IEEE TRANSACTIONS ON AUMMATIC CONTROL, VOL. 40,NO. 7.

JULY I995 1323

Each column of Dp is full-rank for all s E U. One choice for Ysl REFERENCES
is [i2:]. A decoupling controller for P is S . P. Boyd and C. H. Barratt, Linear Controller Design: Limits of
Pelfonnance. Englewood Cliffs, NJ:Prentice-Hall, 1991.
- - C. A. Desoer and M. Vidyasagar, Feedback Systems: Input-Outpur
C ~ Z=, DqALV,5p Properiies. New York Academic, 1975.
C. A. Desoer and M. G. Kabuli, “Linear stable unity-feedback system:
Necessary and sufficient conditions for stability under nonlinear plant
perturbations,” IEEE Trans. Auromr. Contr., vol. 34, no. 2, pp. 187-191,
1989.
__, “Factorization approach to nonlinear feedback systems,” F’h.D.
dissertation, Electronics Res. Lab., Memo. No. UCBERL M89/64,
Univ. California, Berkeley, 1989.
M. Fujita and E. Shimemura, “Integrity against arbitrary feedback-loop
failure in linear multivariable control systems,” Auromatica, vol. 24, p.
765, 1988.
A. N. Giindeg and C. A. Desoer, Algebraic Theory of Linear Feedbnck
Sysrems Wirh Full and Decentralized Compensators, Lecture Notes in
Control and Information Sciences, vol. 142, Springer-Verlag, 1990.

[A :I)
A. N. GiindeS, “Stability of feedback systems with sensor or actuator
- -3(s - 3) failures: Analysis,” Inr. J. Contr., vol. 56, no. 4, pp. 135-153, 1992.
----(I-s + l __ , “Parametrization of decoupling controllers in the unity-feedback
system,” IEEE Trans. Automat. Control, vol. 31, no. 10, pp. 1572-1576,
has zero diagonal entries. 1992.
C. A. Lin and T. F. Hsieh, “Decoupling controller design for linear mul-
tivariable plants,’’ IEEE Trans. Automat. Contr., vol. 36, pp. 485-489,
N. CONCLUSION 1991.
M. Vidyasagar, Conrrol System Synthesis: A Factorization Approach.
We considered the stability of LTI,MIMO systems under stable Cambridge, MA: M.I.T. Ress, 1985.
NLTV diagonal post-multiplicative perturbations of the plant (sensor
failures) and pre-multiplicative perturbations of the plant (actuator
failures). Assuming that any one of the sensors or actuators may
fail one at a time, without prior knowledge of the NLTV failure
and its location, we obtained necessary and sufficient conditions for
stabilization of the LTI plant using LTI controllers. We developed
Design of Observers for Descriptor Systems
controller design methods for two classes of plants and explicitly M. Darouach and M. Boutayeb
derived a family of controllers achieving closed-loop stability under
an unknown stable NLTV failure of at most one loop.
The closed-loop stability condition for all possible unrestricted Absbact-In this note simple and straightforward methods to design
stable NLTV perturbations of one sensor or actuator is that all full- and reducedader observers for linear time-invariant descriptor
diagonal entries of certain transfer functions of the nominal LTI systems are presented. The approach for the reducedader observer
design m based on the generazed Sylvester equation. sufficient eonditiolrs
,
system are exactly zero. Note that such a strict condition is due to
for the existence of the observers are given.-An illustrative erampk is
the general unconstrained nature of the NLTV perturbations. If this included.
condition could not be met and the diagonal entries were not zero
(but their gains were bounded by sufficiently small E), then as in
standard robustness results based on the small-gain theorem, closed- I. INTRODUCTION
loop stability is still guaranteed for a restricted class of perturbations The problem of designing observers for descriptor systems has
(whose gains are bounded by l / ~ ) . received considerable attention in the last two decades [ 11-[7]. Many
The design goal stated above can easily be expressed as a convex approaches exist to design observerg for descriptor systems. In [l],a
design specification [ 11. Despite the obvious advantages of convex method based on the singular value decomposition and the concept
problem formulations, optimization based design approaches cannot of matrix generalized inverse to design a reduced-order observer has
bring an explicit answer in cases where there does not exist such a been proposed. In [2], the generalized Sylvester equation was used
controller (in general, not for a particular finite parametrization only) to develop a procedure for designing reduced-order observers. In [3],
or make explicit use of the plant properties in successive redesigns a method based on the generalized inverse, which extend the method
(each new plant will be treated as a new design problem). Hence, developed in [SI,was presented. Full- and reduced-order observers for
whenever applicable, analytical solutions should be used to weed discrete-time descriptor systems have been presented in [5] and [7].
out infeasible constraints and initialize feasible ones. The results All these works have been done under the assumptions of regularity
in this note are intended for shch complementary utilization. The and generalized observability. Theae conditions are very restrictive
analysis results provide valuable necessary conditions on plants that and, as can be seen in [6], the design of the observer can be done
admit such controllers. For two classes of plants, the explicit design under less restrictive conditions. In [6], the conditions of regularity
approach developed here guarantees a desired nominal controller. and modal observability have been assumed. In [lo], a new method
The particular design approach complements the convex optimization
Manuscript received August 26, 1993; revised July 6, 1994.
based control design approach since it explicitly generates a family The authors are with the CRAN-EAIkAL-CNRS UA 821 - Universitd de
of feasible stable parameters for the controller. The freedom can then Nancy I, 54400 Cosnes et Romain, France.
be used to satisfy other design specifications. IEEE Log Number 9412366.

