Disturbance attenuation for a class of distributed parameter systems
-
Upload
alberto-de-santis -
Category
Documents
-
view
213 -
download
0
Transcript of Disturbance attenuation for a class of distributed parameter systems
INTERNATIONAL JOURNAL OF ROBUST AND NONLINEAR CONTROL, VOL. 7, 759—775 (1997)
DISTURBANCE ATTENUATION FOR A CLASS OFDISTRIBUTED PARAMETER SYSTEMS
ALBERTO DE SANTIS* AND LEONARDO LANARI
Dipartimento di Informatica e Sistemistica, Universita di Roma ‘La Sapienza’, Via Eudossiana 18, 00184 Roma, Italy
SUMMARY
This paper deals with the problem of robust stabilization and disturbance attenuation via measuredfeedback, for a class of dissipative collocated distributed systems with disturbances affecting both the inputand the measured output. The proposed solution is based on a direct L
2-gain characterization which avoids
the usual Riccati equation argument. For this purpose only strong stabilizability of the semigroup isrequired. A flexible slewing link is chosen as an application example. ( 1997 by John Wiley & Sons, Ltd.
Int. J. Robust Nonlinear Control, 7, 759—775 (1997)No. of Figures: 1 No. of Tables: 0 No. of References: 22
Key words: H= control; distributed parameter system; strong stability; flexible beam
1. INTRODUCTION
In the last decade much research effort has been devoted to the study of linear H= control.The problem consists of finding a feedback control law ensuring closed-loop stability anddisturbance attenuation. This last requirement is achieved by preventing the energy transferbetween the undesired inputs (disturbances) and suitable penalty variables from exceeding anassigned bound. This is also known as the suboptimal H= control problem, which consists ofrendering the L
2-gain from the disturbances to the penalty variables less than a given value c2.1
Penalty variables definition usually involves the energy of either regulation or tracking errorsof the variables (state or output) to be controlled, as well as the energy related to the control inputs.
Since the relevant literature is quite vast, we mention only a few significant contributionsreferring to the references therein. Initially this problem has been studied in a frequency domainsetting.2,3 State-space formulations were then presented in Reference 4, based on non-standardsteady-state Riccati equation (SSRE) arguments, and in Reference 5 where differential gametheory was used. In the infinite-dimensional setting, see e.g., Reference 6 for a survey, thestate-space approach for the full information case (i.e., full state available for feedback) wasrecently introduced in References 7 and 8, thus extending known results from the finite-dimen-sional setting.5,9 In References 7 and 8 algebraic operator Riccati equations are encountered asopposite to the integral version dealt with in Reference 10, where a very general problem is
This paper was recommended for publication by editor M. J. Grimble
*Correspondence to: A. de Santis, Dipartimento di Informatica e Sistemistica, Universita di Roma ‘La Sapienza’, ViaEndossiana 18, 00184 Roma, Italy
CCC 1049-8923/97/080759—17 $17.50 Received 8 August 1994( 1997 by John Wiley & Sons, Ltd. Revised 4 December 1995
studied, including also time-varying systems. It is worthwhile noting that the exponentialstabilizability of the semigroup is always assumed in these works.
Robust stabilization of distributed positive real systems is addressed in Reference 11 based ona transfer function coprime factorization. In Reference 12 such a factorization is expressed interms of the strong stabilizing solution of the SSRE.
For practical purposes a great deal of work has been devoted to designing output feedbackcompensators. In References 13 and 14 the existence of a static output controller is studied for thefinite dimensional case, while in Reference 15 the measurement feedback H= control problem issolved for the class of Pritchard—Salomon distributed systems.
In this paper we consider a class of infinite-dimensional dissipative systems in the case ofcollocated input/output, with additive disturbances corrupting both the input and the measuredoutput. This class can be adopted in modelling flexible structures with collocated sen-sors/actuators.16 We show that an output static control law can be designed in order toasymptotically stabilize the system and attenuate the effect of the disturbances on some well-defined penalty variable. Moreover, only strong stabilizability is required and the L
2-gain
characterization is achieved without the SSRE argument.In Section 2 basic definitions and assumptions are given and the problem formulation is stated.
The main result is presented in Section 3, and some interesting particular cases are discussed. Theflexible slewing link is chosen as an application example in Section 4.
2. PRELIMINARIES AND PROBLEM FORMULATION
Consider the following infinite-dimensional linear system on the separable Hilbert space H
xR (t)"Ax(t)#Bu(t) (1a)
y(t)"B*x(t) (1b)
with u (t)3Hu, H
ubeing another separable Hilbert space, and y(t)3H
y"H
u, under the follow-
ing assumptions:
(P1) A is a closed densely operator on D (A)LH generating a strongly continuous semi-group (or C
0-semigroup) S (t). Moreover the resolvent R(j,A)"(jI!A)1, j'0, is
compact, i.e., A has a pure point spectrum.(P2) A and A* are dissipative operators i.e.,17
[Ax, x]#[x,Ax])0 x3D (A)
[A*z, z]#[z, A*z])0 z3D (A*)
(P3) B :HuPH is a linear bounded operator and the pair (A,B) is controllable i.e.,17
B*S* (t)v"0 Nv"0, v3H, t*0
Several systems arising in applications belong to this class. Appealing examples are continuummodels of flexible space structures with collocated sensors/actuators.16
The mild solution17 of (1) is
x (t)"S (t)x (0)#Pt
0
S(t!q)Bu(q) dq, x (0)3H (2)
y(t)"B*x (t) (3)
760 A. DE SANTIS AND L. LANARI
Int. J. Robust Nonlinear Control, 7, 759—775 (1997) ( 1997 by John Wiley & Sons, Ltd.
