Appliation of the
H ∞-norm for the identiation of linear time-invariant and linear parameter-varying
models
Ph.D. dissertation
prepared under the joint supervisor sheme between
University of Poitiers
and
Budapest University of Tehnology and Eonomis
Author :
Dániel Vizer
Advisors :
Guillaume Merère, Ph.D., HdR.
Laboratoired'Informatique etd'Automatique pour les Systèmes,LIAS
(Université de Poitiers)
and
Bálint Kiss, Ph.D.
Irányítástehnika ésInformatikaTanszék
(Budapesti M¶szaki ésGazdaságtudományi Egyetem)
2015
Université de Poitiers
Travaux Sientifiques
présenté
l'Universitéde Poitiers en vue de l'obtention du
Diplme de Dotorat
par
Dániel Vizer
Appliation of the
H ∞-norm for the
identiation of linear time-invariant and
linear parameter-varying models
Soutenuepubliquement le10 déembre2015
Jury
Président :
Rapporteurs : Xavier Bombois Direteur dereherhe
Laboratoire Ampère, Group ACM
Levente Adalbert Kovás Maître de onférenes, Hdr.
Óbudai Egyetem,
Examinateurs : Marion Gilson Professeur des universités
Université de Lorraine,CRAN
Thierry Poinot Professeur des universités
Université de Poitiers, LIAS
József Kázmér Tar Professeur des universités
Université d'Óbuda
Direteurs dethèse : Guillaume Merère Maître de onférenes, Hdr.
Université de Poitiers, LIAS
BálintKiss Maître de onférenes
Université des Sienes Tehniques et
Éonomiques de Budapest,IIT
"A warrior does not give up what he loves, he nds the love in
what he does."
Dan Millman - The way of the peaeful warrior
Aknowledgements
Here, I would like to mention several people in my life who have ontributed
somehow, eitherdiretly orindiretly,tothe birthof thismanusript. Letme todo
this by using the orrespondinglanguages. But rst of all, thanks for the existene
of the oee that has been the mostly used drug during these years of researh. It
has gave melotsof exellentideas and motivation.
Je voudrais tout d'abord remerier le Dr. Guillaume Merère qui m'avait na-
lement onvainu de hoisir la thèse au lieu de travaillerdans l'industrie et qui est
devenu mon direteur de thèse. Puis, au l du temps, ave ses onseils et ses avis,
ilm'a faitprogresser professionnellementainsi quepersonnellement.Il fautpartiu-
lièrement souligner ses eorts pour me pousser à rédiger orretement. Celui était
un vrai dé pour moi et également pour lui. Enn, je pense que je ommençais à
aimer ladoumentation aussi. Tu es un supermentor pour la reherhe. C'était un
vraiplaisir de travaillerave toi.J'espère que nous pourrons ontinuer lareherhe
ensemble n'importe oùje me trouverai dans l'avenir.
Jedois égalementmentionerleprof.EdouardLarohequim'avaitaueillitelle-
menthaleureusement haque fois sur le site de l'ICUBE etl'IRCAD à Strasbourg
et qui m'avait beauoup aidé ave ses onseils et ave nos disussions très inté-
ressantes lors de mon travail. Il ne faut pas oublier non plus le Dr. Olivier Port,
étantmathématiien, gràeàquije peuxdésormaisvoirlesproblèmesposés un peu
diéremment.
Ctéoeur,quoiquel'onseonnaisseilyapeudetemps,ilfautquejesolliitema
famillefrançaise"aussi.Plus préisément,ChantaletGeorgesquim'avaitaueilli
hezeuxommeleurlsetave quij'aipasséde bons moments.Jene trouvepasde
mots à exprimer mes sentiments vers vous. Meri d'être! J'espère que nous restons
en ontat enore pendant quelques dizaines d'années.
Jevoudraiségalement iterles trentaines de personnes que j'airenontrées pen-
dant mon séjour à Poitiers et qui ont tous ontribué en quelque sorte à mon déve-
loppement personnel. Vous êtes tous dans mon oeur (ainsi que dans mon livredes
mémoires) pour toujours!
Etbien sur,je parlede toiapart espèede Cédri.Alors,tuétais toujourslàau
ldu temps pour m'aider, disuter, s'amuser,uisiner etm'embêter... Tout simple-
ment,meripour tout. Lemondeest sansn et ela e quiest génial.Ne t'inquiéte
pas, onsera toujoursdans le oin.
ésMS. tanulmányaimatvégigkísérveszakmailagelvezetettadoktoritanulmányaim
küszöbéig. Ezen felül azelmúlt három évben is mindig számíthattama tanásaira,
segítségére. Köszönet illeti továbbá Dr. Kiss Bálintot, aki konzulensemként sokat
tett azért, hogy ajelen munkaelkészülhessen.
Szeretném megköszönni a saládomnak, hogy felneveltek, hogy id®r®l-id®re fel-
töltenek energiával és akik miatt olyan vagyok, amilyen. Különösen szerensésnek
érzem magam, mert úgy gondolhatok rátok, mint a nagyon közeli barátaimra. Itt
szeretnémkülönkiemelniaszüleimet,agénekért,azérthogybeneveztekazÉletnev¶
játékba, ésazért hogy támogattakmindenben és tudat alatt igazgattákaz utamat,
ami nem kis teljeseítmény tekintve, hogy mennyire önfej® voltam és vagyok. Nem
tudom, hogyansináltátok, de sak hafele olyan sokattudok adni agyerekeimnek,
mintTinekem, akkorisavilágegyiklegjobbszül®jeleszek. Köszönömagyökereket.
Köszönöm a szárnyakat. Szeretlek titeket! Maradjatok ilyenek még nagyon sokáig
nekünk!
Végül, de nem utolsósorban, hálás köszönet az életem Párjának, Feleségmnek,
Böbinek
♥
is. Köszönöm, hogy vagy és leszel nekem, nekünk. Mert hiszen tudod,hogy legértékesebb dolog, amit egy emberadhat a másiknak azazideje.
Boldogvagyok. Itt és most. Jesuis heureux. Ii etmaintenant.
