Figure 5.1:Designproedure basedon theV-Model
Therefore,additional steps have to be inorporated into this model tomake
it possible to derive quantitative requirements of subsystems from overall
re-quirements ofa large-salesystem[78 ℄.
In the rst step, it should be laried, whih design variables of eah
sub-systeminuenewhihobjetivequantities,possiblyinterat withotherdesign
variables of other subsystems, and aet the overall performanes of a
large-sale system. This is done by reating a so-alled dependeny graph. In the
seond step, so-alled bottom-upmappings, heresurrogate models,have to be
established, whih evaluate thequantitative performanes of a large-sale
sys-tem, depending on the design variables of the subsystems. Surrogate models
mustbeassimpleaspossibleandasomplexasneessary. Inthethirdstep,the
so-alled top-downmappings,permissiblerangesofdesignvariablesarederived
from quantitative requirements formulated on the large-sale system. Figure
5.2 showsthe overviewof allthese three steps.
Figure 5.2:Three-enablersforderivingthequantitativerequirementsof
subsys-tems from theoverall requirements ofa large-sale system
Werstlarifywhatadependenygraphisandhowitisonstruted. Seond,
weexplainthebuttom-upmapping proedureandneessarymathematialand
physial surrogate models for ontrol systems, atuators, and vehiles. Third,
we desribe the proedure of top-down mapping and give a brief review of
the optimization method whih delivers permissible ranges of design variables
of atuators and ontrol systems with respet to unertainties and objetive
driving dynamis performanemeasures.
5.1 Dependeny graph
In the proess of designing an atuator and parametrization of a ontrol
sys-tem,a struturedmodelisrequired whih doumentsthe dependeniesamong
thedesign variables ofthesubsystemsand theobjetivedrivingdynamis
per-formane measures. This model is important, as it an manage know-how in
the proess ofthe development. A so-alled dependeny graph is usedfor this
purpose.
As explained in hapter 2, the vehile performane is identied in dierent
dynamis domains, namely longitudinal, lateral and vertial domain. Eah
domain isassessed by dierent assessment indies (AI). These AIsare targets
on thehighestleveloftheV-model,i.e. on thesystemLeveldepited ingure
5.1 . EahAIisharaterized bydierentmaneuversandtherelevant objetive
driving dynamisperformanemeasuresofthese maneuvers. Forexample,"AI
ornering" is one of the assessment indies in the domain of vehile lateral
dynamis. This AI is evaluated by dierent maneuvers suh as aeleration
while ornering (ACWC), brake while ornering (BRWC), quasi steady state
ornering(QSSC),Rampsteer,et. Refereningagaintothehapter2,eahof
thesemaneuvershasdierentobjetivedrivingdynamisperformanemeasures
whih are altered by the hassis' properties, namely subsystem properties in
the V-model. These subsystem properties are also aeted by the properties
ofomponentsomposingthesubsystem. Theexamplesofsuhomponentsin
thehassis areontrolsystems andthe atuators.
So, a dependeny graph is a graph in whih the measurable system
(AIs),objetivedrivingdynamisperformanemeasuresandomponent
prop-erties(ontrol systemparametersandatuatorsdesign variables)aredisplayed
asverties. Their dependenies areexpressedasanarrow.
5.2 Bottom-up mapping
For bottom-up mapping, quantitative evaluation of outputs as a funtion of
inputs, amodelisrequired whihrepresentsthe physial relations between
in-puts and outputs, i.e. design variables of atuators and ontrol systems and
objetive driving dynamis performane measures. The physial model of the
vehile, funtional model of the atuators, and the model of the ontrol
sys-temhavebeenpointedoutintheprevioushapters. Simulationofthesemodels
andlateroptimizationoftheoutputsofthe simulation areusuallyrestritedby
verylargeomputationaltime. Asinglesimulationofouromplex modeltakes
several hours or even a day. Therefore,the optimization problem, nding
per-missible ranges of design variables,is very time-onsuming, beauseit usually
needs morethan thousandsimulation iterations. Therefore,substitutemodels,
based on mathematial funtions, an be applied inorder to approximate and
preditthe realmodelbehaviorwithextremely loweralulation time[31 ℄,but
at a prie of yielding a small error on thenal result. Asmentioned before, a
meta-model an also predit theoutput of the physial modelwith respet to
new sets ofdesign variableswithout a need forfurther simulations.
Meta-models usedinthis work areassoiated withlassiation and
regres-sion methods of the supervised learning tehnique of mahine learning, whih
are trained byalreadyknowninput andoutputdata and arethenable to
pre-dit the output under new ombination of input parameters [62 ℄. The used
regression modelsinthis work areartiial neural networks. Thesupport
ve-tor mahine is then used for the lassiation. These meta-models need an
adequate numberof inputs andoutputs. Asa result, we introdue inthenext
of inputs and outputs. Afterwards, we give a brief reviewof howthe artiial
neural networks(ANN) andthe supportvetor mahine(SVM) workand why
weneed both ofthem.
5.2.1 Design of experiments
InordertotrainANNandSVMproperly,wehavetoonsidertworequirements
on thedesign spaeof alldesign variables (inputsto ANNandSVM):
•
The sampleof the designspae mustbewelldistributed.•
Thedesignspaemustbeoveredasmuhaspossiblewithminimalnum-berof sample pointsin order to aelerate theproess of the simulation
and optimization.
Theonsequeneoftherealizationofthesetworequirementsisthatthe
meta-modelan bewellttedto the realmodelandlater predit theoutputwellfor
a newombinationof inputvariables.
There aredierent kindsof samplingmethods [69℄. Someof them are
•
Sobolsequene•
SrambleSobolsequene•
Full fatorial sequene•
Halton sequene•
SrambleHalton sequeneThequestionarises,whihmethodfulllstheabove-mentioned requirements
most properly. Therefore, we introdue a term, alled disrepany, whih
dif-ferentiates between sampling methods. Imagining that the number of design
variables is
P
,the design spae islabeledasΩ ds
withN
samples,andz
isanarbitrarypointfromthedesign spae
Ω ds
,b[0, z]
isdened asaboxontainingall points loated inthe boxfrom origin to
z
.n(b)
isthe number of points inthis boxand
V ol(b)
showsthevolumeofthebox. Aordingly,thedisrepanyof thepoint
z
is omputed asfollows:Discrepancy(z) =
n(b)
N − V ol(b)
(5.1)
Anotherterm is dened asfollows:
D ∞ = max
z∈Ω ds
(Discrepancy(z))
(5.2)The smaller
D ∞
is , the better the method is in terms of the denedre-quirements. It means, the method with the smallest
D ∞
an sample theP-dimensional design spae
Ω ds
eiently by overing most of its orners andedges. In this way, the data will be more informative by minimum possible
number of samples and aordingly simulations. [72℄ investigates all possible
methods ofsamplingbasedon theintrodued terms. Theresultisthat
Sram-bleSobolsequeneandSrambleHaltonsequenearethebestoptionsinterms
of