Proceedings of the 26th Conference
FP3 3:OO
ROBUST CONTROL WITH
STRUCTURED PERTURBATIONS
L.H. Keel
Department of Electrical Engineering
Tennessee State University
Nashville, Tennessee 37209
ABSTRACT
This paper considers the problem of robustification of a
given stabilizingcontrollerto make the closed loop system
remain stable for prescribed ranges of variations of a set of
physicalparameters in theplant.Theproblem is treated
in thestatespaceandtransferfunctiondomains.Inthe
state space domain a stability hypersphere is determined in
theparameterspaceusingLyapunovtheory.Theradius
of this hypersphere is iteratively increased by adjusting the
controller parameters until the prescribed perturbation ranges
arecontained in thestabilityhypersphere.Inthetransfer
funct.ion domain a corresponding stability margin is defined
and optimized based on the recently introduced concept of the
largest stability hypersphere in the space of coefficients of the
closed loop characteristic polynomial. The design algorithms
are illustrated by examples.
1. INTRODUCTION
In this paper we deal with the problem of designing a
feedback controller for alineartimeinvariantplantthat is
subject to perturbations of a physical parameter vector. This
type of problemarisesfrequentlywhenagoodqualitative
model of theplant is available butuncertaintyexistswith
respect to the numerical values of various parameters which
may be changing during operation or are unknown or difficult
to measure. Aerodynamic coefficients, reaction rates, inetias,
masses, spring constants and friction coefficients are common
examples of such parameters. The controller must preserve
closedloop stability forknown ranges of parameter excur-
sions. Sometimes the controller must be designed to enlarge
the stability region in parameter space to the largest extent
possible.
Therobustcontrolliterature does not deal withthis
kind of problemsincetheperturbation class treated there
is unstructured [ I ] . Somerecentresultsthatdodealwith
structured perturbations are those of Kharitonov (21, Barmish
[3], Soh,BergerandDabke [4] andBiernacki,Hwangand
Bhattacharyya [SI. The result of [1]deals with the same kind
of problem treated here but for the special case in which the
characteristic polynomial coefficients are linear functions of
the parameters. For this case the largest stability hypersphere
in parameterspace is determined in [SI andanalgorithm
is given for enlarging this hypersphere by iterating over the
controller parameter space.
Theproblem is treated in thetransferfunctionand
statespacedomains.Ineachcaseastabilityhypersphere
in parameterspace is determinedandtheradius of this
hypersphere serves as a stability margin. In each domain a
robustfication algorithm is given which increases this stability
margin by iteratively readjusting the controller parameters.
In the transfer function domain the robustification algorithm
utilizesacalculation of theradius of thelargeststability
hypersphereintroduced in (41. Thestatespacealgorithm
utilizes Lyapunov theory and is based on an extension of the
results of Pate1 and Toda [6] and of Yedavalli [7] to consider
structured perturbations.
This research was supported by the National Science Foundation under Grant
no. ECS - 8309792.
CH2505-618710000-2064$1.00 0 1987 IEEE
on Declslon and Control
Lor Angeles, CA December 1987
S.P. BhattacharyyaandJ.W. Howze
Department of Electrical Engineering
Texas A&M University
College Station, Texas 77843
Thealgorithmsareappliedtosomeillustrative design
examples.Thepaperendswitha discussion of theresults
presented and of the many open research problems.
2 . CHARACTERISTIC POLYNOMIAL
CALCULATION
This section we give the formula for computing the char-
acteristicequation of the closed loop system.Considerthe
multivariable feedback system of Figure 2.1 with the plant
transfer matrix G(s) and the feedback controller C(s).
Figure 2.1 Multivariable Feedback System
Let p be a vector of real physical plant parameters that
is subjecttouncertainty.Ingeneralthe coefficients of the
entries of the transfer matrix G ( s ) will be functions of the
parameter p. The nominal values of the parameter vector p
is denoted by po and it is assumed that the given controller
stabilizesthe closedloop system for p = P O . Our problem
is to redesignthecontroller so that closed loop stability is
maintained for a given range of parameter excursions. Our
procedure for doing this will assume that the controller order
is fixed and the numerical values of the controller parameters
are to be readjusted.
To proceed we display a special form of the closed loop
characteristic polynomial on which the transfer function de-
signprocedure is based. Let n betheorder of theplant
(Mcmillan degree of G ( s ) ) , t the order of the controller C(s)
and let
+ +
6 ( ~=) c ~ , , + ~ s " + ~ ......+ 6,s 60 (2.1)
denote the closed loop characteristic polynomial. We refer to
as the characteristic vector and for convenience say that 6 is
Hurwitz if and only if 6(s) is Hurwitz.
In the following we show that the closed loop character-
istic vector 6 satisfies the equation
where M(p) is a matrix containingonly plant parameters and
C ( X ) is a vector containingonly the controller parameter vector
x. Thevector x is aquantitythatcompletely defines the
t t h ordercontrollerand conversely any t t h ordercontroller
determines x.
2 . 1 Single Input Multioutput(SIM0) systems
2064
 Let function
linear a 6(s) as of C ( X ) with coefficients which are
polynomials in z . Since z is afunctiononly of theplant
G ( 5 )= ( G l ( s ) G,(s) . .. . . Gm 1' (2.4) parameters p we have the
result given in ( 2 . 3 ) .

