Sie sind auf Seite 1von 5

Proceedings of the American Control Conference Arlington, VA June 25-27, 2001

H , Control of Differential Algebraic Equation Systems: The Linearizing Change of Variables Approach Revisited
A. Rehmt
and

F. Allgowed

Institut fur Systemdynamik und Regelungstechnik, Universitat Stuttgart, Pfaffenwaldring 9, D-70550 Stuttgart, Germany. Email: ansgarQisr.uni-Stuttgart .de, Fax: +49-711-6856371.

$ Institut fur Systemtheorie technischer Prozesse, Universitat Stuttgart, Pfaffenwaldring 9, D-70550 Stuttgart, Germany. Email: a l l g o w e r Q i s t.u n i - S t u t t g a r t . d e , Fax: +49-711-6857735.
and non-holonomic constraints [lo]. Also in elec. tronics and even in economics DAE descriptions are encountered [5]. DAE systems are able to describe system behaviors that cannot be captured by systems only governed by differential equations [l]. Therefore index reduction techniques (i.e. reduction of a DAE system to an ODE system) necessarily are connected to a loss of information for high index systems. Due to this fact much work has been focused on analysis and design techniques for DAE systems in recent years (see [4] an overview). for For linear systems many of the standard design techniques for state-space systems have been extended to DAE systems. Especially there has been a focus on LMI synthesis techniques which guarantee bounds on induced vector norms (e.g. Hz, , Hnorm) for input-output descriptions of the form

Abstract
In this paper control of high index or even nonregular linear differential-algebraic-equation systems subject to H , performance constraints is considered. In a previous paper [8] this problem was addressed by a linearizing change of variables approach under a certain assumption on the structure of the algebraic equations within the system description. This restriction is overcome in the paper at hand. By means of a modified linearizing change of variables approach to the H , control problem, the synthesis conditions (given as linear matrix inequalities (LMIs)) not only become structurally simpler) but, in contrast to [8], also include the case without algebraic constraints.

1.

Introduction

Differential-algebraic equation (DAE) systems (sometimes also referred to as singular, semistate or descriptor systems) describe a broad class of systems which are not only of theoretical interest but also have great practical significance. Models of chemical processes for example typically consist of differential equations describing the dynamic balances of mass and energy while additional algebraic equations account for thermodynamic equilibrium relations, steady-state assumptions, empirical correlations, etc. [3]. In mechanical engineering DAE system descriptions result from holonomic

E i ( t ) = A<(t) Bw(t), t 2 0, c(O-) = <{ %(t) C<(t) Dw(t). = (1)


Here [ ( t ) 6 E"' denote the descriptor variables, w ( t ) E lRnw the external input variables, and z ( t ) E Elnzthe external output variables. E , A, B , C, D are constant system matrices of appropiate dimensions with E being a possibly singular nt x nt matrix with nt 2 rank(E) =: T. Usually the LMI approaches to this kind of problems (e.g.

+ +

0-7803-6495-3/01/$10.00 0 2001 AACC

2948

[6, 81) assume an E-matrix in SVD form, i.e.

E=

[E i],

C = C T ER'".

(2)

Proposition 2 1 (Generalized . bounded real A lemma, 'GBRL) system ( E ,A, B, C ) is a stable index one system with
((GI(, < y,
G(s) := C ( s E - A ) - l B

(3)

Theoretically there is no loss of generality connected. to this assumption since a transformation to an E matrix of the form (2) is always possible. However, this transformation may be ill conditioned. This is especially the case for mechanical DAE descriptions where point masses of extremely different magnitudes are involved. Furthermore the approaches based on (2) results in synthesis LMIs with all occurring system matrices partitioned according to (2). This is not only notational inconvenient but in fact means that the standard case (regular E matrix) is not included. It is shown in this paper that these drawbacks easily can be avoided when no partioning of E is considered. The modified linearizing change of variables approach' (extensively used in [9] for state space problems) employed in this paper reveals that the standard result is nothing more than a special case in the DAE setup. The paper is structured as follows: in the next section a generalization of the bounded real lemma is presented. This result is the basis for the linearizing change of variables approach in the following main part of the paper. Finally we briefly consider the computational implications of the derived synthesis result.

ifl there exists a matrix X with


ETX = XTE 2 0

(4)

<
0

0.