Résumé
Depuis près de 20 ans, les modèles linéaires à paramètres variants (LPV) font
l'objet d'une étude très poussée en identiationpuisqu'ils peuvent être onsidérés
omme le hénon manquant entre la modélisation linéaire à temps invariant et la
modélisationnonlinéaireet/ouàparamètrestempsvariant. Cettethèse,réaliséeen
o-tutelle entre l'Université de Poitiers et l'Université Siene Tehnique de Buda-
pest, s'insrit dans ette mouvane et herhe à développer de nouvelles méthodes
de modélisationet d'identiation des modèles linéaires à paramètres variants (et,
par extension, des modèles linéaires à temps invariant (LTI)) sous forme de rep-
résentation d'état, à l'aide de la norme
H ∞
, ave une attention partiulière aux strutures méatroniquesexibles. Ene quionerne lesmodèles LPV, dans ettethèse, l'approhe dite loale est plus partiulièrement onsidérée. Elle onsisteplus
préisément (i) à obtenir des données d'entrées-sorties dites loales aquises en un
ertainnombrede pointsde fontionnement hoisispar l'utilisateur, (ii)àidentier
desmodèlesloauxboite noires,(iii)àinterpoleresmodèlesloauxan dealuler
lemodèleLPVnal. Dansettethèse,de nouvellessolutionsontétéproposéespour
améliorer haune des étapes sus-itées. Le hoix optimal des points de fontion-
nement a ainsi été étudié et une nouvelle méthode utilisant la nu-gap métrique a
été proposée. Leproblème de ohérene des basesinhérentà l'approhe loaleaété
onsidéré etdeux approhes utilisantlanorme
H ∞
ont ainsi été développées. Dans un premier temps, en supposant l'aès, loalement, aux strutures boîte grise desmodèles LTI à identier, le problème de restruturation des modèles loaux boite
noire en modèle boite grise est résolu à l'aide de l'algorithme "proximity ontrol"
avant d'appliquer une interpolation, de type moindres-arrés, des paramètres lo-
aux restruturés. Dans un deuxième temps,une interpolationdirete des modèles
loaux boite noire est développée en minimisant, à l'aide toujours de l'algorithme
proximityontrol,ladistane entre leomportemententrée-sortiedu modèle LPV
à identier et les modèles boite noire loaux pré-estimés. Cette dernière tehnique
est développée pour fontionner ave des modèles LPV ayant une représentation
linéairefrationnaire(LFR).Lesdiérentes tehniques développées dansette thèse
sontnalementtestéesenutilisantàlafoisdesdonnéesdesimulationetdesdonnées
expérimentales réelles d'un robotexible.
Kivonat
A lineáris paraméterváltozós (LPV) modellek identikáiója kiemelt gyelmet
kapott az elmúlt 20 esztend® során. Ezek a modellek lényegében lánszemnek te-
kinthet®k a lineárisid®invariáns(LTI)ésa nemlineáris vagylineáris id®benváltozó
(LTV) modellek között. Jelen doktori disszertáió egy kett®s-témavezetés¶, nem-
zetköziegyüttm¶ködés (o-tutelle)kereténbelülvalósult mega Budapesti M¶szaki
ésGazdaságtudományiEgyetemésaPoitiers-iEgyetemközött. Akutatásf® irány-
vonala az LPV és LTI modellek állapottérben történ® identifkáiója volt, exiblis
mehatronikai rendszerek számára, mely a
H ∞
norma bevonásával valósult meg.Az LPV estben, az ún. lokális eljárást vizsgáltam behatóbban, mely a következ®
lépésekb®lépül fel : (i) ki- ésbemenetiadatsorok gy¶jtése az identikáióhoz el®re
meghatározottmunkapontokban,(ii)lokálisfeketedobozmodellekidentikáiójaaz
el®z®lépésbenmértadatokalapján,(iii)alokálismodellekinterpoláiójaazütemez®
változóm¶ködésitartományamentén. Doktoridisszertáiómban, afentebbemlített
lépésekmindegyikére javaslokjavításilehet®ségeket. A lokálisidentikáióbanrészt
vev®munkapontokoptimáliskiválasztásaérdekébenegyújalgoritmustdolgoztamki
a nu-gap mérték alkalmazásával. A lokális modellek interpoláiója során felmerül®
koherens bázis problémát két
H ∞
norma alapú módszer kifejlesztésével, oldottam meg. Els®ként a fekete-szürke doboz áttérést valósítottam meg, lokálisan, az ún.proximity ontrol algoritmussegítségével. Ezt azátstruktúrálást követ®en a loká-
lismodellekinterpoláiójátalegkisebbnégyzetekmódszersegítségévelvalósítottam
meg. A második módszer a lokális fekete doboz modellek direkt interpoláióját
végzi elszinténaproximityontrol algoritmusalkalmazásával. Ezutóbbimódszer
lineáris törtalakban felírt LPV modelleket ad eredményül. A disszertáióban ki-
dolgozott módszerek hatékonyságát szimulált ésvalósadatok együttes bevonásával
demonstráltam.
Abstrat
During the last two deades, the linear parameter varying (LPV) models have
been inthe main fousin the eld ofsystem identiationtheory sinethey an be
onsideredas themissing linkbetween the linear time-invariant(LTI)and thenon-
linear and/or parameter time-varying modeling approahes. This thesis has been
performed in the adre of a bi-direted researh projet o-tutelle between the
University ofTehnologyand Eonomisandthe University ofPoitiers. Duringthis
researh the main fous has been plaed to the identiation of state spae LPV
and LTI models for exible non-linear strutures by using the
H ∞
-norm. As faras the LPV models are onerned, the so-alled loal approah is, more preisely,
onsidered whihontains the following steps: (i) I/O data sequenes are gathered
in ertain working points. (ii) loal blak-box LTI models are estimated in eah
working point, (iii) then, the loal models are interpolated in order to obtain the
nal LPV model. In this thesis, novel methods have been proposed for eah of the
above ited steps. A method able to determine a set of operating points by ap-
plying the nu-gap metri non-linearity measurement has been developed. In order
toensurethe oherent basis representation, whih isneessary forthe interpolation
step, two
H ∞
-norm-based methods have been developed. First, the loally esti- mated blak-box state-spae models are re-strutured into gray-box ones by usingthe proximityontrolalgorithmfollowedby the least-squares-based interpolation of
the obtained loal gray-box models. Seond, a behavioral interpolation (BI) has
been performed by minimizing the frequeny domain distane between eah or-
responding loal blak-box LTI and frozen LPV models by using one global ost
funtion. This latter tehnique is designed to operate with LPV models written as
linear frationalrepresentations (LFRs). Finally,the performane of the developed
methodshasbeentestedby using,atsametime,simulatedandreal datasequenes.
