You are on page 1of 4

8/15/2016

IntroductiontotheStatisticsofNDT

NDT.netMay2006Vol.11No.5

IntroductiontotheStatisticsofNDT
E.GinzelMaterialsResearchInstitute

Abstract
NDTinspectorsareincreasinglybeingrequiredtoproducequantitativeresultsusingtoolsthat
arenotoptimumforthetask.Asaresult,uncertaintiesinthequantitiestheyprovidecanbe
expected.Theseuncertaintiescanbebroadlygroupedintotwocategoriesdetectionuncertainty
anddimensionaluncertainty.

DETECTIONUNCERTAINTY
Detectionuncertaintyrelatestowhetherornotaflawhasbeenfound.Thisisinitselfisa
controversialtopic.Whatconstitutesaflaw?Ifweconsiderthatanydeviationfromapure
conditionconstitutesa"flawed"itemtheneventhesmallestcrystaldislocationinametal
structurecouldbeinterpretedasa"flaw".Notmany(probablynotany)componentswouldbe
considered"unflawed"ifthedefinitionwastakentothisextreme.Flawdetectionmusttherefore
takeonsomepracticalconsiderations.DetectionofflawsinNDTisbyindirectmeansand
whetherornotaflawisdetectedthereforedependsonavarietyoffactors.Thesefactorsinclude
thenatureoftheflawitself,thetestmethodused,thesensitivityofthetestmethodusedand
humanfactors.
Havingagreeduponthetestmethodandtestprotocol,therearefourpossibleoutcomesinan
inspectionofacomponent.Thesefouroptionsconstitutetheprobabilitymatrixofdetection.
1.AnitemisflawedandtheNDTmethoddetectsit(TruePositive)
2.NoflawexistsandtheNDTmethodindicatesaflawpresent(FalsePositive)
3.AnitemisflawedandtheNDTmethoddoesnotdetectit(FalseNegative)
4.NoflawexistsandtheNDTmethodhasnoindicationofaflaw(TrueNegative)
Thesearethefoundationsoftheconceptof"ProbabilityofDetection"orPOD.InNDT,this
conceptwasdevelopedmainlyinNASAintheUSAduringthe1970s.Ithassinceexpandedtoa
fewothervenuesofNDTinspectionsbutisnotyetaswidelyacceptedasmightbeexpected.
PODstrytoassessaminimumflawsizethatwillbereliablydetectedbytheNDTtechnique.This
isbestdonebyplottingtheaccumulationofflawsdetectedagainsttheflawsizeofalltheflaws
"detected"orthatproducearesponseoversomethreshold.Ideallyallflawsoversomecritical
sizewillbedetectedandflawssmallerthanthatarenot"detected".Thetoolmostcommonly
usedforPODdescriptionisthePODcurve.Figure1illustratesaPODcurvethatindicatesthereis
an80%chanceofdetectingaflaw2.2mmhighandwecanstatewith90%confidencethatit
wouldnotbegreaterthan3.3mmhigh.Orconverselywemaystatethatwehavea90%
confidencethataflaw2.2mmhighwillbedetectedwillnothaveaPODlessthan65%.

http://www.ndt.net/article/v11n05/ginzel2/ginzel2.htm

1/4

8/15/2016

IntroductiontotheStatisticsofNDT

Fig1:SamplePODcurve

Thebestwaytodescribeaconfidencecurve(e.g.95%)istostatethat:IftheactualPODcurve
weretobereconstructedoverandoverusingthesamemethodanddata,then95%ofthose
constructedcurveswouldbeabovetheconfidencecurve(i.e.5%wouldbebelow).Inother
words,weare95%surethattheREALPODcurveisabovetheconfidencecurve.
TheconfidencelevelmustalsoconsidertheeffectsofthefullmatrixforPODwherewehavethe
potentialforfalsecalls.Inasetofspecimenswithknownflawsovertheminimumrequiredsize
fordetectionwecanseethatamissofoneitemcoulddestroyourconfidenceinthesystemifa
highconfidenceisrequired.E.g.iftherewasaselectionof30specimensitwouldrequirethat28
samplesbetestedinordertoprovideassurancethatwehavewecouldhavea95%confidence
thattheresultsareindicativeofthedetectioncapabilitiesofthetestsystem.However,ifwe
missedaflawinoneofthe28sampleswewouldnolongerhave95%confidenceofinthe
probabilityofdetection.
Anexampleofthepopulationsizerequiredtobetestedtoassesstheprobabilityreliabilityis
foundatthewebsitehttp://www.surveysystem.com/sscalc.htmTheretheyexplainthatthe
confidenceleveltellsyouthelargeryoursample,themoresureyoucanbethatthetestresults
trulyreflectthepopulationdistribution.Thisindicatesthatforagivenconfidencelevel,thelarger
yoursamplesize,thesmalleryourconfidenceintervalcanbe.However,therelationshipisnot
linear(i.e.,doublingthesamplesizedoesnothalvetheconfidenceinterval).
Conversely,ifwehavealargepopulationweneedtomakeatestofmanysamplestoensure
thattheresultsareindicativeoftheentirepopulation(thecalculatoratthereferencedwebsite
indicates278samplesarerequiredinapopulationof1000toensurea95%confidencelevel).