-yI
(5)

Proof. See [7]. 0 Remark 1. The consideration of the case D = 0 in the previous proposition is not restrictive since every DAE system (1) can be reformulated as a DAE system with D = 0 if additional descriptor variables with &d(t) := Dw(t) are introduced. Remark 2. The LMI (4)is non-strict. The key towards a strict inequality is the symmetry constraint ETX = X T E expressed in (4). All X fulfilling this constraint can be parameterized in terms of the fundamental subspaces of E as

-.

2.

The Generalized Bounded Real Lemma

In contrast to state space system descriptions a DAE system may allow non-unique solutions which possibly contain impulses. This certainly does not fit into the internal stability requirement which goes along with the H,-norm bound requirement in the standard Hw control problem. As a generalization one therefore considers regular (i.e. DAE systems with a unique solution) and impulse-free DAE systems. DAE systems which additionally are stable are termed admissible [ 6 ] . An LMI based characterization of admissible DAE systems ( E , A , B , C ) (i.e. DAE systems (1) with D = 0) which are H,-norm bounded is given in the following proposition:

with E l denoting a full rank matrix such that ETEL = 0 and r?, U being matrices of appropriate dimensions. The parameterization ( 6 ) in X , W is valid since we may write (4) as VETUTU-TXVT = VXTU-lUEVT with Esvd := UEVT being a SVD decomposition of E. With X' := U-TXVT we get E,T,dX'= XITEsvd , i.e. X' = with a block structure corre-

[z a]

sponding to Esvd. This X' clearly can be parameterized as in ( 6 ) . Finally we observe that the (1,l)element in ( 5 ) implies the regularity of X . In view of (4)the parameterization ( 6 ) can be strengthen by r? > 0. A strict inequality characterization of a H,-norm bound y then can be derived by substituting (6) into (5) and replacing (4)by X > 0. Note that the matrix X is over-parameterized by (6) with respect to the variables not affected by the positive definiteness requirement in (4).This may be used to put further constraints on X in (6). The previous remark is helpful for H,-norm bound

2949

computations if only a strict LMI solver is available. However, the main importance of this remark will become clear in the context of the corresponding Hw controller synthesis problem for DAE systems which is addressed in the next section.

Then all controllers K E solving the H , control problem for DAE systems are characterized by the following theorem:

Theorem 3.1 Consider a plant (7) and a controller (10). There exists a controller parameterization A K , B K , C K , DK such that the undisturbed closed loop system (11) is admissible with I I G c r l l ~ Y (with Gcz(s) := Ccz(SEc1- &)-'BCi) < if and only if the LMIs (a), (9) at the top of the pagel-admit a solution {R, S , W y , Wx, K , BK, A C K , DK}.
Proof. Necessity: Application of the generalized bounded real lemma (Proposition 2.1) to the closed loop system matrices (12) renders the necessary and sufficient LMI/BMI conditions

3.

The H , Control Problem for Linear DAE Systems

Consider a generalized plant C E that is a DAE system


CE

E k ( t ) = A z ( t ) Blw(t) B2u(t) %(t) = clZ(t)

dt) =

C24t)

(7)

where z ( t )E Rnz denotes the descriptor variables, u(t)E. Rnu control input, w ( t ) E Rnw exthe the ternal input, z ( t ) E RnX external output, and the y ( t ) E R"u the measured output. A , Bi, Ci are constant matrices of appropriate dimension and E is a possibly singular matrix having the same dimension as A. Notice that there is no loss of generality in the descriptor setup in neglecting a direct fed-through of control/external input to the measured/external output since such a dependency also can be expressed by means of an augmented descriptor vector 2 [6]. . The control problem is to find a linear output feedback controller such that the undisturbed closed loop (w = 0 ) is an admissible system and such that the transfer matrix from the external input w to the external output z is H,-norm bounded by a prescribed number y > 0. With a controller K E ,

E,TX = XTEci 2 0 ,

(13)

< 0.