Contents
1 General Introdution 1
I Theoretial developments 7
2 Appliation of the linear frational representation for model de-
sription 9
2.1 Introdution . . . 9
2.2 Notions and denitions . . . 10
2.3 Linear system modelingby usinglinear frational representations . . 11
2.3.1 Linear time-invariantmodels . . . 11
2.3.2 Multi-dimensionallineartime-invariantmodels . . . 12
2.3.3 Linear parameter-varying models . . . 13
2.3.4 Comments onwell-posedness . . . 15
2.4 Blak and gray-box state-spae models . . . 15
2.4.1 Identiability . . . 17
2.5 Conlusion . . . 19
3 Non-smooth and non-onvex optimization: the proximity ontrol algorithm 21 3.1 Introdution . . . 21
3.2 The proximity ontrol algorithm . . . 22
3.2.1 Denitions . . . 22
3.2.2 Basiidea . . . 23
3.2.3 Outerloop. . . 25
3.2.4 Inner Loop . . . 26
3.2.5 The algorithm . . . 29
3.3 Conlusion . . . 30
4 Gray-boxstate-spaeLTImodelidentiationbyusingre-struturing tehniques 31 4.1 Introdution . . . 31
4.2 Problem statement . . . 32
4.3 Output-error framework . . . 33
4.4 From blak-box state-spae models to gray-box ones . . . 35
4.4.1 An
H ∞
-norm-based approah . . . 374.4.2 Convergene of the proximity ontrolalgorithm . . . 40
4.5 Conlusion . . . 42
5 State-spae LPV model identiation from loal experiments 43 5.1 Introdution . . . 43
5.2 Problem statement . . . 44
5.3 Operatingpointseletionfor LPV modelidentiation . . . 45
5.3.1 Nu-gap metri . . . 47
5.3.2 Seletion algorithm . . . 48
5.4
H ∞
-norm-basedloalre-struturing method . . . 495.5
H ∞
-norm-basedbehavioral interpolation (H ∞
-BI) . . . . . . . . . . 515.6 Conlusion . . . 53
II Appliations 55 6 Simulation examples 57 6.1 Introdution . . . 57
6.2 Gray-boxlineartime-invariantmodelidentiationoftheprinter-belt system . . . 57
6.2.1 Printer-belt . . . 57
6.2.2 Data Generation . . . 60
6.2.3 Identiationproedure . . . 60
6.2.4 Results . . . 61
6.2.5 Conlusion. . . 76
6.3 LPV modelidentiationof the mass-spring-dampersystem . . . 76
6.3.1 Mass-spring-dampersystem . . . 76
6.3.2 Working point seletion . . . 80
6.3.3 Estimationof the nalLPV model . . . 84
6.3.4 Conlusion. . . 91
7 Real data driven examples 93 7.1 2-DoF exible roboti manipulator . . . 94
7.1.1 Nonlinear and linearized dynami models . . . 96
7.2 Experimentdesign . . . 101
7.2.1 Loal experiments. . . 101
7.2.2 Global experiments . . . 101
7.2.3 To sum up . . . 103
7.3 Blak-box linear time-invariantmodelidentiation . . . 107
7.4 Linear parameter-varying modelidentiationand validation . . . 107
7.4.1 Gray-box and blak-box LPV modelidentiation . . . 110
7.4.2 LPV modelvalidation . . . 110
7.5 Conlusion . . . 112
8 Summary of the obtained results and future researh objetives 115 8.1 Thesis Points . . . 115
8.2 Possible researh diretions . . . 116
8.2.1 Improvements of the null-spae-based tehnique . . . 117
8.2.3 Handlingthe possible blak-box modelestimation errors . . . 118
8.2.4 ImprovementsoftheLPVmodelidentiationtehniquesbased onloalexperiments . . . 118
8.3 Furtherindustrialappliations . . . 119
9 Appendix 121 9.1 Nonlinear dynami modelof a2-DOF roboti arm . . . 121
9.2 Summary of the PO-MOESP algorithm . . . 123
9.3 Summary of the SRIVCalgorithm. . . 125
9.4 Norms onmatries, signals and systems . . . 126
9.4.1 Matrixnorms . . . 127
9.4.2 Signalnorms . . . 127
9.4.3 System norms . . . 128
Chapter 1
General Introdution
In the rapidly developing world, not only the available omputational apa-
ity,but alsothe demands for morereliableteleommuniation onnetions,optimal
and intelligentenergy distribution and generation systems, better robot pilots, au-
tonomous systems to explore new plaes, for seurity and resue purposes or for
military appliations..., et. just to name a few, have ontinuously been growing
fromday-to-day. Theseappliations,one-by-one, doneessitatethe adequateappli-
ationofthemostreentresultsofseveraldierentelds,suhas,appliednumerial
mathematis, physis, hemistry, omputer sienes..., et. The olletive applia-
tion of the previously enumerated sienti priniples, with the intention to reah
better and autonomous behaviour of a ertain proess, are joined together in the
so-alled ontrol theory. On top of that, in order to be able to develop reliable
ontrollerswhih meet the strit requirementsstated by the onsidered appliation
eld,itisindispensabletohaveaorretmodeloftheproessorsystemunderstudy.
Thisisthe momentwhensystem identiationenters intothe piture. Beingavery
important omponent of the modern ontrol theory, system identiation aims at
derivingmodelsforphysialproessesandestimatingitsparameters. Averyimpor-
tant requirement at the end of the identiation proedure is that the parametri
modelestimationistratablebynumerialproessingunits,suhas,e.g.,embedded
omputers,PCs ormorereentlyeven by smartphones. Despite the fatthat there
is anenormous amount of ontributions in this eld whih an be overed by some
benhmark publiations, for instane, [78, 113, 103, 47, 135, 81℄ (see as well the
referenes therein), there remain lots hallenging open questions waiting to be an-
swered, mainlyonerning the determination of reliable linear models of nonlinear
and/ortime-varyingproesses. Thislatter pursuitinvolves thedevelopmentof spe-
i model strutures in whih the nonlinear behavior has been broken down into
several loallinear models, in order to obtain amore tratable modelstruture for
the original nonlinear system under study. This interest an mainly be explained
by the following reasons. On the one hand, the resulting model is often designed
to be lose to the standard linear time-invariant (LTI) one but with an embedded
strutural exibility able to ope with time-varying, even highly nonlinear behav-
iors. Ontheotherhand,thedevelopmentofsuhasetofmultiplemodelsisdiretly
linked toontrolengineering, wherea ontroller must bedesigned inorder to meet
thelosed-looprequirementsforagivenplantindierentoperatingonditions,e.g.,
and onquer basiideaisthegainshedulingapproah[3,124,75℄(see alsotheref-
erenes therein for further informationabout gain sheduling) whihan be briey
summarized as follows:
nd one or more sheduling variables whih an ompletely parameterize the
operating spae of interest forthe system toontrol,
dene a parametri family of linearized models for the plant assoiated with
the set of operatingpointsof interest,
design a parametri ontroller whih an both ensure the desired ontrol ob-
jetivesineahoperatingpointandanaeptablebehaviorduring(slow)tran-
sients between one operatingondition and the other.
