C H A P T E R T H I R T E E N

RobustDesi n

Courtesyof Ford Motor Co.

13-1
EXHTBIT
modelt0 exploremany design
Rearseatbelt experiment.This experimentwas nm on a simulation
andnoiseconditions.
parameters

265
266 ChapratiS

F'ordMotor Cornpany sal'etyerrgineerswereworkingwith a supplierto betterunderstand

the performance of rear seatbelts.In any conventionalseatbelt systemwith lap and
shoulder belts,if thelapportionof thebeltridesupward,thepassenger may slidebeneath
it. potenttally
resulting in abdorninal injury.This phenomenon, callecl"submarining,"is
relatedto a largenumberof factors, including thenatureof thecollision,
thedesignof the
vehicle,theproperties of the seatsandseatbelts,andotherconditions. Basedon experi-
mentation. simulation, and analysis, Ford engineershopedto determinewhich of the
trany factorswerentostcriticalto passenger safetyand to avoiclingsubmarining.The
ttnageshou'nin Exhibit13-l depicts themodelused in Ford'ssimulation analysis.
This chapterpresents a methodfor designingand conductingexperiments to improve
the performance of productsevenin the presence of uncontroliable variations.This
nlethodis knownasrobustdesisn.

What ls RobustDesign?
We define a robttstproduct(or process)as one that performsas intendedeven under non-
ldeal conditionssuch as manufacturingprocessvariationsor a range of operatingsitua-
t i o n s W c u s e t h e t e r m n o i s e t o d e s c r i b eu n c o n t r o l l e dv a r i a t i o n st h a t m a y a f f e c t p e r f o r -
rnance,and we say that a quality product shouldbe robust to noise factors.
Robust design is the product developmentactivity of improving the desired perfor_
mance of the product while minimizing the effects of noise. In robust design *"
,r. .^-
',/ , perlmentsand data analysisto identify robust setpointsfor the designparameters
we can
c o t r t r o l .A r t t b t t s ts e t p o i n ti s a c o m b i n a t i o no f d e s i g n p a r a m e t e rv a l u e s f o r w h i c h
the
product perfotmanceis as desiredunder a rangeof operatingconditionsand
manufactur-
ing variatrons,
Conceptually,robust design is simple to understand.For a given performancetarget
(sat'elyrestrainingrear-seatpassengers,for example),there rnay
be many combinatigns
o i p a r u m c t e r ' " ' a l u etsh a t r i , i l l l , i e l c it h e d e s i r . ' dr e s u l t .I { o \ \ , r . \ e r s. r ) m eo f i h e s r :
cornbirra-
t i o n s a r e m o r e s e n s i t i v et o u n c o n t r o l l a b l ev a r i a t i o nt h a n o t h e r s .S i n c e t h e p r o d u c t
will
likely oper:atein the presenceof various rroisefactors.we would likp to
choosethe com-
bination of parametervaluesthat is leastsensitiveto uncontrollablevariation.
The robusr
- designprocessusesan experimentalapproachto finding theserobust setpoints.
To understand the concept of robust setpoints, consider two hypothetical factors
af-
fecting some measureof seat belt performance,as shown in Exhibit l3-2.
Assume that
factor A has a linear effect, fA, on performanceand factor B has a nonlinear
effect, fs.
F'urtherconsiderthat we can choosesetpointsfor each factor: Al or A.2 for factor
A, and
B1 or 82 for factor B. Assuming that the effects of fa and fs are additive,
a combination
o f A l a n d B 2 w i l l p r o v i d e a p p r o x i m a t e l yt h e s a m e l e v e l o f o v e r a l l p e r f o r m a n c e
as a
combinationof 42 and Bl. Manufacturingvariationswill be presentat
any chosen set-
polnt, so that the actual value may not be exactly as specified.
By choosing the value of
BlforfactorB,wherethesensitivityoftheresponsetofactorBisrelativelysmall,unin-
t c n d e dr , a r i a t i o ni n f a c t o r B h a s a r e l a t i v e l ys m a l l i n f l u e n c eo n o v e r a i l p r o d u c t
perfor-
mance. Therefore,the choice of Bl and A2 is a more robust combination
of seipoints
than the combinationof 82 and Al.
The robttst design processcan be used at severalstagesof the product
development
process.As with most product developmentissues,the earlierthat
robustnesscan be con-
sidered in the product development process,the better the robustnessresults
can be. Ro-
 RobustDesign 267

R e s p o n s et o
Factor B

A2 81 82
A1
1.
relationships to identifysetpointswheretheproductperformance
EXHIBIT 13-2 noUuiidesignexploitsnonlinear
the
Inlhis e>iarnple,
to variations. chosenvalue for the factorA setpointdoesnot affectrobustness'
is lesssensitive
Choosing Bl minimizesthe effectof variationin factorB on overallperformance'
whereasthat of factorB does.

canbe usedwithin the conceptdevelopment phaseas a way to

bustdesignexperiments
targets.While it is benefic.ial to
refinethe spe;ificationsand set realisticperformance
as earlyas the conceptstage, experiments for robust design
considerproductrobustness
are usednrost frequently during the detaildesignphaseas a way to ensurethe desired
productperfonnance und.l.u varietyof conditions. In detaildesign,therobustdesignac-
the right set-
iiuity iS also known asparameterdesign,as this is a matterof choosing
underour control. These include the product's materials'
pointsfor thedesignparameters
dimensiOns, tolerances, manufacturing processes,and Operating instructtotrs
designproblems, equations basedon fuudamental physicaiprin-
For manyengineering
foi robuit parameter choices.However, engineers generally cannot
ciplescanbe so-lved
variations,and noise factors that arise under real
fuily modelthe kinds of uncertainties,
the ability to developaccLlrate mathematical models is limited
conditions.Furthermore,
for many engineering problems. For example,considerthe difficulty of accuratelymodel-
problemundera wide varietyof conditions' In suchsitua-
ing the seatUettsubmarining
experillentsis
clesignerl necessary' SLrchexpert-
tio=ns, empiricalinvestigationlhrough
be usedto improve
..nt, .un be usedto directlysupportdecisionmakingand can also
theaccuracy of mathematical models'
In the case of the seat belt designproblem,Ford'sengineers wishedto testa rangeot
seatbelt designparameters and collisionconditions. However,crashtestingis vety ex-
modelwhich
pensive,so Fird^workedwith its seatbelt supplierto developa simulation
crashdata.Considering the hundreds of possiblede-
was caliUraled usingexperimental
combinaiions, collisionconditions'andotherfactors of interest' theengi-
signparameter
planned exlreriment' Al-
n"-..s chor. to explorethe simuiationmodel using a carefully
effort,the simulation model still
thoughsimulationrequiresa greatdealof computational
alloied Fordengineers to rui dozens of experiments under a wide varietyof conditions'
whichwouldnolhayebeenpossibleusingphysicalcrashtesting'
268 Cirapter13

