You are on page 1of 6

Proceedingsof the 34th Conference on Decision & Control New Orleans, LA December 1995

FA15 9:10

Separation Structure of Output Feedback Controllers and Its Application to Low-order Controller Design
X. Ding
Department of Electrical Engineering, FH Lausitz, PF 1538 D-01958 Senf t enberg, Germany ding@etechnik.fh-lausit z.d400.de

L. Guo ITT Automative Europe GmbH D-60488 Frankfurt a.M., Germany


Abstract
It is demonstrated that an output feedback controller internally stabilizing a linear system can be considered as an observer which delivers an estimation for a state feedback law. This result can be interpreted as <general form of separation principle, based on which a scheme of designing low-order stabilizing controllers is proposed. The results are finally illustrated by an example.
1. Introduction and problem formulation

where c ( t ) E R" is the state vector and matrices A , B , C, D are known and of appropriate dimensions. It is one of the most known results of control theory in the 1960's that an H2 (also known a LQG) ops timal controller has a separation structure, i.e. the output feedback controller can be considered as consisting of a state feedback control law and a state estimator, as shown in Fig.2. In their significant work
W

U G i ..&f!.'.

Observer

......._.._.._.._.._
Figure 2: Known separation structure Figure 1: Block diagram of output feedback control Fig.1 shows the block diagram of standard output feedback control problem, in which G represents a linear and time-invariant system, u ( t ) E R p is the control signal vector; w(t) E R is the exogenous q ) input vector; ~ ( t E Rk is the output vector to be controlled; and y(t) E R" is the measured output vector. The task is to choose controller R ( s ) suitably to provide internal stability, i.e. the states of the closed-loop system go to zero from all initial values when U) = 0, and moreover to fulfil some specifications, e.g. robustness against disturbances and sensor noises. In this contribution, we focus on designing stabilizing output controllers which ensure internal stability of the closed-loop system. For this purpose, we set w(t) = 0 and consider systems of the form [3], Doyle et.al. have extended this result and shown that an H , optimal controller, in certain sense, also possesses this property. Recently, a series of studies on the separation structure of optimal controllers has been reported, see e.g. [7]. These results motivated our study presented in this contribution. We shall demonstrate that every stabilizing output feedback controller is an estimator for K c ( t ) , as it is shown in Fig.3. This property reveals that a stabilizing output controller can be in general split up into a state feedback control law ( K ) and an estimator for K z ( t ) . In this sense, we say that every stabilizing output feedback controller possesses separation structure. In our opinion, this result may not only help us to gain a new insight into synthesizing feedback control1 systems but also be useful for constructing output feedback controllers. One of possible applications is designing low-order stabilizing controllers. Since full-order controllers are often of limited use in

0-7803-2685-7/95 $4.00 0 1995 IEEE

3203

A?f-l(s)fi(s)lthe set of all (linear) observers for P z ( t ) is parameterized by

F ( s ) = P ( s ) Y ( s )- Q o ( s ) f i ( s ) , H ( s ) = P ( s ) X ( s ) QO(s)fi(s),

....................
Figure 3: Separation structure of stabilizing controllers practice when the order of augmented plants higher than that of permitted in actual controller implementation, the issure of reducing the order of output feedback controllers has attracted considerable attention in the past few years [5], [6]. In this contribution, we shall propose a scheme to construct low-order controller stabilizing controllers. The idea of this scheme is the utilization of the separation structure shown in Fig.3, following which the design of a low-order controller is reduced to the problem of constructing a low-order observer. Since there exists a large number of well-known algorithms of designing low-order observers [9]] such a scheme may simplify the procedure of designing low-order controllers.
2. Preliminaries

where Y ( s ) , ( s ) satisfy the Bezout identity corX responding the coprime factorization of G ( s ) and Qo(s) RH, is the parameterization transfer funcE tion.

3. A Stabilizing Output Controller is an Estimator for K z


In this section, we prove the claim that every stabilizing output feedback controller is an estimator for K z ( t ) as shown in Fig.3. T h e main result is summarized into the following theorem.

Theorem 2: Given system G ( s )with double coprime factorization G ( S ) = N ( s ) M - ~ ( s ) &-1(s)fi(s), = then the following equalities hold
U(.)

= R(s)y(s) =
- QCliT)-'(X

-(Y

The basis of our study is the well-known parameterization of linear stabilizing controllers (the Youla parameterization) [4]l [ll]and the parameterization of linear observers [l]. The mathematical tool used for our purpose is the factorization technique [4], [ll].It is well known that a transfer function can be factorized as follows:

(PY with Q o ( s ) = - X ( s ) - QC(s)

+ Q C k ) y ( s )= QOfi)u(s) + ( P X + Q0G)y(s)
T(S)

where P ( s ) = K ( s 1 - Ak)-'B E RH, and matrix K is defined by factorizing G(s).