(14)

The (1,l)-element in (14) implies, that X is a nonsingular matrix. Consider the following block partition (corresponding to the partition of Ec1 in (12)) of X and its inverse Y := X-l:

x. y.E p x x n x . a
a7

(15)

Due to (14), together with the usual perturbation argument for matrix inequalities, there exists a solution X to (13)) (14) such that the sub-matrices X3, Y are non-singular. Therefore also the matri3 ces II1, I I 2 with

KE :

E~(~)=AKc(~)+BKY(~)

~ ( ~ ) = C K C ' ( ~ ) + D KC ( t()~ ) Rnx Y E,


(10) parametrized by A K , B K , C K ,DK the closed loop system is given by are non-singular. Furthermore XlI1 = I I 2 holds true. Since lI1 is non-singular, a non-singular congruence transformation

ECli(4 = AClW + BClW(t) (11) <(t) lR2n,, E 4 4 = CClS(t),

IITEzXII1 = IITXTEciII1 2 0 (17) XTBci C:] -TI 0 \kn,<O Ccr 0 -TI with \knl := diag(II1, I,I ) (18)
'Here E+ denotes any generalized inverse with the prop erty EE+E = E.

2950

x1 :=

Y 1

:=

RET+ETLWy, R > 0 , S E + E ~ W ~ , s>o,

(9)

of (13), (14) is possible. The matrix inequality (18) together with the linearizing changes of variables

D K :=D K Cjy : = C K Y & D K C ~ Y ~


BK :=X?BK + X T B ~ D K

>

(19)

The later equations constitute coupling conditions for (22), (23). With Esvd = UEVT being a singular value decomposition of E we may write (22) equivalently as

AK := x T ( A f B ~ D K C ~ ) X ? A K G + Y~+

---r T x l v TV-TYlUT -

+ rTx2vTV-TYsUT
=:x2

=I.

=:XI

=:Y1

=:Y3

+X ? B K C ~ ~ ~ - I - X ~ B ~ C K ~
leads to (8). Inequality (17) becomes

(25)
For X 1 , Y we have ETXl = X T E and EY1 = 1 YFET due to (9). Equivalently these equations can be written as

Esvdx1 T or equivalently (see also Remark 2)

= ZyEsvd,

y1 T EsvdTl = -T Esvd-

(26)

From (26) we conclude the following structure

[t iT] [

R E+ ET 0 [ET+ S ] 0 E ]

'

O1

(21)
of XI, 1 (with block partition corresponding to 7 (2)). Here x11, y 1 1 are positive definite matrices due to the expression for X I , YI in (9). Now we fix the structure of f ? 2 , F3 to be the same as the one of w , in (27), i.e. 1

with R > 0, S > 0. The strict inequality in (9) finally can be ensured by means of the degrees of freedom in R, S (see Remark 2). Suficiency: We have to show that the congruence transformation (17), (18) leading to the LMI conditions (8), (9) can be reversed, i.e. we have to establish the validity of XIIl = IIz with non-singular matrices II1, IIz.a.9 in (16). In terms of the block matrices of X , Y in (15) this requirement breaks down to fulfill the equations

X1Y1+ X2Y3 = I
x3Y1 f x4Y3 = 0

(22) (23)

with non-singular matrices X3, Y3. Additional the matrices Xi, yi have to suffice the symmetry constraints imposed by (13), i.e. we have

ETX2 = X Z E , EY2 = Y:ET, ETX4 = XTE. (24)

