12/30/2016

Chapter2IntroductiontoTaguchiMethod

CHAPTER2
INTRODUCTIONTOTAGUCHIMETHOD
2.1Background
Thetechniqueoflayingouttheconditionsofexperiments[6]involvingmultiplefactorswasfirstproposed
bytheEnglishman,SirR.A.Fisher.Themethodispopularlyknownasthefactorialdesignofexperiments.A
fullfactorialdesignwillidentifyallpossiblecombinationsforagivensetoffactors.Sincemostindustrial
experimentsusuallyinvolveasignificantnumberoffactors,afullfactorialdesignresultsinalargenumber
ofexperiments.Toreducethenumberofexperimentstoapracticallevel,onlyasmallsetfromallthe
possibilitiesisselected.Themethodofselectingalimitednumberofexperimentswhichproducesthemost
informationisknownasapartialfractionexperiment.Althoughthismethodiswellknown,thereareno
generalguidelinesforitsapplicationortheanalysisoftheresultsobtainedbyperformingtheexperiments.
Taguchiconstructedaspecialsetofgeneraldesignguidelinesforfactorialexperimentsthatcovermany
applications.

2.2Basicconcepts
2.2.1Definition
Taguchihasenvisagedanewmethodofconductingthedesignofexperimentswhicharebasedonwell
definedguidelines.Thismethodusesaspecialsetofarrayscalledorthogonalarrays.Thesestandardarrays
stipulatesthewayofconductingtheminimalnumberofexperimentswhichcouldgivethefullinformation
ofallthefactorsthataffecttheperformanceparameter.Thecruxoftheorthogonalarraysmethodliesin
choosingthelevelcombinationsoftheinputdesignvariablesforeachexperiment.
2.2.2Atypicalorthogonalarray
Whiletherearemanystandardorthogonalarraysavailable,eachofthearraysismeantforaspecificnumber
ofindependentdesignvariablesandlevels.Forexample,ifonewantstoconductanexperimentto
understandtheinfluenceof4differentindependentvariableswitheachvariablehaving3setvalues(level
values),thenanL9orthogonalarraymightbetherightchoice.TheL9orthogonalarrayismeantfor
understandingtheeffectof4independentfactorseachhaving3factorlevelvalues.Thisarrayassumesthat
thereisnointeractionbetweenanytwofactor.Whileinmanycases,nointeractionmodelassumptionis
valid,therearesomecaseswherethereisaclearevidenceofinteraction.Atypicalcaseofinteractionwould
betheinteractionbetweenthematerialpropertiesandtemperature.

L9(34)Orthogonalarray
Performance

IndependentVariables
Experiment#
1
2
3
4
5
6
7
8

Variable1
1
1
1
2
2
2
3
3

Variable2
1
2
3
1
2
3
1
2

http://www.ecs.umass.edu/mie/labs/mda/fea/sankar/chap2.html

Variable3
1
2
3
2
3
1
3
1

ParameterValue
Variable4
1
2
3
3
1
2
2
3

p1
p2
p3
p4
p5
p6
p7
p8
1/4

12/30/2016

Chapter2IntroductiontoTaguchiMethod

p9

Table2.1LayoutofL9orthogonalarray.
TheTable2.1showsanL9orthogonalarray.Therearetotally9experimentstobeconductedandeach
experimentisbasedonthecombinationoflevelvaluesasshowninthetable.Forexample,thethird
experimentisconductedbykeepingtheindependentdesignvariable1atlevel1,variable2atlevel3,
variable3atlevel3,andvariable4atlevel3.
2.2.3Propertiesofanorthogonalarray
Theorthogonalarrayshasthefollowingspecialpropertiesthatreducesthenumberofexperimentstobe
conducted.
1.Theverticalcolumnundereachindependentvariablesoftheabovetablehasaspecialcombinationof
levelsettings.Allthelevelsettingsappearsanequalnumberoftimes.ForL9arrayundervariable4,
level1,level2andlevel3appearsthrice.Thisiscalledthebalancingpropertyoforthogonalarrays.
2.Allthelevelvaluesofindependentvariablesareusedforconductingtheexperiments.
3.Thesequenceoflevelvaluesforconductingtheexperimentsshallnotbechanged.Thismeansonecan
notconductexperiment1withvariable1,level2setupandexperiment4withvariable1,level1
setup.Thereasonforthisisthatthearrayofeachfactorcolumnsaremutuallyorthogonaltoanyother
columnoflevelvalues.Theinnerproductofvectorscorrespondingtoweightsiszero.Iftheabove3
levelsarenormalizedbetween1and1,thentheweighingfactorsforlevel1,level2,level3are1,0
,1respectively.Hencetheinnerproductofweighingfactorsofindependentvariable1and
independentvariable3wouldbe
(1*1+1*0+1*1)+(0*0+0*1+0*1)+(1*0+1*1+1*1)=0
2.2.4Minimumnumberofexperimentstobeconducted
Thedesignofexperimentsusingtheorthogonalarrayis,inmostcases,efficientwhencomparedtomany
otherstatisticaldesigns.TheminimumnumberofexperimentsthatarerequiredtoconducttheTaguchi
methodcanbecalculatedbasedonthedegreesoffreedomapproach.

