Professional Documents
Culture Documents
jw*(f)w(r)dz <
0
-V(x(t))
d
dt
(1.3)
- w(t)]*[D*X'x(t)
-~ ( t ) ]
+ w*(t)w(t)
(2.6)
- w(t)]*[D*X'x(t) -WO)]
20 ,
Eq.
(2.6)
d
-V(x(t)) s w'(t)w(t) .
(2.7)
dt
Note that (2.7) describes the tightest possible bound on
d
-V(x(t)), since a w(t) exists that will lead to an equality in
dt
(2.7). This input is w(t) = D*X'x(t).
Integrating both sides of (2.7) and using definition (2.3) yields
(for x(0) = 0)
a 1
CH2642-7/89/0000-2667$1.00
@ 1989 IEEE
= -[D'X'x(t)
Since [D*X'x(t)
implies that
x = (A + AA) x +Dw
(1.1)
with parameter disturbance matrix AA. The work of Yedavalli [l]
and more recently Khargonekar [2] show perhaps the most detailed
analysis of robustness with respect to structured parameter
variations requiring only
1
IAAijlmax$ E < (1.2)
x*(t)X-'x(t) 5 rw'(o)w(o)do
Since for any t 5 -,
jotw*(o)w(o)do 5 j"w'(o)w(o)do
(2.8)
(2.9)
2667
(2.10)
o[X'l= (O[W)-'
3.
Theorem 1.
For the asymptotically stable and controllable system
(2.1), the output covariance matrix is Y CXC* where X is defined
by (2.2). Let wc) in (2.1) be any h disturbance
(llw(.)ll; Aiaw*(z)w(r)dz < 4 and Ily(.)llf ={;
y*(t)y(t)be
(2.12)
and
llx(.)112 5 3xIllw(.)ll% .
(2.13)
Proof:
(T*
h'*)
hence
llCX112 s O[Y]X*X'X
Recalling (2.10), we have
llY(t)ll2= IlCx(t)l12
5 G[Ylx*(t)X-'x(t)
O~Ylllw(.)ll;
Thus
g;;
IlY(.)IIf =
llY(t)1I2 5 ~[Ylllw(.)ll;
G[@WD')--'A AAX(DWD*)-'A']
Example
Let system (2.1) be
s [a[DWD*]"]-'o[X@WD*)-"*]a[AA]
The
conservative by 20% =
Theorem 2.
For the asymptotically stable nominal system (3.1), assume
that D is nonsingular and define X to be the state covariance matrix
solved by (3.4) and AA to be the perturbation matrix, then (3.1) is
asymptoticallystable if
Y = [I, Olx
disturbance is defined by
w(t) =
0
t>l
Use (2.2) to get the state covariance matrix
x=
(3.9)
(3.10)
0.i671
and
o[V=0.0833
kww*(~)w(~)ciz
= J '0.25d~= 0.25
Using (2.12),
2o[x@wD*)-"*]
= 2G[P-'(P/&)] =
After substituting (3.11) into (3.10),
-o[DWD*]" = - .1
Ily(.)ll?, s 0.0208 .
This means that the output L, norm will be less than 0.1443. The
actual output L, norm is
~ G ~ X ~ W D * ) - ~ * Orpi
I
2668
__
6.
(3.11)
lo -21
A=
1 -3
L
then
0.2613 0.9860A
L
=0.382 .
o[DWD*]%
cI= -=0.4111
2a[X(DWD*)-"*]
The example obtains a better upper bound than (1.12) by an
appropriate choice of matrix W. The best choice for W remains an
open problem, but for this example an improvement of 8% is
obtained.
a[P1
AA= Z piAAi
(3.12)
i=l
K=
P2 1,
(3.21)
PmL
[o -2+,,]
"1 1i)]
A + B G C = [ o1 -3
o[xex:]i=lZ p ? < 1
=A,+AA.