3. STABILITY MARGIN

Let po denote the nominal value of the plant parameter

vector and po iAp a perturbation. Then

M(p) := M(p' - Ap) = M(p') t AM(p', Ap) (3.1)

defines the perturbation of M(p). Let x denote the controller
parameter vector. Then the nominal characteristic vector is
M(po)c(x):= 6'. Under perturbations of po the characteris-
tic vector 6' suffers perturbations given by

M(po - AP)c(x) := 6 := 6
' + A6. (3.2)

In this setting the problemof robust stabilization is to en-

sure that the characteristic vector 6 remains strictly Hurwitz
as the' controller parameter vector we getfor the closed loop characteristic vector ( at )

J
- E

MlP)

The multiinput single output caseis the dual of the single

input multioutput case with G ( s ) in (2.4) replaced by G(s)'
and C(s) in (2.5) replaced by C ( s ) = . The resulting equation for the given class of perturbations denoted by {Ap} := A.
for the characteristic vector is identical t o (2.9). If for example the class of perturbations consists of perturbed
ranges of parameter excursion then
2 . 2 Multiinput - Multioutputs(MIM0) systems
A:={Ap~q<
, Ap; < ~ , , i =
l,..*,k}. (3.3)
Let G ( s ) denote the plant transfer function matrix and
( A ,B , C) a minimal realization so that
Let p ( 6 ) denote the radius of the stability hypersphere
G(s) = C ( s 1 - A ) - ' B (2.10) centered at 6' in the space 6,i.e.6' +
A6 is strictly Hurwitz
for all IjA611, < ~(6') and there exists =with (l=[l, = p(6')
Kow let C ( s ) be a t t h order proper feedback with minimal suchthat 6' + is notstrictlyHurwitz. An algorithm for
realization ( A c ,B,, C,, Dcj.Let calculating ~ ( 6 ' )has been given in 14:. From (3.1) and (3.2)

Kt := ( B, A,")
Dc (2.11)
it follows that robust stability is achieved if

/IAM(P'3 AP)C(X)~, < P ( 6 O ) VAp E A (3.4)

A, := (t Bt := ( B OI , ) Ct := ( c o
Let !I . / ~ denote
p theFrobeniusnorm.Theinequality (3.4)
(2.12) can be satisfied if
and note that the closed loop system has the characteristic
polynomial p(60) > /IAM(p',Ap)jlp VAp E A. (3.5)
lic(x)ll2
6(s) = Det[s1 - A, - BtKtC,j. (2.13)