For theFo'r'ci
seatbeltriesignteani,thegoalsof thisdesigneo
experiment
wei:eio learn:
" What cotnbination of seat,seatbelt,'an<l parameters
attaehnrent minimizerirear-seat
passengersubmnrlning during& $r6$n,
' How submarinlng ls afflctedby uncontrollable
conditions,What oonrbination
of de-
stgtlparantctefr
is rDostrobustto suchnoiselactors,?

Design of Experiments
Tlreapproachto robustdesignpresented in this chapteris basedon a methodcalledde-
sign Qf experiztt,zls (DOE). In this method, the teamidentifiesthe parameters that can
,,
'./' be controlledand the noisefactorsit wishesto investigate. The teamthendesigns,con-
ducts,andanalyzes experiments to helpdetermine theparameter setpoints
to achievero-
bustperformance.
In Japanduringthe 1950sand I960s,Dr. GenichiTaguchidevelopecl techniques to
applyDOE to improvethe qualityof productsand manufacturing processes. Beginning
with the qualitymovementof the 1980s,Taguchi'sapproachto experimental design
startedto havean impacton engineering practicein the UnitedStates;particularlyat Ford
Motor Cotnpany,Xerox Corporation,AT&T Bell Laboratories,and throughthe Ameri-
canSupplierInstitute(whichwascreatedby Ford).
Taguchireceivescreditfor promotingseveralkey ideasof experimentaldesignfor the
developmentof robustproductsand processes. Thesecontributionsinclude introducing
noise factorsinto experimentsto observetheseeffectsand the use of a signal-to-noise
ratio metric includingboth the desiredperformance(signal)and the undesiredeffects
(noise).While statisticians had beenshowingengineeishow to run experimentsfor
decades, it rvasnot until Taguchi'smethodswerewidely explainedto the U.S. manufac-
turing industryduringthe 1990sthat experimentsbecamecommonlyutilized to achieve
robustdesign.
DOE is not a substitutefor technicalknowledgeof the systemunderinvestigation.
In fact, the team shoulduse its understanding of the productand how it operatesto
choosethe right parameters to investigateby experiment. The experimental risults can
be usedin conjunctionwith technicalknowledgeof the system'inorder to make the
bestchoicesof parameter setpoints,Furthermore, the experimental resultscan be used
to build bettermathematical modelsof the product'sfunction.In this rvay,experimen-
tation complements technicalknowledge.For example,Ford engineershave basic
mathematical modelsof seatbelt performance as a functionof passenger sizesand col-
lisiontypes.Thesemodelsallow Fordto sizethe mechanical elementsandto determine
the belt attachment geometry.Basedon empiricaland simulationdata,Ford's analyti-
cal modelsandseatbelt designguidelines gainprecisionovertime,reclucing the need
for time'consurningempiricaland simulationstudies,Eventually,this technicalknowl-
edgemay improveto the point whereonly confirming testsof new seatbelt configura-
tions arerequired.
Basicexperimentaldesignand analysisfor productdevelopmentcan be successfully
plannedandexecutedby the development team.However,the held of DOE hasmanyad.
vancedmethodsto addressa numberof complicatingfactorsand yield more useful ex-
perimentalresults.Developmentteamsthus can benefitfrom consultingwith a statisti-
cian or DOE expertwho can assistin designingthe experimentand choosingthe best
analyticalapproach.
-------iFt-

DesiPn 269
Robust

The RobustDesignProcess
process:
thisseven-step
To developa robustproductthroughDOE,we suggest
l.Identifycontrolfactors,noisefactors,andperformancemetrics.
2. Formulatean objectivefunction'
3. DeveloPthe exPerimentalPlan'
4. Run the exPeriment.
5. ConducttheanalYsis.
6. Selectandconfirmfactorsetpoints'
. ReflectandrePeat.

Step 1: ldentifyControlFactors,NoiseFactors,
and FedormanceMetrics
factors,
The.,rpbustdesignprocedurebeginswith identificationof threelists: control
metricsfor
noisefdctors,andperformance the experiment:

, Control factors:Theseare the designvariablesto be variedin a controlledmanner

uuderthe many
duringthe experiment,in orderto explorethe product'sperformance
Experiments are generally run at two or three
combinationsof pararnetersetpoints.
discretelevels (setpointvalues)of each factor. These parameters are calledcontrol
variables that can be specified tbr productiorl
factorsbecausethey are amongthe
the webbing stiffnessand coefficientof
and/oroperationof ihe product.lor example,
frictionarecontrolfactorsofinterestfortlreexperlment.
. Noisefactors..Noise fa'ctorsare variablesthat cannotbe explicitly controlledduring
manufac-
the minufacturingand operationof the product.Noise factorsmay include
or operating
turing variances,changesin materialspioperlies,multipleuserscenarios
or misuseof the product. if through specialtech-
condltio,,s,and evenJeterioration
(but not in produc-
niquesthe teamcancontrolthenoisefactorsduringthe experiment
the experimentto
tion or operation),thenvariancecairdeliberatelybe inducedduping
j'l''J
,'i[":';;:L*,
Tli,i;,:li[:T,:TL::h.iilffi
:ill;:,ff
m,n'fi
mize the effectsof,this variation.For seatbelts to be usedwith a range
of seats,the
The goal is to
shapeof the seatand the seatfabric must be considerednoisefactors'
designa seatbelt systemthatworkswell regardless of thevaluesof thesefactors'
Performancemetrics:Theseare the productspecifications of interestin the experi-
productspecifications
ment.Usually the experimentis analyzedwith one or two key
this
as the perforrnancemetricsin order to find control factor setpolntsto optimize
Thesemefics may be deriveddirectlyfrom key spee ificationswherc rtl'
performance.
iSee Chapter 5, Product Specifications')i"or example,
bo.tn.r, is of criticalconcern.
how far the passenger's back or buttocksmove fonvardduringthe collisionwould be
possibleperformance metricsfor the seatbelt experiment'
thecotltrolfactors,noise
For the seatbelt designproblem,theteamhelda meetingto list
factors,andperformance metrics,As Taguchiteaches, they placedtheselists into a single
13-3"
graphic,cal|eda paranleterdiagram\ar p.diagrai?''':, shownin Exhibit
as