Remarks:
0

According to Lemma 1, the linear system


~ ( s )=

G(s) = C(s1- A)-lB = k-l(s)fi(s),

+ D = N(s)M-l(s)

( P ( s ) Y ( s ) &o(~)fi(s))u(~) (P(S)X(S) Q O ( S ) ~ ( S ) ) Y ( S ) +

where N ( s ) ,M ( s ) ,d ( s ) and fi(s) are right and left coprime RH,-matrices of G ( s ) , respectively. For this double coprime factorization there exist RH,transfer functions Y ( s ) X ( s ) ,Y ( s ) and X ( s ) satisfy, ing the Bezout identity [ll]. It is well known [4], that all R(s) stabilizing sy[ll] stem (1)-(2) are parameterized by

is an observer that delivers an estimation for K z ( s ) in the classical sense [l],[9]:


t "

lim ( K z ( t )- ~ ( t ) )0 =

(4)

for w = 0, all ~ ( t ) initial states. and


0

This theorem means


-

-(Y(s)- Qcfi(s))-'(X(s)
where Q C ( s ) E RH, transfer function.

+ QcG(s))

(3)

denotes the parameterization

Recently, Ding et. al. [l]have demonstrated a parameterization of linear observers.

every linear controllers that internally stabilizes the system under consideration can be expressed in terms of a n observer for K z ( t ) . The corresponding state feedback gain K is equal to the one adopted by the system factorization;

Lemma 1 [l]: Given system G ( s ) with double coprime factorization G ( s ) = N ( s ) M - ' ( s ) =

- every observer that estimates

K z ( t ) with K stabilizing A + B K is a stabilizing controller

3204

Proof of Theorem 2: From Lemma 1 we know that


r(s)

= ( P ( s ) Y ( s ) Qo(s)fi(s))u(s) -

+
(5)

( J ' ( s ) X ( s ) QO(s)n;r(s>)y(s) +
with P ( s ) = K ( s I - A s ) - l B for K z ( t ) . Notice that

E RH, is an estimation

P ( s )= K ( s I - Ak)-lB = M ( s ) - I
According to the Bezout identity we have

(6)

I t follows (11) that the eigenvalues of the closed-loop system are separately dependent on the ones of ( A B K ) , ( A - LC) and A,. If (A, B ) is stabilizable and ( C ,A ) is detectable, then there exist K and L so that the system states z ( t ) ,Az(t), x c ( t ) go t o zero for all initial values. Thus, the eigenvalues of the closedloop system can be assigned by choosing K , L and Q c ( s )separately. This is an improtant feature of the separation structure.

M ( s ) Y ( s )+ X ( s ) f v ( s ) = I , M ( s ) X ( s )= X ( s ) n ; r ( s )
(7) Applying identities (6) and (7) to (5) yields
r(s)

Remark: It can be shown [l]that all (linear) observers for I<z(t)can be written as

+(Xu +
r(s)

= U(.) - (XIV Y Q o I V ) u ( ~ ) - X QoAk)y(s)

+ +

Defining Qo(s) = - X ( s ) - Q c ( s ) gives

L ( y ( t ) - y(t)), $(t)= C q t ) + Du(t) (12) ~ ( s= K i ( s ) Q ~ ( s ) ( ~ ; ~ ( s ) Y ( s ) ) - fi(s)u(s)) = Kf(s) + Q~(s)(Y(s)- S ( S ) ) = Kf(s) - Qc(s)(y(s) - Y(s)) (13)

E(t) = A f ( t ) B u ( t )

~ ( s - ( Y ( s ) Q c ( ~ ) f v ( ~ ) ) u -~ ) ) (

( X ( S )+ Q~(s>&(s>)Y(s)
Substituting U(.) = - ( Y ( s ) - Q,(s)I?(s))-l(X(s) Q,(s)A?(s))y(s) into it results in
~ ( s )

(8)

4 s ) + ( X ( S )+ Q ~ ( s > ~ ~ ( s ) ) Y ( s > ( X ( s )+ Q c ( s > k ( s ) ) ~ ( s= Nslds) ) (9)


This completes the proof.