A wide body of eient ontroller design tehniques sharing this basi idea is now
available in the literature (see, e.g., [126, 64, 2℄), whih an be solved reliably,
provided that a suitable model in a parameter-dependent form has been derived
beforehand. By onstrution, the reliability of these tehniques highly depends on
the availabilityof a suitable and onsistentmodel. In order tosupportthis kind of
multiple model-based ontroller design, several methods have been developed dur-
ing the past two deades, in order to derive reliable multiple model strutures for
nonlinear dynamialsystems, suhas,e.g.,theswithedlinearsystems [110,111℄or
the multiple-model adaptive estimation (MMAE) tehniques [55℄. One of the mul-
tiple model-basedsolutionsfound inthe literatureisthe so-alledlinearparameter-
varying(LPV) model struture. More preisely, the modelingand estimation prob-
lem of LPV models are reently among the most popular researh topis in system
identiation theory [135, 81, 83℄. This is the main reason why the identiation
of (state-spae) LPV models is prinipallyonsidered in this thesis. These spei
strutures anbeseenasaombinationofloalmodelswithparametersevolvingas
a funtion of measurable time-varying signals, alled the sheduling variables, sig-
nals whihan berelated todierent operatingpointsof the system toidentify. As
far asthe determinationof LPV modelsis onerned,twobroad lassesof methods
an befound inthe literature [24℄:
(i) rst, the analyti methods onsisting exlusively in onverting the available
nonlinear equations governing the behavior of the system into an LPV rep-
resentation by resorting to extensions of the familiar notions of linearization
[58, 118, 123, 128, 127, 125, 74, 10, 89, 66,112,135,1℄;
(ii) seond, the experimentalmethodsinlinedtodetermining LPV models of the
plantunder study from the available input-outputdata [84, 30,32,133, 23℄.
The rst lass gathers the solutions using rst-priniples modeling onsiderations.
They more preisely try to transform the original nonlinear model into a reliable
LPV representation. Aording to the literature, more speially, the following
lasses of methodsan be distinguished:
extended linearization[123℄;
pseudolinearization[118℄;
global linearization[58℄;
Jaobian linerization[128, 125℄;
state transformation [127℄;
funtion substitution [89℄;
veloity-based approah[74℄;
tensor-produt polytopi deomposition [10, 112℄;
dierentialgeometri approah [1℄.
Theseondfamilyofmethodsan besplitup intotwofurthersub-lasses, generally
alled, the global approah and the loal approah,respetively. On the one hand,
theglobalapproahfousesontheglobalproedureandassumesthatoneglobalex-
perimentanbeperformedduringwhihthe ontrolinputsaswellasthesheduling
variablesan be both exited[73, 140℄. Hereby, allthe non-linearitiesofthe system
are exited simultaneously by passing through a large number of operating points.
On the other hand, reent methods [51, 84, 30, 32, 23℄ are based on a multi-step
proedurewhere
1. loalexperimentsarearriedoutinwhihtheoperatingpoints(orresponding
to xed values of the sheduling variables) are held onstant and the ontrol
inputsare (persistently) exited,
2. loalLTImodelsareestimatedbyusingthesesetsofloalinput/output(I/O)
measurements,
3. aninterpolationphaseisperformedinordertoderiveanalglobalparameter-
dependent model.
It isobvious that this multi-stepapproah isa lotloserto the standard proedure
used for nonlinear system identiation or the one dediated to gain sheduling.
Both lasses of methods have advantages and drawbaks (see e.g., [135, Chapter
1℄ for more details about the state-of-the-art of LPV modeling and identiation).
Fromapratialpointofview, theglobalapproahmay suerfromthe diultyto
satisfytherih exitationof theontrolinputsand theshedulingvariablessimul-
taneously. Itisobviousthatsuhanexperimentalproeduremay notbereasonable
for spei appliations mainly for safety or eonomi reasons. On the ontrary,
applying small variations around partiular operating points, as onsidered by the
loalapproah,ismore oneivablein many pratialases. Furthermore,the iden-
tiationof LTI modelsis wellestablishedand implementationsare availableinthe
Matlab environment [78, 79℄. That is the reason why during this thesis, the de-
velopmentsfous on the loal approah. Notie also that, inthis thesis, duringthe
identiationproedure, state-spae LPV models are sought to be estimated. This
hoie an be explained by the following reasons. First, when ompared with the
standard input-output LPV representations, a state-spae model often provides a
moreparsimoniousrepresentationofthesystem(fewerparametersandlowerMMil-
landegree)withonlyonetuningparameter: theorderofthesystem. Seond,thanks
totheintrodution(withautionintheLPVframework(see[136,65,138,23℄foran
importantdisussionabouttheoherentbasisissue))ofspeiuser-denedsimilar-
itytransformation,itispossibletoyieldanumeriallybetteronditionedframework
for parameter estimation. Last, but not least, state-spae representations are often
favored when ontroller designis the reason why the modelis built.
Remark 1.1. In many pratial ases, the systems to identify are modeled by a
so-alled quasi-LPV model [24℄. In this ase, the sheduling variables are somehow
relatedtothe statevariablesof themodelrepresentation. Thus,undersuhpratial
onditions, it is diult to guarantee that the sheduling variables are kept exatly
onstant during the rst step of the loalapproah. However, as shown hereafterin
Chapter 7, this diulty an be handled and reliable LPV models an be estimated
In this thesis, the loal approah is revisited by suggesting, for step 1 and 3
mainly, innovative solutions. As far as the seletion of the loal working points
is onerned, it is important to point out rst that, in the literature, most of the
time, the Authorshoose equidistantworkingpointsby seletinga user-dened but
xed step between twoloalexperimentsin terms of the shedulingvariables. It is
pointed out in Chapter 5 that suh a simple hoie may lead to a set of working
points where the loal information about the nonlinear system under study is far
frombeingoptimal. Inorder toaddressthis problem,anovel approahisproposed
in this thesis to selet a better established set of loal operating points [148, 149℄.
Asfarastheinterpolationstepisonerned,itisnowwell-known thatinterpolating
properlyloalonsistent LTIstate-spaemodelsisfarfrombeingeasy,espeiallyin
the blak-box framework, beause, for many LPV state-spae representations, the
similarity transformation onverting the initial, even stati-dependent state-spae
models, yields equivalent LPV models depending on the time-shifted or derivative
versionsof the shedulingvariables[135,65,137℄. Thus, asfarasinterpolationstep
is onerned, the loal blak-box state-spae LTI models must be transformed into
a ommonoherent basis representation. Interesting solutions have been proposed
intheblak-boxframework,i.e.,whennopriorinformationaboutthe systemunder
study isavailable[139,54, 84, 105, 31,99,32℄. Among them, the most numerially
reliable solutions are the ones published in [84, 32℄. In a nutshell, as pointed out
rst in[136℄, then in [65℄, the interpolation step involved inthe loalapproah an
lead to a global LPV model with an inaurate dynami behavior even if the loal
LTI models are onsistent. These diulties are takled hereafter by onsidering
two omplementarysolutions.
When spei strutural prior information about the system to identify is
available,i.e.,whengray-boxLPVmodelsarehandled,thesolutiondeveloped
in this thesis onsists in resorting to the knowledge available from the study
of the non-linear equations governing the system behavior in order to x the
strutureof theglobalLPVmodel,thenusingthe availableexperimentaldata
sets in order to estimate the unknown parameters and to rene the analyti
model omposed of unknown values. More speially, we onvert the loal
input-output data-sets, translated initially into (fully-parameterized) blak-
boxmodels,intore-parameterizedgray-boxstate-spaeformsderivedfromthe
frozen struture of the LPV representation and alulated for the onsidered
working points.