&
$70 Chapter13

EXHIBIT,I3.3
Parameter Control Factors Performance Metrics
diagram used to
Belt webbing stiffness Back angle
designthe seat Belt webbing friction Slip of buttocks
belt experirnent. Lap belt force limiter Hip rotation
Bold text Upper anchorage stiffness Noise Factors
Forward knee motion
indicatesthe Buckle cable stiffness
Front seatback bolster Shapeof rearseat
pcrfonnance Typeof seatfabric
Tongue friction
metric usedand A t t a c h m e n tg e o m e t r y Severityof collision
the control W e a ro f c o m p o n e n t s
factorsand noise P o s i t i o q i nogf p a s s e n g e r
P o s i t i o n i nogf b e l t so n b o d v
lactorschosen S i z eo f p a s s e n g e r
for exploration. T y p eo f c l o t h i n gf a b r i c
W e bm a n u f a c t u r i n va gr i a t i o n s
L a t c hm a n u f a c t u r i nvga r i a t i o n s

After listingthevariousfactors,theteamnrustdecidewhich
oneswill be exploredby
experlment'when a largenumberof parameters are suspected of potentiallylffecting
performance, theselectionof criticaluuriubl.,canbe subsiantiallynarroweduy J\ing un-
alyticalmodelsand/orby runninga sctreening ,xp"ri^ent with two levelsfbr eachof
many factors'Thena finer experimentis run with two
or more levelsof the few parame-
tersbelievedto affectperformance.
Fordengineers consitleredthelistsshownin ExhibitJ3-3.Theychoseto fbcusthe
perimenton exploration ex-
of sevenseatbeltparameters, holdingconstant the geometriclo-
cationsof the threeaftachment points.They decidedto use;back angleat peak,,
as the
outputmetric,the anglethat the passenger's backmakeswith respectto vertlcalat the
momentof maximumrestraint'Back angleis a smaller-is-better
performance metrc,
measured in radians.
A primaryconcemin this experiment was the effectof threeparticularnoisefactor.s:
seatshape,fabrictype,and severityof collision.Through
prelimrnaryanalysis,the teanr
foundthe bestandworstcombinations of thesenois. cJnditionswith respectto the sub-
mariningeffect.Thesethreenoisefactorswerethereblr
combined,i"r"i*J ._,r;n;;-;,r.
conditionsfbr the of the experiment. This
]lurposes approach, known as contpounded
noise'canbe helpful when many noisefactorsmust be considered. (SeeTestingNoise
Factorsin Step3.)

Step 2: Formulate an Objective Function

rust be transformedinto an objective
/itnctrcn
nce. Severalobjectivefunctionsare usefulin
ranceconcems. They canbe forniulatedeitner
, andtheyinclude;
' Maximizing;This type of functionis usedfor performance
dimensions wherelarger
valuesarebetter,suchas maximumdeceleration
beforebelt slippage,common forms
of this objectivefunctionq are = p or :
n n Liz,wherepri, tn. ,n.un of the experimen-
tal observationsundera giventestcondition.
--q-

RobustDesisn 271

Minintixing:This type of functionis usedfor perfomrance dimensions wheresmaller

valuesarebetter,suchas backangleat peakdeceleration, Commonformsof thisob-
jectivefunctionareI = p or 1'l= 02, whereo2 is the varianceof theexperimental ob-
servations undera given test condition.Altematively,suchminimizationobjectives
canbe formulated aslunclions suchas4 = ll p or | ,=I o:.
to be uraxintized,
, Targetvulue:'lhis type oi'flnction is uscdfor perfbrrnlnce dirncnsious u'licrevalues
closestto a desiredsetpointor targetarebest,suchasamountof beltslackening before
restraint.A commorrmaximizingform of this obtectivefunctionis q = lfp - 02"
r ulrtc,
* hcrct rs tltet;-rrg.'t
' Signal-to-rtoise ratio: This type of functionis usedparticularlyto measurerobust-
ness.Taguchiformulates this metricas a ratiowith the desiredresponse in the nu-
meratorand the variance'inthe response as the denominator. Generailythe niean
valueof the desiredresponse, suchas the meanback angleat peak,is not difficult
to adjustby changingcontl'olfactors.In the denominator, we placethe varianceof
this response (thenoiseresponse), which is to be minirnized,suchas the variance
in back angleresultingfrom noiseconditions.In practice,reducingvariauceis
rnoredifficultthanchangingthe mean.By computingthis ratio,we can highlight
robustfactorsettingsfor whichthe noiseresponse is relativelylbw as compared to
the signalresponse. A commonmaximizingtor'niof this ob.lective function is tl :
l0 log (ltzlo').
consultingwith the teamsuggested
The Ford statistician theav-
two objectivef'unctrotls:
eragebackangleat peakandtherangeofthe backangleat peak(thedrffereuce bctween
the maximumandrninimumbackangleat peakat thetwo noiseconditions to be tested).
to be minimized.Togetherthesetwo metricswouldprovide
Both of theseareobjectives
deeperinsightirttothebehavior thaneitheronealone.
of thes5,51911r
I

Step 3: Developthe ExperimentalPlan

havedeveloped
Statisticians rnanytypesof'efficientexperiurentalplans.Theseplanslay
(valuesof the controlfactorsandpossiblyalsosourcof
out how to vary thefaetor/evcrls
the noisefactors)in a seriesof experiments in orderto explorethe system'sbehavior,
SorheDOE plansare more efficientfor cliaracterizing certaintypesof systems,while
otliersprovidemorecompleteanalysis.

