文档名：基于迁移学习和降噪自编码器长短时间记忆的锂离子电池剩余寿命预测
摘要：针对锂离子电池退化数据噪声大、数据量少以及不同生命时期的退化趋势不同而导致的模型预测精度低、泛化能力差等问题,从数据预处理、预测模型的构建与训练三方面展开研究:首先结合变分自编码器(VAE)和生成对抗网络模型(GAN)构建VAE-GAN模型生成多组数据,实现电池的退化数据增强;然后结合降噪自编码器(DAE)和长短时记忆(LSTM)神经网络构建DAE-LSTM模型进行数据降噪和容量预测,为了降低模型参数,此过程中的数据降噪和预测共享同一个损失函数;最后先利用生成数据对DAE-LSTM模型进行预训练,再利用真实数据对其进行迁移训练.在CACLE和NASA公开数据集进行性能测试,实验结果表明该文所提方法精度高、鲁棒性强,能够有效提高锂离子电池剩余寿命的预测效果.
Abstract：Degradationdataofbatterycapacitycanbeusedtopredictthebatteryremainingusinglife(RUL),butthereexistnumerousnoisedatainthebatterydegradationprocesscausedbyfactorssuchasambienttemperature,charge/dischargeprocessandcapacityrecoveryphenomenon.Itmakespredictionofdata-drivenlifelithium-ionbatterychallenging.ToimprovethepredictionaccuracyandgeneralizationabilityofbatteriesRUL,weproposedamethodbasedontransferlearninganddenoisingautoencoder-longshorttermmemory(DAE-LSTM).Firstly,thevariationalautoencoder-generativeadversarialnetwork(VAE-GAN)methodwasconstructeddesigned.Encodingnetworkwasusedtoestimatethedistributionofinputdata,andgeneratingnetworkanddiscriminantnetworkwereusedfordataregeneration.ItimprovedthereliabilityofgenerateddatabyVAEmethod,andsolvedtheproblemthatGANmethodhadbeendifficulttotrain.Secondly,theDAE-LSTMmethodwasconstructedfordatadenoisingandcapacityprediction.TheDAEcanreconstructtheinputdataanditsencoderimprovedtherobustnessofthemethodbyaddingGaussiannoise.LSTMlayercananalyzethetemporalcharacteristicsofdataforcapacityprediction.Duetothesmallamountofdata,thenetworklayeroftheoverallmethodwaslesstoavoidoverfitting.Toreducetheparameters,thesamelossfunctionwasusedinbothdatadenoisingandcapacityprediction.Finally,theoptimaltrainingschemewasdeterminedthroughthedifferentexperiment:ThedatageneratedbyVAE-GANwasusedformethodpre-training,thenallnetworklayersofthebasicmethodwerefine-tunedbyactualdata.Thiswouldeffectivelyimprovethepredictionaccuracyofthemethod,andensurethereliabilityofthepredictionresults.Experimentalresultsshowedthattheproposedmethodhasbetterpredictiveperformance,anddegradationtrendofmostbatteriescanbewellpredicted.MAEandRMSEwerecontrolledwithin2.46%and3.76%respectively,andthelowestwas0.95%and1.06%.Experimentalresultswithdifferentpredictionstartingpointsshowedthatthepredictionweremoreaccuratewhenthepredictionstartingpointwasclosertothefailurethreshold.Thisindicatesthatthemethodcanaccuratelypredictthedegradationtrendinlaterstagesofbatterylife.Experimentalresultswithotherdatasetsshowedthattheproposedmethodhasstrongadaptabilityandgeneralizationability.Itcaneffectivelypredictthelithium-ionbatteryRULinsmalldatasamples.The90%confidenceintervalofthepredictionresultswithNASAdatasetisnarrow,indicatingthatthemethodhasstrongrobustness.Inaddition,wecountedthetimetakentocompletetheRULpredictionfordifferentdatasets.AstheRULpredictionofbatteriesisofflinepredictionwithlowreal-timerequirement,thetrainingandtestingtimeofthemethodmeetstheofflinepredictionrequirement.Thefollowingconclusionscanbedrawnfromthesimulationresults:(1)TheDAE-LSTMmethodcaneffectivelydenoisingthedegradationdataoflithium-ionbatteries,andmakingthepredictionresultmoreaccurate.(2)VAE-GANmethodcangeneratemultiplegroupsofdegradationdataconformingtotherealdegradationtoachievethepurposeofdataenhancement.(3)TransferLearningcanensurethattheeffectiveinformationofgenerateddataandrealdataisfullyutilized,sothatthepredictionmodelhashigheraccuracyandbettergeneralizationability.Bycomparingthepredictionresultsofotherliteratures,itisprovedtheproposedmethodhashigherPreandcanbeusedtopredicttheRULoflithium-ionbatteries.
作者：尹杰 刘博 孙国兵 钱湘伟 Author：YinJie LiuBo Sunguobing Qianxiangwei
作者单位：哈尔滨理工大学测控技术与通信工程学院哈尔滨150080黑龙江大学电子工程学院哈尔滨150006
刊名：电工技术学报
Journal：TransactionsofChinaElectrotechnicalSociety
年，卷(期)：2024, 39(1)
分类号：TM912TP206+.3
关键词：锂离子电池 剩余寿命预测 降噪 自编码器 长短时记忆神经网络 迁移学习
Keywords：Li-ionbattery remainingusefullife denoisingautoencoder longshorttermmemory transferlearning
机标分类号：TP391.41TP181TM561
在线出版日期：2024年1月18日
基金项目：基于迁移学习和降噪自编码器-长短时间记忆的锂离子电池剩余寿命预测[
期刊论文] 电工技术学报--2024, 39(1)尹杰 刘博 孙国兵 钱湘伟针对锂离子电池退化数据噪声大、数据量少以及不同生命时期的退化趋势不同而导致的模型预测精度低、泛化能力差等问题,从数据预处理、预测模型的构建与训练三方面展开研究:首先结合变分自编码器(VAE)和生成对抗网络模型(GA...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
基于迁移学习和降噪自编码器-长短时间记忆的锂离子电池剩余寿命预测 Transfer Learning Denoising Autoencoder-Long Short Term Memory for Remaining Useful Life Prediction of Li-Ion Batteries
基于迁移学习和降噪自编码器-长短时间记忆的锂离子电池剩余寿命预测.pdf
- 文件大小:
- 2.33 MB
- 下载次数:
- 60
-
高速下载
|
|
