-
SELECTED TOPICS IN IDENTIFICATION,
MODELLING AND CONTROL
Progress Report on Research Activities in the
Mechanical Engineering Systems and Control Group
Edited by G.H. Bosgra and P.M.J. Van d
en
Hof
Volume 7, December 1994
Mechanical Engineering Systems and Control Group
Delft University of Technology
Published and Distributed by Delft University Press Stevinweg 1 2628 CN Delft The Netherlands Tel.: (0}15-783254 Telefax: (0}15 -781661 By order of
Mechanical Engineering Systems and Cont ro lGro up Delft University of Technology
Mekelweg 2,2628 CD Delft The Netherlands
Tel.: +3 1-15-786400; Telefax: +31-15-784717 email:arkesteijn@tudw03.tudelft.nl
CIP-GEGEVENS KONINKLIJKE BIBLIOT HEEK, DEN HAAG Selected
Selected topics in identification, modelling and control:
progress report on research activities in the mechanical engineering
systems and control group.-Delft:Mechanical Engineering Systems and Control Group, Del ft Universityof Technology,Vol. 7-ed.by O.H.Bosgra and
P.M.J.Van den Hof.-ill. Metlit.opg. ISBN 90-407-1077-5
SISO656 UDC 531.7+681.5NUGI841
Cover desig nbyRuudSch ram a
©1994 Copy rightDelftUniv ersityPress.All rights reserved.No part of this jou rna I may be reproduced, in any formor byany means, without written permission from the publisher.
Contents
Volume 7, December 1994
Robust performanceH2/Hoo optimal control M. Ste in buch and G.H.Bosgra
A unified approach to stability robustnessfor uncertainty descriptions based on fractional model representations
R.A . de Cal/ af on, P.M.J. Van den Hof and P.M.M. Bongers Solvab ility tests for the Lyapunov inequality
C.W. Schere r
An iter ative algorit hm for freque ncy- weighte d H2-nor m optim al redu ction and eente ring P.M.R. Wortelboer and G. H. Bosgra
Filtering and para met rization issuesin feed back relevant iden tification based on fractionalmodel representations
R.A . de Cal/ afon an dP.M. J. Van den Hof
1
9
17
23
35
Aninstrument al vari abiepro cedurefor theidentifica tion of probabilistic frequ ency responseuncertaintyregion sR.G. Hakvoort and P. M.J. Van den Hof 45
Identifi cation and robust cont rol design of an industrialglass tube manufacturing process
D. de Roover and R. G. Hakvoort 53
Closed loop system identification of an industrial wind turbinesystem and a preliminary validation result
G.E. van Baars 63
Closed-loop identifi cationof a continuous crystaIIization process
R.A. Eek, J. Both and P.M.J. Van den Hof 71
Inner loop design and analysis for hydraulic actuators, with application to impedance con trol
J. Heintzeand A.J.J. van der Weiden 83
Stability analys is of a hydraulicservo-systemincluding transmission line effects
G. van Schothorst, P.C. Teerhuis and A.JJ van der Weid en 93
Zer o-ripple torque cont ro l in brushless DC motors
H. Huisman 101
Predictive servo control through optimal feedforward com pensat ion
search project isa collaborative effor t betw een three university researchgroupsof Delft University, indus-try and the national aerospace research laboratory NLR. Finally,the paper by Henk Huisman discuss-es a new approach to torque ripple minimization in brushless DC motors. In precision drive systerns such as video scanners, it is of utmostim por ta nceto have mass produced, cheap motors disp laying high servo controlperformance.
Robustcontrolissues are discussed in several paper-s. Some preliminary results on an uncompromised 11.2/1ioo optimal con trol problem are given in the
paper by Maarten Steinbuch and Okko Bosgra. As a contrib ution to the recent interest in LMI-based robusi controlproblem formulations, the paper by Gars ten Scherer has results on the solvability of Lya punov inequ al iti es. The issue of model reduc-tion for control design is considered in the paper by Pepijn Wortelboerand Okko Bosgra, where fre-quency weighted versions of the11.2optimal reduc-tion prob lem are studied. Some unifications of ro-buststability theoryfor fractionaluncertainty mod-elsare presen ted in thepaperby Raym on d de Galla-fon, Paul Van den Hofand Peter Bongers. Finally, in the pa per by Dick de Roover the ser vomechanis-m problem with fini te hori zon is cons idere d, usi ng an approach in which ideas of feedfor ward three-degrees-of-fr eedom cont ro l, predictive control, the int ern al model principle,an d rep et itive control, are com bine d .
In the present issu e there finally are two cont ribu-tions on the metho dology of syste m identific atio n. The pap er by Ramond de Gall afo n and Paul Van den Hofdiscusses a feed back relevan t identifi ca t ion approach using coprime fract ion al model represen-tations . The assessment of frequency domain un-certainty bounds fromsystem ident ifica ti on ex peri-ments is discussed in the paper by Richard Hakvoort and Paul Van den Hof
We wish to acknowledg e the important contribu-tions to the work in the group made by colleagues from industry. In this issue we have contributions by Maarten Steinbu ch and Pepijn Wortelboerfrom Philips Research, and by Pet er Bongers from U -nilev er Research.
Ifyou wish to react to any ofthe papers in this vol-ume, please do not hesitate to contact us.
Edit
orial
This is the sevent h volume in the seri es Se/eet ed Topics in Identificati on , M odelling, and Control re-por ting oncur rent resear ch proj ects in theM echani-calEngi nee ri ngSyste ms and Control Groupat Delft Univers ity of Technology, Delft , The Netherlands . This volume cont a ins 13 papers ,of which 6 address eng inee ring modellingand control problemsand the ot her 7 consider methodsand techniquesfor con trol design and system identi ficat ion. This reflects the gen eral pattern of activity in our group,where the research is both ad dressing theoretical issuesin con-trol theory and system ident ificat ion , as weilas con -trolenginee ringissues in thefieldsof process control and motion serv o systems .
An industrial glass tube manufacuring process is considered in the paper by Richard Hakvoort and Dick de Roover, where issues of system identification-based model uncertainty are studied in their relevanee to robust cont ro!. A full scale wind turbine power generatingsystem is considered by Gregor van Baars, where the validation of dy-nam ic system models on the basis of closed-Ioop system identification experiments is discussed. A la rge pilot-scale industrial crystallizeris considered in the paper by RobEek, Jaap Both and Paul Van den Hof The research repor ted hereis partof a col-la bora ti ve project,jointlywith theParticIeTechno l-ogy, Process Equip ment,andCrystallizationgro ups at Delft Univers ity, and co-sponso red by European and US industria l process industries. Here, syste m ident ificat ion experiments are used not only to ob-tain dynamic modelsfor control purposes, but the ex pe rim ent alevidence also reveals im por t ant infor-mation on the actual physical mechanisms taking placein a continuous crystallizer .
The paper by Hans Heintzeand Ton Van der Wei -den considers thefeedbackdesign of impedance con-trol for a hydraulically actuated robot. The project is part of a collaborativ e effort with industry to de-velop a brick-laying robot. The expe rime nt al work rep orted has been performed on a la rgescale proto-type3-DOF SCARArobot ava ila ble in our laborato-ry. The pape r by Ger ivan Schothorst. Piet Teerhuis and Ton van der Weidenconsiders a hydraulicservo system whichisdesigned as an act uat ion m echanis-m for the SI M ONA Flight Si m ulat or to be locat-ed at the Simona Resear ch Cente r operated by the Dep artment of Aerospace Engineering at Delft U-niversity. This simulator is a large 6-DOF Stewart platform,now under construction and plannedto be availab leby theen d of1995. The
re-Okko Bosgra Paul Van den Hof Editors
bosgra@tudw03.tudelft.nl vdhof@tudw03 .tudelft.nl
Selected Topicsin Identification,Modelling andCon t rol Vol. 7, December1994
Robust performance
H
2 /H
oooptimal controll
Maarten Steinbuch § and ükko H. Bosgr-a!