ExperimentalDesigns
A criticalconcernin designingexperiments is the costof settingup and runningthe ex-
perimentaltrials.Insituationswherethiscostis low, runninga largenumberof trialsand
usingan experimental designwith resolutionhigh enoughto exploremorefactors,factor
combinations,and interactions may be feasible.On the otherhand,whenthe costof ex-
perimentationis high, efficientDOE planscan be usedthat simuitaneously changesev-
eral factors at once. Some of tlte most popular experinlentd designs ate listed below and
depictedin Exhibit 13-4.Each one has importantuses,
. Full factorial: This design involves the systematicexploration of every combittationof
levels of each factor.This allows the team to identify all of the multifactorinteraction
272 Chuprer'
13

Full-Factorial Matrix 1/2 Fractional Factorial Matrix

A2 A1 A2
B1 a2 B1 B2 B1 B2 81 82
c 1 l c 2 c 1 l c 2 c 1 c2 c1 c2 c 1 l c 2 c2 ? 1 c2 c1 c2
D1 )' o2 ) 1 D z l D l l D1' D1 D2IDl )'1 >2 ) l D 2 I D l irlor ) 1 )2
1n
X X X X x 31
,,,1
X X X X X

-l FI

FI
tt

G1 x
X

X v. X
X
X
t1

x x X x I X' x x
61 x x x X x x 31 x X x
x x x x x x x x x x x
G1 x x
ez
Irz X

X
X
X x

1/4 FractionalFastorial Matrix 1./8 Fractional Factorial Matrix

A1 A2
B1 82 B1 s2
cllcz cl c 2 c 1 c2 c l c2
)1 ) 2 l D 1 l D ') 1 o2 D l I D : D I I D i D I I D i D l I D 2 I D

L8 Orthogonal Array
(1/16 FractionaF
l a c t o r i a lMatrix) One Factor at a Time

EXHIBIT 13'4 S e v e r aal l t e m a t i v ee x p e r n n e n t apll a n si b r s e v e nf a c t o r s( A , t s , C , D , E , F , a n d G ) a t r w o l e v e l s

c a c h .T h e f t r l l - l a c t o r i ael r p e r i r n e nct o n t a i n s2 1= 1 2 8t r i a l s ,w h r l e t h e L 8 o r t h o g o n aal r r a yd e s i g nc o n t a i n so n l y t r i a l s ,
8
denotedby the x llrarksitt the nratrices.The L8 ofthogonalarray plan rs the one Lrsedfor the seatbelt cxperinrentand ;s
shorvnin conventionalror.v,'column format in Exhibit l3-5.
S o r r r c e :[ r r l c l i o n r r l t i l c l o ! i a l l o y o u t 5 n d t p r c r l t i o n r R o s s ( 1 9 9 6 )
 RobttsrDesign 273

effects, i1 addition to the prirrary (main) et'fect of each factor on perfoimance. This
type of experimentis generallypracticalonly for a small numberof factorsand levels
and when experimentsare inexpensive(as with fast software-basedsimulationsor very
flexible hardware).F-oran investigarionof k factc-rrs at n let,elsbach,the nunrberof trl-
is typically in-
als in the full-factorial experirnentis lA. Full factorialexlreritnentatior'
feasiblefor an experimentwith grcaterthan four to llve ftlctot's'
. F r a c t i o n a lf a c t o r i a l : T h i s d e s i g nu s e so n l y a s m a l l t i a c t i o no i t h e c o m b i r r a t i o nuss e d
above. In exchangefor this efficiency.the ability to computethe magnitudesof all the
i n t e r a c t i o ne f f e c t si s s a c r i l r c e dl n s t e a dt.h e i n t e r a c t i o nas r e c o l l t b u n d c du i t h o l i ) c i l l l -
teractionsor with some of the main f'actoreffects.Note that the fractionalfactoriallay-
'Ihis
out still maintainsbalance within the experimentalplan. meansthat for the sev-
e r a l t r i a l sa t a n y g i v e n factor l e v e l .e a c ho f t h e o t h e r f ' a c t o r i
s s t e s t e dx t e v e r yl e v e lt h c
s a m en u m t r e ro f t i t n e s .
. Orthogonal array: This design is the smallest fractional factorial plan t[at still allows
the team tcr identify the main effects of each factor'.However, theseurain effectsare
-^coufounded
*id.ly with many interactioneft'ects.Nevertheless,orthogonalarray layoutsare
u t i l i z e d i n t e c h n i c a l i n v e s t i g a t i o n sb e c a u s et h e y a r e e x t r e m e l y e f f i c i e n t .
taguctrl popularized the orthogonal array DOE approach,even though statisticials
had developedsuch plans severaldecadesearlierand the roots of thesedesignscarrbe
-
t-. traceclback many centuries.Orthogonalarray plaus are rtautedaccordilrgto the num'
b e r o f r o w s ( e x p e r i m e n t si)n t h e a r r a y : I 4 , L 8 , L 9 , L 2 1 , a n d s o o n . T h e a p p e n d i xt o
thrs chaptershows severalorthogol)alarray erperittretltaiplans
, One factor at a tinte: This is an unbalancedexperimentalplan becauseeach trial is
conductedwith all but one of the factorsat nominal levels (and the first trialhaving all
consideredto be an ineffectivervay
the factorsat the nominal level). This i-sgenerall-'-
-
to explore the factor space,even though the nunber of trials is srnall, I + ft (n l).
However, for parameteroptimizationin systemswith significantillteractions,an adap-
tive versron of the one-at-a-timeexperirnentalpl-anhas been shown lr be generally
l more efficient than orthogonalarray plans (Frey et al., 2003)

The Ford team choseto use the L8 orthogonalarray experimentdesignbecausethis plan

would be an efficient way to exploresevenf-actorsat two levels each.Subsequeutrout]ds
of experimentationcould later be used to explore additionallevels of key parametersas
well as interactioneffects if necessary.The orlhogonalarray experimentalplan is shown
i n E x h i b i tl 3 - 5 .