Now the existence of non-singular matrices x21, Y31 such that (25) holds for the (1,l)-block entries can be shown as in the standard case [9]. It is obvious that for any non-singular F 3 4 the remaining undetermined matrices in (28) can be chosen that (25) holds. With (28) also (22) holds true and Y is given as Y = V T F 3 V T . 3 3 Analogously to the construction of a non-singular solution Y to (22) it is possible to establish (23) 3 with non-singular X3. The important point is to

2951

use x l in the parameterization of x (due to the 2 3 symmetry constraint in (24)). With non-singular matrices X3, Y the equations (19) can be solved 3 for D K , CK,B K , AK (in this order). Since II1 is non-singular the congruence transformation (17), (18) can be reversed and clearly X = I1211T1 solves the generalized bounded real lemma inequalities 0 for the closed loop system.

control problem but also can be applied to the catalogue of problems given in ([9]) when additional algebraic equations in the problem description are taken into account.

References
[l] L. Dai. Singular Control Systems, volume 118 of Lecture Notes in Control and Information Sciences. Springer, Berlin, 1989.
[2] P. Gahinet, A. Nemirowskii, A.J. Laub, and M. Chilali. The LMI control toolbox. In Proc. 33rd I E E E Conf. Decision Contr., pages 2038-2041, Lake Buena Vista, FL, 1994.

4.

Conceptual Controller Computation

The controller computation consists of three steps:


0

Solution of the LMIs (8), (9). This is possible via effective numerical tools tailored for LMI problems arising from control theoretic problem setups [2]. Computation of non-singular matrices X 3 , Y 3 such that (22), (23) hold together with the coupling condition ETX2 = XFE. This is a essentially a factorization problem on the range of E and may be solved as outlined in the sufficiency part of Theorem 3.1. It is always possible provided (8), (9) have a solution. Solution of the linear equations (19) for the CK controller matrices D K , BK, AK.

[3] A. Kumar and P. Daoutidis. Control of nonlinear differential algebraic equation systems. N A T O Advanced Study Institute on Nonlinear Model Based Process Control, Preprint, 1997.

[4]F. L. Lewis. A tutorial on the geometric analysis of linear time-invariant implicit systems. Automatica, 28( 1):119-137, 1992.

[5] D. J. Luenberger. Nonlinear descriptor systems. J. Economics Dynamics and Control, 1:219-242, 1979.

[SI I. Masubuchi, Y. Kamitane, A. Ohara, and


N. Suda. H , control for descriptor systems: A matrix inequalities approach. Automatica, 33(4):669-673, 1997.

5.

Conclusion

We considered H , control problem for DAE systems which are allowed to be of high index and even can be non-regular. Based on the generalized bounded real lemma for DAE systems we provided necessary and sufficient strict LMI conditions for the solution of the H , control problem for DAE systems. These conditions are constructive: only basic linear algebra operations are required to compute the controller parameterization from the solution of the synthesis LMIs. In contrast to a previous solution [8] to the H , control problem for DAE systems we do not require an explicit decomposition of the problem equations into an algebraic and a differential part. Therefore it is possible to include the standard problem (without algebraic equation) without any additional notational overhead. The presented approach is not limited to the Hm

[7] A. Rehm and F. Allgower. H , control of descriptor systems with high index. In IFAC World Congess, Beijing, July 5-9,volume D, pages 31-37, 1999. [8] A. Rehm and F. Allgower. An LMI approach towards H , control of descriptor systems. In IFAC International Symposium on Advanced Control of Chemical Processes (ADCHEM ZOOO), June 14-16, Pisa, Italy, volume 1, pages 57-62, 2000. [9] C. W. Scherer, P. Gahinet, and M. Chilali. Multiobjective output - feedback control via LMI optimization. IEEE Trans. Automat. Contr., AC-42(7):896-911, 1997.
[lo] R. Schupphaus. Regelungstechnische Analyse und Synthese von Mehrkorpersystemen in Deskriptorform. Fortschr.-Ber. VDI Reihe 8 Nr. 478, VDI Verlag, Dusseldorf, 1995.

2952

Das könnte Ihnen auch gefallen