0018-9286/95$04,00 0 1995 IEEE

1324 B E E TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 40, NO. 7, JULY 1995

is developed for the reduced-order observer design for descriptor Then, using the above transformations, we can easily prove that
systems where the regularity assumption (A*,E* are square and
det (XE' - A*) # 0) is not required and a condition less restrictive
(5.1)
than the impulse observability is used.
In this note we present simple new methods to design full- and
reduced-order observers for continuous-time descriptor systems. For or equivalently
the full-order observer the approach is based on a method developed
recently in [9], and extends it to descriptor systems, while the = n
reduced-order observer design method is based on a new resolution
method of the generalized Sylvester equation. These approaches do since E is a full row rank matrix.
not require any coordinate transformation. As in [lo], the regularity Now since rank [E]
= n , there exists a nonsingular matrix [z j]
assumption is not required; however, a condition which generalizes such that
the impulse observability is assumed as in [12]. +
aE bC = I ,
It is well known that a solution to nonregular descriptor systems cE+dC=O.
does not exist or may not be unique for given initial conditions and
time piecewise continuously differentiable input [lo]. According to Our aim is to design an observer of fixed order n satisfying
[ 111, one can note that even if the descriptor vector of a nonregular
i =Nz+L~~+LZ~+GU (7.1)
system cannot be uniquely determined, it is possible to obtain a
unique input-output description. As can be seen in [lo], the observers + +
2 = t by K d y (7.2)
design for nonregular descriptor systems can be realized, since where N . L1, Lz,G, and A- are unknown matrices of appropriate
measurements provide additional information to the descriptor vector dimensions, which must be determined such that 2 will asymptotically
equation. One can note that the descriptor vector can be uniquely converge to T .
determined from the descriptor vector equation and the measurement From (4), (6), and (7) the estimate error
vector if and only if the matrix pencil [""',l"*] is of full column
rank. e=x-z
=(a +Kc)Ez - 2 (8)
II. FULL-ORDEROBSERVER DESIGN satisfies the linear state equation
Consider the linear time-invariant descriptor system
P =Ne + [ ( a + h-c)A- N(a + K c ) E - L I C - L z C ] ~
E'X = A * x + B*u (1.1) + + K c ) B - G]u.
[(U (9)
y* =c+x (1.2)
Quation (9) reduces to the homogeneous equation
where r E R". U E R k , and y * E Rp are the state vector, the
input vector, and the output vector, respectively. E* E R" n , A* E i = Me
R" ", B' E R" ', and C 8 E RP are known constant matrices. provided the matrices G, K , L1. L z , and N satisfy
We assume that rank E* = T < n, and without loss of generality
rank C* = p. G = (a + Kc)B
In the sequel we assume that and

(2)
(a + Kc)A - N ( a + 1ic)E - L1C - LzC = 0 (11)
then 3 will asymptotically converge to z if and only if N is a stability
Then, since rank E* = T , there exists a nonsingular matrix P such matrix.
that The problem of designing a full-order observer of the singular
system (1) is reduced to find matrices K , L1, Lz,and a stability
PE' = [f] matrix N such that (11) is satisfied.
Using (6), (1 1) can be written as
PA' =.[.1] (a + K c ) A - N ( I n - bC - K d C ) - LiC - LzC = 0
and or

PB* = [i] (3)

Take
X = (U + K c ) A - LzC + [ N ( b+ K d ) - Ll]C. (12)

with E E R'." and rank E = I T .