where S (t)x(0)"xh(t) is the solution of the homogeneous version of (1). The behaviour of x
h(t), as
t goes to infinity, gives rise to different types of stability listed below:
(i) the semigroup S (t) is exponentially stable if
ES(t)E)Me~at, M, a'0
(ii) it is strongly stable if, for every z3H,
limt?=
ES (t)zE"0
(iii) it is weakly stable if, for every z, v3H,
limt?=
[S (t)z, v]"0
For the links between these different kinds of stability the reader is referred to Reference 17.Prior to any control problem is the feedback stabilization issue. It consists of finding a control
law us(t)"!Px(t), with P linear bounded, such that the closed-loop operator (A—BP) generates
a stable semigroup according to any one of (i)—(iii). Note that a semigroup cannot be, in general,exponentially stabilized when the input operator B is compact,18 as it often occurs in practicalapplications where B is even finite-dimensional.16 Therefore, weak and strong stability representa more general framework for distributed parameter systems control design. Moreover, sucha control law u
s(t) requires full-state measurement, whereas only the measured output is available
in most situations.For the considered class of systems, it can be easily verified that the output feedback
u(t)"!ky (t), for any k'0, strongly stabilizes system (1). This is a direct consequence of theresults in Reference 19. Indeed by assumption (P2), S (t) is a contraction semigroup,17 i.e.,ES(t)E)1, t*0. Because of (P3), it is also weakly stabilizable by the feedback !kB*x(t),k'0, i.e., the closed-loop operator A
k"(A!kBB*) generates a weakly stable semigroup
Sk(t). Since the resolvent of A is compact, so is the resolvent of A
k, and therefore S
k(t) is strongly
stable.17Besides stabilizability, another interesting concept for control design purposes is the system’s
L2-gain, i.e., the ratio between the L
2-norms of both the forced system’s response and the input.
For these quantities to be well defined let u3¼u"¸
2[[0,R),H
u]. If S (t) is exponentially stable,
then the solution of (1) belongs to ¼x"¸
2[[0,R),H] and consequently
y3¼y"¸
2[[0,R),H
y].17 Note that in this case the whole system’s response y belongs to ¼
y.
Then, the forced component yf
of the output y, obtained from equations (2), (3) by settingx(0)"0, is given by
yf(t)"B* P
t
0
S(t!q)Bu(q) dq
and establishes a linear bounded input—output map
G :¼uP¼
y, y
f"Gu
with
EGE" supEu E
L2/1EGuE
L2*0,=)with EuE2
L2*0,=)"P
=
0
[u (t), u (t)]Hudt
The quantity EGE is defined as the L2-gain of system (1).8
DISTURBANCE ATTENUATION 761
( 1997 by John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 7, 759—775 (1997)
If the semigroup S (t) is only strongly stable, the state evolution (2) no longer belongs to ¼x, so
that the quantity EGE cannot be generally defined. However, Lemma 2 in the next section willshow that for the considered class of systems the input—output mapping G, involving theclosed-loop semigroup S
k(t), is bounded. Therefore the L
2-gain can still be defined without the
exponential stability property.Having introduced the concepts of stabilization and L
2-gain, we can now state the problem to
be solved. We want to achieve closed-loop stability via output feedback and attenuate theinfluence of an exogenous input d(t)"(w1 (t)
w2 (t)) on a suitably defined penalty variable z(t). More
precisely, given the system
xR (t)"Ax(t)#Bu(t)#Bw1(t) (4a)
y(t)"B*x(t)#w2(t) (4b)
satisfying (P1)—(P3), with w13¼
uand w
23¼
y, find a static output control law
u(t)"!ky (t)
such that, given the penalty variable
z (t)"Ay (t)
u (t)Bthe closed-loop system is strongly stable and the L
2-gain from the signal d to z is less than a
given value c'0, i.e., EzE2L2*0,=)
)c2EdE2L2*0,=)
. This results in designing a finite energycontrol u(t) to minimize the energy transfer from the disturbances to the output regulationerror.
The solution can be achieved through the direct L2-gain characterization as shown in the next
section.
3. MAIN RESULT
In this section we find an upper bound for the L2-gain from the disturbances to the penalty
variable for the given class of systems, via direct computation. We first state two lemmas to beused in the next theorem.