§Philips Research Laboratories,Prof. Holstlaan4,5656AA Eindhoven ,TheNetherlands . ttMechanical Engineering Syst em sandContral Group,
Delft Uni versi ty of Technology,Mek elweg 2,2628 CDDelft,The Netherlands.
Abstract . This pap er considers therobust performance mixedH
2
/H
00 optimal cont rol probl em formulation. An ex plicit param etrization of H00 norm bounded , causal, r eal-rational uncert ain ties is used , based on LMIs. This leads to a const rained H2 optimal cont ro l prob lem. In case the uncertainty can be considered to lieon its bound,a newpar am etri zationfor lossless bounded real functions can be used. Using this param
etriza-tion ,it ispossibl etoformulatean unconstrain edoptimization problemfor thesolution of
the robu st performance mixed H2/ Hoo optimal control problem. The theory is applied to a Comp act Discrobust control problem .
Keyword s. H2/H00 control;optimalcontrol;robust control; compact disc player; uncer-tainty;opt im iza t ion .
1
Introduction
In the sixties the opt im al LQG control approach provid ed a systemati c solution to the control d
e-sign problem of multivariable systems , see for
in-stance Anderson and Moore (1971), Kwakernaak and Sivan (1972) and Levin e and Athans (1970).
Despi te of itsrelevan eein theformulationof p erfor-mance req uir em en ts,LQoptimal control was shown
to possess na guara nteed robustness margins if ap -plied in conjunction with an observer or Kalman filter, see Doyle and Stein (1981). The robustness issue has attracted a lot ofatte nt ion during the las t decade, resul t ing in the developme nt of H00 co n-trol theory, see for instanee Fra ncis (1987), Doyle and Glover (1988), Maciejows ki (1989) and Doyle et al. (1989). Robust cont rol based on the Hoo norm of a weighted closed-Ioop transfer fun ct ion is capableof handling the design problem for systems with uncert ainties. However , the use ofone single
number, albeitfrequency dependent, to address r o-bustness and perform an cefor multivariabl esyst ems
is rather restrictive. The design problem of stating lThis paper ispresentedat the 33rd IEEECon ferenceon Decision and Con trol, Orlando , FL, USA , December14-1 6, 1994. Copyrig htof this paper remainswith IEEE .
performance and robustness objectives in both the H2 and Hoo framework has recently been stated by several authors, see Bernstein and Haddad (1989),
Doyle et al. (1989), Mustafa (1989), Rotea and Khargonekar (1991), Scherer (1995), Steinbuch and Bosgra (1991a,1991b),Stoorvogel (1993),Paganini
et al. (1994), Yeh et al. (1992),Zhou et al. (1994).
In general one can define the problem as follows.
Suppose the plant is given by its transfer function G(s)with threesets of inputs and outputs.
W2 Z2 Wl G(s) Zl U y
I
F(s)I
I
I
The manipulated vari abl es are u, the measured outp ut s are y. The signa l sets (Wl, Zl ) are r e-lated to H2 or LQ type of performan ce crite ria,
whereas (W2, Z2) are rela ted to Hoo norm r equire-ments. Doyle and Zhou (1989) and Zhou et al. (1994) posed theprobl em Zl = Z2, Wl =FW2and gave a veryniceinterpret ation for an induced semi-norm on the transfer functi on. Their work showed that the H
2
/
H00 cont ro l probl em definiti on can bestated precisely in termsof signalsets . Bernstein and Had-dad (1989) have considered thedual mixed H2 /Hooprobl em for the situation where Wl
=
W2 ,Zl =F Z2· They used a performan ce crite rion relating w(=
Wl
=
W2) with Zl, under the const raint of an H00 norm bound on the tran sfer function from w to Z2.They deriv ed necessar y condit ions forredu ced order
controllers and necessa ry and sufficient condit ions for the fullorde r case. Must afa (1989) showed that
if Wl
=
W2 and Zl=
Z2 the auxilia ry performan ceindex of Bernst ein and Hadda d can be interpret ed
nicely as an entropyexp ression, yielding the cent ra l
H00 controller for the full orde r case.
In all these cases, the probl em solved does not add ress the true robust perform an ce control prob
-lem relevan t in practi cal applications : only a very
restri cted class of problem sdohave thesame inputs (or dua lly outpu ts) for the H2 performance m ea-sure s as for the H00 norm- bounded uncertain ti es.
The mixed H2/ Hoo problem with minimizing the transfer functi on in the H2 sense from Wl to Zl, while constraining the H00 norm of the transfer function from W2 to Z2 to some bound is unsolved . However , this problem does not address robust H2
performance.
In contrast to the aforem entioned results, in this paper a robust performance mixed H
2/
H00optimal control problem (or 'robust H2 problem ' as calledby Stoorvogel (1993) and Paganini et al. (1994)) is
considered including a parametrizationof theworst
case H00norm boundeduncertaintyrelatingthesig -nals W2 and Z2 (Steinbuch and Bosgra (1991b)),see Fig. 2.
Inthisproblem ,theworst- casenorm-bounded
un-certainty represents unmod eled dynamicsof thesy s-tem for which we are abIe to formul ate both their
magni tude bound and the ir structural inter conn ec-tion with the system dynamics.
The robusi performan ce mixed H2/Hoo control
probl em is to minimize the H2 norm of the tr
ans-fer functi on from Wl to Zl usin gthe feedback Kts), while maximizingthe H2 norm of thesam etransfer
functi on over the allowableuncertainties:
sup min11TW1 - Z1(K,L),)
11
2
(1) 11.::\11""<1'K( s )In the liter ature there exist vari ous formulations
for performance optimization problem s involving
worst-case disturban ces , resulting in a minimax
formulati on , e.g. Mills and Bryson (1994) and Sweriduk and Calise (1993).
From an application poin t of view this rolrust p
er-formancemixed H2/Hoo cont rol problem add resses many design problems in which H2 performan ce is
theoptimizationcrite rion, su bject to (un)s t ructu red H00 norrn-bounded uncertainties. This problem is
hard to solve, but becau se of the possibl e impli -cat ions for use in practi ce, it seems worthwhile to
further investigate the problem . Not e also the in-teresting work of Paganini et al. (1994) in which
the behavioral framework is used to asym p to tica lly
calculate the worst-case uncertainty for the robust
performan ce
lh/ H
00 cont rol probl em. In thispap erwe will explore the validity and usefuln ess of this
mixed H dH00 probl em formul ation using num eri-cal too!s. In earl ier work we considered the calc u-lati on of the worst- case uncer t ain ty, see Stein bu ch and Bosgr a (1994). In this paper we will exte nd the calcula tionofthe wor st case uncer t ain ty and in-vesti gate the related control design . A numeri cal
approa ch will be presented and applied to a mixed H2/Hoo Com pac t Disc control prob lem.
I
L),(s)I
I
Z2 W2 Wl G(s) Zl U yI
K(s)I
I
I
Fig.2: Robust performan ce H2 /Hoo cont rol
prob-Iem.
2
Pr
e lim in a rie s
Given a stabiestrictly proper transfer function ma-trixG(s) with statespace realizationC(sI - A)-lB,
then thefollowing performancemeasures can bed
(i) FTX
+
XF+ [
X G HT1
a:
1 [C
:
X ]<
0 Definition2.1 TheH2 norm of atransferfunct ionGts )is defined as:
(
1
J
OO
)
1/211 C 112
=
271" -
00
t7-(CT (- jw )C(jw))dw (2)untli tr(. ) the trace mat rix operat or.