Forexample,incaseof8independentvariablesstudyhaving1independentvariablewith2levelsand
remaining7independentvariableswith3levels(L18orthogonalarray),theminimumnumberof
experimentsrequiredbasedontheaboveequationis16.Becauseofthebalancingpropertyoftheorthogonal
arrays,thetotalnumberofexperimentsshallbemultipleof2and3.Hencethenumberofexperimentsfor
theabovecaseis18.

2.3AssumptionsoftheTaguchimethod
Theadditiveassumptionimpliesthattheindividualormaineffectsoftheindependentvariableson
performanceparameterareseparable.Underthisassumption,theeffectofeachfactorcanbelinear,quadratic
orofhigherorder,butthemodelassumesthatthereexistsnocrossproducteffects(interactions)amongthe
individualfactors.Thatmeanstheeffectofindependentvariable1onperformanceparameterdoesnot
dependonthedifferentlevelsettingsofanyotherindependentvariablesandviceversa.Ifatanytime,this
assumptionisviolated,thentheadditivityofthemaineffectsdoesnothold,andthevariablesinteract.

2.4Designinganexperiment
Thedesignofanexperimentinvolvesthefollowingsteps
1.Selectionofindependentvariables
2.Selectionofnumberoflevelsettingsforeachindependentvariable
http://www.ecs.umass.edu/mie/labs/mda/fea/sankar/chap2.html

2/4

12/30/2016

Chapter2IntroductiontoTaguchiMethod

3.Selectionoforthogonalarray
4.Assigningtheindependentvariablestoeachcolumn
5.Conductingtheexperiments
6.Analyzingthedata
7.Inference
Thedetailsoftheabovestepsaregivenbelow.
2.4.1Selectionoftheindependentvariables
Beforeconductingtheexperiment,theknowledgeoftheproduct/processunderinvestigationisofprime
importanceforidentifyingthefactorslikelytoinfluencetheoutcome.Inordertocompileacomprehensive
listoffactors,theinputtotheexperimentisgenerallyobtainedfromallthepeopleinvolvedintheproject.
2.4.2Decidingthenumberoflevels
Oncetheindependentvariablesaredecided,thenumberoflevelsforeachvariableisdecided.Theselection
ofnumberoflevelsdependsonhowtheperformanceparameterisaffectedduetodifferentlevelsettings.If
theperformanceparameterisalinearfunctionoftheindependentvariable,thenthenumberoflevelsetting
shallbe2.However,iftheindependentvariableisnotlinearlyrelated,thenonecouldgofor3,4orhigher
levelsdependingonwhethertherelationshipisquadratic,cubicorhigherorder.
Intheabsenceofexactnatureofrelationshipbetweentheindependentvariableandtheperformance
parameter,onecouldchoose2levelsettings.Afteranalyzingtheexperimentaldata,onecandecidewhether
theassumptionoflevelsettingisrightornotbasedonthepercentcontributionandtheerrorcalculations.
2.4.3Selectionofanorthogonalarray
Beforeselectingtheorthogonalarray,theminimumnumberofexperimentstobeconductedshallbefixed
basedonthetotalnumberofdegreesoffreedom[5]presentinthestudy.Theminimumnumberof
experimentsthatmustberuntostudythefactorsshallbemorethanthetotaldegreesoffreedomavailable.In
countingthetotaldegreesoffreedomtheinvestigatorcommits1degreeoffreedomtotheoverallmeanof
theresponseunderstudy.Thenumberofdegreesoffreedomassociatedwitheachfactorunderstudyequals
onelessthanthenumberoflevelsavailableforthatfactor.Hencethetotaldegreesoffreedomwithout
interactioneffectis1+asalreadygivenbyequation2.1.Forexample,incaseof11independentvariables,
eachhaving2levels,thetotaldegreesoffreedomis12.Hencetheselectedorthogonalarrayshallhaveat
least12experiments.AnL12orthogonalsatisfiesthisrequirement.
Oncetheminimumnumberofexperimentsisdecided,thefurtherselectionoforthogonalarrayisbasedon
thenumberofindependentvariablesandnumberoffactorlevelsforeachindependentvariable.
2.4.4Assigningtheindependentvariablestocolumns
Theorderinwhichtheindependentvariablesareassignedtotheverticalcolumnisveryessential.Incaseof
mixedlevelvariablesandinteractionbetweenvariables,thevariablesaretobeassignedatrightcolumnsas
stipulatedbytheorthogonalarray[3].
Finally,beforeconductingtheexperiment,theactuallevelvaluesofeachdesignvariableshallbedecided.It
shallbenotedthatthesignificanceandthepercentcontributionoftheindependentvariableschanges
dependingonthelevelvaluesassigned.Itisthedesignersresponsibilitytosetproperlevelvalues.
2.4.5Conductingtheexperiment
Oncetheorthogonalarrayisselected,theexperimentsareconductedasperthelevelcombinations.Itis
necessarythatalltheexperimentsbeconducted.Theinteractioncolumnsanddummyvariablecolumnsshall
notbeconsideredforconductingtheexperiment,butareneededwhileanalyzingthedatatounderstandthe
interactioneffect.Theperformanceparameterunderstudyisnoteddownforeachexperimenttoconductthe
sensitivityanalysis.
2.4.6Analysisofthedata
http://www.ecs.umass.edu/mie/labs/mda/fea/sankar/chap2.html