Following [?I. system (1) is a restricted system equivalent (r.s.e) to L1 = N ( b + Kd) (13)
Ei = A x + Bu then

where
y =Cx N = ail + KcA - LzC = aA + [ K - Lz] . ['."I (14)

which must be a stability matrix.

Now Y is a stability matrix if and only if the pair ( a A , rg])
is
detectable. In this case, matrices K and LZ can be determined from
the standard observer design problem.
T. From the definition of matrices E, A , and C it is easy to prove
the following lemma.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 40,NO. 7, JULY 1995 1325

Lemma 2.1: For systems (1) and (3) we have Our aim is to design an observer in the form

rankpEL:A'] =rank[ S E - A 1, V s E C. (15)

6 =nu, + Ly + H u
i=Mw+Fy
(19.1)
(19.2)

Now we can give the sufficient conditions for the existence of the
where U' E R n P q .
observer (7) for system (1).
Theorem 2.1: Sufficient conditions for the existence of the ob- The problem of the observer design is reduced to find matrices
server (7) are II, L , H, M . and F such that the estimate j: converges acymptot-
i c d y to s.

a) rank
:[ 0
3
E* = n + r a n k E * ,
The solution of this problem is given in the following theorem.
Theorem 3.1: Let T be an ( n - q ) x r matrix such that

T A - IITE = LC

Proof: From (2)-(5) condition a) is equivalent to rank [ g ] = n,

which yields (6), (8), and (12).
where
det r:] f0.

Now, note that Then for

rank "'1 =rank [SE- '4] H =TB (21)

[:; I:;- [ $ ]
and
0 0 SE-A
=rank I,
0 0

=rank[ 5]
SE- A
(17)
we have ? ( t )- ~ ( t=) bfe"'[w(O) - T E z ( O ) ] .
The convergence of the reduced-order observer is obtained when
II is a stability matrix. This theorem is easy to prove, hence its proof
is omitted.
then using (a) or equivalently (6) we obtain We will next give a new method to find the matrix T.

rank[
SE-A
2 ] =rank[+ i][ 2 ]
--: b 0 SE-A
Define the following nonsingular matrix:

I:[ K
piq I q ] [ c ]
TE

=rank[
SI, - n A

Vs
1,
E C, Re(s) 2 0
=

where K is an ( T I - q).q arbitrary matrix and R is an ( n - q ) . n

matrix of full row rank.
(18) Then, we have

or (ad. ['e])
is detectable. (T h-) [E] =R
Remark I: Matrices a, b , c, and d such that (6) is verified can
be found by using the singular value decomposition (S.V.D.)of the
matrix [F]. In fact, since rank [E] = R there exist two orthogonal since rank [E] = n, (23) has the following solution
matrices U and 1' such that

where
[:] P:.]=

C =diag(al? U Z , a,)
where [:I+ = A ( E T C T ) : A = ( E T E +CTC)-l, and 2 is an
arbitrary matrix of appropriate dimension.
and From (24), we have
I
IT, >0 ( i = 1, n )
+Z
then
T = RAET
['.- -CAET

and from (20) we obtain

11 = T A M
111. REDUCED-ORDER DE~IGN
OBSERVER L =TAF.
In this section we present a new method to design a reduced-
order observer for singular system (1). We assume without loss of The problem is reduced to find the matrix T such that II is a
generality that rank C = q. stability matrix.
 1326 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 40. NO. 7. JULY 1995

Substituting (25) into (26) gives We can now propose the following design procedure:
1) Choose an (n - q ) x n matrix R such that [g]
is nonsingular,
this can be always done, since C is of full row rank.
2) The matrix A4 can be obtained from (22) and (23)

[a]-' ["F'].
where
M =
r =R A E ~ A M
and
3) If the pair (I?, 12) is detectable, we can find 2 such that the
observer is asymptotically stable, then we can deduce T from
(25).
If the pair (I?, a) is detectable, one can design a reduced-order 4) The matrix F can be obtained from (22)
observer from the standard methods, in the form (19).
In the sequel, we will give the sufficient conditions for the existence
of the reduced-order observer.
Theorem 3.2: For system (l), the reduced-order observer (19)
exists if conditions a) and b) in Theorem 2.1 are satisfied. 5 ) Matrices H and L can be obtained from (21) and (27).
Pro08 Condition a) can be obtained from (23), (2), and (5).
Now define the following matrices IV. EXAMPLE
The following example illustrates the above design methods.
Example: Consider the singular system (1) described by
1 0 0 0

r
E'
=lxj
=[-a H 4 y]
- AIF
RAET(~E 1 0 0 1 0
-CAET(~E -Ay
-(rr - E A E ~ ) ( -~ AE) F A*
- C A E ~ ( ~-EA ) F