(ii)
.J
>
0 with.J
= [
I _JT ]-J
I (10) (11) ( 12) The2-norm ean be eomp uted with Lyapunov equa-tions:
whereS is the eont ro lla bilit yGram ian and P is the observa bilit y Gram ia nsolving:
AT p + P A +CTC = 0
(4
)
Proof: Follows direet1y from the Bounded Real Lemmaas formulatedin Pet ersen et al. (1991). 0 In the sequel wewill denotetheset of all transfer functions .6.(s) with 11.6.
11
00
<
1 asV.This result direetly leads to the fol1owin g parametrization whieh eha racte rizes all real rati o-nal causal stable transfer functions .6.(8)of order n having 11 .6.
11
00
<
1:1. Choose J sueh tha t Definition 2.2 The H
00
norm of a transfer func-tion C(s) is defi n ed.as:
.J
>
0 (13)11 C
1100
=
maxë (C(jw)) wERuntliü the maximum singular value.
(5) 2. Let C and H be matrices of appropriate di-mensions eontaining free parameters, and let F
=
F.+
Fk withF.
=
Ft
and Fk=
-Fr ,sueh tha t3
Parametrization
of
H
oonorm
bounded transfer funetions
(14) and
3.1 Inequality formulation
Theorem 3.1 Let (F ,C,H,J) be an asymptoti-call yst abie min imalrealizat ion of thetransfer [utic-tion .6.(s )
=
H(sI- F)-lC+
J. Then thefol/owing sta te me nts are equiv alent.In Steinbueh and Bosgr a (1991b) a parametriz ati on wasint roducedfor stab lestrict1y prop er Hoo n
orm-bounded uneer t ainty mod eIs. In this paper wewill extend this parametrization to proper H
00
norm-boun deduneertainties using Linear Matrix In equal-iti es(LMIs),seeSection3.1. Basedon numerical e x-peri enees with this parametrization it seems worth-whileto ex ploit thebound ary eonditions, i.e. when the worst- cas eperturbation is losslessbounded real , see Section 3.2.. 1
1=1,2 "" ''2n
(15)
(16) for neven , and
3.2 Lossless bounded real formulation When the inequality const raints stated in thepr evi-ous sectionareall activethentheworst-case p ertur-bation is lossless bounded real (all-pass property). This means that the perturbation is on its bound at all frequ eneies and for all its singular values. In order to investigatethemixedH
2/
H00
problemalso for this case, wedevelop a suitable parametrization . Let us first eonsider losslesspositive real and lossless bounded real transfer funetions, see also Anderson and Vongpanitlerd (1973).for n uneven.
These equat ions follow from thetheorem aboveby . selecting a coordinate fram e sueh that X
=
l and by utilizing an additional orthogonal transformation that brings Fk into modal from.11
.6. 1100
< 1(6)
3X>
0 sueh that (7)[FTX+XF
XC HT]
CTX - I JT <0 (8) H J - I 3X>
0 sueh that (9) 1. 2. 3.Proof: foll ows by direct application of D
efini-ti on 3.2. 0
Lemma 3.4 Let.6.(s )
=
(I - f(s»)(I + f(s »)- l , then.6.(s ) is lossless bound ed realiffI'(s ) is /oss/ess posi-tivereal.
(1
8)
Ax+
Bçui;+
B2lU2 Cl x + DI2lU2 C2x+
D2llUi xr
.6.(s)1
I
r
lU2 z2 lUl M(s) zlFig.3: H2 worst case uncertainty.
Consi de r Fig. 3. The noise dist urbances lUI ha ve unit noise int ensi ty, and the uncertainty .6.(s) r ep-resen ts unstru ctured H00 norm bou nde d pert u rba-tions of the nominal closed-loo p system (M(s) i n-cludesG(s )and /( s»).
4
Worst case perturbations
and let .6.(s )
=
H(sI - F)-IG+ J. The n the c losed-loop syste m is:In cont ras t to Stoorvogel (1993), causality of the perturbation is assume d as an implicit and n eces-sary ingredi ent to pose the tr ue probl em. Thus we consider a setV ofreal-ration al ca usal stable t rans-fer function matrices .6.(s ) sat isfying 11.6. 1100
<
IJandVa if.6.T (- s ).6.(s)= J. As performanceindi ca-tor we use the varianee of the sign al Zl, i.e. the H2 norm of the closed- loo p transfer funct ion from lUl to Zl. Let Mt.s) have the statespace reali za ti on
(17
)
F - flT(1+
J) - Ifl-V2
fl
T(
1
+
J)-IV2(
1
+
J)-lfl (I - J)(I + J)- I F G H J Lemma 3.5 Let I'(s )=
fl(sl - F)-l(; + J, with F+
FT=
0 ,o
=
flT andJ
+
JT=
0, withF E IRn x n and
J
E IRm xm, and with fl and (;of com pat ible dimensions. Then the real matrices F, fl and
J
parametrizes all /oss/ess positive realtransferfunctions I' wit h st ate dimension n.
Lemma 3.6 Let .6.(s )
=
(I - f(s»)(I + f(s») -l withr(s)
=
fl(sl - F)-l(;+
J,
with F, fl andJ
as dejined in the previ ous lemm a. Then a stat e spacerealization for .6.(s)
=
H(sl - F)-IG+
J is givenby: Proof: .6.T(-s ).6.(s )
=
(I+ rT(_ s »)-l(1- rT(- s »)(1 - f(s»)(1+ f(S»)-1= (I + rT(_s »)-l(1+ rT (-s )f(s»)(1+ f(S»)-1=
(I + rT(- s »)-l(l+rT (-s»)(I+f(s»)(I+f(s»)-l=
I, and conversely. 0Defini tion 3.2 The realrationa / fun cti onI'(s ),sE <C, is loss/ess positive rea/iff(s)
+
rT (- s )= O. Definition3.3 The real rationol fun ct ion .6.(s),s E <C, is loss/ess bounded rea/ if .6.T (-s).6.(s)
=
I.The set of allsuch .6.(s ) is den oi edVa. or max tr[Cf[C]S (21) (F. ,F k, G, H ,J ) sup I1 T
W' _
Z
'(
.6.)112=
suptr[Cf [C]S (20) 6EV 6EV(
19)
( A+ B2JC2 GC2(
s
,
+
GBD2JD21) 21 lUl ( Cl+ D12JC 2 D12H ) ( ; )(
~)
=
denoted as [A], [B],
[
Cl.
Notice th at D 12JD2l=
0 for 11TW1 - Z1(.6.)112<
00.Let the system
(19)
be stabie and consider theconsirainedoptimization problem
A nd this isa param et rization[or allst able loss/ess
bounded real.6.(s ).
Theorem 3.7 Defin e the matrices0 and if; as up-per triangu/ar realmatrices, with zero on their di-agonal, and with appropriate dimensions, such that
F
=
0 - OT, andJ
=
if; - if;T , then the triple (0,if;, fl) parametrizes all stabie lossless bounded rea/ transfer functions .6.(s)
=
H(sI - F)-lG+
J, withH,F,G and J dejin ed by(17).
Proof: follows by some matrix manipula ti ons and
is omitted her e. 0
Since ma tri ces
F
and J are defined as s kew-sym met rie (see Lemm a 3.5) ,we further reduce the number of free variables and end this section with the main result.withF., Fç ;C,H,J according to (13)-(16),and with the state-varianee matrixS
=
sr
the solution to:and let ~(s)
=
H(sI - F)-IC+
J,then the p er-turbed system is given by:[AJS
+
S[Af+
[B][Bf=
0 (22) In case of ~ being lossless bounded real , i.e.~ EVa the optimization problem can be
reformu-lated as an unconstrainedoptimization problem: with y Apxp
+
B plWI+
Bpuc
.;», +
Dplu Cpx p+
DlpWI (27)(23)
max tr[Cf[CJs (8,,p,H )with S the solution to (22), and with (F,C,H, J) defined by (17), where
ft'
=
0 - OT,J
=
</J - </JT.5
Numerical exploration of the
con-trol design problem
In the preceding section we have considered worst-case per turbations ~(s) for the fixed system Mte ).