When blak-boxstatidependentLPVmodelsareonerned,aspeiatten-
tion is paid to the preservation of the input-output dynamis of the models
instead of doing a diret interpolation of the system matries. More spei-
ally,by onsidering a user-dened tting measure (hereafterthe
H ∞
-norm),given a set of loalLTI models as well as a fully-parameterized stati depen-
dentLPVmodelstruture,our solutiononsistsinestimatingtheLPV model
parameters so that the nal LPV isoptimal with respet toa global measure
of the error between the loalmodels and the LPV representation. As shown
inChapter5,suhaninput-outputbehaviorapproahallowsustoirumvent
the diult oherent basis issue.
Thedissertationisbuiltup asfollows. The rstpart ofthismanusriptpresents
the maintheoreti developments. InChapter 2,the LFRs,usedoriginallyinthe ro-
resentations anbe appliedfor systemmodelingand system identiation. Chapter
3introduesanon-smoothandnon-onvexoptimizationtehniqueusedhereafterin
thisthesistominimize
H ∞
-norm-basedostfuntions. Notieindeedthat,asshown inChapter 4,the solutionsdeveloped tobypass the diultproblemofLTImodelsinterpolationare mainlybased on
H ∞
-norm-baseddisrepany measurements. The identiationofLPV models from loalexperiments istakled inChapter 5 where,more speially, a novel method is proposed to determine the set of loal operat-
ingpointsinvolved in the loalapproah followed by the introdutionof two novel
methods aiming at estimating a reliable LPV state spae model. The developed
methods support the identiation of both blak- and gray-box LPV models. The
rst Part is then losed by Chapter 8 where, rst, the developed novel tehniques
areenumerated intheformofthesis points. Then,severalideas have beengathered
so that tomotivate and support further researh pursuits. The seond Part of this
thesis demonstratesthe eetiveness of the developed methods. To reahthis goal,
rst,simulationsbyonsideringdierentkindsof systemsareperformedinChapter
6. Seond, the proposed LPV identiation approahes have been tested on a real
testbenhaswell,onsistingina2-DoF exiblerobotimanipulatorusedinardia
surgery. Finally, in the Appendix, mathematial preliminaries, the desriptions of
several existing identiationtehniques, system modeling aspets are provided.
Part I
Theoretial developments
Chapter 2
Appliation of the linear frational
representation for model desription
2.1 Introdution
The main purpose of this Chapter istoexplain howa largelass of linear state-
spae models an be transformed into a linear frational representation (LFR) in
order toset the stage for the system identiation. Originally,this modelrepresen-
tation isintrodued inrobust ontrol theory [164℄ asa speial formof the so-alled
Redheer star produt [119℄. More preisely, LFRs are widely used to represent
any feedbak interonneted losed-loop model having dierent kind of unertain-
ties during the ontrollersynthesis [164℄.
Asfarasitsappearane insystemidentiationisonerned,LFRshavealready
been employed in[71,70,73℄,toestimateblak-box linearparameter-varying state-
spaemodelswhereanoutput-error(OE)set-upisonsideredandtheresultingost
funtion is minimized by applying nonlinear programming. More reently, among
others, LFRs have been used in [108, 141, 142, 102℄ (see the referenes therein as
wellforotherreentappliations)toestimatemodelshavinginnerstrutured stati
non-linearities. In [108℄, more preisely, a pieewise ane proedure is developed,
while in [141, 142, 102℄, the so-alled Best-Linear-Approximation (BLA) approah
is onsidered. On top of that, in [34℄, the Modelia software tool is applied to
generate LFRs, in order to model nonlinear systems having unertain parameters
foridentiationpurposes. AsitisshownhereafterinthisChapter,LFRshavesome
spei properties whih make its appliation in the identiation of LPV models
reallyinteresting. InthefollowingChapters,theLFRsareutilized,morespeially,
for the identiation of LPV models by using a spei
H ∞
-norm optimization framework.Thus, in the sequel, some neessary notations and denitions are given followed
by denitions related to the identiability onept of the LFRs. Then, it is shown
howalarge lassof linearstate-spae models anbetransformed intoanLFR.The
urrentChapterisonludedbyadisussiononthewell-posednessissueofthelinear
frationalrepresentations.
M
∆ u ✛
✲
✲
✲
u 12 u 11
y 12 y 11
(a)The upperLFR.
∆ l
M
✛
✲
✲
✲
u 21
u 22
y 21
y 22
(b)The lowerLFR.
Figure2.1: Thelinearfrationalrepresentationbyusinganupperandalowerlinear
frational transformation (
F u (M, ∆ u )
andF l (M, ∆ l )
) of two matriesM
and∆ u
or
∆ l
having appropriatedimensions.2.2 Notions and denitions
Asmentionedabove, inthis Setion, someimportantdenitionsare introdued.
So, let us rst dene what a linear frational representation is and how it an be
alulated [164℄.
Denition 2.2.1. A linear frational representation is a feedbak interonnetion
of two appropriately partitioned matries denoted hereafter by
∆ u ∈ R n u 11 ×n y 11
or∆ l ∈ R n u22 ×n y22
andM
(see Fig. 2.1),M =
M 11 M 12 M 21 M 22
∈ R n M ×n M .
(2.1)If
(I − M 11 ∆ u )
or(I − M 22 ∆ l )
is invertible, the so-alled upper or lower LFR an be omputed byF u (M, ∆ u ) = M 22 + M 21 ∆ u (I − M 11 ∆ u ) −1 M 12 ,
(2.2a)F l (M, ∆ l ) = M 11 + M 12 ∆ l (I − M 22 ∆ l ) −1 M 21 .
(2.2b)Eah form omposing Eq. (2.2) is, in fat, the input/output representation be-
tween
u 12
andy 12
, when an upper LFR is handled, or betweenu 21
andy 21
whena lower LFR is onsidered (see Fig. 2.1). The operation denoted by
F u ( • , ⋆)
andF l ( • , ⋆)
isthe soalledupperand lowerlinear frationaltransformation(LFT) of•
and
⋆
,respetively. To larifythe ambiguoususageof the notationsLFT and LFR, by LFT, the operation itself is meant, while LFR stands for the model representa-tion. Now, in order toset the stage for a generi LFR-based modeling framework,
let usonsider a spei blok-matrixstruture by followingthe lines found in [70,
Chapter 8℄,
K ∆ = { diag(δ 1 I n 1 , · · · , δ S I n S ), δ i ∈ O Lin , i = { 1, · · · , S }}
(2.3)and a vetor
Υ =
δ 1 · · · δ S
⊤
∈ O S Lin ,
(2.4a)where,
O Lin
denotes the set of linear operators [22℄ (for instane, the derivativeoperator
d
dt
ortheforwardshiftoperatorq
)whiletheK ∆
blok-diagonalsetontainsn
of them, wheren = Σ S i=1 n i
isthe sum of the blok dimensions given by Eq. (2.3).[70℄ beause, in this thesis, only blok-diagonal
∆ ∈ K ∆
matries are employed.Furthermore,onsider
d =
n 1 · · · n S
⊤
∈ N S .