Let the vector x equal the ordered list of entries of Kt and And let
z denote the elements of A , , B,Ct in some order. It follows max //AM(p',AP) I I F := P. (3.6)
A P € P
from (2.13) that b ( s ) is a polynomial in the variables x, z and
s. Therefore each coefficient of s in 6(s) is a polynomial in the We have now stated the following result.
variables x and z . Thus each coefficient in 6(s) can be written
Theorem 3.1
as a polynomial in z with coefficients which are polynomials
in x. If C ( X ) denotesavectorcontaining all monomials in Let x be a stabilizing controller when p = p o . Then x
the variables x which occur here we have each coefficient of stabilizes the closed loop system for all perturbations Ap E A
2065
 if
(3.7) n
RemarkTheformula(3.7)indicatesthatthecontroller x
must not only enlarge the stability hypersphere but do
so with
,i2.
small "gain" i~c(x)
The quantity I I
+
ul
+
u2
is now proposed as astabilitymargin for thesystemwiththeFigure4.1Twomass - twospringMultivariableSystem
given controller x. Moreover if x1 and x, are two stabilizing
controllersand p ( x l ) > p(x,) then the level of perturbations The h a m f e r function Of this system is
p1 and p2 that can be tolerated by the respective closed loop
related are systems

P1
P (x1)
= &J-P-Z (3.9) w
(il)
Y
s2 2k,

= (
+
s2 -I- 2k2 k

s3 t ( ; k 2 s
A ( 4
s3

Girl
v
Ai.)
+ ( 2 k , - k:)s
A ( 8 )
+
1 (ti)
,w
-

so that the controller x1 is more robust than x , . The quantity

P is independent of thecontrollerandcan often be easily where
calculated as the example shows.

4. ROBUSTIFICATION PROCEDURE
A ( s ) := m l m , s 4 + (k,m, + k l m 2 $ k , m , ) s 2 + k , k ,
Using thestabilitymargin defined in theprevioussectionregard
an algorithm for controller design can be designed that itera- P = (m, mz k1 k2)T
tively upgrades the vector x of adjustable controller parame-
ters to increase the stability margin. Therefore if xk denotes as the physical parameter vector subject to perturbation, ~~d
the choice of controllerthe at k t h iteration let we get
6(s) =

Our objective is t o choose x k + lso that

CL(Xk+l) >4Xk)
" 2
and x k + is stabilizing. From the equations M(P)

M(p0)c(xk+,)
= gk+l (4.3)

M(p')c(x,) = (4.4)
we see that if

xi;+l is guaranteed to be stabilizing. The correction Ax, =

xi;+ -xk is chosen via a gradient method based on numerical
evaluation of the gradient of p(x). po :=(l 2 1 2)
Notethatsincethe convexity of thefunction is not
established we can make no guarantee that a global minimum
will be found. A common procedure in such cases is to choose
Aml IAmll 5 0.01
severalinitial guesses and to select thebestanswer as the A := Am, IArn,l 5 0.01
global minimum. Akl iAkll 5 0.01
Ak2 lAk,l 5 0.01
Example Then
2 0 0 0 0
0 1 1 0 0
Consider the following two mass - two spring multivari-
able system.

and
y : sensors which measure sum of displacement and
- P = max lIAM(po,Ap)1IF = 0.112099
velocity. A P € P
u : actuators which provide damping in the system.
-
Choosing the initial stabilizing controller as
2066
 2.2649934 -
-4 B K C - pi E, forgiven matrices E, which prescribe

13.898785 the structure of the perturbation.

