Professional Documents
Culture Documents
AComparisonofMeasurementSystemAnalysisMetrics:
Part2of2
NickFinstrom, JeremyStrief, TarekHaddad, EricMaassand KarenHulting 0
Theprecisionofameasurementsystemiscommonlyassessedusingagagerepeatabilityandreproducibility(GR&R)study.Part1ofthisarticlediscussed
metricsusedinmeasurementsystemanalysis.Here,Part2comparescommonlyusedGR&Rmetricswithprobabilitiesofmisclassification.
ComparisonofGR&RMetricswithProbabilityofMisclassificationUsingNumeric
Simulation
Asimulationstudywasundertakentoquantifytherelationshipbetweenguardbanding,percenttolerance(alsoknownastheprecisiontotolerance[P/T]ratio),
andtheprobabilityofmisclassificationallinthepresenceofvaryingdistributionsforbothpartvaluesandgageerror.Aresponsesurfacedesignedexperiment
wasutilizedtogenerateabalancedsetoffactorlevelcombinations.Thefollowingfourfactorswereusedtosummarizetheimportantcharacteristicsofthegage
andpartdistributions:
1.Truevaluecapabilityindex,Ppk.Thisfactordescribesvarianceofthetruevaluepopulationwithrespecttospecificationlimits,includingthecenteringofthe
truevaluemean,,withinthespecificationlimits.
iSixSigma CreateaniSixSigmaAccount
WhereCPL=thedifferenceofthecenterlineandtheLSL(lowerspecificationlimit)andCPU=thedifferenceoftheUSL(upperspecificationlimit)andthe Login
centerline:
www.iSixSigma.com
iSixSigmaJobShop
iSixSigmaMarketplace
2.Theratio
WherePpisdescribedbyequation(9b).Thisfactordescribesthecenterednessofthetruevaluepopulationwithinthespecificationlimits.
3.Theratio
Thisfactordescribesgagevariancewithrespecttotruevaluevariance.Thisiseffectivelytheinverseofpercentprocess[seeequation(4)inPart1]
dividedby100percent.
4.Guardbandcounttakenask,wherek(11),isthenumberofgagestandarddeviations(g)takenwithineachspecificationlimittoestablishguardbanded
specifications.
Assumingnormaldistributionsforgagevarianceandtruevaluepopulations,eachofthesefourfactorscanbeusedtoestablishprobabilitydensityfunctionsfor
gagevariabilityandtruevaluepopulationwithrespecttospecificationlimits,butwithouthavingabsolutevaluesofgagevarianceandtruevaluepopulation.
(Note:Inreality,thereisnosuchthingasatruevaluesinceGR&Ronlyseekstounderstandtheprecisionofthemeasurementsversussomemeasurement
space[eitherthetoleranceortheobservedspreadofthepartswhichalsoincludesthegagevariation].Referencestoatruevalueimplythataccuracywas
studiedbutthatisnotincludedinthisarticle.)
Theseprobabilityfunctionscanbeusedtocalculatepercenttoleranceandprobabilityofmisclassification.Thefourfactorscanbecombinedtoestablisha
spaceoftypicalgagevariance,truevaluepopulationvarianceandguardbanding,whicharethenusedtomappercenttoleranceandprobabilityof
misclassificationoverthevariouscombinationsofthesefactors.Oncethesetwogagemetricsaremappedovercombinationsofthesefactors,relationships
betweenthetwometricscanbeestablishedoverthesamespace.Ifa1sidedspecificationequationisnotavailable,however,thesameprobabilitydistribution
functionscanbeestablishedandthesamemappingisstillpossible.
Thesefourfactorsarenotindependentfromoneanother,andwhencombinedtomapandcomparegagemetricstheywillnotformanorthogonalcomparison
grid,wheregridlinesareperpendiculartointersection.However,comparisonbetweengagemetricsoverrangesofeachfactorprovidesameanstodraw
generalconclusionsabouttheeffectivenessofeachmetricinvariouscircumstanceswithouthavingabsolutemeasuredvalues.Thelackoforthogonalitymust
betakenintoaccountwhenmakingthiscomparison,butitdoesnotprecludeobtainingsomeusefulinformationasaresultofthestudy.
Percenttoleranceandprobabilityofmisclassificationcanbemodeledovercombinationsofthefourfactorsusingresponsesurfaceanalysis,inwhichmodels
canberepresentedusingcontourplotsofthefittedsurfaces.Thetypeofresponsesurfacedesignofexperiment(DOE)chosenforstudyisacentralcomposite
design(CCD).1Inthistypeofdesign,factorialcombinationsoffactorsalongwithcenterpointsandaxialpointsareusedtostructurestudyinputs.The
resultingdatacanbeusedtofitmodelsinvolvingprimaryfactors,theirinteractionsandsecondorderpolynomialterms.Theratiodefinedin(10)canvaryover
multipleordersofmagnitudeforthissimulationtheinputwasconvertedtoanaturallogscalethatforcedarangeovermultipleordersofmagnitudetobeinput
onalinearscale.Experimentalresultscanbeusedtoestimatemultifactorregressionequations,whichcanthenbeusedtonumericallypredictresponsesover
thedesignspace.
