基于神经网络的维修物料需求预测系统的研究与开发

上海交通大学硕士学位论文基于神经网络的维修物料需求预测系统的研究与开发姓名：薛锋申请学位级别：硕士专业：电子与通信工程指导教师：姜丽红;谢晨20060112IB/SC/STerminalB/SC/STerminalIIBPBPBPABSTRACTIIIREARCHANDDEVELOPMENTOFREPAIRMATERIALREQUESTFORECASTSYSTEMBYNEURALNETWORKThesubjectofthispaperisthematerialrequestforecastsystembyneuralnetworkforrepairingbusiness.Firststageoftheresearchworkistobaseonthecharacteristicofrepairingbusinesscompany,pointouttheimportanceofrepairmaterialintherepairingbusinesscompany.Thencarryoutthetheoreticalanalysisofcurrentmaterialrequestforecastsituation,describethedifficultlevelandthereasonsforsuchkindofphe-nomena.Becausetheinformationsystemsaredividedinthisrepairingbusi-nesscompany,duringtheresearchprocessitisrequiredtocollectthedatawhichisstoreddifferentsystems,includingtheB/Sbasedrepairingrec-ords,theC/SbasedmanufactureengineeringinformationandalsotheTerminalbasedfinancialdata.Allthesedataneedtobecentralizedandstoredaccordingtorepairingmaterialforecastrequirements.Thispaperproposestheactivedatacollectingmethod,designsthespecificB/S,C/SandTerminalbaseddataintegrationprocedures.Afterdataintegrationisfinished,thispaperfiltersanddigitalizesthekeyinputsrelatedtotheforecastcalculationandthendesignsappropriateforecastcalculationprocess.Bycomparingtheoutputandhistoricaldata,repeattheneuralnetworkparameterstrainingprocedureuntiltheforecastoutputmeetstheexpectation.Thisresearchanddevelopmentoftherepairmaterialforecastsystemisfinishedandthissystemisimplementedforactualworking.Byusingthisrepairmaterialforecastsystem,wecanhavetheaccuraterepairmate-rialforecastintime,sothiscanbeusedasthedecisionmakingaidformorereliablerepairmaterialpurchaseplanandinventorycontrol.Finallythissystemassiststhecompanytolowerthecapitaldepositforrepairmaterialinventory,reducetheoperationcostandcontroltheoperationrisk.ABSTRACTIVTheinnovationofthesubjectliesinthatbasedondeepanalysisofneuralnetworkcomputingtheory,combiningitwiththerequirementsofmaterialrequestforecastforrepairingbusiness,thispaperintroducestheneuralnetworkcomputingtheory.Consideringthecharacteristicsofalltheavailableneuralnetworkmodules,thispaperchoosestheBPneuralnetworkasthemajorfocus.Basedonthoroughconsiderationofrepairmaterial’sfeaturesandthevariousimpactingfactorsoftherepairmaterialforecast,filteranddigitalizeofinputsourceswithprioritization,sothattheycanbefitintoBPneuralnetworkparametertraining.ThisactioncanfullymakeuseoftheadvantageofBPneuralnetworkcomputingtheoryandsufficientlybaseonhugeamountofhistoricaldata,byanalysisandintegrationprocessmakeitfitforforecastcalculation,greatlyincreasetheaccuracyoftheforecastresults.Thevalueofthispaperisthattheneuralnetworkbasedforecastmethodofrepairmaterialrequestforrepairingbusinesscanbeusedastheassistingguidanceofrepairmaterialpurchaseandinventorycontrolforrepairingbusinesscompanies.Keywords:repairingbusiness,materialrequestforecast,neuralnetworkcomputingtheory,trainparameters20060112120%40%[26]2Fig.1-1RepairingMaterialCapitalOccupationChart314Fig.1-2RequestForecastErrorAnalysis2[27]345[28]1-111-=MTSdbSbTMdb6b123457nC/SnB/SnTerminal891943McCullochPittsMP[1]1944HebbHebb[37]Hebb50601957RosenblattPerceptronMP1962WidraAdalineAl[36]MinskyPaper1969Perceptron10[38]MinskyVonNeumannGrossbergKohonenFukushimaAmariAndersonBSBWebosBP[39]70VonNeumann1982HopfieldHNN[40]FeldmannBallardHintonSejnowskjBoltzmanRumelhartMcClellandPDP[33]BPKosko[35]Hecht-NielsenBPHolland1988L.0.Chua[34]11PEProcessingElement[45][21]2.1y∑sqΜixxx21w1w2wiFig.2-1ArtificialNeuralNetworkInputOutputDiagram2.1),,2,1(nixiΛ=wi∑qs∑=-=niiixws1q)(sys=1.=)(ss0001≥ss2.()ssy==s3.Sigmoid12()()011-=-BexfBxtanh()sssseeees--+-=s[18][7]A.FeedforwardNN2.213Fig.2-2ForwardingFeedbackNeuralNetworkTopologyDiagramB.FeedbackNN2.3[43]NN[42]Hopfield[41]apunovHopfield[19]ΛΛΛΛΛΛFig.2-3HopfieldNeuralNetworkTopologyDiagram14C.SupervisedLearning[31]UnsupervisedLearningReinforcementLeaming[11]1BP[20]2Hebb[44]Hebb[3]D.BPBPSigmoid01[2]Kolmogorov[32]Hecht-Nielsen1987ffx=YU[01]fBPBP15ijk12QΜΜΜ12M12LFig.2-4BPNeuralNetworkTopologyDiagramBP2.4MQLijwjkw[13]BPBP1yjk2x=),,,(21mxxxΛD=()lddd,,,10Λ3N[15]j)12(1-=∑=MilijjOnetwj)(jjnetfO=f*S)22(11)(-+=-jnetjenetf)32()(1)[()(--=′jjjnetfnetfnetfjOjjkwkk)42(1-=∑=qjjjkkOnetwqk)52()(-=kknetfO416()622121--=∑=LkpkpkpOdE5Ep[23]i.)72(-∂∂-=ΔjkjkEwhwh0h)82(-∂∂∂∂=∂∂jkkkjknetnetEEwwkd)92(-∂∂∂∂=∂∂-=kkkkknetOOEnetEdjjqjjkjkjkkkkkkkkkkOOnetnetfnetfnetnetOOdOE=⎟⎟⎠⎞⎜⎜⎝⎛∂∂=∂∂′=∂∂=∂∂--=∂∂∑=1)()()(=′-=Δ)()(ii.[4]ijjjjjjijjjijijOOnetfOEOnetEnetnetEEhdhwhwhw=′⎟⎟⎠⎞⎜⎜⎝⎛∂∂-=∂∂-=∂∂∂∂-=∂∂-=Δ)(jkLkkjjnetfwdd∑=′=1)(kjkkkjkkkjkOOdOOOdnetf)())((-=-′=ΔhhwjiLkjkkjjiLkjkkjijOOOOnetf))(1())((11∑∑==-=′=ΔwdhwdhwOkkOjjOii17[14]62EpBP2.5e≤EFig.2-5BPLearningAlgorithmWorkflowE.GeneticAlgorithmGA[12]BP[30][5]GAchromosome[45]genePopulationGAGAfitnessfunction[6]Coding18FitnessSelectionCrossover[46]Mutationbit[29]2.6Fig.2-6GeneticAlgorithmWorkflow1NS2S3K1KN/2S'4S'5KS196S7S+SKS8SS3[16]BPBPBP[47]BP[27]GA[51]BPGAGABPGANNBP[48]GABPBPPBPBP[17]5120[49]2GABPNBP[50]3BPN4()iE()if5BP636NiyiyˆmymyˆySˆySeiiiiyye-=ˆi1RMSEE)102(112-⎟⎠⎞⎜⎝⎛=∑=NiiRMSEe