简介:Weldshapecontrolisafundamentalissueinautomaticwelding.Inthispaper,adoublesidevisualsystemisestablishedforpulsedgasmetalarcwelding(P-GMAW),andbothtopsideandbacksideweldpoolimagescanbecapturedandstoredcontinuouslyinrealtime.Byanalyzingtheweldshaperegulationwiththemoltenmetalvolume,sometopsideweldpoolcharacterizedparameters(WPCPs)areproposedfordeterminingpenetrationinbuttweldingofthinmildsteel.Moreover,someBPnetworkmodelsareestablishedtopredictbacksideweldpoolwidthwithweldingparametersandWPCPsasinputs.
简介:Basedonthemethodofartificialneuralnetwork,anewapproachhasbeendevisedtopredictthemechanicalpropertyofE4303electrode.Theoutlinedpredicationmodelfordeterminingthemechanicalpropertyofelectrodewasbuiltupontheproductiondata.Theresearchleveragesabackpropagationalgorithmastheneuralnetwork'slearningrule.Theresultindicatesthattherearepositivecorrelationsbetweenthepredictedresultsandthepracticalproductiondata.Hence,usingtheneuralnetwork,predicationofelectrodepropertycanberealized.Forthefirsttime,thisresearchprovidesamorescientificmethodfordesigningelectrode.
简介:Inthepresentstudy,artificialneuralnetwork(ANN)approachwasusedtopredictthestress–straincurveofnearbetatitaniumalloyasafunctionofvolumefractionsofaandb.Thisapproachistodevelopthebestpossiblecombinationorneuralnetwork(NN)topredictthestress–straincurve.Inordertoachievethis,threedifferentNNarchitectures(feed-forwardback-propagationnetwork,cascade-forwardback-propagationnetwork,andlayerrecurrentnetwork),threedifferenttransferfunctions(purelin,Log-Sigmoid,andTan-Sigmoid),numberofhiddenlayers(1and2),numberofneuronsinthehiddenlayer(s),anddifferenttrainingalgorithmswereemployed.ANNtrainingmodules,theloadintermsofstrain,andvolumefractionofaaretheinputsandthestressasanoutput.ANNsystemwastrainedusingthepreparedtrainingset(a,16%a,40%a,andbstress–straincurves).Aftertrainingprocess,testdatawereusedtochecksystemaccuracy.Itisobservedthatfeed-forwardback-propagationnetworkisthefastest,andLog-Sigmoidtransferfunctionisgivingthebestresults.Finally,layerrecurrentNNwithasinglehiddenlayerconsistsof11neurons,andLog-Sigmoidtransferfunctionusingtrainlmastrainingalgorithmisgivinggoodresult,andaveragerelativeerroris1.27±1.45%.Intwohiddenlayers,layerrecurrentNNconsistsof7neuronsineachhiddenlayerwithtrainrpasthetrainingalgorithmhavingthetransferfunctionofLogSigmoidwhichgivesbetterresults.Asaresult,theNNisfoundedsuccessfulforthepredictionofstress–straincurveofnearbtitaniumalloy.
简介:Laserblankweldingisbecomingmoreandmoreimportantintheautomotiveindustryandthequalityoftheweldiscriticalforasuccessfulapplication.Afullyautomatedsolutionisrequiredtoinspectthequalityoftheblanks.ThispaperpresentsavisioninspectionsystemwithaCMOScamerawhichusesART2networktoinspectthedefectson-linetoobtainthegeometryandthequalityoftheweldseam.TheneuralnetworkART2hasthecapabilityofself-learningfromtheenvironment.Itcandistinguishthedefectsthathavebeenlearnedbeforeandgivenewoutputsfornewdefects.SoART2networkissuitableforweldqualityinspectioninlaserblankwelding.Additionally,aCO_2laserisusedfortheblankbutt-welding.
简介:Thenon-linearrelationshipbetweenparametersofrapidlysolidifiedagingprocessesandmechancalandelectricalpropertiesofCu-Cr-Zralloyisavailablebyusingasupervisedartificialneuralnetwork(ANN).Aknowledgerepositoryofrapidlysolidifiedagingprocessesisestablishedviasufficientdatalearningbythenetwork.Thepredictedvaluesoftheneuralnetworkareinaccordancewiththetesteddata.Soaneffectivemeasureforforeseeingandcontrollingthepropertiesoftheprocessingisprovided.