As Mt.s) includes the feedb ack cont roller J«s) , we have to investiga te how ~( s) and Kt.s )are related .
Recall the set V of allstabIe norm-bounded un-certainties. Let ~k (S) EV be the worst- case dis tur-bance, i.e. a disturb an ce attaining
X Ax
+
B1WI+
B2W2+
B 3u ZI CIx+ D12W2+
D 13U(26)
Z2 C 2x+
D21WI + D23U y C3x+
D3IWI+
D32W2 su p 11TWt-Z t( ](' ~) 112(24)
6E"Dfor a certain feedb a ck cont roller Kt;s ). Althou gh the underlying probl em may show a much more com plicatedst ruct ure, we assume herethat (at least locally) there exists a unique maximizing ~k(S).
Nowconsider all such~k(S) , and determinethef eed-backlaw JC thatwouldbe H2-optimalincase~k(S) indeed would qualify as worst-case uncertainty. By assigning an H2-optimal K" to each ~k(S), we it er-ate over ~k(S) until it satisfies the conditions for a worst-case disturbance, i.e. the gradients with resp ect to its paramet ers are zero for a fixed H
2-optim al](*.
Using the parametrization of the worst case un-certaint y we are now able to rewritethe control d e-sign problem in a numerically tractabIe way. Con-side r Fig. 2,with the noise disturbances WI having unit noise intensity and uncertainty ~(s) E V or
~(s) EVaresp ectively. The design problem is
(A; - CJIDpIBJ)X
+
X(A p -BpD;ICpt)-XBpBJX
+
CJI(I - DpID;I)CpI=
0(29)(A p - BpIDîpCp)Y
+
Y(A; CJ DIpBJI) -YCJCpY+
Bpl(I - DîpDIP)BJI=
0(30) and the H2 optimal control law u=
J(*(s)y is defined byDIp
= (
D31+
D32JD2I )Notice that D I2JD2I
=
0 for 11TWt-Zt(~) 112<
00 and that D32JD23
=
0 is assumed for simplicity. Thefeedback cont rollerKts)connects the measured variables y with the inputs u.Let the system (27) be stabIe and assume for the moment that the uncertainty paramet ers (F,C,H,J) are fixed. Then the optimization prob-lem
min 11 Twt_zt(u
=
J«s)y)Ib
(28)K
can be solved as a standard H2 or LQG type of problem:
(25) sup min 11TWt-Zt(](,~) 112
6E"D(. )K(. )
Fig . 5: Configuration of the CD MIMO control loop . Xrad xJoe f-L---.-o--.--spot L . . - _ - ' trac k- - -o--o-l
the controller uses position-error information pro-vided by four photo diodes. As input to thesystem the controller generates contro! curre nts to the r a-dial and focus actuator, which both are p ermanent-magnet/ coil systems.
In Fig.5 a block-diagram of the control loop is shown .Thedifference between theradia l(Xrad) and vert ical (Xjoc)spot positionand the referen ce track is detected by an optica ! piek-up which generates a radia !erro r signa l (erad) and a focus error sig-na! (ejoc)' A controller J(( s) feeds the system with the currents iradand ijoc,see also Steinbuch ei al.
(1994).
In thenumericalexperimentsa MIMO 2x2 mode! (G)of orde r 10 is used . We will consider two types of uncertainties . The first is a mu!tiplicativeoutput uncertainty. The uncertainty ~ is 50
%
.
The noise disturbances w] acting on the multivariablecontrolloop enter the loop at the input of the system and measurement noise act on the feed ba ck loop. The performance measures ZI are the inpu t to the plant
and the outp ut ofthe plant ('spot'). Our interest is in how the uncert ainty ~( s) can disturb the H2 p
er-formance from w ] to Z], and how the H2 optima! controller Kts) counteracts this. By using a g en-eral purpose optimization program the results are obtained, using the formulation of the previous s ec-tion. The inputs w] and W2 have been scaled such
that the nominal performance equals 1 and such that " ~ 1100 = 1 corresponds to 50%modelunc er-tainty.The results are presented in the tablebelow, and have been obtained with a first order (2 x 2) uncertainty; see Steinbuch and Bosgra (1994) for investigations with respect to the number of states in a worst-case ~. The results obtained using the LMIformul at ion showed that alltheboundary con-ditionsweremet, andafterapplicationof thelossless bounded real formulati on it appeared that identical solut ionswere obtained. Itsho uldbe noted thatthis resul t holds for this example. Itisnot clearwhet her more generalstateme nts can be mad e. This is s ub-ject of future resear ch.
Theresults aresum m arized in Tabl e1. In the Ta-bie the column'nom ina l,J(n om ,~
=
0'den ot es the valu eof the objec t ive functi on for the unperturbed max 11 Tw,_z,(u=
J(*(s)y) 112 (33)(8, q" H )
with J(* the so!ution to (29)-(3 1).
The optim ization prob!em inclu din g the un cer-tainty ~ E V can now be form u!ate d as a co n-sirai nedopt im iza t ion prob! em over a standard H2
optima!cont ro!prob!em:
6
E
xample:
Compact Disc player
ro-bust control problem
with IC the so!ut ion to (29) - (31) , and F.,Fk,G,H, J according to (13)- (16) .
Ir ~ E Va the optim iza tio n prob!em can be f or-mulated as an unconstrainedopt im ization prob!em:
Fig.4: Schematic view ofa rotatingarm Compact Disc mechanism. max 11 Tw,_ z,(u
=
IC (s )y) 112 (32) (F.,Fk,G, H,J) V (A p- Bp(BJ X+
DJ]CP]»v+
(YCJ+
Bp]D[p)Y (31) u.=
- (BJ X+
DJ]CP])vBoth the radi a! and the verti cal (focus) position of the laser spot, relative to the track of the disc, have to be con t ro lled active!y. To accom plish this , In FigA a schematic viewof a Compact Disc mec h-anism is shown. The mechanism is composed of a turn-tab!e DC-motor,and a ba!ancedradia!arm for track-following. An optica l elem ent is mou nted at the end of theradia!arm. A diode!ocated in thisel -eme nt generates a laser beamthat passesthroug h a seriesof optica!!enses to givea spot on the inform a-tion !ayer of the discoAn objective lens,suspended by two paralIe! !eaf springs, can be actuated ver ti-cally for focussing.
11Tw ,_ z , 112 nominal:J(n om , .6.
=
0 1.00 nominal: J(* , .6.=
0 1.06 perturb ed : J(* , .6.k
1.38 per turbed: J(n om , .6.k
1.48Table 1: H2performan cefor CD player with output
uncer t ain ty
case with the st anda rd H2 opt im al control [(nom .
The second colum n 'no m inal, J(*, .6.
=
0' means the robust performan ce opt im al controller solving (32) or (33) and with which the H2 perform an ceisevaluated wit hout a perturbation . The third co l-um n 'pe rt urbe d ,[(*, .6.
k'
means therobust perfor-man ce optimal cont roller solving (32) or (33) and with which the H2 performance is evaluated with the worst case perturbation , i.e. th is number isthe valueof (32) or (33). Finally ,thelast column 'pe r-tur be d ,J(n om ,.6.
k'
meansthenominal performance optim alcont roller [(n om for which the H2 p erfor-man ceisevaluate d for theworst case perturbation, i.e. this nu mber is the value of (32) or (33) with J(=
J(n om fixed.The results shown in theTable indicate that the worst case varianeeof theperformance variables Zl increases with a factor 1.48 if thenominal optimal H2 cont roller is used. In caseof the optimized con-troller for theuncertainty,thenominal performance degrades with 6%. The worst-case performance is 1.38,which is about 10% better than with the nom-inal feedback . In the following figure the nominal and perturbed transfer functions are shown from the input i-: «to the radial spot position.