Thedesignspaceischosentorepresenttypicalrangesofequations(8),(9),(10)and(11),whileavoidingconditionswherecombinationsof(10)and(11)
wouldsatisfy(7),therebyresultinginnullvaluesforprobabilityofmisclassification.Thevaluesofaxialpointsareselectedas1.1timeslargerthantheextentof
primaryfactorialpointsawayfromthecenterpointforsimilarreasons.Theloweraxialpointfor(11)issettozerotoavoidanegativevalue.Thevaluesofthe
factorialpoints,centerpointsandaxialpointsforthethreeremaininginputstothedesignareshowninTable1.
Table1:DOEInputsforResponseSurfaceMappingPercentToleranceandProbabilityofMisclassification
TheDOEanalysisofvariance(ANOVA)tableprovidesinformationregardingsignificanttermsandlackoffitforeachoftheoutputsstudied.Significantterms
aretakenashavingapvaluelessthan0.05.1ThepredictedR2ischosentodeterminelackoffitandusefulnessofthemodeltopredictresults.PredictedR2
capturesthepercentageofaresponsevariationexplainedbyrelationshipswithinputsusingpredictedmodeloutputversusobservedoutputtoquantifylackof
fit.Thesecondorderpolynomialtermfor(9a)wasfoundtobesignificanttotheprobabilityofgoodobservedbadandpercenttolerance.Theinfluenceof(9a)on
percenttoleranceisduetononorthogonalityofinputfactors.Basedonobservationofinsignificantfirstordertermsandinteractiontermsfor(9a),theDOEinput
factorswerereducedandinteractiontermsincluding(9a)wereremoved.Thefirstorderandsecondordertermassociatedwith(9a)wereleftinsubsequent
analysisduetothesignificanceofthesecondorderterminthemodelforprobabilityofgoodobservedbad.Goodnessofmodelfitandfactorsignificancefor
eachofthethreemeasurementsystemanalysismetricsusingthereducedmodeltermsareshowninTable2.
Table2:GoodnessofModelFitandFactorSignificance
pvalues
Ppk/Pp 1 1 1
0.005 0 0
Ppk*Ppk 0 0 0
Ppk/Pk*Ppk/Pk 0.057 0.002 0
0.67 0 0.01
*
0.007 0 1
Ppk*
Ppk*GuardBandk 0 0 1
0.028 0.109 1
*GuardBandk
ResponsesurfacecontourplotsfortheP/TratioandtheprobabilityofbadmisclassifiedasgoodareoverlaidinFigure4overthestudiedrangeof
andPpk.Twooverlaidcontourplotsaredrawnforguardbandk=0and2respectively.BothplotshaveafixedvaluePpkPp=1.
Figure4:PlotofP/TRatioandProbabilityofBadMisclassifiedasGood(ClicktoEnlarge)
Thebandsdefinedbytheadjacentcontourlinesindicatesensitivityofeachoutputtotheinputfactorsoneachaxis.Theprobabilityofmisclassifyingbadas
goodshowsthemostsensitivitytoPpkandisrelativelyinsensitiveto .TheoppositeistrueforP/Tratio.Thistrendholdstrueforbothplotsattwo
iSixSigma CreateaniSixSigmaAccount
differentguardbandvalues.Thissensitivityanalysisestablishesthattheprobabilityofmisclassificationismoredependentontheprobabilitythatavalueisbad Login
orgood,asopposedtotheprobabilitythatthemeasuredvalueisdifferentfromthetruevalue.CurvatureisshownintheP/Tratioresponse,whichindicates
www.iSixSigma.com
iSixSigmaJobShop
sensitivitytoPpkand thiscurvatureisduetononorthogonallyoftheinputfactors.Accordingtoequation(5),theP/Tratioisnotdependenton
iSixSigmaMarketplace
processstandarddeviationandisonlydependentonprocessmeanforonesidedspecifications.Inthismodel,however,gagestandarddeviationis
establishedbasedonaratiowithprocessstandarddeviation,andprocessstandarddeviationisaninputtothefactorsoneachplotaxis.
TheinfluenceofguardbandingoneachofthetwooutputsisestablishedbycomparingthetwoplotsinFigure4.P/Tratiodoesnotchangeasafunctionof
guardbanding,whichisexpectedaccordingtoequation(5).Theprobabilityofmisclassifyingbadasgoodchangessuchthattheprobabilityisreducedforlower
valuesofprocesscapability.Guardbandinghasmoreinfluenceontheprobabilityofmisclassifyingbadasgoodwhentheprobabilityisgreaterthan1in
1,000,000forvaluesequaltoorlessthanthislevel,guardbandinghasasmallerinfluenceonreducingtheprobabilityofmisclassification(i.e.,atahigher
processcapability).