DIMENSIONINGUNCERTAINTY
QuiteadifferentsetofuncertaintiesexistswhenanNDToperatorisrequiredtoprovide
dimensionalinformationabouttheflawsdetected.Dimensionsusuallyrequestedincludelength,
maximumverticalextent,depthbelowatestsurfaceandpositionwithrespecttosomeXY
coordinatesystem.
http://www.ndt.net/article/v11n05/ginzel2/ginzel2.htm

2/4

8/15/2016

IntroductiontotheStatisticsofNDT

NDTusesindirectmeasurementmethodstoassessflawdimensions.Magneticparticleinspection
usesthepatternofsmallparticlesentrappedinamagneticfield.Eddycurrentusescoil
impedancephaseandamplitudeinformationtogaugeflawdepthandsize.Ultrasonictestinguses
arrivaltimes,amplitudeandphaseinformationtoprovidedepth,areaandnatureofaflaw.
However,evenmoredirectmethodscansufferfromvariationinestimatesofdimensions.
Handingaverniercallipertotenpeopleandaskingthemtoprovidethethreedimensions(length,
width,height)ofasteelblockwillstillresultinascatteringofdimensionsbeingrecorded.
Variationswillresultduetodifferentskilllevelswiththecalliper,measurementsatdifferent
placesalongtheblocksurfaces,differentpressuresusedonthecalliperactuatorandeven
roundingofreadingsifthevernierismechanicalinsteadofdigital.
AnalysisofthevariationsofsizinginNDTbecomesimportantwhentheresultsareusedwith
fracturemechanicstoassessifaflawcanbeallowedtoremaininthepartornot.Materials'
propertiesarestudiedinmechanicallaboratoriestodeterminethemaximumflawsizethatwill
notresultinthecomponentfailing.ThisisanimportantconceptforNDTusersasitstressesthe
oppositetowhatmostpeopleaskofNDT.Insteadof"Whatisthesmallestflawatestmethod
canfind?"therealconcernis,"WhatisthelargestflawthatcouldbemissedbytheNDT
method?"
WhenanNDToperatorisaskedtomakeajudgmentaboutthesizeoftheflawtheresultisthen
comparedtoanallowabledimensionchart.ButiftheNDTmethodisnotexactlyaccuratethen
thereissomepossibilitythatthedimensionestimatedbytheNDToperatorislessthanactually
exists,i.e.theoperatorundersizestheflaw.TheNDToperator'sestimatethenresultsinan
underestimateoftheseverityoftheflaw.Byananalysisofthesizingtechnique,thefracture
mechanicsengineercanincorporateasafetyfactorinthecalculationsthatprovidesforthe
potentialthatthesizingtechniqueusedmightundersizeaflaw.
AswithPODcalculations,thesizingestimatesareanalysedusingacomparisontoresults
obtainedfromdestructivetesting.OnesuchanalysisisillustratedinFigure2.Figure2isaplot
oftheverticalextentofflawsthatweresizedbydestructivelyexaminingthetestsampleand
comparingthosevaluestotheNDTestimateofverticalextent.Destructivetestinginvolves
cuttingandpolishingorsplittingthesamplebymeansofa"nickbreak"andexaminingtheflaws
underamicroscopetodeterminethemaximumflawsize.Thismethodtoohassomevariability
associatedwithit.Itisthemaximumoverallvaluethatisgenerallyconsideredthereferenceto
whichtheNDTsizeiscompared.AnidealNDTmethodwouldplotalltheNDTestimatesalongthe
linethatequalsthedestructivetestsizes.However,ascatterofestimatesfromtheidealismore
typicalofNDTmethodsassignedthetaskofsizing.ThisscatterisindicatedinFigure2.
Figure2alsoincorporatessizingtolerancelines.Thesearelinesparalleltotheidealindicating
theerrorinthesizeasestimatedbyNDTcomparedtotheidealor"assumed"correctsize.

http://www.ndt.net/article/v11n05/ginzel2/ginzel2.htm

3/4

8/15/2016

IntroductiontotheStatisticsofNDT

Fig2:PlotofNDTSizeversusDestructiveSizeofFlaws

SUMMARY
Probabilityofdetectionandsizingtolerancesarenowcommonlyrequiredinseveralindustries.
Theaircraftindustry,andinparticularthemilitary,usethistypeofinformationfordamage
toleranceanalysisoftheircomponentsandschedulingofinspectionintervals.Nuclearindustries
areadaptingthisformofanalysistoassessthereliabilityofNDTtodetectflawsincomponents
duringinserviceinspections.Morerecentlythepipelineindustryhasreliedonthistypeof
analysistodevelopfitnessforpurposeacceptancecriteriafortheconstructionofpipelines.
ThestatisticalandprobabilisticassessmentsoftheNDTtechniqueareusefultools.However,the
processistimeconsuming,costlyandnotabsolute.Statisticalassessmentreliesonthe
comparisonoflimitedsamplesthatmustbeverifiedusingdestructivemethods.Althoughitis
assumedthatthedestructivemethodsarethe"true"values,heretootherecanbeuncertainty
andvariation.Wasthemacrosectionmadeatthepeakverticalextent?Howprecisewasthe
measurementreportedbythedestructivetestlab?Iftheflawwaslongandtherewasdepth
variationtotheupperedgedoesthedestructivetestreflectthis?Whenlengthisreportedwhatis
thelimittowhichthedestructivetestcanassesstheflaw?
Probabilityofdetectionisalsoafunctionofwhatisdefinedasaflaw,andsincevariationsin
determining"actual"sizeofaflawcanexistduetothedestructivetestingmethods,itfollows
thatthePODresultsofNDTassessmentsarenotsolelyafunctionoftheNDTtechnique.
NDT.net

http://www.ndt.net/article/v11n05/ginzel2/ginzel2.htm

|Top|

4/4