Testing Noise Factors

Severalmethodsareusedto explorethe effectsof ttoisefactorsin experintents. If some
noisefactorscanbe controlledfor thepurposeof the experiment, then it may be possible
noisefactorscannotbe controlled
theeffectof thesenoiset'actors.ll'the
to tlirectlyassess
duringthe experiment. we allow the noiseto vary naturallyand sirnpl5,assesstheprod'
uct'sperfbrmance in thepresenceof noise,Somecomnron ways to test lllctors
lltllse are:

. Assignadditionatcolumnsin theorthogonal an'ayor fractional la,vout

factorial to the
noisefactors,essentially This
treatingthe noiseas anothervariable. allowsthe effects
of thenoisefactorsto be deternrinedalongrviththecontrolfactors. *i
274 Cl+dpter13

EXHIBIT
13.5
Factor Description
Factor
usstgntnents Belt webblng stiffnessrcompliancecharacteristic of the webbing measuredin a
ondthoLg t e n s r l el o a d m a c h i n e
orthogonul unoy B Edt wrbblng filctlonr Coofflcirntof frlction,whlch ie a functlono{ tho b6lt weeve
expeflmsnt a n d s u r f a c ec o 6 t i n g
clesign usedfbr c Lop bclt forco llmltor: Allowscontrolledreleaspof the eet belt at a cenainforce rever
D Upper anchoregestlffness:cgmpliancecharacteristiof the structureto which the
theseatbelt
u p p e ra n c h o r a g e( D - l o o p i)s m o u n t e d
experiment. This Buckle cable stiffness:compliancecharacteristic of the cablesby which the bucrre
DOEplantests i s a t t a c h e dt o t h e v e h i c l eb o d y
sevenfactorsat F Front seatbackbolster: Profi and stiffnessof seat backwhere the kneesmav conracr
two levelseach. G Tongue friction: coefficient frictionfor the bearingarea of the tongue which
Eachrowwas s l i d e sa l o n gt h e w e b b i n g
replic1Lcrl tw,ice.
underthehvo
cornpounded
noiseconditions,
y i e l d i n gl 6 t e s t 1 1 1 1 1 1 1 1
datapointsfor
analysis, 2 1 1 1 2 2 2 2
3 1 2 2 1 1 2 2
4 1 2 2 2 2 1 1
5 2 1 2 1 2 1 2
6 \ 2 1 2 2 j 2 1
7 2 2 1 1 2 2 1
8 2 2 1 2 1 1 2

' Use anouter array for the noise factors.This method testsseveralconrbinations
of the
noire factors for each rorv in the main (inner) anay. An example
of this approachis
shown tn the appenclix,where the outer array consistsof an L4
design,testing combi-
nationsof threenoise ractorsby replicatingeach row four times.
' Run replicatesof eachrow, allowing the noise to vary in a natural,
uncontrolledman-
n e r t h r o u g h o u tt h e e x p e r i m e n t r. e s u l t i n gi n m e a s u r a b l ev a r i a n c e p e r f o r m a n c e
in for
e a c hr o n ' W i t h t h i s a p p r o a c hi,t i s p a r t i c u l a r i yi m p o r t a n t o r a n d o m i z e
t h e o r d e ro f t h e
trials so that any trends in the noise are unlikely to be correlatedwith
the systematic
changesin the control factors.(SeeStep4.)
' Run replicatesof each row with compottndecl nQise.ln this method, selectednoise fac-
tors are combined to createseveralrepresentativenoise
conditions or extreme noise
conditions.This approachalso yields measurablevariancefor
each row, which can be
attributedto the effect of'noise.

The Ford team chosEto tttilize the compoundednoise

approachin the seat belt experi-
ment' The team testedeachrow using the two combinations
of the threenoise factorsrep-
resentingthe best- and worst-caseconditions.This resulted
in l6 experimentalruns for
the L8 DOE plan, as shown in Exhibit lJ-5.
-T-

Design 275
Rctl.tust

A B c D E F G N- N+ Avg Range

1 1 1 1 1 1 1 1 0.3403 0.2915 0.3159 0.0488

2 1 1 1 2 2 2 2 0 4608 0 . 3 9 8 4 0.4296 0,0624
3 1 2 2 1 1 2 2 0 3682 0 3627 0 3655 0 0055
4 " 1 2 2 2 2 1 1 o 2961 0,2647 0 . 2 8 0 4 0 0 3 1 4
5 2 1 2 1 2 , 1 2 0 , 4 4 5 0 0 . 4 3 9 8 o 4424 0.0052
6 2 1 2 2 1 2 1 0.3517 0.3538 0.3528 0.0021
7 2 2 1 1 2 2 1 0.3758 0 3580 0.3669 0 0178
8 2 2 1 2 1 ' , \ 2 0 4504 0.407
6 0.4290 0.0428

fromtheseatbeltexperiment.
EXHIBIT 13-6 Dataobtained

Step4: Runtfra e*periment

To executethe experiment,the product is testedunder the various treatmentconditions
describedby each row in the experimentalplan. Randomizingthe sequenceof the experi-
mental runs ensuresthat any systematictrend over the duration of the experimentis not
correlatedwith the systerhaticchangesto the levels of the factors.For example,if the ex-
perimentsof the L8 plan are not randomized.and the test\nditions drift over time, this
e f f e c t m a y b e i n c o r r e c t l y a t t r i b u t e dt o f a c t o r A s i n c e t h i b c o l u m n c h a n g e sh a l f i v a y
through the experiment.For some experiments.changingcertain factorsntay be so diffi-
cult that all trials at each level of that f actor are rur.ltogetherand only partial ratrdorrriza-
t i o n u r a y b e a c h i e v e dI.n p r a c t i c e r, a n d o r n i z et h e t r i a l sr v h e n e v epr r a c t i c a l a
. n dw h e n n o t
p o s s i b l ev, a l i d a t et h e r e s u l t sr v i t h a c o n f i r m a t i o nr t t n ,( S e eS t e p6 . )
I n t h e s e a tb e l t e x p e r i m e n te, a c ho f t h e e i g h t f a c t o rc o m b i n a t i o n isn t h e L 8 d e s i g nw a s
t e s t e du n d e rt h e n v o c o m p o u n d e d n o i s e , c o n d i t i o nTs h , e 1 6 d a t ap o i n t sc o n t a i n i n ltrh e
b a c k a n u l ed a t aa r e s h o r v ni n E x h i b i t i 3 - 6 r n t h e c o l u m n st i t l e dN - a n d N - r .