(2.4b)Then, by using Eq. (2.2), by fousing on the upper LFT 1
(see Fig. 2.1a), the fol-
lowing input-output (between
u 12
andy 12
(see Fig. 2.1a)) representation an be alulatedF u ( M , ∆(d, Υ)) = D + C ∆(d, Υ)(I n − A ∆(d, Υ)) −1 B ,
(2.5)where
∆(d, Υ) ∈ K ∆ ,
(2.6)isastruturedblok-diagonalmatrix parameterizedby the vetors
d
andΥ
respe-tively, and with
M =
A B C D
∈ R (n+n y )×(n+n u ) ,
(2.7)where
n u
andn y
are the dimensions of the input and output hannels involved inFig. 2.1a while
n = Σ S i=1 n i
. In Eq. (2.7), the system matries( A , B , C , D )
an bepartitioned asfollows [70℄
A =
" A
1,1 ··· A 1,S
.
.
. .
.
.
A S,1 ··· A S,S
# , B =
B 1
.
.
.
B S
,
(2.8a)C = [ C 1 ··· C S ] , D = D .
(2.8b)NotiethatthenotationsoftheinnermatriesdenedabovebyEq.(2.8)aregeneri.
In the next Setion, the above introdued tool, namely the LFR, is applied to
represent some spei lassesof linear models.
2.3 Linear system modeling by using linear fra-
tional representations
In this Setion, it is shown how a large lass of linear models an be trans-
formed into an LFR [70, Chapter 8℄. From Eq. (2.5)-(2.7), it an beseen that any
linear state-spae form an be represented aording to the operators plaed into
the blok-diagonal struture denoted by
∆(d, Υ)
. This results, basially, in theI/Orepresentation orthetransferfuntion formof onsideredthelinear state-spae
models.
2.3.1 Linear time-invariant models
Let us rst onsider an
n x
order LTI model given by the following state-spaeform
γx(t) y (t)
=
A B C D
x(t) u (t)
,
(2.9)1. Notie that, hereafter in this Chapter, the upper LFR is applied. However, this an be
replaed bylowerLFRs as well by modifying theinner struture of thematrix
M
aordingtowhere
u(t) ∈ R n u
is the input signal vetor,y(t) ∈ R n y
is the outputsignal vetor,x(t) ∈ R n x
isthe statevetorandt ∈ R
orZ
. Herein,γ
standsfor theforward shiftoperator (
q
) when disrete-time systems are onsidered or for the dierential (d dt
)operator when ontinuous-time systems are handled. This modelan be desribed
by anLFR satisfying
S = 1, d = n x ,
Υ = 1
η , ∆(d, Υ) = 1
η I n x , M LT I =
A B C D
,
where
η
istheLaplaetransform(ontinuous-timease)orthez
transform(disrete-time ase) of
γ
. Indeed, the transfer funtion form of the onsidered linear state-spae modelan be alulatedby using anupperLFT as follows,
G LT I (η) = D + C (ηI n x − A ) −1 B = D + C 1
η (I n x − A 1
η ) −1 B = F u
A B C D
, 1 η I n x
= F u ( M LT I , ∆(d, Υ)),
(2.10)where the system matriesdenoted by
( A , B , C , D )
have appropriatedimensions.2.3.2 Multi-dimensional linear time-invariant models
Theproposedmodelstrutureanalsodealwithaspeimulti-dimensionalLTI
(MDLTI) model, namely,a disrete-time
2D
Roesser model[121℄, whih satises
x 1 (k 1 + 1, k 2 ) x 2 (k 1 , k 2 + 1)
y(k 1 , k 2 )
=
A 1,1 A 1,2 B 1 A 2,1 A 2,2 B 2 C 1 C 2 D
x 1 (k 1 , k 2 ) x 2 (k 1 , k 2 ) u(k 1 , k 2 )
.
(2.11)This 2D model an be writtenas anLFR with
S = 2, d =
n x 1 n x 2
⊤
, Υ =
z 1 −1 z 2 −1 ⊤
, ∆(d, Υ) =
z 1 −1 I n x1
z 2 −1 I n x2
,
as follows
G M D (z 1 , z 2 ) = F u ( M M D , ∆(d, Υ)),
(2.12a)M M D =
A 1,1 A 1,2 B 1 A 2,1 A 2,2 B 2 C 1 C 2 D
,
(2.12b)∆(d, Υ) =
z 1 −1 I n x1 0 0 z 2 −1 I n x2
.
(2.12)Here,
z 1
andz 2
are thez
transform variables. Again, the system matries foundin
M M D
have appropriate dimensions. The identiation problem of this speial model written as anLFR is addressed in [40℄. Notie that this formulation an beextended toany Roesser
nD
-models without any restritions.2.3.3 Linear parameter-varying models
As farasLPVmodelsare onerned,throughoutthis thesis,the followingstate-
spae representation is onsidered
γx(t) = A (p(t))x(t) + B (p(t))u(t),
(2.13a)y(t) = C (p(t))x(t) + D (p(t))u(t),
(2.13b)where
u(t) ∈ R n u
is the input signalvetor,y(t) ∈ R n y
is the output signal vetor,x(t) ∈ R n x
is the state vetor andt ∈ R
orZ
. In the LPV framework, the system matries( A , B , C , D )
are rational or ane funtions of measurable time-varying signals asA : P 7−→ R n x ×n x B : P 7−→ R n x ×n u C : P 7−→ R n y ×n x D : P 7−→ R n y ×n u .
The shedulingvariablesare, more preisely,gathered intothe vetor
p(t) ∈ P ⊆ R n p ,
where
P =
p (t) ∈ R n p
p i (t) ≤ p i (t) ≤ p i (t)
˙
p i (t) ≤ p ˙ i (t) ≤ p ˙ i (t) ∀ i ∈ { 1, · · · , n p }
,
(2.14)isthe so-alledsheduling"spae"[134℄whihisaompatset [22℄. Inthe sequel,it
isfurthermoreassumed thatthesystemmatriessatisfyastatidependenyon
p(t)
[135℄, i.e., they do not depend on the time-shifted versions or the time-derivatives
of the sheduling variables (
γp(t), γ 2 p(t), · · ·
) when disrete or ontinuous-time modelsareonsidered,respetively. Notiehoweverthat,ifthe LPVmodeldependsalso on the derivatives of some sheduling variables whih are measurable, then
the sheduling variable vetor an be augmented by onsidering these signals as
shedulingvariables as well.
Now, by using again the upperlinear frationaltransformation presented previ-
ously (see Fig. 2.1a), any arbitrarily LPV model an be transformed into an LFR
intwosteps.
