Research On The Method Of Estimating The Health State Of Lithium-ion Battery Based On Transfer Learning

Lithium-ion batteries are widely used in electric vehicles.Since electric vehicles cannot accurately obtain the SOH value in the actual driving process,the reliable labels required for data-driven modeling can only be collected in the laboratory environment,but the battery aging cycle is long,resulting in a high cost of time for data collection.Therefore,it is difficult to obtain sufficient data.There are many types of lithium-ion batteries,but different lithium-ion batteries have different charging and discharging strategies,which will cause a certain degree of data distribution difference.If a pure data-driven approach is adopted,it cannot meet the requirements of traditional machine learning models for the same distribution of training data and test data.Since different lithium-ion batteries contain common features related to SOH prediction,for scenarios where there are many open source data sets of different types of batteries available,this paper applies the Transfer Learning algorithm to the battery SOH estimation.There are few researches on the Transfer Learning algorithm for SOH estimation of lithium-ion batteries.Therefore,it is of great application value to study the knowledge transfer algorithm across battery types to achieve a reliable SOH estimation model.The training data of the Transfer Learning model consists of source domain data and a small amount of target domain data.This paper fully considers the data distribution differences between different lithium-ion batteries,and uses two open source data sets of NASA and CALCE to conduct single-source and multi-source domain knowledge transfer experiments,and different transfer learning methods are used in different migration scenarios.Summarized as follows:(1)Due to the data distribution differences between different lithium-ion batteries,this paper proposes a data distribution difference measurement index m MMD,which is used to measure the difference between the source and the target domain data in the training data.Based on the classic transfer learning algorithm: Twostage-Tr Adaboost.R2 algorithm,the weight initialization method and weight update strategy are optimized,and the improved Twostage-Tr Adaboost.R2 algorithm realizes the knowledge transfer between two different battery data,and predicts the target battery SOH estimation result.(2)Aiming at the scenario where the information provided by a single source domain battery is relatively limited,this article expands the single source domain knowledge transfer to the multi-source domain knowledge transfer scenario,the purpose is to enable the model to automatically select useful information from multiple source domain batteries to help the target domain battery SOH estimation.Therefore,a Multi-source domain Transfer Gaussian Process Regression algorithm(MSTR-GPR)is proposed and theoretically proved,and an integrated weighting method based on prediction uncertainty is designed to integrate multiple TR-GPR models,and finally obtain the target battery The SOH estimation result.This paper verifies the effectiveness of the algorithm based on MATLAB language and PYTHON environment.The results show that even if only the first 20% of the target battery data is used as the target domain training data,the Improved TwostageTr Adaboost.R2,MSTR-GPR algorithm can still accurately track the reference SOH of the target battery.Verified the effectiveness of the algorithm.In summary,under the condition of only a small amount of target data,the traditional machine learning model can use less relevant information,while the transfer learning model can transfer useful knowledge in the source field,thereby achieving a reliable estimation of the target battery SOH.Advances the practicality of machine learning in battery SOH estimation.