement of d l 1 in
regions of the complex plane. Typically the poles may
It is argued in the Appendix that if Q > 0 is any matrix be required to lie in
satisfying (38) then Q > Q,,,. Hence the requirement
of minimising trace(Q) follows from the desire to ap-
proach the true minimal cost given by trace(&,,,). Of 0 conic sector centred at (0,O) with inner angle ea
course a particular sub-optimal cost is enforced by the
0 disc of radius T, and centre (ya,O)
requirement that P := Q-l is in the structure of (19).
vertical strip aa < x < ba
In inequality (38), W is the performance weighting ma-
trix for the observer, and V is the co-variance ma-
trix of the system's sensor noise. As in classical Lin- Chilali and Gahinet [2] have proven that the following
ear Quadratic Gaussian theory the choice of W and V inequalities will describe these regions
trades-off performance and noise amplification.
< 0 (45)

5 Design of the Sliding Motion System Matrix

< 0 (46)
A consequence of the design procedure proposed in $4 is < 0 (47)
that the dynamics of the sliding motion, although guar- < 0 (48)
anteed to be stable, are designed somewhat implicitly.

2590
where s = sin 36a and c = cos 36, To obtain a con-
vex optimization problem, write PIA11 = P11.411
P121A211 and substitute into (45) - (48). This results
in a well-posed convex problem as the inequalities (45)
- (48) are affine in 9 1 and P121. Thus the new opti-
mization problem can be stated as : Minimize trace(X)
subject to the LMIs (32), (33) and (45) - (48).
The weighting matrices W and V associated with the
+ Ricatti-like inequality (30) were chosen respectively as
W = 0.0117 and V = 0.2514.
6.2 Synthesis results
The following matrices were obtained for the canonical
form described in $2 -

-2.0722 5.0994 1.6893 * * * *

6 AnExample 0.0000 -0.0000 -0.0000 * * * *
0.0000 0.0000 0.0000 * * * *
The new design method proposed in this paper will now A= -0.0000 0.4962 0.0226 * * * *
he demonstrated by an example. This example is a 7th 0.0000 0.0122 -0.9159 * * * *
order aircraft model taken from [8]. The states are :- * * * * * * *
* * * * * *
bank angle (rad) 0
= 0 0.0000 16.3353
yaw rate (radls)
roll rate (radls)
DT
[ 00 0
0
0
0 0 0 25.8356 -10.8242
0 0 0 0 0 -0.4126 -0.9109
X = sideslip angle (rad)
0 0 0 0 0 -0.9109 0.4126
washout filter state
0 0 0 0 1 0
rudder deflection (rad) 0 0 0 1 0 0 0
aileron deflection (rad)
the inputs

and the outputs

rudder command(rad)
aileron command(rad)
From this representation

A11 =
[ 0 4 2
] = [
T I -2.0722 5.0994 1.6893

0.0000 0.0000 0.0000

Ta roll acceleration (rads/s2) and

.-[ 21 yaw acceleration(rads J s2)

bank angle (rad)
washout filter state
A211 = [ 0 AZl ] =
[ 0.0000 I
0.0000 0.4962
0.0122 -0.9159
o.0226 1
The system matrices are given in (49)-(50).
Notice the system has an invariant zero at -2.0722.
6.1 Regions and weighting matrices _.
The eigenvalues of the sliding motion represented by Following the synthesis procedure in 54 and the impos-
the system matrix dl1 were required to lie in the inter- ing of the constraints in $6.1, the following matrices
section of the following regions - were obtained -

1
a circle of centre (0,O) and radius 5 -9.5598 1.5465
L= -4.0743 -0.0894
0 a vertical upper bound at x = -2 0.5161 2.1753

1
a conic sector symmetric about the real axis, with - 7.2296 -15.6675 -5.7861
inner angle 6 = 80" PI = -15.6675 49.1586 13.6032
-5.7861 13.6032 5.1419

r
0 0 1.000 0 0 0 0 . 0 0 -
0 -0.154 -0.004 1.540 0 -0.744 -0.032 0 0
0 0.249 -1.000 -5.200 0 0.337 -1.120 0 0
A = 0.0386 -0.996 -0.000 -2.117 0 0.020 0 0 0 (49)
0 0.500 0 0 -4.000 0 0 0 0
0 0 0 0 0 -20.000 0 20 0
0 0 0 0 0 0 -25.000 , 0 25-
-0.744 -0.032
0.249 -1.000 -5.200 0 0.337 -1.120
0 0 0 0
0 0 0 1 0 O I
0