First, the parameter-dependent state-spae matriesintrodued by Eq. (2.13)an
be derived by
A (p(t)) B (p(t)) C (p(t)) D (p(t))
=
A 0 B 0
C 0 D 0
+ B w
D yw
∆ p(t) (I n ∆ − D zw ∆ p(t) ) −1
C z D zu
= F u ( M LP V , ∆ p(t) ),
(2.15a)with
∆ p(t) =
p 1 (t)I r 1
.
.
.
p n p (t)I r np
∈ R n ∆ ,
(2.15b)n ∆ =
n p
X
i=1
r i ,
(2.15)and
M LP V =
D zw C z D zu
B w A 0 B 0 D yw C 0 D 0
,
(2.15d)and where
A 0 , B w , · · · , D 0
are time-invariantmatrieshavingappropri-
ate dimensions. In this partiular ase,
S = n p , d = [r 1 , · · · , r n p ] ⊤ , Υ = [p 1 (t), · · · , p n p (t)] ⊤ .
Seond, thedynamialstate-spae LPVmodeldened by Eq.(2.13)an bealu-
lated by applyingone again the upperLFT, i.e.,
F u
A ( p (t)) B ( p (t)) C (p(t)) D (p(t))
, 1 η I n x
= F u ( F u ( M LP V , ∆ p(t) ), 1
η I n x ).
(2.16)If the sheduling variables are xed in a ertain working point denoted here-
afterby
p i
,thentheloalfrozentransferfuntionsanbealulatedasfollowsG LP V (η, p i ) = F u
A (p i ) B (p i ) C (p i ) D (p i )
, 1 η I n x
(2.17)
= F u ( F u ( M LP V , ∆ p i ), 1
η I n x ),
(2.18)where
i
standsfor thei th
xed valueof the shedulingvariablevetor. Thisisbasially a standard LTI transfer funtion. Aording tothe notations intro-
dued by Eq. (2.4), similarlytothe LTI ase,
S = 1, d = n x ,
Υ = 1
η , ∆(d, Υ) = 1
η I n x .
As pointed out in [70℄, the obtained modelisI/O equivalent withthe one dened
as follows
γx(t)
z(t) y(t)
=
A 0 B w B 0 C z D zw D zu C 0 D yw D 0
x(t) w(t) u(t)
,
(2.19)w(t) = ∆ p(t) z(t).
(2.20)The resulting LPV model is referred hereafter as an LPV/LFR one. In order to
highlightthegenerinatureoftheLFRsappliedduringthemodelingofLPVmodels,
it an be onluded fromEq. (2.15) that the parameter dependeny an be
(i) ane when
D zw = 0
;(ii) rationalwhen
D zw 6 = 0
.This property is a very important feature of the LFR when it is ompared to the
other LPV model representations found in the literature (see [135, Chapter 3℄ for
a reent overview of the existing LPV modeling approahes), beause, most of the
by Eq. (2.13) depend on the sheduling variable vetor in an ane or polynomial
manneri.e.,
♣ (p(t)) = ♣ 0 +
n p
X
i=1
♣ i p i ,
(2.21a)♣ (p(t)) = ♣ 0 +
n p
X
i=1 s i
X
j=1
♣ i,j p s i i
(2.21b)where
♣ •
is a onstant matrix standing forA •
,B •
,C •
andD •
and wherep = p 1 · · · p n p
⊤
.
Thanks to the examples presented above, it an be seen that a wide range of
linear systems an be desribed by using LFRs. This is the main reason why the
developments of this thesis utilizethis modelrepresentation tool.
2.3.4 Comments on well-posedness
The LFR an be dened and omputed if and only if it is a well-posed LFR,
i.e., for our study, if
I n − A∆(d, Υ)
is invertible (see Eq. (2.5)). More preisely,in the
1D
or2D
LTI ases, this well-posedness ondition an be related to the standard stability onditions available, e.g., in [39, 35℄. Thus, if the system toidentify is assumed to be stable, the LFR should be well-posed. However, in the
LPV framework, the problem is muh more ompliated beause, in this ase, the
matrix
∆(d, Υ)
expliitly depends on the sheduling variablep(t)
. Therefore, thewell-posedness onditions of the problem should be related to the evolution range
of
p(t)
dened on the ompat setP
(see Eq. (2.14)).Remark 2.1. When the sheduling variable
p(t)
is xed, whih is the ase in theloal approah, the resulting frozen LPV model is basially an LTI one, where the
well-posedness an easily be veried.
Aording tothe Author'sknowledge, noneessary and suientonditions are
availableforthisproblemintheliterature. Nevertheless, inthegray-boxframework,
whenthemodelstrutureisobtained fromthe physiallaws governingthe behavior
of the system, the struture of
∆(d, Υ)
an be xed a priori and suient LMIonditions ould be obtained to ensure the well-posedness of the LFR by applying
a S-proedure tool [21℄. This problem is not solved in this thesis and is referred
to future researh. At the same time, this study, more preisely the link between,
e.g., the ondition number of
I n − A ∆(d, Υ)
and thep(t)
-trajetory, an be seenasaninteresting startingpoint forananalysis ofthe optimal
p(t)
-trajetorydesignensuring a onsistentidentiation.
2.4 Blak and gray-box state-spae models
As desribed in the previous Setion, the value of
S
inEq. (2.8) is user-denedand an be hosen aording to the onsidered linear model (linear time-invariant
(LTI), multi-dimensional linear time-invariant (MDLTI), linear parameter-varying
(LPV), et). In order to emphasize the generi nature of this representation, the
matriesfound in Eq. (2.7) an be blak-box or, aording to the prior knowledge,
is meant that the struture of the model to identify is a parameterized struture
derived from some prior knowledge governing the behavior of the system under
study. In other words, the parameters to estimate are unknown funtions (e.g.,
+, − , ÷ , ×
)ofthe realphysialones whilesomematrixentriesanbexed as0or1. On theotherhand,theunknownparametersfound inlight-gray-boxmodels have
real physialmeaningbased onrst priniples modelingonsiderations. In orderto
illustrate thesedenitions, let usonsider the following simplemass-spring-damper
system, i.e.,
m¨ z + k z ˙ + dz = F,
where
z
denotes the position,m, k, d
denotethe mass, stinessand damperoe-ients, respetively. This simple system an then betransformed intoa state-spae
form by onsidering
F
andz
as input and output,respetively,z ˙
¨ z
=
0 1
− m d − m k z
˙ z
+ 0
1 m
F,
(2.22a)y =
1 0 z
˙ z
,
(2.22b)while the state variablesare
x = z
˙ z
.
(2.23)In the blak-box ase, the modelis fully-parametrizedand has the followingform
ξ ˙ ξ ¨
=
θ 1 b θ 2 b θ 3 b θ 4 b
ξ ξ ˙
+ θ b 5
θ b 6
F, y =
θ 7 b θ b 8 ξ ξ ˙
+ [θ 9 b ]F.