Lemma 1
Consider the system
xR (t)"(A!kBB* )x (t)#Bu (t) (5)
with u3¼uand the pair (A,B) satisfying (P1)—(P3) so that (A!kBB*) generates a strongly
stable contraction semigroup Sk(t). For some finite ¹'0 choose a sequence
MunN3C1[[0, ¹ ],H
u] (i.e., the space of functions strongly continuously differentiable in (0, ¹ ),
with derivative continuous in [0, ¹ ]) converging to u in the ¼u-norm. Let
v(t)"Pt
0
Sk(t!q)Bu(q) dq, t3[0, ¹ ]
762 A. DE SANTIS AND L. LANARI
Int. J. Robust Nonlinear Control, 7, 759—775 (1997) ( 1997 by John Wiley & Sons, Ltd.
be the mild solution of (5) with x (0)"0, and vn(t) be the analogous quantity with u"u
n. Then
limn?=
E v(t)!vn(t)E
L2*0,T +"0 (6a)
limn?=
Ev(t)!vn(t)EH"0 (6b)
limn?=
D[u, v]L2*0,T +
![un, v
n]L2*0,T +
D"0 (6c)
Proof. See Appendix.
We can now show that the input—output map of the closed-loop system has a well-definedL
2-gain.
Lemma 2
Consider the system
xR (t)"(A!kBB* )x (t)#Bu(t)(7)
y(t)"B*x (t), x (0)"0
with u3¼u, k'0, and the pair (A,B) satisfying (P1)—(P3); then the system’s forced response
y belongs to ¼y"¸
2[[0, R), H
y].
Proof. Choose a sequence MunN3C1[[0, ¹ ],H
u] converging to u in the ¼
u-norm. Define
v(t)"Pt
0
Sk(t!q)Bu(q) dq, v
n(t)"P
t
0
Sk(t!q)Bu
n(q) dq (8)
By Theorem 4.8.2 in Reference 17, function vn(t) is strongly continuously differentiable in (0, ¹ ) so
that we can write, using also property (P2),
d
dtEv
n(t)E2H"[A
kvn(t), v
n(t)]H#[v
n(t),A
kvn(t)]H#2[Bu
n(t), v
n(t)]H
)!2k[BB*vn(t), v
n(t)]H#2[Bu
n(t), v
n(t)]H (9)
Integrating equation (9) from 0 to ¹, one obtains
Evn(¹ )E2H)2[Bu
n(t), v
n(t)]
L2*0,T +!2kEB*v
n(t)E2
L2*0,T +
and therefore
EB*vn(t)E2
L2*0,T +)
1
k[Bu
n(t), v
n(t)]
L2*0,T +)
1
kEu
n(t)E
L2*0,T +EB*v
n(t)E
L2*0,T +
from which
EB*vn(t)E
L2*0,T +)
1
kEu
n(t)E
L2*0,T +)
1
kEu
n(t)E
L2*0,=)
implying that B*vn3¼
y. Now note that
EB*v(t)EL2*0,T +
)EB*(v (t)!vn(t) )E
L2*0,T +#
1
kEu
n(t)E
L2*0,=)
DISTURBANCE ATTENUATION 763
( 1997 by John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 7, 759—775 (1997)
where the first r.h.s. term tends to zero as nPR by Lemma 1, while the second one is in ¼ufor
any n. Therefore, we can now let ¹ go to infinity which leads to B*v3¼y. K
Lemma 2 states that system (7) has a well defined L2-gain.
Remark 1
As argued in Section 2, the state trajectories of system (7) do not belong to ¼x
for any x (0),since this would require
P=
0
ESk(t)x (0)E2 dt(R, ∀x (0)3H
which, according to Datko’s theorem,20 is true if and only if Sk(t) is exponentially stable.
However, it can be shown that not only the systems’s forced response but the whole outputtrajectory belongs to ¼
yfor any x (0)3H. We indeed have
Ey (t)E2L2*0,=)
)2EB*Sk(t)x (0)E2
L2*0,=)#2EB*v(t)E2
L2*0,=)
and only the boundedness of the first term in the r.h.s. need be shown. Owing to the denseness ofD(A
k), for any x (0)3H we can find a sequence Mx
nN3D (A
k) converging to x (0). Now, being
Sk(t)x
ndifferentiable for any n, we can use the same argument as in Lemma 2, obtaining
d
dtES
k(t)x
nE2H)!2kEB*S
k(t)x
nE2H
from which
ESk(t)x
nE2H!Ex
nE2H)!2kEB*S
k(t)x
nE2L2*0,T+
By the strong stability of Sk(t) we eventually obtain that
EB*Sk(t)x
nE2L2*0,=)
)ExnE2H
A limiting argument can now be used to substitute x (0) to xnin the above inequality.
We now state the following theorem.
Theorem 1
Given the system
xR (t)"Ax (t)#Bu(t)#Bw1
(t) (10a)
y(t)"B*x (t)#w2(t) (10b)
satisfying (P1)—(P3) with w13¼
u"¸
2[[0, R), H
u] and w
23¼
y"¸
2[[0, R), H
y], the output
feedback law
u (t)"!ky(t), k'0
strongly stabilizes the system and the L2-gain from d (t)"(w1 (t)
w2 (t)) to z (t)"(y(t)
u(t)) is bounded by
c2"(1#k2)maxG1,1
k2H (11)
764 A. DE SANTIS AND L. LANARI
Int. J. Robust Nonlinear Control, 7, 759—775 (1997) ( 1997 by John Wiley & Sons, Ltd.