Use (3.12),
(3.23)
m
AA = Z pi AAi
note that
to see that
o[XeKl 5 z[.z I ki I
I=
Ai 8 @")-%xi@")-%*
x, 8 E l , 22, ...,im]
(3.24)
i=l
i=1
(i=l, 2, ...m)
. IPi I1 5
"zl I
m
VX
I pj I
Ai
I1
(3.25)
DWD'=
vl
;1 I:
a) l/o[XeX: J = 1.6
c)
l/E[,i1 I & I ]
= 1.0
I=
= 1.8065
q = 2 b)
q = 2.55 c)
i= 1
ACTUATOR
'
'
'
i+
Fig. 4.1
where the plant is described by
mq + dq + kq= b(y2 + w)
(4.la)
1y1
SENSOR
Fig. 4.2
The given parameters of the controlled system are:
~ = 0 . 1 , m = 1 . 0 , d = 0 . 1 , k = 1 . 0 ,b=1.0,
pY -- 1.0, pz = 1.0, r,, = 0.5, r, = 1.0.
The design disturbances for requirement (4.2~)are described
by the white noise processes,
E[~(t)]=O;E[~(t)w(~)]=W8(t--~);Ew(t)v(z)=OW=
1.0
d
k
Y 2 w
(4.3b)
q=-(-q+-q)+-+-.
m
m
m
m
Using the above data for Z,d, m, and k, we let x1 = ql, x2 = q
and x3 = y2. Then the state space description of the system is
x = Ax+ Bu + Dw
(4.4a)
y= cx
(4.4b)
z=Mx+v
(4.4c)
;I:{ 11;
After substituting~-~
the given parameters into A, B, D, C and M we
obtain
0
1
0
0
A = 1-i.0 -:.l
B=
D=
[i]
(4.lb)
= Py9 + ryq
0.5
(4.1~)
= Pz9 + rzil
l;],M=H
1
01
The design goal of the closed loop system is to keep the sensor and
actuator signals within their saturation bounds under the
c=
disturbances w(-) and v(-) in Fig. 4.2 when r = 0. So the
Suppose there exists some parameter errors in Azl, A z , then,
requirements of performance are as follows:
Y1
1;
Y3
llY3(9ll- 2
4 E3
Ily2e)ll- 2 0.8 4 E2
4-2CT1=4
(4.2a)
(4.2b)
(4.2~)
4.L.
is stated as follows
03
t+-
y
z
= E3/llW(.)llZ
0 2 =~ 2 / l l W I l 2
Y, = (CXC')ii
then
=Ax+Bu+Dw
=
cx
Mx+v
=
(4.5)
Y11
0:
Y22
03
= 2
(4.12a)
= 0.3200
(4.12b)
Y 3 3 5 0: = 2.5000
(4.12c)
The existence of a controller with variances satisfying (4.12) is
demonstrated by calculating maximal accuracies (from step 1 and
eq. (4.7)):
(4.5)
-
This problem was shown in [24] to have an LQG solution for some
diagonal positive semidefinite matrix Q in the cost function
lim E(y*Qy + u'Ru). Furthermore the diagonal elements of Q are
t+-
(CXC'),
03
QE
Jy?
- =0 .
(4.6)
Hence, the following algorithm from [22] produces Q, = 0
whenever constraint i is not binding (Ey' < 0:) (see step 4 of
the OVC algorithm below). The solution of this problem (from
[22]) follows.
The OVC algorithm: (LQG weight selection)
Given data : (A, B, R, D, W, V, M, C, oi. E )
1. Solve
for
=o
0:
< of
I:!';]
o=XA* + A X - X M * V - ' ~ + D W D *
(4.7)
X o.
[cXC'I,
s 0: ?
F = XM*V-'
(4.13b)
Q, = diag [... 0
;
' ...I, set k = 0 .
2. Solve 0 = KkA + A*& - KkBR-'B'Kk
for Kk> 0 .