Disc player MIMO control loop, but now with an uncertainty affecting the resonance at 860 Hz. The uncertainty perturbs especially the dampingof the resonance,with an amountalmostequa l to 100%I. Again,the noisedisturbances Wl acting on the mul-tivariablecontrolloopente r theloop at the input of thesystem and measurem ent noise act on the f eed-back loop. The performance measures ZI are the input to theplantand theoutput of theplant . The results of the ca lcula t ions are presented in Tabl e 2.
I1Tw , _
z
,
112 nominal: J(n om, .6.=
0 1.00 nominal: J(* ,.6.=
0 1.01 perturbed : IC ,.6.k
1.01 perturbed : [{n om ,Ók
17.90Table2: H2 performancefor CD player with r eso-nanceuncertainty
The varianeeof the performance variables Zl in -creases with a factor of almost 18(!) if thenominal optimal H2 cont roller is used , while in case of the
optimized controller for the uncertainty,the robust performance is only 1
%
less than the nominal p er-formance. Analysis of the worst-case uncertainty revealed that it is equal to a proportional term (1) driving the dam ping of the system close to zero. The H2 /H00 optimal controller counteracts this ef-fectively. In the following figure the nominal and perturbed transfer functions areshown.no~-:per ... . '!e!~!_ ~~.. __..~.. .__ IO-' ~ - -__ !oom_per ~\.opf.J)Cr
.,.
~nol 10.2 10-) oom "'W'" ,.•• • -,'.;\..... ~..'.'.....-. 10-) IO~ frequency(Hz ) 10' 10' Irequency(Hz)10' Fig.7: Nominal (-) and worst case perturbed (--, ... ) transfer function from WI to ZI
Fig. 6: Nominal (-) and worst case perturb ed (- -, ... ) transfer function from Wl to Zl
Thesecond exam ple considers thesameCompact
1Although this ca n be described as a pararnetri c un-certainty (real-valued) it is treated here as a complex uncertainty.
Conclusions
A new formulation of the robust performance
Hz!H00 opt im al con trol problem has been pro-posed in this paper , and an ex plicit parametriza-tion for a worst-casenorrn-bounded uncertainty has been used ,yieldingan (un)constrainedoptimization problem . A Compact Discsystem with an unstruc
-tured uncertainty has been discussed. Using a nu-merical algorithmwe haveshown th at it is possible to calculate worst case uncertainties and to rob us-tify the controllerto counteract with the worst case perturbation. The robust performance obtained in these examples shows that it is worthwhile to fur
-ther explore the theory of the robust performance mixed H2/Hoo control problem .
References
Anderson , B.D.O . and J.B. Moore (1971) . Lin
-ear Optima! Control. Pren ti ce-Hall , Englewood Cliffs,N.J.
Ande rson, B.D.O . and S. Vongpanitlerd (1973). Neiuiork An a!ysis and Synthesis' . Prentice Hall, Englewood Cliffs.
Bernstein , D.S. and W.M. Haddad (1989). LQG control with an H00perfor m an cebound: aRicca ti equation approach. IEE E Trans. Auto m. C on-trol, A C-:14, 293-305.
Doyle, J.C . and K. Glover (1988). State space for-mulaefor all stabilizing controllers that satisfy an Hoo-norm bound and relationsto risk sensitivity. Syst. Cont rol Leti., 11, 167-172 .
Doyle, .1.C., K. Glover , P. Khargonekar and B.A.
Fran cis (1989). State-sp acesoluti ons to standard H2 and Hoo cont ro l probl ems . IEEE Trans.
Au-tom. Con trol. AC-34 , 83 1-847 .
Doyle ,.1.C.andG. Stein (1981). Multivariable feed-back design: concept for a classical/modern syn-thesis. IEEE Trans. Autom. Control. AC-25 , 4-16.
Doyle, .1.C ., K. Zhou and B. Bodenh eim er (1989). Optimal con trol with mixed H2 and Hoo
perfor-mance obje ct ives. Proc. 1989 Amer. Contra! Conf.,pp. 2065-2070.
Francis, B. (1987). A Course in Hoo Control The
-ory. Lect . Notes Contr. Inform. Sciences, vo!. 88, Springer Verlag , Berlin.
Kwakernaak. H. and R. Sivan (1972). Linear Op-tima!Contra!Systems. Wiley-Interscience, New
-York.
Levin e, W.S. and M. Athans (1970). On the de-termination of the optimalconst ant outp ut f eed-backgains forlinear multivariablesystems. IE E E Trans. Autom. Control. AC- 15,44-48.
Maciejowski, .1.M. (1989). Multivariab!e Feedback Design. Addison-Wesley Pub!. Comp ., Working-ham, UK.
Mills, R.A. and A.E . Bryson (1994). Calculus of variations derivation of the rninirnax linear-quadratic (H
00
)
controller. 1. Guidance Contra ! and Dynamics, 17, 153-160.Mustafa, D. (1989). Relations betw een maximum ent ropy/H00 control and com bined H00/LQG control. Syst. Contra!Leii., 12, 193-203.
Paganini, F., R. D'Andrea and J.C . Doyle (1994).
Behavior al approach to robustnessanalysis. PT'OC.
1994 Amer. Contra!Conf., pp. 2782-2786.
Petersen , LR., B.D .O. Anderson and E.A.
Jonck-heere (1991). A first princip!es solution to the non-singular H00 control problem. Intern. J. R o-bust and NonluieerControl. 1, 171-185.
Rotea, M.A. and P.P . Khargonekar (1991). H2
-optimalcontrol with an Hoo-const raint, the state-feedback case . A utomatica, 27, 307-316.
Scherer, C.W. (1995). Multiobj ective Hz!
n.;
con-trol. Preprint 1993, to appear in IEE E Trans. A utom. Cont rol.Steinbu ch, M. and O.H. Bosgra (1991a) . Neces-sary conditions for stat ic and fixed order dy-namicmixedH2/Hooopt imalcont rol. Proc. 1991
Amer. Control Conf., pp.1137-1143.
Steinbu ch, M. and O.H. Bosgr a (l991b) . Robust performance in Hz] H00 opti m al cont rol. Proc.
31s t IEEE Conf. Decisi on and Cont rol.Brighton, UK, pp. 539-550.
Steinbuch, M. and O.H . Bosgra (1994). H2 per-formance for unstructured uncertainties. Proc. IM A CS Symp. Mathem. Modelling '1. MA TH-MOD VIENNA" Feb 2-4,1994, pp. 265-268. Steinbuch , M.,P.J.M.van Groos , G.Schootstraand
O.H. Bosgra (1994). Multivariable cont rol of a compact disc player using DSPs. Proc. 1994 Amer. Control Conf., pp. 2434-2438.
Stoorvogel, A.A. (1993). The robust H2 cont rol problem: a worst case design . IEEE Trans. Au-tom. Control. A C-38, 1358-1370 .
Sweriduk, G.D. and A.J. Calise (1993). Robust fixed order dynamic compensation: a differential game approach. Proc. 1st IEEE Regiona! Conf. Aerosp. Contra! Syst., Westlake Village. CA,pp.
458-462.
Yeh, H.H.,S.S.Banda and B.C.Chang (1992). Nec-essary and sufficient conditions for mixed H2 and Hoo optimalcontrol. IEEE Trans. Autom.
Con-trol, AC-37,355-358.
Zhou, K., K. Glover, B. Bodenheimer and J.C. Doyle (1994) . Mixed H2 and Hoo performance
objectives I+II. IE E E Tra ns . Autom. Control, AC-39, 1564-1 587.
