基于Unet卷积神经网络的电磁场快速计算方法
文档名：基于Unet卷积神经网络的电磁场快速计算方法
摘要：有限元法(FEM)是物理场分析常用的方法,但庞大的求解自由度导致FEM计算成本很大.针对FEM计算时间长的问题,构建一种基于U-net卷积神经网络的物理场快速计算方法,将样本数据通过栅格化或点云化处理后作为神经网络的输入和标签数据,通过网络训练实现物理场的快速计算并研究该方法在电磁场计算中的应用.结果表明,该方法能准确有效地预测电势、电场强度、磁感应强度等物理量的分布,且预测时间较FEM仿真计算时间大幅缩短.同时,通过合理选择数据集大小,即使在小数据集下也能有较高的预测精度.
Abstract：Duringtheengineeringcasedesignsession,FiniteElementMethod(FEM)analysisisessentialtoaddressthephysicalfieldsofthestudyobject.However,thesignificantnumberofdegreesoffreedomoftenresultsinlengthycomputationtimesforFEM,particularlywhensolving3Dproblemsinvolvingcoupledmulti-physicalfields.Indesign-orientedproblemslikebatterystructureoptimization,findingtheoptimalsolutiontypicallyrequiresthousandsofiterations.IftheFEMmethodisemployed,thisprocesscouldtakemonthstocomplete,significantlydelayingprojectschedules.Furthermore,itiscrucialtoassessequipmentstatusbasedonsensordatasuchasmotordigitaltwinmodelingandtransformerhotspottemperaturereal-timecalculation.Utilizingthetime-consumingfiniteelementmethodinthesescenariosmayleadtodelayedidentificationandresolutionoffailures,potentiallycausingfurtherharm.Adeeplearning-basedfastcomputationmethodforphysicalfieldsisproposedbycombiningFEMwiththeU-netconvolutionalneuralnetwork(U-netCNN),studyingtheeffectivenessofthisapproachinelectromagneticfields.Firstly,weemploythefiniteelementmethodtomodelandsimulatetheresearchobjectbasedontheactualphysicalmodel.Secondly,weexporttheobtainedsimulationresultsandtransformthemodelbydividing,rasterizing,andconvertingitintoapointcloud.Thesemodificationsarebasedonthegeometricparameters,boundaryconditions,andphysicalfieldsolutionresults.Thirdly,wetraintheU-netCNNbyoptimizingitsnetworkparametersandutilizethetrainedmodelforswiftphysicalfieldcalculationsoftheresearchobject.Thetrainedmodelallowsforrapidcomputationofthephysicalfield.Toassessthepracticalityofourmethod,weselectatwo-dimensionalinsulatoruniformpressureringmodelandathree-dimensionaltransformermodeltocalculatetheelectrostaticandmagneticfields.Inthetwo-dimensionalmodel,weemploybothsingle-channelandmulti-channelinputs.Thesingle-channelinputsolelyconsistsofthegeometricmodelparameters,whilethemulti-channelinputincludesadditionaldataincludingtheradiusandupliftofthesizeequalizingring.Tominimizememoryspace,wemergethemulti-channelinputsusingaGaussiandistribution.Theresultsdemonstratethatthesingle-channelU-netCNNachievestheaccuracyof99.88%forpotentialand99.52%forelectricfieldstrength.Meanwhile,theFCNs-16modelachievesapotentialaccuracyof99.15%andanelectricfieldstrengthof98.33%.Thepotentialaccuracyattainedwiththemulti-channelinputis99.93%,withanelectricfieldstrengthof99.52%.ThepotentialaccuracyachievedwiththeFCNs-16modelis99.80%,alongwithanelectricfieldstrengthof99.24%.Additionally,thecomputationtimefortheU-netCNNis0.017sforthesingle-channelinputand0.016sforthemulti-channelinputbothsignificantlyfasterthanthefiniteelementmethod.Finally,byreducingthesizeofthedataset,thenetwork'spredictionaccuracyremainsabove90%with306groupsofdata,andevenwithjust203groups,itmaintainsanaccuracyofover85%.Inthethree-dimensionalmodel,weaddressthemagneticfieldofthetransformeratt=0.05sthroughfield-pathcoupling.TheinputsforthismodelconsistofSamplingpointcoordinatesandvoltage.TheresultsrevealthattheU-netCNNachievesthehighestaccuracyformagneticinductionintensityat99.26%,whiletheFCNs-16modelachievesanaccuracyof98.91%.Byreducingthedatasetsize,themodelcanstillmaintainahighpredictionaccuracy,evenwithonly193groupsofdata.Themethodpossessesthecapabilitytobeemployedinreal-timecalculationsforequipmentconditionassessment,aswellasindesignsessionsnecessitatingmultipleiterations.
作者：张宇娇赵志涛徐斌孙宏达黄雄峰Author：ZhangYujiaoZhaoZhitaoXuBinSunHongdaHuangXiongfeng
作者单位：合肥工业大学电气与自动化工程学院合肥230009
刊名：电工技术学报
Journal：TransactionsofChinaElectrotechnicalSociety
年，卷(期)：2024, 39(9)
分类号：TM15
关键词：电磁场卷积神经网络快速计算有限元法
Keywords：Electromagneticfieldsconvolutionalneuralnetwork(CNN)fastcalculationfiniteelementmethod(FEM)
机标分类号：TP301.6P458.121.1O441.4
在线出版日期：2024年5月13日
基金项目：国家自然科学基金基于U-net卷积神经网络的电磁场快速计算方法[
期刊论文]电工技术学报--2024, 39(9)张宇娇赵志涛徐斌孙宏达黄雄峰有限元法(FEM)是物理场分析常用的方法,但庞大的求解自由度导致FEM计算成本很大.针对FEM计算时间长的问题,构建一种基于U-net卷积神经网络的物理场快速计算方法,将样本数据通过栅格化或点云化处理后作为神经网络的输入...参考文献和引证文献
参考文献
引证文献
本文读者也读过
相似文献
相关博文
基于U-net卷积神经网络的电磁场快速计算方法Fast Calculation Method of Electromagnetic Field Based on U-Net Convolutional Neural Network
基于U-net卷积神经网络的电磁场快速计算方法.pdf
页:
[1]