纺织专业 人工神经网络 中英文 外文 资料 文献 原文和翻译.docx
- 文档编号:9358029
- 上传时间:2023-05-18
- 格式:DOCX
- 页数:24
- 大小:426.97KB
纺织专业 人工神经网络 中英文 外文 资料 文献 原文和翻译.docx
《纺织专业 人工神经网络 中英文 外文 资料 文献 原文和翻译.docx》由会员分享,可在线阅读,更多相关《纺织专业 人工神经网络 中英文 外文 资料 文献 原文和翻译.docx(24页珍藏版)》请在冰点文库上搜索。
纺织专业人工神经网络中英文外文资料文献原文和翻译
TextileResearchJournalArticle
UseofArtificialNeuralNetworksforDeterminingtheLeveling
ActionPointattheAuto-levelingDrawFrame
AssadFarooq1andChokriCherif
InstituteofTextileandClothingTechnology,Technische
UniversitätDresden.Dresden,Germany
Abstract
Artificialneuralnetworkswiththeirabilityoflearningfromdatahavebeensuccessfullyappliedinthetextileindustry.Thelevelingactionpointisoneoftheimportantauto-levelingparametersofthedrawingframeandstronglyinfluencesthequalityofthemanufacturedyarn.Thispaperreportsamethodofpredictingthelevelingactionpointusingartificialneuralnetworks.Variouslevelingactionpointaffectingvariableswereselectedasinputsfortrainingtheartificialneuralnetworkswiththeaimtooptimizetheauto-levelingbylimitingthelevelingactionpointsearchrange.TheLevenbergMarquardtalgorithmisincorporatedintotheback-propagationtoacceleratethetrainingandBayesianregularizationisappliedtoimprovethegeneralizationofthenetworks.Theresultsobtainedarequitepromising.
Keywords:
artificialneuralnetworks,auto-lev-eling,drawframe,levelingactionpoint。
Theevennessoftheyarnplaysanincreasinglysignificantroleinthetextileindustry,whilethesliverevennessisoneofthecriticalfactorswhenproducingqualityyarn.Thesliverevennessisalsothemajorcriteriafortheassessmentoftheoperationofthedrawframe.Inprinciple,therearetwoapproachestoreducethesliverirregularities.Oneistostudythedraftingmechanismandrecognizethecausesforirregularities,sothatmeansmaybefoundtoreducethem.Theothermorevaluableapproachistouseauto-levelers[1],sinceinmostcasesthedoublingisinadequatetocorrectthevariationsinsliver.Thecontrolofsliverirregularitiescanlowerthedependenceoncardsliveruniformity,ambientconditions,andframeparameters.
Attheauto-levelerdrawframe(RSB-D40)thethicknessvariationsinthefedsliverarecontinuallymonitoredbyamechanicaldevice(atongue-grooveroll)andsubsequentlyconvertedintoelectricalsignals.Themeasuredvaluesaretransmittedtoanelectronicmemorywithavariable,thetimedelayedresponse.Thetimedelayallowsthedraftbetweenthemid-rollandthedeliveryrollofthedrawframetoadjustexactlyatthatmomentwhenthedefectivesliverpiece,whichhadbeenmeasuredbyapairofscanningrollers,findsitselfatapointofdraft.Atthispoint,aservomotoroperatesdependingupontheamountofvariationdetectedinthesliverpiece.Thedistancethatseparatesthescanningrollerspairandthepointofdraftiscalledthezeropointofregulationorthelevelingactionpoint(LAP)asshowninFigure1.Thisleadstothecalculatedcorrectiononthecorrespondingdefectivematerial[2,3].Inauto-levelingdrawframes,especiallyinthecaseofachangeoffibermaterial,orbatchesthemachinesettingsandprocesscontrollingparametersmustbeoptimized.TheLAPisthemostimportantauto-levelingparameterwhichisinfluencedbyvariousparameterssuchasfeedingspeed,material,breakdraftgauge,maindraftgauge,feedingtension,breakdraft,andsettingofthesliverguidingrollersetc.
UseofArtificialNeuralNetworksforDeterminingtheLevelingActionPointA.FarooqandC.Cherif
Figure1Schematicdiagramofanauto-levelerdrawingframe.
Previously,thesliversampleshadtobeproducedwithdifferentsettings,takentothelaboratory,andexaminedontheevennesstesteruntiltheoptimumLAPwasfound(manualsearch).Auto-levelerdrawframeRSB-D40implementsanautomaticsearchfunctionfortheoptimumdeterminationoftheLAP.Duringthisfunction,thesliverisautomaticallyscannedbyadjustingthedifferentLAPstemporarilyandtheresultedvaluesarerecorded.Duringthisprocess,thequalityparametersareconstantlymonitoredandanalgorithmautomaticallycalculatestheoptimumLAPbyselectingthepointwiththeminimumsliverCV%.Atpresentasearchrangeof120mmisscanned,i.e.21pointsareexaminedusing100mofsliverineachcase;therefore2100mofsliverisnecessarytocarryoutthesearchfunction.Thisisaverytime-consumingmethodaccompaniedbythematerialandproductionlosses,andhencedirectlyaffectingthecostparameters.Inthiswork,wehavetriedtofindoutthepossibilityofpredictingtheLAP,usingartificialneuralnet-works,tolimittheautomaticsearchspanandtoreducetheabove-mentioneddisadvantages.