172 = [ M F ] E R"'" 0
and

and
B'
= [i]
where U 2 is nonsingular matrix and rank S = n + q , since P,. = C' = [0 1 0 01.'
Ker[F] and rank Cil = T 9 . +
Then we have In this case

and
SE-A
rank [E'
0
0
A*
E * ] - r d E ' =4.
c*
s~,,-q - R AE~AM o If we choose the transformation matrix P = 1 4 , we obtain the r.s.e
-CA AM singular system (4)described by

I
E=[;
0 0
: x x] 1 0
which gives .=[1 0 0 01
- 1 0 0 1
.I=[ 0 1 1 11

Re(s) 2 0 ( V S E C).

Remarks:
. 1) If rank [$:I = n the above results can be applied directly to
system (1) without any transformation.
2) The reduced order p of the observer is given by: j i = n -
rank C.
- 1 0 0 1
C = [ " 1 1
0 1 0 0
'1
EEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 40, NO. 7, JULY 1995 1321

[ s’i
and and
y=[-Ftf],
I:[
thenwehaverank =1.

The above result can then be applied to the transformed system.

IL=
0

a) Full-order observer: Matrices a, b. c, and d are

-112

%] For II = -4,
T = [l -31
we can choose 2 = [0 -3 0 0 31, then

i]
1 / = [ -1 0

/_[ -1 ; ;]
1 -1

and
F = [-1!

L=[-l
1 -4

1 -131
c = [0 11
and H =O.
d = [O 0 -11. Thus the reduced-order observer is

Now, let the observer eigenvalues be specified by -4. -3, -2, and ii = -4w - 13y’ + U
-1, then we can find

V. CONCLUSION
We have presented systematic design methods of full- and reduced-
order observers for linear descriptor systems. The reduced-order
observer design method is based on the new resolution method of
the constrained generalized Sylvester equation. It was shown that the
N = diag (-4, -3, -2, -1) existence conditions of the observer generalize those adopted in [12]
for the observer design of square descriptors systems problem. An
extension to a less restrictive conditions is under study.
m
ACKh’OWLEDGMENT
and
The authors would like to thank the anonymous reviewers for their
G =O. helpful and constructive comments.

Thus the full-order observer is REFERENCES

[l] M. El-Tohanu, V. Lovass-Nagy, and R. Mukundan, “On the design
of observers for generalized state space systems using singular value
decomposition,” Int. J. Contr., vol. 38, pp. 673-683, 1983.
[Z] M. Verhaegen and P. Vandooren, “A reduced observer for descriptor
systems,” Syst. Contr. Lett., vol. 8, no. 5, pp. 29-37, 1986.
[3] B. Shafai and R. L. Caroll, “Design of minimal order observers for
singular systems,” Int. J. Contr., vol. 45, no. 3, pp. 1075-1081, 1987.
[4] M. M. Fahmy and J. O’Reilly, “Observer for descriptor systems,” Int.
J. Contr,, vol. 49, pp. 2013-2028. 1989.
[5] L. Dai, “Observer for discrete singular systems,” IEEE Trans. Automat.
Contr.. vol. 33, pp. 187-191, 1988.
I
[6] P. N. Paraskevopoulos and F. N. Kounboulis, “Observers for singular
systems,” IEEE Trans. Automat. Contr., vol. 37, pp. 1211-1215, 1992.

M = [-;I
b) Reduced-order observer: Choose I3 = [l 0 0 01, then

1
[7] L.Dai, Singular Control Systems. Berlin: Springer, 1989.
[8] G. Das and T. K. Goshal, “Reduced-order observer construction by
generalized matrix,” Int. J. Contr., vol. 33, pp. 371-378, 1981.
[9] M. Darouach, M. Zasadzinski, and S. J. Xu, “Full-order observers for
linear systems with unknown inputs,” IEEE Trans. Automat. Contr., vol.
39, pp. 606609, 1994.
[lo] P. C. Muller and M. Hou, “On the observer design for descriptor
systems,” IEEE Trans. Automat. Contr., vol. 38, pp. 1666-16671, 1993.
0 112 -112 [ l l ] M. Hou and P. C. Muller “A singular matrix pencil theoty for linear
-1 -112 descriptor systems,” in Symp. on Implicit and Nonlinear Systems, Ft.
7/2 Worth,TX, Dec. 14-15, 1992.
11 0 -2 21 [12] D. N. Shields, “Observers for descriptor systems,” Inr. J. Contr., vol.
r=-1 55, pp. 249-256, 1992.