Data-driven Prediction Of Remaining Service Life Of Lithium-ion Batterie

New energy electric vehicle industry has achieved vigorous development,and lithium-ion batteries have become the main source of power for electric vehicles.With the cyclic use of the battery,its performance and the remaining useful life(RUL)gradually degrade,which seriously affects the safe driving of electric vehicles.Since the data-driven method does not need to understand the degradation mechanism or build a complex battery model,this article will be based on the fusion method with deep learning and transfer learning to mine the internal laws of the historical operation information of lithium-ion batteries.The main research contents are as follows:First,the working principle and aging characteristics of lithium-ion batteries are introduced.By analyzing the changes in voltage and current during charging,health factors that can reflect the degradation of battery performance are extracted.It visualized the battery degradation data sets published by NASA and CACLE,and explained the influence of self-recovery phenomenon on the degradation trend.Since the degradation process of lithium-ion batteries is a timing issue,Long-Short Term Memory(LSTM)is selected as the basic method.Aiming at the shortcomings of LSTM in extracting spatial features from multi-dimensional data,an improved SLSTM network is proposed.Aiming at the problem of capacity self-recovery,an attention mechanism is proposed to make the model focus more on the learning of important information.Finally,deep mining of the time and space information in the historical data,and the establishment of a spatial-temporal RUL(Spatial-temporal Remain Useful Life Forecasting,STRULF)model.By comparing the experimental results,it proves that the STRULF model can fully extract the features in the data,accurately estimate the data changes and eliminate the influence of the capacity self-recovery on the overall predict results.Taking into account the long degradation time of lithium-ion batteries,the difficulty of data acquisition and the change of data distribution over time,a transfer learning method is introduced to assist the construction of the target domain model using source domain knowledge.A fusion method of covariance difference and maximum mean difference is proposed to accurately measure the difference between source and target domain.Meanwhile,a timing distribution matching strategy is designed to solve the problem of time dependence.Finally,combining the idea of pretraining and fine-tuning methods,a Time Series Deep Transfer Network(TSDTN)model based on time series is proposed to predict the RUL value.The experimental results after and without transfer learning are compared to prove the proposed method.The effectiveness of the transfer learning method when data is scarce.Aiming at the complicated driving conditions of electric vehicles in practice,the data collected in the form of data streams will dynamically change.Offline learning cannot meet the demand,online learning is researched.Based on the STRULF model,an online learning strategy based on changes in data distribution is designed,and an Auto Dropout mechanism is proposed to enhance the generalization ability.In addition,based on the TSDTN model,using the idea of integrated learning,multi-source integrated online transfer learning is proposed.By evaluating the correlation between the target domain and different source domains,the amount of knowledge transfer in each source domain is determined.Finally,through comparative experiments of different groups,it is verified that the proposed online learning method can be applied in practice and obtain better prediction results.