Proof. The closed-loop system
xR (t)"(A!kBB*)x (t)#B (w1(t)!kw
2(t))
is strongly stable, as already discussed in Section 2. For the L2-gain characterization, we need to
find an upper bound for
EzE2L2*0,=)
"EyE2L2*0,=)
#EuE2L2*0,=)
"(1#k2)EyE2L2*0,=)
(12)
Define the intermediate variables v1(t) and v
2(t) as follows
v1(t)"P
t
0
Sk(t!q)Bw
1(q) dq, v
2(t)"P
t
0
Sk(t!q)Bw
2(q) dq
The closed-loop system output, with initial conditions x (0)"0, is given by
y (t)"B* Pt
0
Sk(t!q)B(w
1(q)!kw
2(q)) dq#w
2(t)
therefore
y(t)"B* (v1(t)!kv
2(t))#w
2(t)
and
EyE2L2*0,=)
"EB* (v1!kv
2)E2
L2*0,=)#Ew
2E2L2*0,=)
#2[w2,B* (v
1!kv
2)]
L2*0,=)(13)
where all the terms are well defined by Lemma 2.Consider a sequence of compact time intervals MI
jN"M[t
j~1, t
j]N such that Z=
j/1"[0, R) and
let sIj(t) be the indicating function of I
j(i.e., s
Ij(t)"1 if t3I
j, and s
Ij(t)"0 elsewhere), so that we
can consider the restrictions
wj1(t)"w
1(t)s
Ij(t), wj
2(t)"w
2(t)s
Ij(t)
of w1(t) and w
2(t) to the generic interval I
j. The function
wj (t)"[w1(t)!kw
2(t)]s
Ij(t)"w(t)s
Ij(t)"wj
1(t)!kwj
2(t)
of course belongs to ¼u
for any j. Moreover, with
v(t)"v1(t)!kv
2(t)"P
t
0
Sk(t!q)B(w
1(q)!kw
2(q)) dq
we have
vj (t)"v (t)sIj(t)"P
t
tj~1
Sk(t!q)B(w
1(q)!kw
2(q)) dq
"Pt
tj~1
Sk(t!q)Bwj (q) dq, t3I
j
On any Ijwe can find a sequence Mwj
n(t)N3C1 (I
j) converging to wj(t) in the ¼
u-norm, so that the
function
vjn(t)"P
t
tj~1
Sk(t!q)Bwj
n(q) dq
DISTURBANCE ATTENUATION 765
( 1997 by John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 7, 759—775 (1997)
is strongly continuously differentiable for any t3Ij, obtaining
vR jn(t)"A
kvjn(t)#Bwj
n(t), t3I
j
Let us compute the quantity
d
dtEvj
n(t)E2H"[A
kvjn(t)#Bwj
n(t), vj
n(t)]H#[vj
n(t),A
kvjn(t)#Bwj
n(t)]H
)2[wjn(t),B*vj
n(t)]H!2kEB*vj
n(t)E2H , t3I
j(14)
where we used assumption (P2). Intergrating (14) from tj~1
to tjwe obtain
Evjn(tj)E2H)2[wj
n(t),B*vj
n(t)]
L2*tj~1, tj+!2kEB*vj
n(t)E2
L2*tj~1, tj+and therefore
Evjn(tj)E2H#2kEB*vj
n(t)]2
L2*tj~1, tj+)2[wj
n(t) ,B*vj
n(t)]
L2*tj~1, tj+(15)
Now, by Lemma 1, as nPR both sides in equation (15) converge and it eventually results that
Evj (tj)E2H)2[wj (t),B*vj (t)]
L2*tj~1, tj+!2kEB*vj(t)E2
L2*tj~1, tj+
Now, for any N, we can add up the terms of each interval Ij, j)N, obtaining
N+j/1
Evj (tj)E2H)2[w(t),B*v (t)]
L2 *0, tN+!2kEB*v(t) E2
L2*0, tN+(16)
Here, we used the fact that
Evj(t)E2L2*tj~1, tj+
#Evj~1(t)E2L2*tj~2, tj~1+
"Ev(t)sIjXIj~1
E2L2*tj~2, tj+
since vl(t) and vm (t) are orthogonal in the ¸2-sense for lOm. By Lemma 2 both terms in the r.h.s.
of (16) are bounded as tNPR, so that the l.h.s is bounded for any N; we finally have
2kEB*v(t)E2L2*0,=)
)2[w (t), B*v (t)]L2*0,=)
(17)
In order to find an explicit bound for EyE2L2*0,=)
, we now expand inequality (17) to obtain
kMEB*v1E2L2*0,=)
#k2EB*v2E2L2*0,=)
!2k[B*v2, w
2]L2*0,=)
!2k[B*v1,B*v
2]L2*0,=)
#2[B*v1, w
2]L2*0,=)
N
)!k EB*v1E2L2*0,=)
#2k2[B*v1,B*v
2]L2*0,=)
!k3EB*v2E2L2*0,=)
#2[w1,B*v
1]L2*0,=)
!2k[w1,B*v
2]L2*0,=)
(18)
which used in (13) yields
EyE2L2*0,=)
)!EB*v1E2L2*0,=)
#2k[B*v1,B*v
2]L2*0,=)
!k2EB*v2E2L2*0,=)
#
2
k[w
1,B*v
1]L2*0,=)
!2[w1,B*v
2]L2*0,=)
#Ew2E2L2*0,=)
)!EB* (v1!kv
2)E2
L2*0,=)#
2
k[w
1,B* (v
1!kv
2)]
L2*0,=)#Ew
2E2L2*0,=)
766 A. DE SANTIS AND L. LANARI
Int. J. Robust Nonlinear Control, 7, 759—775 (1997) ( 1997 by John Wiley & Sons, Ltd.