Step5: Conductthe Analysis

Therearemanywaysto analyzethe experimental data.For all but the mostbasicanalysis,
with a DOE expeftor from referingto a goodbookon
the teambenefitsfrom consulting
statistical andexperimental
analysis design.Thebasicanalyicalmethodis srtmmarizedhele.

Computing the Objective Function

Theteamr.rillhavealreadydevisedtheobjective functionsfor theexperiment andwill gen-
erallyhaveen objectivarelstcdto ths meanperfortnanee and the varianeoin perfotmartce,
$gmetimes the meanandvarlanee will be ssnrbined arrdexpressedas n rlugleobjeetivn ln
ratlo,
the form of a signal-to-noise The valuesof the objectivefunctionean lte computed for
eachrow of theexperiment. Fortheseatbeltexperiment. thecolumrson theLiglttsideo1the
tablein Exhibitl3-6 showthecomputed objective values(average
F.rnction backangleand
rangeof backangle)f'creachrcw. Recalithattheser:e bothobjectives to be minrn:rzeC.
 476 Chistter13

Computing Factor Effeqts by Analysis

of Means
The most straightforward analysis
to Jonductrryilt
simply yield the rnain effect of each
factorassignedto a columnin ihr .*p.rir*ii, rn"r. -uin
.rr*ri, arecatteothefactoref-
fecrs' The enal-vsisof'nreansinvorvesri;;t averagingail the computed objective firnc-
t i o n s f o r e a c h f a c t o rl e v e l .I n t h e
L 8 D o E e x a m p t e , i h ee f f e c to f f a c t o r
l e v e l A l ( f a c t o rA
l ] 1 : : ]l l l , l J , . . u : . l u e e o frt,r2i a, 3l s, a n J + . s i , n i r a r l y , r h e e f f e c t oEf2f ai sc t o r l e v e l
7.rheresulrs ofananalysis ofn,eun, areconventionalry
:ffiil:ir;:iX$:;r.;f#d
Exhibit l3-7 presents
the factoreffectschartsfor the seat
areplottedtbr eachof the objective belt example.Theseeffects
functjons,Exhibit l3_7(a)plotsthe overage
ntznceat eachlactorlevet(thefirst per,for_
objectivetunction),rrris cnanshowsrvhich
elscanbe usedto raiseor lowerthe factor
lev_
meanperformance. Recallthatbackangleat peakis to
beminimized, and notethatthe.i,urt,rgg.risuratfactor
tru.t,fn r-nz c2 Er Fl GI I wil

EXHIBTT
13.7
Factor effects
charts for the p o't:
seatbelt
expenment,
E o.*t
e
c -'-* v I vz f-'lftro
rfl-
E n24
a./t'" zf"
6

0.34

{ o.sz

(a)
Control Factors

o
.c 0,04
€
o
6 0,03 n
a
F1

J
AE2 \ Acz
(J
\ f f i t
E 0.02 nF2
o
P o.or
o
tr

0.00

(b)
control Factors
 Design 277
Robust

firean
minimizethe averageback anglemetric (FactorD appearsto haveno effectupon
performance.) achieve
However,theselevelswill not necessaril-r' robust performance'
-Exhibit
13-7(b)is basedontheraitgeof per/brmanceateachfactorlevel(thesecondob-
jectivefunction). thatlevelslA2 82 C2 D1 E'1t'2 Cllrvili nrinirnizc
Thischartsuggests
therangeofbackanglcat Pcak
Taguchirecommends ratiofor eachfactorlevelbe plottedin order
tliatthesignal-to-noise
Sincethe signal-to-noise ratioinch'rdes themeanperfontiance in
to identiforobustsefpoints,
the numeratoranclthe variancein the denominator, it represents a cornbination of thesetivo
obiectivesor a trade-offbetween them. Rather than specificaliy plotting the signal-to-
prefer
andstatisticians to sin4rly interpretthe tw'o ob.jecttvesto-
nolseratio,manyengineers
To
gether,giving rrrorecontrolover the trade-off. do so, the factor effectschafis sltou'nin
ixhibit l3-7 canbe compared in orderto choose a robust setpoint in the next step

Step 6: Select and Confirm Factor Setpoints

which factors
Ana[isis of means and the factor effects'chartshelp the tealn determine
and therefore how to achievero-
havJb'strong effect on mean performanceand variance.
able to reducethe
bust perforrnance.These chartshelp to identify which factorsare best
improve the per-
product,svariance(robustnessfactors;and rvhich factorscan be used to
i n s i g h t s 't h e t e a m
f b r m a n c e ( s c a l i n g f a c t o r s ) .B y c h o o s i n g s e t p o i n t sb a s e d o n t h e s e
should be able to improve the overall robustnessof the product.
r a n g eo f b a c k a n g l :
F o r e x a m p l e , . o n r i , l . , t h e e f f e c t so f f a c t o rA o n b o t h a v e r a g ea n d
w o u l d r n i n i r n i z eb a c k
i u t h e e x p e r i r n e n tT. h e c l i a ( s i n E x h i b i t l 3 - 7 s h o i v t h a t l e v e l A 1
the range of back angle, representing a ttaile-off be-
angle, bui level A2 rvould minimize
trade-off is evident in factor F. lJowever,
tween performanceand robustness,A sirnilar
t r a d e - o f f a
, n d l e v e l s8 2 , C 7 , Dl, El, and
f o r f a c i o r sB , C , D , E , a l d G , t h e r ei s n o s u c h
G I m i n i m i z eb o t h o b . j e c t i v e s '
factorsA and F
Using factorsg, C, D, E, and G to achievethe desiredrobustnessand
t h e s e t p o i n t B 2 C 2 D1 E1 Fl G1]'
t o i n c r e l s ep e r f o r m a r r c eF,o r d e n g i n e e r s e l e c t e d [Al
setpoint is not one of the eight ortl'togonal array row$
As is usualiy the case,the chosen
Given that ihis setpoint has never been tested. a confirmation
tested in the erperimenr.
been achieved'
run should be used to ensurethat the expectedrobustperformancehas