Thedifferenceinsensitivityofeachoutputovertheplotrangeatbothvaluesofguardbandingillustratesfourconditionsforgageprecision,asdefinedbyP/T
ratio,andprobabilityofmisclassification.Theyare:
Condition1: >2.0,Ppk>1.3
P/Tratioiswithintypicalacceptancelimitsandtheprobabilityofmisclassifyingbadasgoodisrelativelysmall.
Condition2: <2.0,Ppk>1.3
P/Tratioislargerthantypicalacceptancelimitshowever,theprobabilityofmisclassifyingbadasgoodisrelativelysmall.
Condition3: <2.0,Ppk<1.3
P/Tratioislargerthantypicalacceptancelimitsandtheprobabilityofmisclassifyingbadasgoodisrelativelylarge.
Condition4: >2.0,Ppk<1.3
P/Tratioiswithintypicalacceptancelimitsandtheprobabilityofmisclassifyingbadasgoodisrelativelylarge.
Forconditions1and3,theP/Tratioandprobabilityofmisclassifyingbadasgoodagreeintheirassessmentofgagesuitabilityfordecisionmaking.Incondition
1,thegageisgenerallyconsideredsuitable.Incondition3thegageisgenerallyconsideredillsuitedfordecisionmaking.Forconditions2and4,theP/Tratio
andprobabilityofmisclassifyingbadandgooddisagreeintheirassessmentofgagesuitabilityfordecisionmaking.Incondition2,thegageisconsidered
imprecisehowever,theunderlyingtruevaluepopulationissufficientlyfarawayfromspecificationvaluesastominimizetheprobabilityofmisclassifyingbadas
good.Thisconditionmayavoidriskoffalseacceptancebymisclassifyingnonconformingvaluesasgood,butadditionalcostmayresideintheprobabilityof
misclassifyinggoodasbad.Incondition4,thegageisconsideredprecise,buttheunderlyingtruevaluepopulationiscloseenoughtospecificationvaluessuch
thattheprobabilityofmisclassifyingbadvaluesasgoodremainshigh.Herethegagemaybepreciseenoughtodifferentiatevalueswithinthespecification
tolerance,butthemagnitudeofmeasurementerrorisstilllargeenoughtowarrantsignificantriskinusingthemeasurementsystemtosortvalues,wherethesort
conditionisbasedonspecificationlimits.
Guardbandingreducestheprobabilityofmisclassifyingbadasgood,therebyincreasingthesuitabilityofameasurementsystemformakingeffectivedecisions
atlowervaluesofprocesscapability.TheprobabilityofmisclassifyingbadasgoodisnotreducedtozeroovertheentirerangeofPpkshown.Evenforguard
bandingat2g,theprobabilityofmisclassifyingbadasgoodcanremainrelativelyhighatlowprocesscapability.
Guardbandinghasbeenshowntoincreasetheprobabilityofmisclassifyinggoodasbad.ThisisillustratedinFigure5inwhichresponsesurfacecontourplots
fortheprobabilityofgoodmisclassifiedasbadareoverlaidwiththesamecontoursshowninFigure4.Plotrangesof andPpkarethesameasin
Figure4.Twooverlaidcontourplotsaredrawnforguardbandk=0and2,respectively.BothplotshaveafixedvaluePpkPp=1.
Figure5:OverlaidContourPlotofP/TRatio,ProbabilityofBadMisclassifiedasGoodandProbability
ofGoodMisclassifiedasBad(ClicktoEnlarge)
www.iSixSigma.com
AsinFigure4,thebandsdefinedbytheadjacentcontourlinesindicatesensitivityofeachoutputtotheinputfactorsoneachaxis.Theprobabilityof
iSixSigmaJobShop
iSixSigmaMarketplace
misclassifyinggoodasbadisnearlyequallysensitivetoPpkand probabilityofmisclassifyinggoodasbadincreasesasPpkor
decrease.TheinfluenceofguardbandingontheprobabilityofmisclassifyinggoodasbadisseenbycomparingthetwoplotsinFigure5.Whennoguard
bandingisapplied,theprobabilityofmisclassifyinggoodasbadislessthan1percentovertherangeofPpkand ,whereP/Tandtheprobabilityof
misclassifyingbadasgoodwouldbeconsideredgenerallyacceptable.Whenguardbandingisapplied,theprobabilityofmisclassifyinggoodasbadisfoundto
tradeoffismostprevalentatlowvaluesofPpkand .
Thefourconditionspreviouslyestablishedcanbesummarizedtoincludeinformationontheprobabilityofmisclassifyinggoodasbad.