!Ye now statearesultthatcalculatestheradius of a
spherical stability region in the parameter space p E R ' . This
7.4233423
result is a useful extension of those given by Patel and Toda
-6.212874
'6; and Yedavalli [ 7 ] .
Let the nominal asymptotically stable system be
p ( x 0 )= 0.0011208

and coefficientsof a robust controller which tolerates thegiven X(t) = Mz(t)= ( A - BKC)z(t) (5.4)
class of perturbations are and the perturbed equation be

X(t) = ( M +Cp,E,)z(t) (5.5)

,=1
where the p , i = 1 , .' . , r are perturbations of parameters of
~

interest and the E , , i = 1, .. ' , r are matrices determined by

the structure of the parameter perturbations. Let Q > 0 be
p ( x ' ) = 0.139547 apositivedefinitesymmetricmatrixand let P denotethe
and the roots of corresponding characteristic polynomial are unique positive definite symmetric solution of

-4.95948E10-? *
j0.30490336 M T P - PA4 +Q =0 (5.6)
-0.190843011 z j1.925854956
Theorem 5.1
5 . STATE SPACE FORMULATION AND DESIGN
The system (5.5) is stable for all p , satisfying
Assume now that the plant equations are derived from
physical considerations in the state space form
(5.7)
X=Ax&Bu
(5.1)
y = Cz +
where p, := ~iE,'P PE; i 2 .
This quantity determines the range of perturbations for
and let the controller of order t be described by which stability is guaranteed and is the radius of the stability
hypersphere in parameter space. Proof is straight forward and
X, = A e z , - B,y omitted here.
(5.2)
u = C c x , -I D,y.
Design Procedure
The closed loop system equations are Using the index obtainedin (5.7):we now give an iterative
design procedure to obtain the optimal controller K' so that
( 5 . 7 ) is as large as possible. For a given K the largest stability
hypersphere we can obtain is

(5.8)
At

and the problem of designing a robust controller with respect

to structured parameter perturbations can be formulated as:
(5.3)
For the given ( A ,B , C )find
, K to minimize
+
Sow (5.2) is a stabilizing controller if and only if A, B, Kt C,
is stable. Since we will consider the compensator order to be
fixed at each stage of the design process we drop the subscript
t henceforth and consider the problem of robustification of subject to
+
A B K C by choice of K when the plant matrices are subject J , := max &(X) < 0.
to perturbation. LEoJA+BKC)

Let p = ( p ,p , . . p , ) denoteaparametervector where

consisting of physical parameters that enter the state space
descriptionlinearly.Thissituationoccursfrequentlysince ( A t B K C ) = P+ P ( A +BKC) = -LTL (5.10)
the state equations are often written based on physical con-
siderations. In anycasecombinations of primaryparame- Note that the positive definite matrix Q hasbeenreplaced
ters can always be defined so that the resulting dependence without loss of generality by L'L. For anysquare full
of A , B , C on p is linear. e%
\' also assume that the nominal
rank matrix L , L T L is positive definite symmetric. Here we
model (5.1) has been determined with the nominal value po consider a slightly more general class of perturbations, i.e.
of p. This allows us to treat p purely as a perturbation with
nominal value po = 0. Finally. since the perturbation enters
at different locations me consider that A - B K C perturbs to

2067
 Starting with this nominal stabilizing controllerwe performed
the robustification procedure. For this step the set of initial
values are

Then we get

M = A. + BoKC, andE, = Ai + BiKC (5.12) Lo= [ 1

0.5
-0.1
0.2
0
1
0.4
0.6
-0.5
-0.03
1
-0.13
0.06
0
0.14
1.5 1
A gradient - based algorithm for enlarging the radius of
the stability hypersphere p ( K , L ) by iterating on ( K , L ) can These nominal values gave the stability margin
be devised using its gradients with respect to K and L. The
closed form expression of its gradients can be easily derived po = 0.02712 < 0.0648 = jjApI1,
and omitted here. This algorithm has been used to solve the
illustrative problem given below. which does not satisfy the requirement.After 84 iterations we
have
Example p’ = 0.12978 > 0.0648 = iJApiI2
which satisfies the given requirement. The robust stabilizing
As an example a control problemof a VTOL helicopter is Oth order controller computed is
considered as in [8]. The nominal stabilizing controller K has
-0.996349
been found from the algorithm describedin 191. The linearized
dynamic equation of the VTOL helicopter is shown below. K* = ( 1.801848 )
(ii (
x4
=
-0.0366
0.0271
0.0482 -1.010
0.1002
0
p,
0
0.0188
0.0024
-0.707
1
-0.4555
-4.0208)
pz
0
(I: and the closed loop eigenvalues are