According to Theorem 2, we know that this description also represents a parameterization of all stabilizing output controllers. I t should be pointed out that this structure has been early given as an interpretation of the Youla parameterization [2], [8]. To our knowledge, however, the recognization that r ( s ) = K f ( s )- Q,(s)(y(s) - y(s)) is just an estimation for K z ( t ) is new. In our opinion, the importance of Theorem 2 and Theorem 3 lies not only on the proof of the equivalence between output feedback controllers and observer structured controllers and the separation structure, but also on the fact that they provide us with a new interpretation for constructing an output controller using the Youla-parameterization. It follows (3) that constructing an output controller means to determine matrices I<,L and transfer function Q C ( s ) . While matrix K is knownly interpreted as a state feedback control law with full information (all state veriables are measured) [2], [8], Theorem 2 and Theorem 3 reveal that the matrix L and the transfer function Qc(s) only influence the performance of the estimator for K z ( t ) . In this sense, we say that all stabilizing output controllers have separation structure.
4. Low-order Stabilizing Controllers

Remark: The above theorem reveals that for a given state feedback matrix I<, the difference among linear controllers is characterized by the selection of observers for K z ( t ) .
To verify the result given in Theorem 2 and show that the closed-loop system is internally stable and does have the separation structure, we observe the dynamics of the closed-loop system. To this end, substitute U(.) = ~ ( s into (l),which yields )

i ( t ) = (A + BI<)z(t)- B ( K z ( t ) - Ht))

(10)

Denote the state vector of estimation error system K z ( s )- r ( s ) with Ax and the state-space realization of Qc(s) with ( A c ,B,, C,, Dc). Then, we have the following theorem.

Theorem 3: Given system (1)-(2) and control law


U(.)

r(s)

( J ' ( s ) Y ( s ) Qo(s)fi(s))u(s) -

(J'(s)X(s) Q o ( s ) ~ ( s ) > ~ ( s )

with Q o ( s ) = - X ( s ) - Q,(s), the dynamics of the closed-loop system is governed by

i = ( A BK)x - B(K A = ( A - LC)AX X xc = BCCAx AcxC

+ DCC)Az

BCcxC (11)

In this section, we present a scheme t o design loworder output feedback controllers. The basic idea is the utilization of the separation structure given in the last section, which allows t o construct a (equivalent) stabilizing controller using the observer structure shown in Fig.3. The problem of designing low-order stabilizing controllers has already received much attention during

3205

last years. A number of significant approaches, from designing special low-order controllers, e.g. H sub, optimal low-order controllers [6], t o charaterizing loworder controllers of general form [5], has been reported. In general, these approaches can be divided into two schemes [5]: designing a low-order controller by reducing the order of a designed controller with the aid of model-reduction methods, for which it is necessary to reanalyze the design to check that any degradation in performance is not too significant; designing a low-order controller directly to meet the design objectives. In this section, we focus our attention on designing low-order controllers directly. We characterize the set of the low-order controllers, designed by the approach proposed here. To this end, we utilize their relationship t o full-order controllers, which, on one hand, allows one to analyze the performance of the low-order controllers directly and, on the other hand, can be used to design a low-order controller by reducing the order of a designed controller without loss of performance. We begin with considering a Luenberger type observer of the form

a low-order controller

$(t) = Gd(t) H u ( t ) L L Y ( ~ ) RP (18) E u(t) = Jqqt) V u ( t ) V y ( t ) (19)

+ +

It is well known in observer theory that the existence of low order observer for K z ( t ) as well as its dimension depends on the system structure and the desired linear combination of state variables namely the matrix K . T h e following two theorems provide sufficient conditions for the existence of observers and so stabilizing controllers of a certain order.

Theorem 4: There exist stabilizing output controllers of order (n-m) that can be expressed by (18)-(19), where n is the dimension of the state variables and m the one of the measured state variables. Theorem 5: Assume p = 1 (single input system). Then there exist stabilizing output controllers of order ( a - 1) that can be expressed by (18)-(19), where a is the observability index of the system. Remarks:
0

$ ( t )= G4(t)+ H u ( t ) L ~ y ( tE Rp (14) ) r ( t ) = Jqqt) U u ( t )+ Vy(t) (15)

The above two theorems are well known 191 in observer theory for constructing observers of order (n-m) as well as a - 1 for K z ( t ) . Thus, following Theorem 2 they also hold for constructing stabilizing controllers of order (n-m) as well as a - 1; Theorem 4 holds for all possible state feedback matrices K that stabilize A B K . Notice that (n-m) may not be the lowest possible order;

3 with r ( t ) as an estimation for K z ( t ) , / stabilizing A B K .