Step 7: Reflectand RePeat

may be sufficientto identif,vappropriatelyrobustsetpoints'
One roundof experiments
however,furtheroptimizationof the product's perfotmance is worthr'vhiie.
Sometimes,
roundsof experimentatton.
andthismay requireseveraladditional
Insubsequentexperimentationandtestittg,theteammaychooseto:
. Rcconsider chosenfor factorsdisplaying
thesetpoints of performance
a trade-otJ ver-
susrob:lstness'
. F . x p l r l r iel t t c r a c t l o nasn t o l t gs o m eo f t h e f a c t o r si n o r d e rt o f i r r t h e ri t t t D r o vtel - t t :
performance.
testedor outsidethis
. Fine-tunethe paratrtetersetpointsusing valuesbetweenthe levels
range.
 Design 279
Robust

fect performance, such as manufacturing variations, operating conditions, and product

deterioration
. We suggestan approachto the developmentof robustproductsbasedon designof ex-
periments(DOE). This seven-stepprocessfor robustdesignis: d.

l. Identify control factors,noise factors,and performailcemetrics.

2. Formulatean objectivefuttction.
3. Develop the experimentalplan.
4 . R u n t h ee x p e r i t n e n t .
5 . C o n d u c tt l i e a n a l y s t s .
6 . S e l e c ta u d c o n f i n n f a c t o l ' s e t p o i n t s .
7. Reflect and repeat.
. Orthogonal array experimental plans provide a very efficient method for exploring the
main effectsof eachfactor chosenfor the experiment.
. T4 achieverobust perfonnance,use of objective functions helps in capruringboth mean
p"rhtman." due to each control factor and varianceof performancedue to noise factors.
. Analysis of means and f'actor effects charts facilitate the choice of robust paralneter
setpornts.
. Becausemany nuancesare involved in successfulDOE, most teams applying these
methodswill benefit from assistance by a DOE expert.

Referencesand Bibliography
areavailableon theIntemetvla
Manycurrentresources
www.u lrich-eppinger.net
Taguchi'smetltods designanddetailsaboutot'thogonal
for experintental array
experimentation plansareexplained texts,includingTaguchi's
in several classichvo-
yolunretexttranslatccl intoEnglishPhadke provides erarnples
r'ttllerous andltrlctical
adviceon application of DOE.Rossemphasizes insights gainedthrough ANOVA analysts.
Tagr"rchi, Genichi,Svstern Design.Engineering
of Exlterintental Methods to Optimize
Quutityundlvlinintize'Costs, two volumes,LoutseWatanabe (trans.)'
Tr"rng White
P l a i n sN, Y , 1 9 8 7 .
Taguchi,,Ge nichi,Introductionto QualityEngineering;DesigningQualityinto
Prodttctsand Processes, AsianProductivity Organization (trans.andpub.),
Tokyo,1986.
Ptradke, Madhav5., Qualitt,Engineering UsingRobustDesign,PrenticeHall,
Englewood Cliffs.Ni, 1989.
Ross,PhillipJ., TaguchiTechniques.for Qualitv- Engineering, McGrarv-Hill, New
York, 1996.

GroveandDavispresenta thoroughexplanation of experimentaidesigntechniques rn

engineering. planning,
inclucling running, artdanalyzingtheexperiments. A diff'erent
analysisof Ford'sseatbeltexperimentis includedin thistext,aswell asrnanymore
automotive of robustdesign.
applications
280 Chapter13

Grot'e,Daniellvl.,andTinrothyP. Davis,Eng,ineering, [i.t1tc't.itttatrtul

Qtrulit.r.und
Design,AddisonWesleyLongman, EdinburghGate,UK, l9r)2
$evernlexcellenttextsprovidecletniledu*plrnatlons
of theuseof statisticulrncrrhods,
tlactional
fnctolialexperimentnlpluns,analytiealandgraphieallnterpretations,0rrcl
, responsesurfacemethods,
Box,GeorgeE,P.,J. StuartHunter,andwilliam G. Hunter,Statisticsfor
Experimenters' An IntrodLrction
to De,sign,
Data Anab)sis,qnclModel Bttilding, John
WileyandSons;New york, 197g.
Box, GeorgeE. P.,andNormanR. Draper,EmpiricalMoetetBuilclingancl Re.sponse
Swtaces,JohnWiley andSons,New york, 19g7.
Montgomery, DouglasC., Designlncl Anabt.sisof Experimenrs,fifth edition,John
WileyandSons,New York,2001.

R c c c nrtc s c a r chhl s r c t t c * c ci nl t c f c si tn o n e - a t - a - t i rDnO

e Ep l a n si.\ n a r l l p t i v o
c nc-
factor-at'a-1ime approach hasbeenshownto yield betterperformarlce optimization than
theconesponding orthogonal arrayclesign for systems wheretheinteraction effectsare
moresignificantthanthenoiseanderrorefl'ecrs.
Frey,DanielD., FredrikEngelhardt, andEdwardM. Greitzer,..A Rolefor one-
Factor-at-a-Time Experimentation in Parameter Design,"Research in Engineering
Design,2003.
DoE canbe usedin rnanyaspects
of productdevelopment.
Almqr.rist andwyner explain
how carefully
plannedexperinrents
areeffective
in evaluating
andtunrngparameters ot
salescanrpaigns.
Almquist,Eric,andGordonwyner, "Boostyour MarketingRol with Experimental
Design,"HervardBu-siness
Review,y or.79,No,9, octob.i zottt, pp. 135-14 1,

Exercises
I' Designan experiment to determine
a robustprocessfor makilg coffee.
2' Explainwhy the l/4-fractional-factorialand
orthogonalanayplan\shownin Exhibitl3-4
arebalanced.
3' Fonnulate an appropriate
signal-to-noiseratiofor the seatbelt expepment.Analyze
theexperinientaldatausingthismetric.Is signal-to-noise
ratioa usefulobjective
func-
tionin thiscase?Why or whv not?