-18.39643
-0.247593 k 1.2501385
-0.0736273 1
0.4422
0.1761
6. CONCLUDING REMARKS
The stability margins defined in this paper may be con-
servative in some cases. This can be improved by developing
y=(o 1 0 0) effective measures for determining the largest stability hyper-
sphere in parameter space as has been done in the linear case
The physical description of the dynamic equations is as !5]. This problem appears to be a difficult one because of the
follows. nonlinearrelationships involved but is anundoubtedly im-
portantopenproblem.Anotherimportantarea for further
x1 horizontal velocity, knots research is the development of techniques to simultaneously
x, vertical velocity, knots handle structured and unstructured perturbations.
x3 pitch
rate,
degree/sec
x4 pitch
angle, degree REFERENCES
u1 collective pitch
control
ZL, longitudinal cyclic pitch
control J.C. Doyle, J.E. Wall and G. Stein,“Performanceand
Robustness Analysisfor UnstructuredUncertainty,” in
The given dynamicequation is computed for typical Proc.Mst IEEE Conf. DecisionContr.,Orlando,FL,
loading and flight conditions of the VTOL helicopter at the 1982.
airspeed of 135 knots [5][8]. As the airspeed changes all the V.L.Kharitonov,“AsymptoticStability of anEquilib-
elements of the first three rows of both matrices also change. rium Position of a Family of Systems of Linear Differ-
The most significant changes take place in the elements p,, ential Equations,”Differentsial Uravnen,vol.l4, pp.2086-
p,, and p,. Therefore, in the following all the other elements 2088, 1978.
are assumed to be constants. The following bounds on the B.R. Barmish,“Invariance of theStrictHurwitzProp-
parameters are given in [8] for linear controls. erty for Polynomials with Perturbed Coefficients,” IEEE
Trans. Automat. Contr., vol.AC-29,pp.935-936,1984.
C.B. Soh, C.S. Berger and
K.P.
Dabke,“On
the
Stability Properties of Polynomials with Perturbed Co-
efficients,”IEEE Trans.Automat.Contr. ,vol. AC-
30,pp.1033-1036, Oct.1985.
R.M Biernacki, H. Hwang, S.P. Bhattacharyya,“Robust
Now we compute Stabilization of Plants Subject to Structured Real Param-
eter Perturbations,” IEEE Trans.Automat.Contr., vol.
max 11 Apll, = 0.0648 AC-32, pp.495-506, June 1987.
R.V. Pate1 and M. Toda,“QuantitativeMeasures of
From 191, the nominal stabilizing controller is given by Robustness for MultivariableSystems,” in Proc.Joint
Automat. Contr. Conf., San Francisco, CA, 1985.
R.K. Yedavalli,“Improved Measures of Stability Robust-
= (-1.63522
1.582236)
ness for LinearStateSpaceModels,”IEEETrans.Au-
tomat. Contr.,vol. AC-30, pp. 577-579, June 1985.

2068
. . S.N. Singhand A.R. Coelho, Tonlinear Control of
18’
Mismatched Uncertain Linear Systems and Application
toControl of Aircraft,” J . of Dynamic Sys. Meas. and
Contr., ~01.106,pp.203-210, Sep. 1984.
[9j S.P. Bhattacharyya, L.H. Keel and J.W. Howze,
“FeedbackStabilizationwith Low OrderRegulators,‘’
TCSL Technical Report #8602, Department of Electrical
Engineering,Texas A&hi University, College Station,
Texas, Aug. 1986.

2069