iden-tiation
As mentioned in the previous Setion, the information about the
parameter-a pratial onstraint makes the working point seletion step ruial in order to
ensurethatthenalLPVisabletoapturethedynamisofthenon-linearsystemto
identify eiently. A naiveapproahwouldbetoinrease, asmuhaspossible, the
numberofloalmodelsby onsidering averytiny step between two loaldata-sets.
In order to avoidthis kind of numerially and pratiallyumbersome approah,it
issuggested hereaftertomeasure the omplementaryinformationbroughtby anew
loaloperatingand deide, aording tothis measurement,
(i)
if this loaldata-setiskept for the identiationofthe LPV model, (
ii
) ifthe user-dened step betweentheurrentmodelandthenextonemustbeinreasedordereased. Intheliterature
dediatedtoLPVmodelidentiationfromloalexperiments,mostofthe time(see
amongotherse.g.,[82,143,135, 23℄),the Authorshoose equidistantworkingpoints
byseletingauser-denedbutxedstep
∆p
betweentwoloalexperimentsinterms of the sheduling variables. More preisely, by applying suh a simple operatingpointseletionapproah, theuser doesnot haveany informationof howthe system
dynamis evolve between the seleted working points. As is shown in Figure 5.1a,
byusing the sameuniedstep interms ofthe shedulingvariable
∆p
between eahworking point, the next loal model may be found too far or too lose in terms
of the system dynamis
∆β
. In the sequel, in order to address this problem, anovelapproah isproposed wherethe seletion isperformedby xing the allowable
dynamidistane
∆β
ratherthanaxedstepsizeintermsoftheshedulingvariablevetor
∆p
. In other words, by xing∆β
between the loaloperating points,it anbe ensured that the estimated loal models are evenly distributed in terms of the
system dynamis along the trajetory of the sheduling variable vetor. This
may result in a set where the xed sheduling variables
p i , i ∈ { 1, · · · , N op }
aredistributed unevenly. This novel approah isdepited inFigure5.1b. So, basially,
the proposedmethodappliesa xed stepsize grid samplingintermsof the system
dynamis
∆β
andseeks theorrespondingxedvaluesofshedulingvariablevetorp i , i ∈ { 1, · · · , N op }
. Now, the next questionwhiharises isthe following: how anwequantify thedynamis of thesystem? Inthisthesis, morepreisely,the nu-gap
metri[146℄has been seletedforthis purpose. Thanks tothis mathmeasurement,
the step between two loal experiments is automatially modied and tuned with
respet to the information brought by the next loal model. For this seletion
proedure, it is assumed that the maximal (
p max = p
) and minimal (p min = p
)possiblevaluesoftheshedulingvariablesareknown(seeEq.(2.14)forthedenition
of
p
andp
, respetively). By further assuming that an initial operating point is available,the next one should be1. far enough from the rst one to derease the omplexity of the LPV model
identiationstep;
2. lose enough tothe rst one to ensurethat the global behaviorof the system
is well-aptured.
Depending on the value of the seleted
∆β
, this rule allows us, on the one hand,to redue the number of loal models in the end (ompared with a standard
teh-nique involving a onstant step), and on the other hand, to aquire more reliable
informationabout the variationof the system dynamisalong the trajetory of the
shedulingvariables. Wemayneedmoreloalexperimentsthannallyseletedloal
modelsbutitistheprietopayiftheuserwantstoseletreliableinformationabout
the nonlinearsystemtoidentifywhenthe shedulingvariableannotbepersistently
(a)Thelassial approah. (b)Our approah.
Figure5.1: OperatingpointseletionapproahesforLPVmodelidentiationbased
onloalexperiments.
Remark 5.1. Hereafter in this Chapter, we fous on ontinuous-time
representa-tions for the sake of oniseness. Notie however that the following developments
anbeextended todisrete-timemodelsdiretly. Inthis ontext,
γ = dt d
andη = s
isthe Laplae variable where
η
is the Laplae transform ofγ
based on the denitionsgiven in Setion 2.3.
5.3.1 Nu-gap metri
In order to desribe the working point seletion tehnique, let us rst
intro-due the measure of model t, the present seletion method is based on. Being a
omplex-topologialtool, the nu-gap metri [146℄, as a dynami distane
measure-mentbetween two LTI models 2
,is alulated in [145℄ by
ν g = δ( G 1 (s), G 2 (s)) =
k (I + G 2 (s) G 2 (s)) − 1 2 ( G 2 (s) − G 1 (s))(I + G 1 (s) G 1 (s)) − 1 2 k ∞ ,
(5.3)where
G i (s)
,i ∈ { 1, 2 }
, stands for the I/O representations of two LTI model while•
,i ∈ { 1, 2 }
, andk • k ∞
stand for the omplex onjugate and theH ∞
-norm of•
,respetively. Thismeasure is bounded between 0 and1 [145℄. In addition,itan be
shown that if
ν g
islose to0, then the two onsideredmodels havesimilar dynamibehavior. Contrarily, when
ν g
is lose to 1, then the two models behavedier-ently [145, 146℄. By this way, the nu-gap metri an be onsidered as a normalized
H ∞
-norm of the dierene between two LTI models. Thanks to its advantageous properties, this tool an eiently be employed during the operating pointsele-tion proedure as adistane measurement,in terms of the dynamisof the system,
denoted by
∆β
inFigure 5.1.2. Here,thegeneral(alligraphi)notationisappliedbeausethismetrianbealulatedfor
5.3.2 Seletion algorithm
First, it is important to notie that the following seletion algorithm an be
used with any identiation method based on loal experiments. For this seletion
proedure, it is assumed that the maximal (
p max = p
) and minimal (p min = p
)possiblevaluesoftheshedulingvariablesareknown(seeEq.(2.14)forthedenition
of
p
andp
, respetively). Byassumingthatthis (weak)assumption issatised, the natural questionswhiharise when dealingwith aloalapproah are the following:Howmany working points doyou need toget a onsistentnal LPVmodel?