Ppk<1.3 Ppk>1.3
Condition4 Condition1
>2.0
TheP/Tratioiswithintypicalacceptance TheP/Tratioiswithintypicalacceptance
limits limits.
Theprobabilitiesofmisclassificationare Theprobabilitiesofmisclassificationare
relativelylarge. relativelysmall.
Condition3 Condition2
<2.0
TheP/Tratioislargerthanthetypical30 TheP/Tratioislargerthanthetypical30
percentacceptancelimits. percentacceptancelimits.
Theprobabilitiesofmisclassificationare Theprobabilityofmisclassifyingbadasgood
relativelylarge. isrelativelysmallandtheprobabilityof
misclassifyinggoodasbadisrelatively
large.
Forconditions1and3,theP/Tratioandprobabilitiesofmisclassificationagreeintheirassessmentofgagesuitabilityfordecisionmaking.Basedonthis
analysis,agageissuitedfordecisionmakingincondition1.Forcondition2,theimprecisionofthegagedoesnotjeopardizetheriskoffalseacceptance.A
significantriskoffalsereject(i.e.,excessscrap),however,mayexist,andthegagessuitabilityfordecisionmakingisconditionalbasedonthecircumstances
surroundingthemeasurement.Forcondition4,thegagesprecisionmaynoteffectivelyeliminatetheriskofmisclassifyinggoodvaluesasbadorbadvaluesas
goodthegagemaynotbeconsideredsuitableforapplicationsofsortingvalueswithinapopulationwherethesortingconditionisbasedonspecificationlimits.
NonnormalDistributions
TheaforementionedDOEsimulationapproachandresultsarebasedonnormaldistributionassumptionsforunderlyingtruevaluepopulationsandgageerror.
Althoughgageerrortypicallyfollowsanormaldistribution,underlyingtruevaluepopulationsmaynotbe.TypicalGR&Rresultsshouldberelativelyinsensitive
tononnormaldatasets,wheretheprobabilityofmisclassificationwillbesensitive.
Lognormaltruevaluepopulationsandanytruevaluepopulationwithsignificantskewwillnothavesymmetricprobabilitydensityfunctionsaboutthedistribution
arithmeticmean.Asaresult,theseDOEsimulationscannotbegeneralizedbasedontheexactsamefactorsresultsneedtobebasedonabsolutedistribution
positionrelativetospecificationvalues.Initialattemptstorepeatthistrialusingnonnormaldistributionssuggestgeneralagreementwiththeseresults,except
thattheresultsaredependentontheabsolutepositionofthepopulationwithinthespecificationlimits.
Conclusions
ComparisonofGR&Rmetrics,namelyP/Tratio,withtheprobabilitiesofmisclassificationrevealsthatprecisemeasurementsystemsmaystillgenerateresults
withsignificantprobabilitiesofmisclassification.ADOEstudyusingnumericsimulationrevealedthatthemostsignificantriskofmisclassifyingmeasuredvalues
existswhenagageisconsideredprecisewithrespecttospecificationtolerance,butthepopulationofvaluesbeingmeasuredresidesnearoroutsideof
specificationtolerance.WheretruevaluepopulationisdefinedbyaPpkcapabilitymetriclessthan0.75,agagemaybeclassifiedassuitablewithaP/Tratio
lessthan10percent,whileprobabilitiesofmisclassifyinggoodvaluesasbadorbadvaluesasgoodmaybegreaterthan1in10,000.
Guardbandinginfluencesprobabilitiesofmisclassification,andcanbeusedtoreducetheprobabilityofmisclassifyingbadvaluesasgoodatthecostof
increasingtheprobabilityofmisclassifyinggoodvaluesasbad.Guardbandinghasthemostinfluenceonprobabilityofmisclassificationforimprecise
measurementsystemsandwhenapopulationofvaluesbeingmeasuredhaslowcapabilityindex.2
References
1.Montgomery,D.C.,andG.C.Runger.GageCapabilityandDesignedExperimentsPartII:ExperimentalDesignMEthodsandVarianceComponent
Estimation.QualityEngineering6(1993b):289305.
2.Taylor,Wayne.GenericSOPStatisticalMethodsforMeasurementSystemVariation,AppendixBGageR&RStudy.WayneTaylorEngerprises,Inc.,n.d
iSixSigma CreateaniSixSigmaAccount Login
IfYouLovedThisArticle,YouMightAlsoLove
www.iSixSigma.com
1.TurningJudgmentCallsintoReliableDatawithGageR&R
iSixSigmaJobShop
2.TrustingtheData:GageR&RinTransactionalProjects
iSixSigmaMarketplace
3.ASimpleWaytoTestDataWithoutDoingaGageR&R
4.MakingSenseofAttributeGageR&RCalculations
Tags:GageR&R,MSA