Thought Q,uestions
t . If you are able to afford a largerexperiment(with more runs),
how might you best uti,
Iize the additionalruns?
2 . when would you choosenot to randomizethe order of the
experiments?How would
you guardagainstbias?
J. Explain the importanceof balancein an experimentalplan.
ry

RobustDesign 281

Appendix

OrthogonalArrays
DOE,textsprovideseveralorthogonalarrayplansfor experitnents. The sitnplestarrays
are for two-leveland three-leveifactorexperiments. Using advancedtcchniques. DOE
planscanalsobe createdfor mixedtwo-.three-.and/orfour-levelfactorexperiments and
manyotherspecialsituations. This appendixshowssomeof the basicorthogonal anays
from Taguchi'stext Introductionto QualityEngineering (1986).Theseplansareshown
in row/columnformat,with the factorlevelassignments in the columnsand the experi-
nrentalrunsin the rows,The numbers1"2. and 3 in eachcell indicatethe f'actorlevels.
(Alternatively, factorlevelscanbe labeled as- and* for trvo-level factorsor *,0, and+
for threelevels.)Recallthatthe orthogonal arraysarenamedaccording to the numberof
rowsin the design. lncluded here are the two-levelanays L4, L8, and Ll6 andthethree-
l 2 7 A l s os h o w ni s a D O E p l a t ru s i r t gt h c L 8 i n n u l l r r a y ' l i r r : e v e r t
l * , e l a r r a y sL 9 a n c L
coritrolfactorsandthe L4 outeranay fbr threenoisefactors.This planallowsanalysis of
theeffectsof thethreenoise factors.

Two-LevelOfthogonal Arrays
L4: 3 Factors at 2 Levels Each

A B c
1 1 1 1
2 1
3 2 1 a

4 2 2

L8: 7 Factorsat 2 LevelsEach

A B c D E F G
1 1 I 1 1
a
2 1 1 1 2 2 2
3 1 2 2 1 2 2
4 1 a 2 2 2 1 1
,1
5 2 4 2 1 z

6 1 I 1 t 1
7 4. 2 1 z z

8 z J 2 1 2
282 Chaprer
13

L16116 Fnctorr il € Levclr Each

A B c D E F G H I J K L M N o
1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 2 )
2 2 2 a 1
2
3 1 1 ) 2 2 2 'l
)
2 2 2
4 1 2 2 a
2 ) 2 1
1
Z 1 1 1
5 1 2 2 I ) ) ,]
2 1 2 2
6 2 2 2 2 2 2 1 2 2 1
7 2 a
2 2 1 1 )
1 2 a a
1
I 2 2 2 2 I 2 2 1 1 1 2 2
9 2 1 2 ,1
1 2 2 1 ) 1 2 1 2 1 2
10
') )
1 1 2 1 a
1 2 , 'l
a 1 2 1
11 1 2 2 1 2 1 a 1 1 2
12 1 I 1
2 1 2 1 z 1
13 z 2 1 2 2 ,l
'14
2 1 2 2
z 1 2 z 2 1 2 1 1 a

15 2 2 1 2 2 2 ') 1 2 1 1 2
16 2 2 1 2 1 ) ,l
2 1 1 a
2 2 1

Three-LevelOrthogonalArrays
L9: 4 Factorsat 3 LevelsEach

A B c D
1 1 1
2 1 a
2 2
5 3 3 3
4 2 1 2 3
1
2 3 1
6 2 3 1 2
7 3 1 ?
8 3 2 1 3
I 3 3 2
 Robu.st
Design 283

L27t13 Factorsat 3 LevelsEach

A B c D E F G H I J K L M
,] 1 1 1
1 1 1 1

2 1 I 1 2 2 2 2 2 2 2 2 2
3 1 1 3 3 3 3 3 3 3 3 3
4 1 2 2 2 1 1 1 2 2 2 3 3 3
5 2 2 2 2 2 2 3 3 3 1, 1 1
6 1 2 2 2 3 3 3 1 1 1 2 2 2
7 1 3 3 3 1 1 1 ) 3 3 2 2 2
I 1 3 3 ) 2 2 1 1 1 3 3 J

9 3 3 3 3 ? 2 2 a

10 1 I ? 1 2 1 2
11 z a, 1 z z 1
"1,2 a
z 3 3 Z 3 1 z 1 I

,l
13 2 2 3 1 2 3 2 3 1 3 1 2
,1
14 2 2 3 1 2 3 1 3 2 2 3
15 a a
3 1 3 1 ) 1 2 3 a
z 3 1
,1
16 2 3 1 2 1 2 3 3 1 2 2 3
) ,1 2
17 2 3 1 2 3 1 2 3 1

18 a
3 1 2 J 1 ) 3 1 2 3
19 3 1 3 2 3 2 1 3 2 3 2
20 3 3 2 2 1 3 2 1 3 ) ?

21 3 3 a 3 2 1 3 2 3 2
22 3 1 3 3 z I 3
23 3 z 2 1 3
,l
24 ?
2 3 3 I 1 3 2 2 J

25 3 J
a
1 1 3 z 3 2 1 2 1 3
') 3 3 3 2 1
26 3 ? 1 1 1

27 3 J 2 1 3 2 2 1
3 1 3 2
284 Chdpter13

CombinedInnerand Outer Arrays

L8 x L4:7 control Factorsand 3 NoiseFaqtorsat 2 LevelsEach

1 1 2 2 Na
1 2 1 2 Nb
A B c D E F G 1 2; 2 1 Nc
1 1 ,l
1 1 1 1
2 1 1 1 2 2 a

3 1 z z 1 z
4 1 z z I z 1 1
5 I a 1 2 1 z
6 2 z 2 1 2 1
7 z 1 1 z 2 1
I Z 1 1 1

i
I

l
I
I