_< n and K
0

It is well known that the Luenberger type observer theory is a power tool to design reduced-order observers. It is therefore applied for our purpose of constructing a low-order observer for K z ( t ) that, following the separation structure shown in Fig.3, equivalently means a low-order controller. According to Luenberger observer theory [9], in order to get an estimate for K x ( t ) , matrices G, H , J , LL, U and V in (14)-( 15) have t o satisfy following conditions

In Theorem 5, the assumption p = 1 means state feedback matrix K E R1 n . It is well known that x CY - 1 5 n - m and it is often the case that a - 1 considerably less than (n-m); In order to construct an observer and so a controller of order (n-m) or ( a - l),we need to solve (16)-(17) for G, H , J , L L , U and V. To this end, there exists a number of algorithms, see e.g. [9].

G E RPxP is a stability matrix, T A - GT = L L C , H L L D = T B U == -VD, = J T + V C Ii;

(16) (17)

where T E RPxndenotes a state transformation. Notice that for p < n we have a reduced order observer. Thus, the problem of designing low-order controllers is reduced to constructing a Luenberger type observer, for which we only need to solve (16)-(17). If these equations are solvable for some ,6 < n , then we have

We now have a n approach to designing low-order stabilizing controllers. Surely, one would like t o know how to analyze and check the performance of such controllers. T h e following theorem answers this question and provides us with an important relationship between the low-order controllers designed using the scheme proposed here and the controllers constructed by using the Youla parameterization.

Theorem 6: Given systems

$ ( t )= G$(t) H u ( t ) L L Y ( ~ )RP (20) E u ( t )= Jqqt) -I-U u ( t ) + V Y ( t ) (21)

3206

with , 5 n , G, H , J , L L ,T , U and V satisfying (16)B (17) and

5 . An Example

?(t) = A s ( t ) Bu(t) L ( y ( t )- Y(t)),

$(t) = C q t ) + D u ( t ) with T L = LL and L stabilizing L


d(0) = T 2 ( 0 )

(22)
-

AC, then for

In the last two sections, we have demonstrated that every stabilizing controller is an observer for K z ( t ) and, based on it, presented a scheme to design loworder stabilizing output feedback controllers. To illustrate these results we give the following example. The system considered is an inverted pendulum. The state vector is:

u ( t ) = Jd(t) U u ( t ) V Y ( q = I/;P(t)+ V(y(t) - y ( t ) ) for all t .

(23)

It is well known [2] and also shown in the last section that the Youla parameterization can be interpreted as follows: all stabilizing controllers can be written as
Bu(t) L ( y ( t )- y(t)), (24) 2(t)= Y(t)= C q t ) + DU(t) (25) 4 s ) = K ~ ( s-) Q~(s)(Y(s) Ys) - () (26)
Comparing this system with the ones given in Theorem 6 obviously shows that a low-order controller

where r is position of base, @ angular position of pendulum and X I , 2 2 , x3 are measured output. The input variable is the input voltage U to the power amplifier which drives the motor. This is a nonlinear unstable system which, after a linearization, can be expressed as follows:

x
U(S)

= A x + bu,
= Fvrref

+ + L L Y ( ~E) Rp u ( t ) = Jd(t) + V u ( t )+ V y ( t )
&t) = G$(t) H u ( t )

(27)
(28)

+ R(s)y(s),
z3IT

(31) (32) (33)

= [zi

$2

has the same performance like a full-order controller described by

4 )= R ( s ) y ( s ) .
= - ( Y ( s ) + v&(s))-l(x(s) V k ( s ) ) y ( s ) so long as Q e ( s ) = -V,$(O) = T2(0) and L solves
T L = LL and ensures the stability of matrix A - LC. This theorem provides us with a relationship between
the set of full-order controllers, for which the parameterization transfer function Qc(s) is set t o be a constant matrix, and the set of low order controllers that are designed using the observer structure and Luenberger observer theory. From this theorem we know that for every low-order controller constructed using the scheme proposed here we can easily find a equivalent full-order controller by just defining

where the input variable U consists of the system input r r , f , the desired position of the moving base, and an output feedback control law R ( s ) y ( s ) .The numerical values of the (4x 4) system matrix A, the (4x 1) input matrix b and the pre-filter FV are given in Appendix. Using the Youla parameterization the feedback control law R(s) can be determined as follows: Choose state feedback matrix k and observation matrix L (the numerical values are given in Appendix A) so that the eigenvalues of matrix -5 A+bk and A-LC are - 3 , -4, -6 and -6, -7, -8, -9, respectively. Set
Q C ( s ) [-3.32 =

9.98

5.741

(34)

= -v,L = T * - ~ L *

(29)

Calculate R ( s ) which gives


- Qc*(s))-'(X(s)

This property can be used for analyzing low-order controllers suggested here. Another application of Theorem 6 is t o construct a low-order controller for a designed controller, for which the parameterization transfer function QC(s) is a constant matrix. In this case, we have to solve equations (16)-(17) for given matrices I / , L and V = -Qe. If they are solvable for some p < n , then we have a low-order controller of the form (20)-(21). Notice that in this case the reduced-order controller has the same performance like the full-order controller.