@Delft Universi ty Press Selected TopicsinIdentification,Modelling and Control Vol. 7, December1994
A unified approach to stability robustness for
uncertainty descriptions based on fractional model
representations
Raymond de Callafonl, Paul Van den Hof and Peter Bongerst
Mechanical Engineering Systems and Control Group
Delft University of Technology, Mekelweg 2, 2628 CD Delft, The Netherlands.
Abstract. The powerful standard representation for uncertainty descriptions in a basic perturbation model as introduced by Doyle can be used to attain necessary and sufficient conditions for stability robustness within various uncertainty descriptions. In this paper these results are employed to formulate necessary and sufficient conditions for stability robustness of several uncertainty sets based on simpIe additive coprime factor uncertainty, gap-metric uncertainty as weil as the recently introduced and less conservative A-gap
uncertainty.
Keywords. Stability robustness,coprime factorizations, gap-metric
1
Introduction
In a model-based control design paradigm, the de-sign is basedon a (necessarily) approximative model
F
of a plant to be controlled. An apparentlysuc-cessful control design leads to a controllerC, having somedesired closed loop propertiesfor thefeedback
cont rolled model
F,
but due to the possible mis-match betw een the actual plant Poand the modelF
,
averificationof thesedesir edclosed loop prop er-ties is preferr ed before implem entingthe controllerCon the act ua l plant Po. In this paper thedi scus-sionisdirected towardstheverificationof oneof the
most importantclosed loop properties: stability. To evaluatest a bility when the cont roller C is being applied to the plan t Po, a characte rizat ion of the discrepancy betw een theplant Poand the mod el
F
can be used (Doyle, 1982; Francis, 1987; Doyle etal., 1992). Since the real plant Pois unknown , the
discrepancy in general ischaracte rized by aso called
uncertainty set , denot ed with P. Typically an
un-cert ainty set
P
is defined by the (nominal) modelP,
which is found by physical modelling or identifi-lThe work of Raymond de Callafonis sponsored by the Dutch Systems and Control Theory Network.
§Now with UnileverResearch, Technical Application Unit, Vla ardingen ,The Netherlands.
cation techniques, and some bounded 'area ' around it (Doyle et al., 1992). The uncertainty set Pitself
reflects all possible perturbations of the (nominal) model
F
that may occur. Some typical examplesof commonly used unstructured uncertainty sets that are norm-bounded by using an Hoo-norm can for example be found in Doyle and Stein (l981),Doyle et al.(1992):• Additive uncertaintyset:
PA(F,V,W,r):= {P
I
P=F+.6.. A, withIIV
LlAWil00 ::;r}.
(1) • Multiplicative (out put) uncertaintyset :PM(P, V,
w
,
r) := {PI
P =[I
+
LlM]F ,with IIVLlMWlloo ::;
r}
.
(2)• Additive (right) coprime fractional uncertainty
set, where
F
has been factorized first in a right- - - - 1
coprimefactorization P := N D and theset
is defined as
PCF(JÎl,ÎJ,VD, VN, W,r):=
{P
I
P= [JÎl+
LlN][ÎJ+ LlD]-1, withClearly, by defining the uncertainty set in such a way that at least the plant Po E P, stability ra-bustness results for the set P will refiect sufficient cond it ions under which the plant Powill be stabi-lized by C,(Doyl e et al., 1992). Inthis persp ective, special attention will be given in this paper to an uncertain ty set PcF which ischa rac te rized by
per-turbations on an additivecoprimefactor description of the nominal model
P,
like in (3). The sp ecif-ie app lica t ion of the uncertainty set description of(3) will be motivated by the favourable properties it has over a standard additive (1)or multiplicative
(2) uncertainty set description.
Stability robustness results for uncertaintysets em-ployin gweighted and unstructured additive pertur-ba tion s on a cop rime factorization ,gap-meti cbased uncert ain ty set s and the recently introduced A-gap uncer t ai nty sets will be sho wn to be closely related to each othe r . The cont ribu t ion of this paper is in the unified treatment of the sit uat ions of the dif-feren t uncer t ainty sets, by em ploy ing the weigh tin g an d the factoriz ation used in an uncertainty PCF
as given in (3). While stability robustness results for uncertainty set susin g addit ive perturbations on norm aliz ed(Jeft) coprime fact oriz ations (Gloverand McFarlan e,1989) andgap-met ric based uncertainty sets (Georg iou and Smit h, 1990a) have separately been derived before, thispaper amp lifiestheir rela-tion, as weil as the extension to a less conservative A-gap (Bongers, 1991; Bonge rs, 1994) uncer t ainty set descript ion.
The out line of this pap er will be as follows . In
section 2 som eprelimin arynotation s an d definitions will be given, while in section 3 the basic stability robustn ess result usin g the powerful perturbation mod el (Doyle , 1982) will be sum m a rized . This per-tu rba t ion mod el gives rise to an unified approach to handle stability robustness for various uncertainty descriptions , including additive weighted perturba-tions on acoprime fa ctorization. Section4contains the results ofap plying this unified approachto ad-ditive uncertainty descriptions on fractional model representations likein (3) and favourable properties areilluminated. Thelink with gap and A-gap based stability robustness results is discussed in sections 5 and 6, the latter one being less conservat ive than the former one , as sho wn in secti on 7. The paper ends with some conclud ing remarks .
2
Preliminaries
Through out th is pap er , the feedback configur at ion of a plant Pand a cont roller C is denoted by T(P,C) and defined by the feedback conneetion st ruct ure depicted in figure 1.
P and Care represented by real rational (discret e
1'1 U
I
P
I
+'
+
Î
I
I
I
- 1' 2Ye
I
Cr
Ue ~+Fig. 1: Feedback conneetio nstruct ure T(P,C)ofa plant Pand a controller C
time) transfer function matrices, and it will be as-sumed that the feedback conneeti on is well-posed, i.e. th at det[I
+
CP]=1= O.Infigure 1 the signals u and y refiect resp ectiveiy the inputs and out puts of the plant P. The signals
Ue and Ye are resp ectively the inputs and out p uts of the cont roller Cp, and 1'1 and 1"2 are external
referen ce signais . The feedb ack system T(P,C) is defined to be internallystabie if the mappingfrom COI(1'2, 1"1) to col(ue,u) is BIBü stabi e , i.e. if the corresp onding transfer function is in IR7-l
oo ,
bein g the I-Iardy space of real rational transfer function matri ceswith bounded 7-loo-n
orm (Franc is, 1987) :II
Glloo :=
sup o-{G(ei w)} (4)wE[D, ")
with 0- the maxim um sing ula rvalue.