→ ξ ˙ = A (θ b )ξ + B (θ b )F,
y = C (θ b )ξ + D (θ b )F,
(2.24)where
θ b =
θ b 1 θ b 2 θ 3 b θ 4 b θ b 5 θ b 6 θ 7 b θ 8 b θ 9 b ⊤
,
and
ξ = ξ ξ ˙ ⊤
.
Notie that
ξ
denotes the blak-boxstate vetor whihan bealulatedfromx
byusing any invertible similarity transformation
T
, having appropriate dimension, as follows,ξ = Tx.
(2.25)Asfarasthedark-gray-boxframeworkisonerned,byusingsomepriorinformation
about the struture ofthe system understudy (see Eq. (2.22)), somematrix entries
an bexed as0 or1, i.e.,
z ˙
¨ z
=
0 1 θ 1 g θ 2 g
z
˙ z
+ 0
θ g 3
F, y =
1 0 z
˙ z
,
→ x ˙ = A(θ g )x + B(θ g )F,
y = C(θ g )x,
(2.26)where
θ g =
θ 1 g θ g 2 θ g 3 ⊤
,
and
x =
x x ˙ ⊤
.
Here, the parameters to identify are unknown funtions of the real physial ones.
Indeed,
θ 1 g = − d
m , θ 2 g = − k
m , θ 3 g = 1 m .
Finally,thelight-gray-box version isequaltothe originalstate-spae representation
given by Eq. (2.22) by onsidering the physial parameters to be estimated, i.e.,
θ g =
d k m ⊤
,
z ˙
¨ z
=
"
0 1
− θ θ g 1 g 3 − θ θ g 2 g 3
# z
˙ z
+ 0
1 θ g 3
F, y =
1 0 z
˙ z
.
→ x ˙ = A(θ g )x + B(θ g )F,
y = C(θ g )x.
(2.27)Throughout this thesis, when the modelidentiationproblem isaddressed, blak,
dark-gray and light-gray-box versions of the above presented matries are denoted
asfollows,
M (ϑ) =
A (ϑ) B (ϑ) C (ϑ) D (ϑ)
−→
M (θ b ) =
A ( θ b ) B ( θ b ) C (θ b ) D (θ b )
,
M( θ g ) =
A(θ g ) B(θ g ) C(θ g ) D(θ g )
,
(2.28)
where
ϑ ∈ R n ϑ
isthe generalnotationof the unknown parameter vetor,θ b ∈ R n θb
isavetor ontainingallthe matrixentrieswhen blak-boxmodels aresoughttobe
estimated, and where
θ g ∈ R n θg
is a vetor ontaining the unknown parameters to identifywhihanbeunknown funtionsoftherealphysialparameters(dark-gray-box) or diretly the physial parameters themselves (light-gray-box). Although, in
the gray-box ases, the same notations are employed, the distintion of these two
types will be apparent during the onrete appliations. For blak- and gray-box
models, the matries found in
M (θ b )
andM(θ g )
are assumed to be of appropriate dimensionsandontinuouslydierentiablew.r.t. theunknownparametersθ b ∈ R n θb
or
θ g ∈ R n θg
. Furthermore,inthe gray-boxframeworkthemodelstruture,i.e.,the waythematriesA(θ g )
,B(θ g )
,C(θ g )
andD(θ g )
dependontheunknownparametervetor
θ g
, is assumed to beknown a priori.2.4.1 Identiability
Beause, in this thesis, a partiular attention is paid to the identiation of
state-spaemodels,it isimportanttoexamine whether the parametersof suhrep-
resentations an be estimated uniquely. In order to do so, some properties of the
involved state-spae mappings must be introdued [22℄. The input-output repre-
sentation
F u ( M , ∆(d, Υ))
, namely the parameterizationM (θ b )
(blak-box ase)or
M ( θ g )
(gray-box ase2), over the parameter setΘ
, is a mapping whih an bedened as follows:
2. Hereafter, when gray-box representations are mentioned, it is meant that both dark-and
Denition 2.4.1. A blak-box model struture
χ (b) d,Υ
is a dierentiable map, for a partiulard
andΥ
, from a parameter setθ b ∈ Θ ⊂ R n Θ
to a set of models (see Eq. (2.28))χ (b) d,Υ : Θ ⊂ R n Θ −→ R n y ×n u
θ b 7−→ F u ( M (θ b ), ∆(d, Υ)).
(2.29)Denition 2.4.2. A gray-box model struture
χ (g) d,Υ
is a dierentiable map, for a partiulard
andΥ
, from a parameter setθ g ∈ Θ ⊂ R n Θ
to a set of models (see Eq. (2.28))χ (g) d,Υ : Θ ⊂ R n Θ −→ R n y ×n u
θ g 7−→ F u (M(θ g ), ∆(d, Υ)).
(2.30)Inordertodenethe mostimportantmappropertiesinagenerimanner, letus
onsider two sets denoted by
X
andY
and a mapf : X → Y
. Then the followingdenitions an beintrodued[22℄.
Denition 2.4.3. The map
f
is surjetive if and only if,∀ y ∈ Y
there exists anf (x), x ∈ X
suh thaty = f (x)
.Denition2.4.4. Themap
f
isinjetiveifandonlyif,∀ x 1 , x 2 ∈ X, f (x 1 ) = f(x 2 )
implies that
x 1 = x 2
.Denition 2.4.5. The map
f
is bijetive if it is both injetive and surjetive.Bylookingloser atthe statementgiven by Denition2.4.4,itan beonluded
that the injetivity property isrelated to the uniqueness of the involved parameter
vetor. So, thisisaveryimportantkeypropertyofagivenmodelstruturebeause,
inthegray-boxase, uniqueparametersaresoughttobeestimatedduringtheiden-
tiationproedure. First,letusfousonblak-box,fully-parameterizedstate-spae
models. In this ase, it is well-known that, in the LTI ase, suh representations
are not unique but they an only be determined from the available I/O data sets
up to a similarity transformation [60℄. The LFR framework does not form an ex-
eption eitherto this rule. In other words, blak-box representations are inherently
not identiable [78℄, sine ablak-box parameterizationis not injetive whih gives
rise toa nonunique relationbetween the parameter vetor
θ b
and the input-output representationF u ( M ( θ b ), ∆ ( d , Υ ))
[144℄. In the blak-box ase, it is interesting to denethe similaritysub-spae involvinginvertible similaritytransformation ma-tries. This goal an be reahed by following the lines found in [71, 72℄. This an
be seen as an extension of the standard LTI ase [78, 33℄. Thus, the similarity or
indistinguishable sub-spae of the LFRs an be more preisely dened as follows
[71℄,
Denition 2.4.6. Given the blok struture
K ∆
andS ∆ = { diag(T 1 , · · · , T S ), T i ∈ R n i ×n i , i ∈ { 1, · · · , S } ,
(2.31)the system realizations
( A 1 , B 1 , C 1 , D 1 )
and( A 2 , B 2 , C 2 , D 2 )
are found in the samesimilarity sub-spae if an invertible similarity transformation matrix
T ∈ S ∆
existssuh that