Research On Robust SOH Estimation For Lithium-Ion Batteries Based On Transfer Learning And Mixture

With the advantages of high energy density,long service life,high rated voltage,light weight,low self-discharge rate,no memory effect and environmental protection,lithium-ion batteries have become the best choice for new energy electric vehicles.However,with the increase of chargedischarge cycles and usage time,the capacity of lithium-ion batteries will gradually decrease and their performance will gradually deteriorate due to aging and operating environment.which will seriously affect the operation and driving of electric vehicles and the safety of human life and property.Therefore,it is especially important to achieve accurate state of health(SOH)estimation of lithium-ion batteries.In this paper,we take lithium-ion battery as the research object.On the one hand,considering that the Battery Management System(BMS)often works under complex environmental conditions and is easily affected by non Gaussian noise and random fluctuations,on the other hand,considering the large distribution differences between different datasets,the difficulty of obtaining lithium-ion battery data,and the small sample size of the studied data,a datadriven SOH estimation model for lithium-ion batteries is constructed by combining Transfer Learning and Mixture Generalized Maximum Correntropy Criterion(MGMCC),and the effects of different noises on the estimation performance of the model are investigated.The main studies are as follows:First,the composition,working principle and performance parameters of lithium-ion batteries are analyzed,and the internal and external factors affecting the SOH decline of lithium-ion batteries are discussed.Then the selected lithium-ion battery data set is analyzed in detail,and according to the actual application scenario,the duration of different charging voltage ranges from the constant current charging voltage curve is selected as the health features to indirectly characterize SOH,and then the correlation analysis of the selected health features is performed using the Spearman correlation coefficient,and the health features with high correlation are retained as the input of the model.Finally,the selected health features were pre-processed.Second,to solve the problem of complex non-Gaussian noise or outliers in the measurement data,MGMCC is introduced,and a new robust estimation model(MGMCC-ELM)is constructed to improve the estimation accuracy and robustness of SOH of lithium-ion batteries by combining MGMCC with Extreme Learning Machine(ELM),and the regularization term is introduced in the substitution function to avoid the numerical instability of the model during the calculation.The simulation is validated with the measurement noise as Gaussian noise and non-Gaussian noise.The simulation results show that the proposed MGMCC-ELM model can effectively suppress the effects of Gaussian noise and non-Gaussian noise,and achieve accurate and robust SOH estimation.Finally,for the problem of large distribution differences between different lithium-ion battery data sets and few available data samples,the idea of migration learning is introduced,and the Deep Extreme Learning Machine(DELM)is combined with MGMCC to constitute a migration learning model based on MGMCC-DELM.Two lithium-ion battery datasets,NASA and CACLE,are selected as the source and target domain data respectively,and the simulation experiments are validated from the perspective of single-source and multi-source domains in the Gaussian and non-Gaussian noise environments.The simulation results show that under the interference of different noises,the Transfer Learning model based on MGMCC-DELM proposed in this paper fully learns the features of source domain data and target domain data after pre-training of source domain data and finetuning of target domain data,and can achieve accurate estimation of SOH in target domain after transfer.