Gk = - R-'B'Kk
0 = &(A + BGk)'
+C*QC
+ (A + BGk)Xk + F W *
(4.8)
(4.9)
(4.10)
= (A + AA)x + Bu + Dw
y=cx
z=Mx+v
X
+ k)c*]ii)
(4.11a)
(4.11b)
(4.14a)
(4.14b)
(4.14~)
i,=(A-FM)x,+Bu+Fz
(4.14d)
u=Gx, .
(4.14e)
So. if we consider the parameter errors, the closed loop system is
(4.15b)
The closed loop system parameter error is
Define D, =
and 3 cannot
in Theorems
used in the
definition of
. Since D,WDT
is singular Theorems 2
example. Hence,
XLAL + ALXL+ DLWDL = 0 , DLWDL = I
where AL is the closed loop "plant" matrix in (4.20a). Using
Theorems 2 and 3 this yields
(from Theorem 2)
p = 0.2804
I p I = 0.4064 (from Theorem 3, (3.15))
Hence, for the closed loop system, if the unstructured parameter
errors AA satisfies
:(AA,) < P
Then, the system will be asymptotically stable. If the structured
parameter errors p1, p2 satisfy
max(lp1 I, I P ~ I ) <I P I
the closed loop system will also be asymptotically stable.
5. Conclusions
This paper shows that the following robustness properties of
continuous systems are directly related to the system covariance
matrix: robustness to uncertain disturbances, to uncertain structured
parameters, to uncertain unstructured parameters, to L,
performance requirements, and to
performance requirements.
Some of the structured and unstructured parameter bounds are new.
Come are improvements over existing bounds [ 1, 21. Several of the
existing parameter bounds for continuous time systems can be
obtainrd as special cases of the bounds developed in this paper.
i-mre research on the controller synthesis is desired to take
further advantage of the covariance controller's explicitness in
terms of the state covariance.
References
R.K. Yedavalli, "perturbation bounds for robust stability in
linear state space models", Int. J. Contr., 1985, Vol. 42, No.
6, 1507-1517.
K. Zhou and P.P. Khargonekar, "Stability Robustness
Bounds for Linear State Space Models with Structured
Uncertainty", IEEE Trans. Auto. Contr., Vol. AC-32, No. 7,
1987.
J.C. Doyle and G. Stein, "Robustness with Observers," IEEE
Trans. Auto. Contr., Vol. AC-24, pp. 607-610, 1979.
M. Tahk and J.L. Speyer, "Modeling of Parameter
Variations and Asymptotic LQG Synthesis," IEEE Trans.
Auto. Contr., Vol. AC-32, No. 9, 1987.
D.J.N. Limebeer and G. Malikias, "A Controller Degree
Bound for H, - Optimal Control Problems of the Second
Kind," Proc. 25th Conference on Decision and Control,
Athens, Greece, 1986.
B.A. Francis, A Course in H , Control Theory, SpnngerVerlag, Berlin, 1987.
M.A. Dahleh and J.B. Pearson, "L' - Optimal Feedback for
MIMO Discrete-time Systems," IEEE Trans. Auto. Contr.,
AC-32, NO. 4, pp. 314-322, 1987.
D.D. Siljak, Large - Scale Dynamic Systems, Stability and
Structure, North-Holland, 1978.
B.R. Barmish, M. Corless and G. Leitmann, "A New Class
of Stabilizing Controllers for Uncertain Dynamical System,"
SIAM J. of Contr. Optimiz., Vol. 21, pp. 246-255, 1983.
M. Corless and G. Leitmann, "Controller Design for
Uncertain Systems Via Lyapunov Functions," Proc. Amer.
Contr. Conf., Atlanta, 1988.
M. Corless and D. Da, "New Criteria for Robust Stability,"
Proc. International Workshop on Robustness in
Identification and Control, Turin, Italy, 1988.
I.R. Peterson, "Disturbance Attenuation and Hp.
Optimization: A Design Method Based on the Algebrac
Riccati Equation," IEEE Trans. Auto. Cont., Vol. AC-32,
pp. 427-429, 1987.