3/4

12/30/2016

Chapter2IntroductiontoTaguchiMethod

Sinceeachexperimentisthecombinationofdifferentfactorlevels,itisessentialtosegregatetheindividual
effectofindependentvariables.Thiscanbedonebysumminguptheperformanceparametervaluesforthe
correspondinglevelsettings.Forexample,inordertofindoutthemaineffectoflevel1settingofthe
independentvariable2(referTable2.1),sumtheperformanceparametervaluesoftheexperiments1,4and
7.Similarlyforlevel2,sumtheexperimentalresultsof2,5and7andsoon.
Oncethemeanvalueofeachlevelofaparticularindependentvariableiscalculated,thesumofsquareof
deviationofeachofthemeanvaluefromthegrandmeanvalueiscalculated.Thissumofsquaredeviationof
aparticularvariableindicateswhethertheperformanceparameterissensitivetothechangeinlevelsetting.
Ifthesumofsquaredeviationisclosetozeroorinsignificant,onemayconcludethatthedesignvariablesis
notinfluencingtheperformanceoftheprocess.Inotherwords,byconductingthesensitivityanalysis,and
performinganalysisofvariance(ANOVA),onecandecidewhichindependentfactordominatesoverother
andthepercentagecontributionofthatparticularindependentvariable.Thedetailsofanalysisofvarianceis
dealtinchapter5.
2.4.7Inference
Fromtheaboveexperimentalanalysis,itisclearthatthehigherthevalueofsumofsquareofanindependent
variable,themoreithasinfluenceontheperformanceparameter.Onecanalsocalculatetheratioof
individualsumofsquareofaparticularindependentvariabletothetotalsumofsquaresofallthevariables.
Thisratiogivesthepercentcontributionoftheindependentvariableontheperformanceparameter.
Inadditiontoabove,onecouldfindthenearoptimalsolutiontotheproblem.Thisnearoptimumvaluemay
notbetheglobaloptimalsolution.However,thesolutioncanbeusedasaninitial/startingvalueforthe
standardoptimizationtechnique.

2.5RobustDesign
Amaincauseofpooryieldinmanufacturingprocessesisthemanufacturingvariation.Thesemanufacturing
variationsincludevariationintemperatureorhumidity,variationinrawmaterials,anddriftofprocess
parameters.Thesesourceofnoise/variationarethevariablesthatareimpossibleorexpensivetocontrol.
Theobjectiveoftherobustdesign[4]istofindthecontrollableprocessparametersettingsforwhichnoiseor
variationhasaminimaleffectontheproduct'sorprocess'sfunctionalcharacteristics.Itistobenotedthatthe
aimisnottofindtheparametersettingsfortheuncontrollablenoisevariables,butthecontrollabledesign
variables.Toattainthisobjective,thecontrolparameters,alsoknownasinnerarrayvariables,are
systematicallyvariedasstipulatedbytheinnerorthogonalarray.Foreachexperimentoftheinnerarray,a
seriesofnewexperimentsareconductedbyvaryingthelevelsettingsoftheuncontrollablenoisevariables.
Thelevelcombinationsofnoisevariablesaredoneusingtheouterorthogonalarray.
Theinfluenceofnoiseontheperformancecharacteristicscanbefoundusingtheratio[4]whereSisthe
standarddeviationoftheperformanceparametersforeachinnerarrayexperimentandNisthetotalnumber
ofexperimentintheouterorthogonalarray.Thisratioindicatesthefunctionalvariationduetonoise.Using
thisresult,itispossibletopredictwhichcontrolparametersettingswillmaketheprocessinsensitiveto
noise.
However,whenthefunctionalcharacteristicsarenotaffectedbytheexternalnoises,thereisnoneedto
conducttheexperimentsusingtheouterorthogonalarrays.Thisistrueincaseofexperimentswhichare
conductedusingthecomputersimulationastherepeatabilityofacomputersimulatedexperimentsisvery
high.

http://www.ecs.umass.edu/mie/labs/mda/fea/sankar/chap2.html

4/4