)!KKB* (v1!kv
2)!
1
kw1 KK
2
L2*0,=)
#
1
k2Ew
1E2L2*0,=)
#Ew2E2L2*0,=)
)
1
k2Ew
1E2L2*0,=)
#Ew2E2L2*0,=)
)maxG1,1
k2H EdE2L2*0,=)
Therefore, by substituting in (12), one obtains
EzE2L2*0,=)
)(1#k2) maxG1,1
k2H EdE2L2*0,=)
Hence the L2-gain is bounded above by (1#k2)maxM1, 1
k2N. K
Note that the L2-gain upper bound attains its minimum at c*2"2 for k"1.
We can derive some interesting particular cases from the previous analysis.
Corollary 1
In the case w1"w
2"w, and of course d"w, the L
2-gain can be bounded as
EzE2L2*0,=)
EdE2L2*0,=)
)(1#k2) A1#D1!k D
k B2
(19)
Proof. From (13) we have
EyE2L2*0,=)
"(1!k)2EB*vE2L2*0,=)
#EwE2L2*0,=)
#2(1!k) [w,B*v]L2*0,=)
(20)
with obvious meaning of the terms involved. From (17), using the Scwartz inequality, we obtain
EB*vE2L2*0,=)
)
1
k[w, B*v]
L2*0,=))
1
kEwE
L2*0,=)EB*vE
L2*0,=)
from which
EB*vE2L2*0,=)
)
1
kEwE2
L2*0,=)(21)
By using (21) in (20), after a little computation, one has
EzE2L2*0,=)
EdE2L2*0,=)
)(1#k2) A1#D1!k D
k B2
which attains its minimum at c*2"2 for k"1. K
Remark 2
Corollary 1 has an interesting interpretation. Consider the closed-loop system withu(t)"!ky (t)
xR (t)"(A!kBB*)x (t)#(1!k)Bw(t)
DISTURBANCE ATTENUATION 767
( 1997 by John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 7, 759—775 (1997)
Setting k"1 the output feedback exactly cancels the effect of the disturbances in the stateequation. Therefore the effect of the disturbances on the output is given only by the directfeedthrough between the disturbances and the output.
Corollary 2
In the case w1"0 (disturbances affecting only the output) the L
2-gain can be bounded as
EzE2L2*0,=)
EdE2L2*0,=)
)(1#k2) (22)
(with d"w"w2).
Proof. From (13) we have
EyE2L2*0,=)
"k2EB*vE2L2*0,=)
#EwE2L2*0,=)
!2k[w,B*v]L2*0,=)
By setting w1"v
1"0 in (18) we obtain
0)2k2[w,B*v]L2*0,=)
!2k3EB*vE2L2*0,=)
0)k (2k[w,B*v]L2*0,=)
!k2EB*vE2L2*0,=)
)!k3 EB*vE2L2*0,=)
and therefore
k2EB*vE2L2*0,=)
!2k[w, B*v]L2*0,=)
)!k2EB*vE2L2*0,=)
)0
which implies
EyE2L2*0,=)
)EwE2L2*0,=)
, EzE2L2*0,=)
)(1#k2)EdE2L2*0,=) K
Remark 3
Corollary 2 states that the minimum L2-gain upper bound achievable with the above
computations is c*2"1 attainable for kP0. In other words the output feedback used to stabilizethe system cannot attenuate the effect of the direct feedthrough from the disturbances to theoutput. A trade-off between two different objectives (stabilization and disturbance attenuation)has to be done.
Corollary 3
In the case w2"0 (disturbances affecting only the input) the L
2-gain can be bounded as
EzE2L2*0,=)
EdE2L2*0,=)
)A1#1
k2B (23)
(with d"w"w1).
Proof. We have
EyE2L2*0,=)
"EB*vE2L2*0,=)
768 A. DE SANTIS AND L. LANARI
Int. J. Robust Nonlinear Control, 7, 759—775 (1997) ( 1997 by John Wiley & Sons, Ltd.
Arguing as in Corollary 1, one obtains the same inequality as (21) which implies
EyE2L2*0,=)
)
1
k2EwE2
L2*0,=), EzE2
L2*0,=))A1#
1
k2B EdE2L2*0,=)
K
Remark 4
The minimum L2-gain upper bound achievable is c*2"1 for kPR. In other words since the
stabilizing controller does not depend upon the disturbances directly, the more the system isstabilized, the better bound can be achieved.
In order to evaluate the obtained results we consider an example referring to a standard H=
control technique, for the chosen class of systems under the stronger assumption of exponentialstabilizability. In order to have an easy solution of the SSRE involved, it will be further assumedthat D(A)"D (A*) and that property (P2) holds with the equality sign
[Ax, x]#[x,Ax]"0, ∀x3D(A)
Then Theorem 4.3.1 in Reference 17 applies and A generates a semigroup S (t) of isometries(ES (t)E"1, t*0) on H. Note that in this case, Corollary 3 can be seen as a special case ofCorollary 4.2 in Reference 8 with D
11"0 which would lead to
c2"k2
k2!1
i.e., the minimum c*2"1 is attained for kPR.