Thedyn ami csoftl~eclosed loop systemT(P ,C) will also be describ ed by the mapping from eol( r2, 1"t}
to col( y, u) which is given by the transfer functi on
matrixT(P,C):
T(P, C) :=
[~]
[I+CP]-l[
C
I] . (5)Using the theory of fractional representations , as e.g. presented in Vidyasagar (1985), a plant P is
expressed as a ratio of two stabie transfer functions
N and D. For two transfer functions N,DE IR7-loo
the pair(N,D) is called right coprimeover IR7-loo if there exist X, Y E IR7-loo such that X N
+
Y D=
J.Thepair(N, D) is a right coprimefactorization (ref) of P if additionally det{D} =1= 0 and P
=
N D-1. A
right coprimefactorization (N,D) is called normal-ized (nref) if it satisfies N* N
+
D*D=
I,wher e * denotes com plex conjugate transp ose. For (normal-ized) leftcoprimefactorizations (lef) dual definition-s exist . With resp ect to internal stability of the feedback system T(P,C) the following lemmawill be used .Lemma 2.1 LetP have ref(N,
D)
atul Icf(D,N),and let C have re] (Ne,De) and lef(De,Ne) . The
t. thefeedback systemT(P,C) given in figure 1is
internallystabie
H. T(P,C) EIR1l00 1 with T(P,C) defined in (5). m. A- I EIR1looJ with A:=
[
'D
e N
e
1[
~
]
IV. Ä-I EIR1l001 with Ä:=
[D N
1
[~:]
Proof: iÇ:} ii: Fram figure 1 and lemma 2.1it can
be seen that
~
~-~
d z rl UY
Pr-
Yr
+
-
r2 Yc C Ue ~+ d z3
Stability robustness
in standard
form
IGiven two topologies Ol andO2 ,Ol is said to be weaker
than 02if Ol is a subcollection of 02, see also Vidyasagar
et al. (1982)
Definition 2.2 Considerthetwo plants PI,P2 with
nrcf's(NI,DI) ofPI and (N2,D2 ) ofP2 . Then the
gap between PI andP2 is expressed by
(7)
(6) U Ue M IIMll~lIoo<
1Fig. 2: (a) Feedback conneetion structure of a
perturbed plant Pt>. and a controller C;
(b) general feedback conneetion structure
K(M,~).
(b) !C(M,~) is BIBO stablefor all~ with 11~1100
::;
"y if and only if
It is weilknown that thefeedbacksystem depictedin
Figure 2(a) can be recasted into the situation given
in Figure 2(b), where the uncertainty ~ has been
pulled out (Doyle et al., 1992) using the artificial
signals d and z. Provided that the transfer M is
BIEO stabie, intern al stability of the configuration
in Figure2(a) isequivalent to BIBO stability of the
feedbacksystemK(M,~)in Figure2(b),wher eM
ll
will denote the map from from d onto Z only.
The small gain theorem can now be applied to
char-acterize stability results for the conneetionstructure
of Figure 2(b) and has been summarized in the
fol-lowing lemma.
Lemma 3.1 Let the stabie transfer functions
M,~ E IR1loo construct a feedback connection
K(M,~). Then
(a) A sufficient condition for BIBO stability of
K(M,~) is given by
max{5(PI,P2 ) ,5(P2,PI)}, with
QEiR~oo
11 [~:
] - [~~
]Qll
oo
5(PI,P2)5(Pi ,Pj)
Fractionalrepresentations havea close relation with
approxim at ion in the graph topology. The graph
topo logy is the weakest top ology' in which a vari
a-tion of the eleme nts of a stabi e feedback c
onfigura-tion aro un d their nominal values,preserv esstability
of that closed loop system (Vidyasagar et al.,1982).
The graph topology is known to be induced by s
ev-era l metri cs, as e.g. the graph metric introduced
in Vidyasagar (1984) or the gap metric introduced
in Zames and EI-Sakkary(1980),being expressed in
the following way (Georgiou, 1988).
For describingthestability robustness of several
un-cert ainty sets based on fractional model
representa-tions, the standard results on stability robustness
for a general situation as depicted in Figure 2 will be used .
Henc e T(P,C) E IR1ioo if and anly if H(P,C) E
IR1ioo which is equ ivalent to T(P, C) being int
er-nall ystabie .
i Ç:} iii, iv: See Vidyasagar (1985), Bangers (1994)
+
with A
=
[beD+
JÎleJÎl] and (De,Ne)a lef ofC. As the weightingfunctions arestabieand stably inv ert-ible, and A-1 is stabie according to Lemma 2.1 itfollows that M1 1 is stabie. Application of Lemma
3.1then proves the result. 0
Fig. 3: Additiv e pert ur bations on a righ t coprime fact oriz ati on
+
.6. r------ --- --- --- -- ----, I I , I I I I I I I I I I I I I l _Denoting col(d1 ,d2 )
=
.6.z with the appropriatesignals defined as indicated in figure 3, the co r-responding stabie transfer functi on Ml l satisfying
z
=
Ml1col(d 1,d2) can be written asIn the next section it will be shown how these r e-sults can be exploited to deriv estability robustness conditions for gap-metri cuncertaintysets as weIlas for uncertaintysets basedon further generalizations of the gap-metric . To this end , the result on the e -quivalencebetwe en several formulations of thesame uncertainty sets wiII be presented first.
Theresult of secti on 3onstability robustness can be applied to various Hoo-norm bounded uncertainty sets byrewritingthe uncertaintydescription into the basi c perturbation modelIC(M, .6.). In this section th is is done for uncer t ain ty sets based on coprime fact or uncertainties.
A crucia l aspect in the result of lemma3.1 is the assum pt ion that .6. E lR1ioo. In case of an add i-tive (1) or multiplicative (2) uncertainty set in the basic per turbation mod el , this assumption implies the condit ion that the locati onof all unstable poles of Pare assume d to be fixed. Additive per turba-tions on coprime factoriza tio ns (3) are mor e fl exi-bie and allowchanges in both the num be r and the locat ions of poles and zeros anywhere in (f; (Chen and Desoer, 1982). Mor eover , fractional r epresen-tations havea close rela tion with approx imationin the graph topology. Dist an ce measure s (or metrics) like thegrap h and gap metricgiven in defini tion 2.2 ind uce this same grap h topo logy and can also be used to evaluatestability robust ness proper ti es ofa closed loop system (Vidyasagar, 1984 ; EI-Sakk ar y, 1985; Vidyasagarand Kirnura ,1985).
First ly, the uncer t ain ty set PCF (3) ,as menti oned in the introdu cti on ,will be discussed.
Corollary 4.1 Consider a plant
ft
with rcf( JÎl ,D),st abilized by a given controller C, and consider ihe
uncertainty set
PCF( JÎl ,D, VD,VN,W,,) :=
{P 1P
=
[JÎl+
.6.N][D+
.6. D]-l, with A proof of this lemma can befound in Doyle et al.(1992) and Maciejowski (1989) .
4
Stability robustness for
uncertain-ty descriptions based on fractional
model representations
for stabie and st ably invertibl e filters VD, VN, W. Theti ihefeedback system T(P,C) is internally sta-biefor all P EPCF if and only if
,
<
IIW-1 [D+
CJÎl]- l[I
C
]
[V~l V~
l]
1[1
Proof: Definin g
.6. :=
[~D ~N] [~~
]
W, such that 11.6.1100 ::; " (9) the basic perturbation structure of the uncertain-ty set PCF ca n be written into a form that cor -resp ond s to IC (M,.6.) for a specific form of MIl.Proposition 4.2
Theuncertaintyset PCF(JÎl,D,VD,VN,W,,) as d
e-finedin Corollary
4.1
can alternativelybewrittenin thefollowing equivalent forms:(a) PCF(JÎl,D, VN, VD, W,,)
=
{PI
(b) PCF(N, ÎJ, VN,VD,W,I)
=
{P
I
P=
N nD;;,1, (Nn, D n) nrcf, and thereexists aQE IR1ioo such that
(a) PCF(N,ÎJ,VN, VD,
w
,
I)=
Pg(P, I);(b) For1
<
1,Pg(p ,I
)
=
P~(p,I)·
Proof: Part (a). According to Proposition 4.2(b)
and taking into account thespecificchoieeofw
eight-ing functions in the lemma,it follows that PCF(N ,ÎJ,VN,VD,W'/)
=
{P 1P
=
NnD;;,l, (Nn,D n) nrc], and there exists aQ
E IR1ioo sueh thatProof: Part (a) follows by simpIe calculation.The
proof of part (b) is more involved. Inthis derivation the freedom in (11)is used to denote:
(NW
+
V;Vl!::iN)=
NnQ (ÎJW+
VD1!::iD)=
DnQwith (Nn,Dn) a nrcfand
Q
E IR1ioo. Such factorsean always be found . It follows then that
Since
(N
,
ÎJ) is chosen to bea nrcfof Pit is straight-forward to verify that PCF=
Pg.Part (b) is proven in Georgiou and Smith (1990a).