Howan you seletthe good working points forthe loalexperiments?
How do you measure the informationbrought by a new working point and a
new loalmodel?
These important questions an guide the user in developing an heuristi tehnique
for seleting the working points inan eient way. They more preisely show that
a distane between two onseutive loal models must be introdued to deide if a
loal model must be kept beause it gives aess to new and reliable information
abouttheglobalbehaviorofthesystemtoidentify. Inthisthesis,ithasbeenhosen
to use the nu-gap metri [146℄ to measure suh a distane beause of its ability to
quantify (between 0and 1)the behavioralsimilaritiesordierenes between toLTI
models. Notie, however, that any other similar metris, suh as e.g., the Martin
distane [91℄, may be applied instead. One this frequeny tting measurement is
introdued, aording to prior knowledge or spei pratial onstraints satised
bythe system toidentify,the user must hoose aspei rangeof admissiblevalues
for this mathing measurement
ν g
(related toa user-dened threshold alled∆β
inthe following,see Figure5.1), rangewhihmust piturethe ondene the user has
in the apability of the seleted loal models to apture the global behavior of the
system. In the sequel, depending onwhether the urrent value of
ν g
is in or out ofthe user-dened range of admissible values, the distane between the loal models
must be updated(inreased or dereased).
This senario being set, it is now neessary to translate it into an iterative
al-gorithm. Hereafter, the atual operating point and the step size are denoted by
p act
andp step
, respetively. The following proedure an be performed in order to determine thebest operatingpointsandtoidentifythe loalblak-boxmodelsfromthe loalinformationavailableatthese points. Notiehere thatby best seletion,it
is meant thatthe nal set of working pointsisthe best interms ofthe user-dened
∆β
whih is a maximal allowable distane, in the nu-gap metri sense [145℄,be-tween eahloalmodel. In the sequel, the algorithmispresented by usingonly one
shedulingvariable. Notiethatthe extensionofthealgorithmtoseveral sheduling
variablesan bearried out straightforwardly.
By onstrution, suh an algorithm is heuristi. A statistial investigation of
the eet of the parameters
∆β
and the eet of noisy data during loal modelidentiationistakled inSetion6.3, inPart II, by resortingtosimulations. Now,
in order to guide the user, some important hints to eiently selet the
hyper-parameters of the tehnique are given. First, the initial value of
p step
should behosen quitesmallw.r.t.
p max
, e.g.,p step ∈ [0.01, 0.05]p max .
(5.4)This hoie an be justied by the fat that the step size should be initially small
will determine (inrease or derease) at eah iteration a new value of
p step
if it isneessary. Seond,asfaras
∆β
isonerned,asanbeseenfromtheresultsobtainedin Setion 6.3, the admissiblehoie of this hyper-parameter depends on the noise
involved inthe loalI/Odata sequenes and on the onsidered framework 3
.
The operatingpoint determinationproedure an be startedfrom the maximal
value(
p max
)orfromanyotherintermediatevalueofthe shedulingvariablesaswell by(only) slightlymodifyingtheafore-desribedalgorithm. Whenthealgorithmannotndanydynamialhangesenrouteto
p max
,twoworkingpointswillbeseletedanyway, the ones orresponding to
p min
andp max
.Notie that this seletion algorithm an be easily implemented by using the
gapmetri funtionof the RobustControlToolbox availablein Matlab. As
men-tioned above, the proposed algorithm an be applied with any identiation
teh-niquebased onloalexperiments(suhmethodsan be,amongothers,[97, 96,157,
156℄). In this thesis, the proedure is embedded into the
H ∞
-norm-based identi-ation method developed rst in [156℄ and also utilized for [157℄. The obtainedsimulationresults are presented inSetion 6.3.
Seletion Algorithm
1. Estimate a blak-box LTI state-spae model
G i (s)
,i = 1
, in theini-tialoperatingpoint(
p act = p min
)by usingany dediatedidentiation method[78,46,45,93℄. Denetheinitialp step
asp step ∈ [0.01, 0.05]p max
andsettheoperatingpointindexas
i = i + 1
. Computethepositionofthenext workingpoint(
p i = p act + p step
).2. Performthefollowingsteps untilitholdstrue that
p act ≤ p max
.Estimateablak-boxLTIstate-spaemodel
( G i (s))
intheatualop-eratingpoint(
p act = p i
)byusinganydediatedidentiationmethod [78,46,45,93℄.Computethenu-gapmetribyusingthedenitiongiveninEq.(5.3),
as
σ = δ( G i−1 (s), G i (s)),
(5.5)If
∆β < σ < ∆β
, thenthe atualmodel(G i
) iskept,i = i + 1
andp i = p act + p step
, where the user-dened threshold∆β
and∆β
arehosen tobeequalto
(∆β − 0.1)
and(∆β + 0.1)
,respetively. Else ifσ > ∆β
, then the atualmodel (G i (s)
)is disarded,p step =
p step − ∆p step
andp i = p act − p step
.Else if
σ < ∆β
, then the atualmodel (G i (s)
)is disarded,p step = p step + ∆p step
andp i = p act + p step
.Theresultingsetontainsthekeptmodels,thedimensionofwhihisdenoted
by