Example
As in Reference 8 consider the case
xR (t)"Ax(t)#Bu(t)#Bw(t) (24a)
y(t)"B*x(t) (24b)
with penalty variable z1(t)"(y(t)`w(t)
u(t)) . Applying Corollary 4.2 of Reference 8, one obtains the
following Riccati equation
CPx, AA#
1
c2!1BB*BxD#CAA#
1
c2!1BB*Bx, PxD
#A1
c2!1!1B [PBB*Px, x]#
c2c2!1
[BB*x, x]"0
Then setting P"kI it results that
c2"2k2
k2!2
Therefore the minimum c*2"2 is attained for kPR. Note that the resulting control lawu(t)"!kB*x (t) belongs to the class of feedback laws considered here. The same result can berecovered, under the loosened assumption of strong stabilizability of A (the context in which all
DISTURBANCE ATTENUATION 769
( 1997 by John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 7, 759—775 (1997)
the presented results are derived), from Corollary 3. The two penalty variables are actually relatedas follows
Ez1E2L2*0,=)
"EzE2L2*0,=)
#EwE2L2*0,=)
#2[y, w]L2*0,=)
)A1#1
k2B EwE2L2*0,=)
#EwE2L2*0,=)
#
1
kEwE2
L2*0,=)(25)
The minimum L2-gain upper bound attains its minimum at c*2"2 for kPR.
This example shows how, for the considered class of systems and under hypotheses(P1)— (P3), the obtained upper bounds for the L
2-gain are quite satisfactory considering that
there is need for neither exponential stability nor full-state measurement (nor dynamic outputcompensator).
4. FLEXIBLE SLEWING LINK EXAMPLE
In this section we briefly introduce a significative case study which belongs to the class of systemsaddressed in this paper. Let us consider the motion of a flexible beam rotating in the horizontalplane where bending occurs.
The rigid-body motion is defined as the rotation of the non-inertial frame (X, ½) in which thebeam deflection d (t,s) is described, with respect to the inertial frame (X
0, ½
0); t and s denote
respectively time and spatial variable. The X-axis is tangent to the beam at the base. Note that therigid body motion h (t) corresponds to the effectively measured joint angular position (Fig. 1).
We will briefly state the beam model, referring the interested reader to Reference 21 fora detailed derivation. Let EI denote the beam flexural rigidity, l its length, q(t) the applied torqueat the hub, M
pand J
pthe mass and inertia of the payload, J
hthe hub’s inertia and, assuming
a uniform beam with linear mass density o, Jbthe beam inertia
Jb"o P
l
0
s2 ds
Furthermore let J be
J"Jh#J
b#J
p#M
pl2
Figure 1. The flexible slewing link
770 A. DE SANTIS AND L. LANARI
Int. J. Robust Nonlinear Control, 7, 759—775 (1997) ( 1997 by John Wiley & Sons, Ltd.
In terms of g (t, s)"d (t, s)#sh(t) and under the small deflection hypothesis, the system behaviouris described by the following PDE
EIg(IV) (t, s)#og(t, s)"0 (26)
with boundary conditions
g (t, 0)"0
Jhg @ (t, 0)!EIgA (t, 0)"q(t)
(27)EIgA(t, l)#J
pg@(t, l)"0
EIg@@@(t, l )!Mpg(t, l)"0
where fQ and f @ denote respectively the time and spatial derivative of a function f. Note that therigid-body motion h (t) is equal to g@(t, 0).
Following the approach in Reference 16, the previous PDE can be put in a state-space form byincluding the boundary variables as part of the state in addition to the functional part describingthe beam deformation. The state vector p can therefore be chosen as
pT"(g( · ) g(t, l) g@(t, l ) g@ (t, 0))T
As a consequence we can write in an abstract setting the following set of equations
Mp(t)#Ap (t)#Bu(t)"0 (28)with u (t)"!q (t), A, M and B such that
Ap"AEIg(IV) ( · )
!EIg@@@ (t, l )
EIgA (t, l )
!EIgA (t, 0) B , p3D(A)
M"Ao 0 0 0
0 Mp
0 0
0 0 Jp
0
0 0 0 JhB , B"A
0
0
0
1 Band D(A) defined as
D(A)"Mg : g@, gA, g@@@, g(IV)3¸2[0, l], and g(0)"0N]R3
On this domain we consider the inner product [ · , · ]
[a, b]"Pl
0
a (s)b(s) ds#a (l )b (l)#a @(l )b @ (l)#a @ (0)b @(0)
Equation (28) is defined on the Hilbert space H obtained as completion of D (A) in the norminduced by the chosen inner product.