The restrietion to 1
<
1 is ca used by the fact thatthese sets with 1
>
1 can not be stabilized by asingle controller. 0
whieh proves the resul t . Note that the factor Q
cance ls in the representation of P. 0
!::iN VN[N nQ- NW] !::iD VD [Dn Q - ÎJW].
for whieh the following relation with the coprime
factor uncertainty sets can be shown, as presented
befor e.
Lemma 5.1 Let a plant Pand a controller
C constitute an internally stab/e feedback sys-tem T
(P,
C). Consider the uncertainty set PCF(N ,ÎJ,VN, VD,W'/) under the additiona/ con-ditions th at(N
,
ÎJ) is a nrcf of P, and VD=
J, VN= J, and W = J. Then(16) 1
<
IIT(P,C)II
;;:,I .
Lemma5.1 relates the set defined by a gap metri c
bound with theset ofcop rime factor purturbation s
by a special choi ee of the weigh ting functions VD,
VN ,Wand the coprime fact oriza t ion(N ,ÎJ) of t he model
P.
This gives rise to an unified approach to handlesets of plants that are bounded by a ga pmetric ,and the stability robustness result for these
sets follows now directly from Corollary 4.1.
Corollary 5.2
Consi der the situativn of Lemma 5.1 with 1
<
1. Then for each of the three sets of plants PcF, Pg and P~, T(P,C) is internallç stable for all PEP if and only ifNote that the result of this corollary is not new. It was shown already in Georgiou and Smith (1990a), where a complete proof of the stability robustness
result is given. It has been shown here that the stability robustness results in the standard form has simply be exploited, as formulated in section
section 3. Restricting attention to the situation that 1
<
1is natural, as Bode's sensitivity integral shows that IIT(P,C)lIoo>
1, (Maciejowski, 1989), show-ing that stability robustness can only be achieved for sets with1<
1.Proof: The proof follows simply by substituting
the specific weightings in the result of Corollary 4.1,
employing the fact that premultiplication of thee
x--T -T
pression within the norm by [N D
jT
leaves thenorm invariant, due to the normalization of the ref.
o
(13) (14){P
I
b(P,
P) ::; I}
{P 16(P,P) ::;/}, Pg(P,I) P~(P,I)
5
Stability robustness based on
dis-tance measures
In this sect ion stability robustness results for g
ap-metric uncertainty sets are formulated. The main
result of this section is not new, but al ready proven
sepa ra te ly in Georgiou and Smith (1990a). The
close relation of the stability robustness result here with the formulation in the previous seetion
eon-cerninggen eral coprimefactor uncertainty sets will
beilluminated .This relation will be employed in the
next section to formulate similar results for
uncer-tainty sets based on theso-called A-gap,as recently introduced in Bongers (1991) and Bongers (1994). The following uncertaintysets are being considered
Finall y it sho uld be noted that the gap and graph metri c are induced by the same topo logy and are
uniformly equiva lent (Georgi ou , 1988;Packard and Helwig , 1989 ). Therefor e stability robustness in the gra ph metricyields a similar result as mentioned in
corollary 5.2.
6
Stability robustness in the A-gap
The resuItsobtainedin theprevioussection for gap-based stability robustness can be further extendedfor uncertaintysetsbasedon the recently introduced A-gap, (Bongers, 1991;Bongers, 1994).
Definition 6.1 Let two plants
Pi
,
P2 have nrcJ's (NI ,DI), (N2 ,D22
re~pectively. Let C be acon-troller with nlcf(De,Ne) such that T(Pl ,C) is in-ternally siable. Then theA-gap between the plonis
Pl, P2 is definedto be expreseed by
unth. A
=
[DeDl+
NeNd .The differen ce between g(P1,P2) and 8t.(P1,P2) is
the addit iona lshaping of the nrcf of PI with A-I int o arcf(N,
iJ)
.
In this wayÄ:= DeÏJ+NeN = I,with
N
=
NIA-1,IJ
=
DIA-1, which is usedto conside r the close d loop ope ra tio n of PI in
-duced by the controller C bein g em ployed. This
mak es the distance betw een PI and P2 dep end en-t on the nrcf of the cont roller C. Note that the
distan ce measure 5A(Pl,P2) is not a metric since 8t.(Pl,P2)
-I-
8t.(P2 ,Pi) due to theinfluenceofthecont ro ller C (Bon gers ,1994).
Accordingly , an uncertainty set bas ed on A-g ap
un-certainty can bedefined as:
PgJP , , ):= {P
I
8t.(p ,P):Ç, }.Thisuncer tainty set can alsobeshown to be e
quiv-alent to an uncertainty set of coprime factor un
cer-tainties , provided appropriate weighting functions
are chosen .
Proof: The proof of (a) is straight forwa rd, along
the same lines as the proof of Lemm a 5.1(a ). R
e-sult (b) thenfollow sdirectlyfromCoro lla ry4.1,e m-ploying the fact that A[b
+
CN]- l [I C]=
[DeNe]having an co-nor m of 1 due to the fact tha t it is a
normalized left cop rime factorization. 0
As said before,in case of the A-gap,the uncertainty
set defined accordingly considers perturbations of the nominal plant
P
that are controller dep end ent.The introduetion of weightings in the gap met
-ric has also been studied in Gedd es and Postleth
-waite (1992), Georgiou and Smith (1990b)or Qui and Davidson (1992). In Gedd es and Post!ethwait
-e (1992) a multiplicativeuncertainty description on thenrcf
(N,
0)
of themodelP
is being used ,leadingto an uncertainty structure ~ having a diagonalf
or-m. Dueto thediagonal form only necessaryands uf-ficient conditions based on the st ru ct u red singular value J.l{.} ca n be obtained. The weightings in the
weighted gap of Gecirgiou and Smith (199Db)have
to be defined a posteriori which makes the cho ice of the weightingfunctions , to access robustness i
s-sues on the basis of a weighted gap, not a trivial
task. Informationon the si ze ofthe coprime fact or
perturbationscan be used in the weighted pointwise gap metricdefined in Qui and Davidson (1992) ,but st ill an efficientcom pu ta t ional method for pointwise gap metricis not availa bleyet.
7
Conservatism issues
All stability robustness results in this paper refiect
necessary and sufficient con dit ions of an un
certain-ty set to be stabilized by a single cont roller. As such no conservatism is introduced in the test for
checking stability robustness itself. However, for a single given cont roller, differ ent of such un cert.ain-ty sets contain a different portion of the set of all
syste ms that is stabilized by the cont roller. In th is
persp ectivethe conce ptofconservatism is an intrin -sic property of the uncertaintyset beingused. As a
result an uncertaintyset will be called morecons
er-vative than another if one controller stabilizes both sets ,while the former set is contained in thelatter.
Theorem 7.1 (Bongers, 1991) Consider a plant Pand a stabilizing controller C with nlcf (De,Ne).
Consider thefollowing two uncertaintysetsresulting
[rom ihe stability robusin ess resulis in the previous
sections:
Lemma 6.2 Let a plant Pand a controllerC with
nlcf (b; Ne) cons tit uie an internally stabie f
eed-back syst em T(P, C). Consider the uncertainty set
PcF( N ,D,VN, VD,
W
,,)
uiuler the additional c on-ditions that (N,D) is a nrcf of P, and VD=
I, VN=
I, and W=
A-I with A=
[DeD+
NeN]. Tlten(a) PCF(N,D,VN, VD,W, ,)
=
PgJP, ,);(b) T(P,C) is internally stabie [or allP EPCF if and only if,
<
1.S6(P,C) SgJP,C) then {u P6(P,b), b