ArtificialNeuralNetworks
Themotivationofusingartificialneuralnetworksliesintheirflexibilityandpowerofinformationprocessingthatconventionalcomputingmethodsdonothave.Theneuralnetworksystemcansolveaproblem“byexperienceandlearning”theinput–outputpatternsprovidedbytheuser.Inthefieldoftextiles,artificialneuralnetworks(mostlyusingback-propagation)havebeenextensivelystudiedduringthelasttwodecades[4–6].Inthefieldofspinningpreviousresearchhasconcentratedonpredictingtheyarnpropertiesandthespinningprocessperformanceusingthefiberpropertiesoracombinationoffiberpropertiesandmachinesettingsastheinputofneuralnetworks[7–12].Back-propagationisasupervisedlearningtechniquemostfrequentlyusedforartificialneuralnetworktraining.Theback-propagationalgorithmisbasedontheWidrow-Hoffdeltalearningruleinwhichtheweightadjustmentiscarriedoutthroughthemeansquareerroroftheoutputresponsetothesampleinput[13].Thesetofthesesamplepatternsisrepeatedlypresentedtothenetworkuntiltheerrorvalueisminimized.Theback-propagationalgorithmusesthesteepestdescentmethod,whichisessentiallyafirst-ordermethodtodetermineasuitabledirectionofgradientmovement.
Overfitting
Thegoalofneuralnetworktrainingistoproduceanetworkwhichproducessmallerrorsonthetrainingset,andwhichalsorespondsproperlytonovelinputs.Whenanetworkperformsaswellonnovelinputsasontrainingsetinputs,thenetworkissaidtobewellgeneralized.Thegeneralizationcapacityofthenetworkislargelygovernedbythenetworkarchitecture(numberofhiddenneurons)andthisplaysavitalroleduringthetraining.Anetworkwhichisnotcomplexenoughtolearnalltheinformationinthedataissaidtobeunderfitted,whileanetworkthatistoocomplextofitthe“noise”inthedataleadstooverfitting.“Noise”meansvariationinthetargetvaluesthatareunpredictablefromtheinputsofaspecificnetwork.Allstandardneuralnetworkarchitecturessuchasthefullyconnectedmulti-layerperceptronarepronetooverfitting.Moreover,itisverydifficulttoacquirethenoisefreedatafromthespinningindustryduetothedependenceofendproductsontheinherentmaterialvariationsandenvironmentalconditions,etc.Earlystoppingisthemostcommonlyusedtechniquetotacklethisproblem.Thisinvolvesthedivisionoftrainingdataintothreesets,i.e.atrainingset,avalidationsetandatestset,withthedrawbackthatalargepartofthedata(validationset)canneverbethepartofthetraining.
Regularization
Theothersolutionofoverfittingisregularization,whichisthemethodofimprovingthegeneralizationbyconstrainingthesizeofthenetworkweights.Mackay[14]discussedapracticalBayesianframeworkforback-propagationnetworks,whichconsistentlyproducednetworkswithgoodgeneralization.
Theinitialobjectiveofthetrainingprocessistomini-mizethesumofsquareerrors:
(1)
Where
arethetargetsand
aretheneuralnetworkresponsestotherespectivetargets.Typically,trainingaimstoreducethesumofsquarederrorsF=Ed.However,regularizationaddsanadditionalterm,theobjectivefunction,
(2)
Inequation
(2),
isthesumofsquaresofthenetworkweights,andαandβareobjectivefunctionparameters.Therelativesizeoftheobjectivefunctionparametersdictatestheemphasisfortraining.Ifα<<β,thenthetrainingalgorithmwilldrivetheerrorssmaller.Ifα>>β,trainingwillemphasizeweightsizereductionattheexpenseofnetworkerrors,thusproducingasmoothernetworkresponse[15].
TheBayesianSchoolofstatisticsisbasedonadifferentviewofwhatitmeanstolearnfromdata,inwhichprobabilityisusedtorepresenttheuncertaintyabouttherelationshipbeinglearned.Beforeseeinganydata,theprioropinionsaboutwhatthetruerelationshipmightbecanbeexpressedinaprobabilitydistributionoverthenetworkweightsthatdefinethisrelationship.Aftertheprogramconceivesthedata,therevisedopinionsarecapturedbyaposteriordistributionovernetworkweights.Networkweightsthatseemedplausiblebefore,butwhichdonotmatchthedataverywell,willnowbeseenasbeingmuchlesslikely,while
theprobabilityforvaluesoftheweightsthatdofitthedatawellwillhaveincreased[16].
IntheBayesianframeworktheweightsofthenetworkareconsideredrandomvariables.Afterthedataistaken,theposteriorprobabilityfunctionfortheweightscanbeupdatedaccordingtoBayes’rule:
(3)
Inequation(3),Drepresentsthedataset,Mistheparticularneuralnetworkmodelused,andwisthevectorofnetworkweights.
isthepriorprobability,whichrepresentsourknowledgeoftheweightsbeforeanydataiscollected.
isthelikelihoodfunction,whichistheprobabilityofdataoccurring,giventheweightsw.
isanormalizationfactor,whichguaranteesthatthetotalprobabilityis1[15].
Inthisstudy,weemployedtheMATLABNeuralNet-worksToolboxfunction“trainbr”whichisanincorporationoftheLevenberg–MarqaurdtalgorithmandtheBayesianregularizationtheorem(orBayesianlearning)intoback-propagationt
- 配套讲稿:
如PPT文件的首页显示word图标,表示该PPT已包含配套word讲稿。双击word图标可打开word文档。
- 特殊限制:
部分文档作品中含有的国旗、国徽等图片,仅作为作品整体效果示例展示,禁止商用。设计者仅对作品中独创性部分享有著作权。
- 关 键 词:
- 纺织专业 人工神经网络 中英文 外文 资料 文献 原文和翻译 纺织 专业 人工 神经网络 原文 翻译