M is a bounded self-adjoint positive definite operator with bounded inverse, A is a self-adjointnon-negative definite operator with compact resolvent and B is a bounded operator. The
DISTURBANCE ATTENUATION 771
( 1997 by John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 7, 759—775 (1997)
eigenvalues uk
and corresponding eigenvectors /k
can be obtained by solving the followingeigenvalue problem
A/k!u2
kM/
k"0, [M/
k, /
k]"1
Note that j"0 is in the spectrum of A due to the presence of the rigid body motion g@(t, 0), andthe /
k’s constitute a basis on H.
In order to cast equation (28) as a well-posed first order abstract differential equation on theHilbert space H]H in terms of p and pR , we need to define a suitable norm. Unfortunately theusual ‘energy norm’ [Ap, p]#[MpR , pR ], i.e., the kinetic plus the potential energy, does not definea norm on H]H due to the zero eigenvalue in A. To avoid this difficulty (see Reference 22) let usfirst close the motor position feedback
u(t)"KpB*p(t)#v(t) , K
p'0
The resulting closed-loop equation is
Mp (t)#App (t)#Bv (t)"0 (29)
with Ap
defined as
App"A
EIg(IV) ( · )
!EIg@@@(t, l )
EIgA (t, l )
!EIgA (t, 0)#Kpg@(t, 0) B
Note that Ap
is a positive definite self-adjoint operator with compact resolvent and boundedinverse, i.e., equation A
pp"q can be solved uniquely for every q3H. In other words A
phas no
zero eigenvalue. Defining
x"Ap
pR B3H]H
we can write the following first order equation
xR (t)"Apx (t)#Bv (t) (30)
with
Ap"A
0
!M~1Ap
I
0B , B"A0
!M~1BBwhich is now well-posed by endowing H]H with the following energy inner product
[½, Z]E"[A
p½
1, Z
1]#[M½
2, Z
2]
½"A½
1½
2B , Z"A
Z1
Z2B
We denote this space by HE. As output y(t) we consider the joint angular velocity gR @(t, 0) which,
according to the state vector definition, can be expressed as
y (t)"Cx(t)"B*x (t) (31)
772 A. DE SANTIS AND L. LANARI
Int. J. Robust Nonlinear Control, 7, 759—775 (1997) ( 1997 by John Wiley & Sons, Ltd.
where the adjoint is computed according to the inner product adopted. We obtain the so called‘collocated case’, i.e., C"B*.
In this setting, Ap
is positive definite with a compact resolvent and zero is in the resolvent set.Moreover
A*p"!A
p(32)
where the adjoint operator is again intended in the energy inner product. Therefore the followingdissipativity equality holds
[Apx, x]#[x, A
px]"0
and, because of (32), the same is true for A*p
. Thus both Ap
and A*p
are dissipative operators sothat A
pgenerates a strongly continuous semigroup S (t) of isometries on H
E. It can also be easily
verified that the pair (Ap,B) is controllable.16 Hence hypotheses (P1)—(P3) are fulfilled; therefore
if disturbances d (t)"(w1(t)w2(t)
) affect system (30), (31) as in (10), Theorem 1 can be applied and thefeedback
v(t)"!KvB*x (t), K
v'0 (33)
can be chosen to strongly stabilize the closed-loop system (30)—(31)—(33) and render the L2-gain
from d(t) to z (t)"(y(t)v(t)
) less than the prescribed value given in (11) with k"Kv. In conclusion, the
proportional plus derivative feedback
u(t)"KpB*p (t)#K
vM~1B*pR (t)
"Kph(t)#
Kv
Jh
h0 (t)
provides both strongly stability and disturbance attenuation for the flexible slewing link (28).
5. CONCLUSIONS
For a class of dissipative distributed parameter systems with finite-dimensional collocatedinput/output we solved the stabilization and disturbance attenuation control problem viameasured feedback under the loosened condition of strong stabilizability. The main point consistof showing that the closed-loop system has a well defined L
2-gain despite of lack of exponential
stability. The solution goes through the direct characterization of the system’s L2-gain, without
invoking the commonly used Riccati equation argument. The obtained control law is a staticoutput feedback. A flexible showing link has been chosen as an application example.
APPENDIX
Since C1[[0, ¹ ],Hu], is dense in ¸
2[[0, R),H
u], there always exists such a sequence Mu
nN mentioned in
Lemma 1. Then we can write
Ev(t)!vn(t)E
L2*0,T+"KK P
t
0
Sk(t!q)B(u(q)!u
n(q)) dq KK
L2*0,T+
)KK Pt
0
ESk(t!q)E EBE Eu(q)!u
n(q)Edq KK
L2*0,T+
DISTURBANCE ATTENUATION 773
( 1997 by John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 7, 759—775 (1997)
Now, by applying the Schwartz inequality to the integral above, we obtain
Pt
0
ESk(t!q)E EBE Eu (q)!u
n(q)Edq)EBECP
t
0
ESk(q)E2 dqD
1@2
CPt
0
Eu(q)!un(q)E2 dqD
1@2
)EBEJ¹Eu!unEL2*0,T+
(34)
where the last step holds because Sk(t) is a contraction. Finally we have
Ev(t)!vn(t)E
L2*0,T+)¹ EBE Eu!u
nEL2*0,T+
)¹ EBE Eu!unEL2*0,=)
and the r.h.s. goes to zero as nPR, which proves (6a).Moreover
Ev (t)!vn(t)EH)P
t
0
ESk(t!q)E EBE Eu(q)!u
n(q)Edq
and by (34) we have
limn?=
Ev (t)!vn(t)EHP0
i.e., we also have pointwise convergence (6b). Finally we can write
D[u, v]L2*0,T +
![un, v
n]L2*0,T +
D"D[u!un, v]
L2*0,T +#[u
n, v!v
n]L2*0,T +
D
)Eu!unEL2*0,T +
EvEL2*0,T +
#EunEL2*0,T +
Ev!vnEL2*0,T +
where we applied the Schwartz inequality. Both terms in the r.h.s. converge to zero as nPR thus proving(6c). K
ACKNOWLEDGEMENTS
The authors would like to thank Professor S. Monaco for his helpful suggestions on an earlyversion of this paper. This work is partially supported by M.P.I. 60% and A.S.I. 94RS17 funds.