+ Qc&f(s))

-3.32~~-97.89~~-1076.61s2-5231.47s-9473.32 s4+471.22s3f31200.63sz~7668 17s+16874.03 9.89s +256.8s f 2 4 4 7 . 9 4 ~ +10240.4s+15847.84

s4~71.22s3+1200.63s~+~668.17s+16874.03 5.74s +165.04s3+1771.78s +8418.78s+14934.42 s4+71.2283+1200.63s2+7668.17s+16874.03

To illustrate und verify the results presented in the last sections, a low-order controller, which has the sa-

me performance like the one given above but is presented in form of a reduced-order observer, will be designed.

3207

For our purpose, the same state feedback matrix le is used. Notice that we have a system with three outputs and one input and the system order is four. According the well-known Luenberger observer theory, the minimal order of the observer for Rz is one. To design such a n observer and moreover the corresponding low-order output feedback controller, the following steps have been carried out:
e

[4]B.A. Francis, A Course in H , Control Theory,


Berlin
-

New York: Springer-Verlag, 1987.

[5] D.-W. Gu, B.W. Choi and 1. Postlethwaite, "Loworder stabilizing controllers" , IEEE Trans. Autom. Contr., AC-38, pp.1713-1717, 1993. [6] C.S. Hsu, X. Yu, H.-H. Yeh and S.S. Banda, " H , compensator design with minimal order observers", IEEE Trans. Autom. Contr., AC-39, pp.1679-1681, 1994. [7] P.A. Iglesias and D. Mustafa, "State-space solution of the discret-time minimum entropy control problem via separation", IEEE Trans. Autom. Contr., AC-38, pp.1525-1530, 1993. [8] J. M. Maciejowski, Multivariable Feedback Design, Addison-Wesley Publishing Company, 1989. [9] J. O'Reilly, Observer for Linear Systems, London, Academic Press Inc., 1983.

for given R and V = -Qc, solve equations (16)(17) for G, N , , LL and T ; J Substitute G , H , J , LL and V into (20)-(21), which gives

$ ( t ) = -9#(t)

- 41.22r(t)

[-1.67 42.68 7.23]y(t) (35) r ( t ) = # ( t )+ [3.32 - 9.98 - 5.74]y(t) Set u ( t ) = Fvr,,f(t)

+ r(t)

To compare the above two controllers, simulation has been carried out. The results show that the controllers have the same performance. Thus, we now have a controller of minimal order and it offers the same performance like a n full-order output feedback controller.
6. Conclusions

[lo] A. Saberi and P. Sannuti, "Observer design for loop transfer recovery and for uncertain dynamical systems", IEEE Trans. Autom. Contr., AC-35,
pp.878-897, 1990.
[ll] M. Vidyasagar, Control System Sythesis: A Factorization Approach, Cambridge, MA, M.I.T. Press, 1985. Appendix

The proof that a stabilizing feedback controller is in fact an observer that estimates a state feedback control law reveals that the well known separation principle is true for all output feedback controllers internally stabilizing linear systems. On this basis, we have presented a scheme to construct low order controllers using reduced-order observer theory. These results have been illustrated using a n example of an inverted pendulum control system.

T h e numerical values of the system matrices of the inverted pendulum example:

A=

0 0 0 0 0 -0.14
0
21.17

-1.40

0
-0.55 6.24 47.82
- 4.36

0.0020
-0.29 l l

b = [ 0 0 -4.19 k = [3.13 - 3.46

1'
- 0.751

References
[l] X . Ding, L. Guo and P.M. Frank, "Parameterization of linear observers and its application to observer design, "IEEE Trans. Autom. Contr., 1994. [a] J.C. Doyle, Matrix Interpolation Theory and Optimal Control, Ph.D. Diss., Univ. Calif., 1984 [3] J.C. Doyle, K. Glover, P.P. Khargonekar and B.A. Francis, "State-space solutions to standard Hz and H , control systems," IEEE Trans. Autom. Contr., AC-34, pp.831-847, 1989.

8.00 -0.14 L= -0.07 -1.13 FV= -0.565

-0.0037 15.2654 0.013 75.92

-1.46 3.66 5.91 33.47

3208

You might also like