2591
 [12] B.L. Walcott, S.H. Zak. Combined Observer Con-

A= [ . 476.7394
-1.5752
-5.9434
-12.3227
-1.5752
1.6130
-0.5957
-0.7175
-5.9434
-0.5957
1.4526
2.0924
-12.3227
-0.7175
2.0924
3.6798 1
troller Synthesis for Uncertain Dynamical Systems with A;?-
plications. IEEE Transactions on Systems, Man and Cyber-
netics, Vol 18, No 1: 88-104, 1988
[13] J.C. Willems. Least Squares Optimal Control and the
Algebraic Ricatti Equation. IEEE Transactions on Auto-
matic Control, Vol AC-16, No 6: 621-634, 1971
I :nce the eigenvalues of Jlo are { -44.7043, -21.135
-1.4057f 0.88883, -4.3773, -1.4793, -4.0039) and the
eigenvalues of dl1 which govern the sliding motion are 8 Appendix
(-2.0722, -2.0016 f 0.04903)
In 54 is was argued that the design problem should be
one of minimising t r a c e ( Q ) subject t o the inequality
7 Conclusion given in (38). The motivation for this comes from the
following:
This paper demonstrates how Linear Matrix Inequal-
ities can be used t o synthesize the gains of a sliding Lemma 1: Let Qlmi be any symmetric positive definite
mode observer. A formulation has been presented in matrix satisfying
a way that the linear component of the observer re-
sembles a sub-optimal version of the classical Linear
Quadratic Gaussian observer and two design weight-
and let QaTebe the stabilising solution t o the ARE (39).
ing matrices allow a trade-off between performance and
Then Qlmi > Qa,e and trace(Qlmi) > tmce(Q,,,).
sensor noise. Additional LMIs can be augmented to fa-
cilitate the design of the reduced order sliding motion Proof
whilst still retaining a convex optimization problem. Inequality (51) can be expressed as:-

References
[l] S. Boyd, L. El Ghaoui, E. Feron, V. Balakrishnan.
Linear Matrix Inequalities in System and Control T h e o y . for some symmetric positive definite matrix A. Sub-
SIAM, 1994 tracting (39) from (52) and defining Q = Qlmi - &
,,
[2] M. Chilali, P. Gahinet. H , design with Pole Place- implies
ment constraints : An L M I Approach. IEEE Transactions
on Automatic Control, 41(3):358-367, March 1996
[3] C. Edwards, S.K. Spurgeon. O n the development of
discontinuous observers. International Journal of Control,
25:1211-1229, 1994
(41 C. Edwards, S.K. Spurgeon. Sliding Mode Control:
T h e o y and Applications. Taylor & Francis, 1998
[5] C. Edwards, S.K. Spurgeon, R.J. Patton. Sliding
Mode Observers for Fault Detection and Isolation. Auto- ( A - QlmiCTv-lC)& + &(A- Qlmi(yV-1C)T
matica, 36(4), 2000 + A +QC'V-'CQ = O(54)
[6] P. Gahinet, A. Nemirovski, A. Laub, M. Chilali. An
L M I Control Toolbox, For Use with Matlab. The Math- In equality (51) can be written as
Works, Inc., 1995
[7] S. Gutman, E. Jury. A General Theory for Matrix
Root-Clustering in Subregions of the Complex Plane. IEEE
Transactions on Automatic Control, Vol AC-26, No 4, 1981
~ [8] B.S. Heck, S.V. Yallapragada, .M.K.H. Fan. Numeri-
cal methods to design the reaching phase of output feedback
variable structure control. Automatica, 31:275-279, 1995 and hence, since Qlmi > 0, it follows that the ma-
[9] J.M. Maciejowski. Multivariable Feedback Design. trix (2 - &lmiCTV-lC) is stable. Therefore, as ar-
Addison-Wesley, 1989 gued in Lemma 3 in [13], equation (54) implies Q > 0
[lo] V.I. Utkin. Sliding Modes in Control Optimization.
and hence Qlmi > QaTe as claimed. The fact that
Springer-Verlag, Berlin, 1992 trace(&l,i) > trace(Q,,,) follcws from the properties
[ll] B.L. Walcott, S.H. Zak. State Observation of Non-
of the trace operator.
linear Uncertain Dynamic Systems. IEEE Transactions on
Automatic Control, Vol AC-32, No 2: 166-170, 1987

2592