REFERENCES
1. Knobloch, H. W., A. Isidori and D. Flockerzi, ¹opics in Control ¹heory, Birkhauser, Boston, 1993.2. Francis, B. A., A Course in H
=-control ¹heory, Lectures Notes Control Inf. Sci. No. 88, Springer-Verlag, Berlin, 1987.
3. Francis, B. A. and J. C. Doyle, ‘Linear control theory with an H=
-optimality criterion’, SIAM J. Control andOptimization, 25, 815—844 (1987).
4. Doyle, J. C., K. Glover, P. Khargonekar and B. A. Francis, ‘State-space solutions to standard H2
and H=
controlproblems’, IEEE ¹rans. on Automatic Control, AC-34, 831—847 (1989).
5. Ichikawa, A., ‘Differential games and H=
-problems’, Proc. of M¹NS Conf., Kobe, Japan, 1991, pp. 115—120.6. Curtain, R. F., ‘H
=-control for distributed parameter systems: a survey’, Proc. of 29th IEEE Conf. on Decision and
Control, Honolulu, Hawaii, 1990, pp. 22—26.7. Barbu, V., ‘H
=boundary control with state feedback; the hyperbolic case’, Int. Series of Num. Mathematics, 107,
141—148 (1992).8. Van Keulen, B., M. Peters and R. F. Curtain, ‘H
=-control with state-feedback: the infinite-dimensional case’, J. Math.
Syst., Estim. and Contr., 3, 1—39 (1993).9. Tadmor, G., ‘Worst-case design in time domain: the maximum principle and the standard H
=problem’, Math. of
Contr. Sign. and Syst., 3, 301—324 (1990).10. Tadmor, G., ‘The standard H
=problem and the maximum principle: the general linear case’, SIAM J. Control and
Optimization, 31, 813—846 (1993).11. Curtain, R. and B. van Keulen, ‘Robust control with respect to coprime factors of infinite-dimensional positive real
systems’, IEEE ¹rans. on Automatic Control, AC-37, 868—871 (1992).
774 A. DE SANTIS AND L. LANARI
Int. J. Robust Nonlinear Control, 7, 759—775 (1997) ( 1997 by John Wiley & Sons, Ltd.
12 Curtain, R. and H. Zwart, ‘Riccati equations and normalized coprime factorizations from strongly stabilizableinfinite-dimensional systems’, Proc. of the 3rd IEEE Mediterranean Symposium on New Directions in Control andAutomation, Limassol, Cyprus, 1995, pp. 13—20.
13. Guoniang Gu., ‘On the existence of linear optimal control with output feedback’, SIAM J. Control and Optimization,28, 711—719 (1990).
14. Trofino-Neto, A. and V. Kucera, ‘Stabilization via static output feedback’, IEEE ¹rans. on Automatic Control, AC-38,764—765 (1993).
15. Van Keulen, B., ‘H=-control measurement-feedback for Pritchard—Salomon systems’, Int. j. Robust and Nonlinear
Control, 4, 521—552 (1994).16. Balakrishnan, A. V., ‘Compensator design for stability enhancement with collocated controllers’, IEEE ¹rans. on
Automatic Control, AC-36, 994—1007 (1991).17. Balakrishnan, A. V., Applied Functional Analysis, 2nd edition, Springer-Verlag, New York, 1980.18. Gibson, J. S. and A. Adamian, ‘Approximation theory for linear-quadratic—Gaussian optimal control of flexible
structures’, SIAM J. Control and Optimization, 29, 1—37 (1991).19. Benchimol, C. D., ‘A note on the weak stabilizability of contractions semigroups’, SIAM J Control and Optimization,
16, 373—379 (1978).20. Datko, R., ‘Extending a theorem of A. M. Liapunov to Hilbert spaces’, J. Math. Analysis and Application, 32, 610—616
(1970).21. Bellezza, F., L. Lanari and G. Ulivi, ‘Exact modeling of the flexible slewing link’, Proc. IEEE Int. Conf. Robotics and
Automation, Cincinnati, OH, 1990, pp. 734—739.22. De Santis, A. and L. Lanari, ‘Stabilization and control of a flexible structure continuum model’, Proc. of the 32nd
IEEE Conf. on Decision and Control, San Antonio, TX, 1993, pp. 3210—3215.
.
DISTURBANCE ATTENUATION 775
( 1997 by John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 7, 759—775 (1997)